From 5d350fc8433182ec206ffe8a7e50eb58db563809 Mon Sep 17 00:00:00 2001
From: Megvii Engine Team <megengine@megvii.com>
Date: Fri, 5 Mar 2021 17:40:25 +0800
Subject: [PATCH] feat(dnn/cuda): add deconv int8 and fix cutlass conv wrapper
 base on modify cutlass 2.4

GitOrigin-RevId: 49e0565e8a882a455387c95a80a8037c29166a76
---
 .gitattributes                                |   1 +
 dnn/src/common/convolution.cpp                |  37 ++-
 .../conv_bias/cutlass_convolution_wrapper.cu  | 103 ++++---
 ...s_int8_implicit_gemm_cutlass_wrapper.cuinl |  30 +-
 ...4a_ncdiv4hw4_128x128x32_64x32x32_hswish.cu | Bin 1915 -> 1848 bytes
 ...m_dp4a_ncdiv4hw4_128x128x32_64x32x32_id.cu | Bin 1909 -> 1842 bytes
 ...dp4a_ncdiv4hw4_128x128x32_64x32x32_relu.cu | Bin 1913 -> 1846 bytes
 ...p4a_ncdiv4hw4_128x32x32_64x32x32_hswish.cu | Bin 1914 -> 1847 bytes
 ...mm_dp4a_ncdiv4hw4_128x32x32_64x32x32_id.cu | Bin 1908 -> 1841 bytes
 ..._dp4a_ncdiv4hw4_128x32x32_64x32x32_relu.cu | Bin 1912 -> 1845 bytes
 ...p4a_ncdiv4hw4_128x64x32_64x32x32_hswish.cu | Bin 1914 -> 1847 bytes
 ...mm_dp4a_ncdiv4hw4_128x64x32_64x32x32_id.cu | Bin 1908 -> 1841 bytes
 ..._dp4a_ncdiv4hw4_128x64x32_64x32x32_relu.cu | Bin 1912 -> 1845 bytes
 ...4a_ncdiv4hw4_16x128x16_16x128x16_hswish.cu | Bin 1914 -> 1847 bytes
 ...m_dp4a_ncdiv4hw4_16x128x16_16x128x16_id.cu | Bin 1908 -> 1841 bytes
 ...dp4a_ncdiv4hw4_16x128x16_16x128x16_relu.cu | Bin 1912 -> 1845 bytes
 ...m_dp4a_ncdiv4hw4_16x64x8_16x64x8_hswish.cu | Bin 1910 -> 1843 bytes
 ..._gemm_dp4a_ncdiv4hw4_16x64x8_16x64x8_id.cu | Bin 1904 -> 1837 bytes
 ...emm_dp4a_ncdiv4hw4_16x64x8_16x64x8_relu.cu | Bin 1908 -> 1841 bytes
 ...cdiv4hw4_1x1_128x128x32_64x32x32_hswish.cu | Bin 1916 -> 1849 bytes
 ...4a_ncdiv4hw4_1x1_128x128x32_64x32x32_id.cu | Bin 1910 -> 1843 bytes
 ..._ncdiv4hw4_1x1_128x128x32_64x32x32_relu.cu | Bin 1914 -> 1847 bytes
 ...ncdiv4hw4_1x1_128x32x32_64x32x32_hswish.cu | Bin 1915 -> 1848 bytes
 ...p4a_ncdiv4hw4_1x1_128x32x32_64x32x32_id.cu | Bin 1909 -> 1842 bytes
 ...a_ncdiv4hw4_1x1_128x32x32_64x32x32_relu.cu | Bin 1913 -> 1846 bytes
 ...ncdiv4hw4_1x1_128x64x32_64x32x32_hswish.cu | Bin 1915 -> 1848 bytes
 ...p4a_ncdiv4hw4_1x1_128x64x32_64x32x32_id.cu | Bin 1909 -> 1842 bytes
 ...a_ncdiv4hw4_1x1_128x64x32_64x32x32_relu.cu | Bin 1913 -> 1846 bytes
 ...cdiv4hw4_1x1_16x128x16_16x128x16_hswish.cu | Bin 1915 -> 1848 bytes
 ...4a_ncdiv4hw4_1x1_16x128x16_16x128x16_id.cu | Bin 1909 -> 1842 bytes
 ..._ncdiv4hw4_1x1_16x128x16_16x128x16_relu.cu | Bin 1913 -> 1846 bytes
 ...4a_ncdiv4hw4_1x1_16x64x8_16x64x8_hswish.cu | Bin 1911 -> 1844 bytes
 ...m_dp4a_ncdiv4hw4_1x1_16x64x8_16x64x8_id.cu | Bin 1905 -> 1838 bytes
 ...dp4a_ncdiv4hw4_1x1_16x64x8_16x64x8_relu.cu | Bin 1909 -> 1842 bytes
 ...ncdiv4hw4_1x1_32x128x32_32x64x32_hswish.cu | Bin 1915 -> 1848 bytes
 ...p4a_ncdiv4hw4_1x1_32x128x32_32x64x32_id.cu | Bin 1909 -> 1842 bytes
 ...a_ncdiv4hw4_1x1_32x128x32_32x64x32_relu.cu | Bin 1913 -> 1846 bytes
 ..._ncdiv4hw4_1x1_32x32x32_32x32x32_hswish.cu | Bin 1914 -> 1847 bytes
 ...dp4a_ncdiv4hw4_1x1_32x32x32_32x32x32_id.cu | Bin 1908 -> 1841 bytes
 ...4a_ncdiv4hw4_1x1_32x32x32_32x32x32_relu.cu | Bin 1912 -> 1845 bytes
 ..._ncdiv4hw4_1x1_32x64x32_32x64x32_hswish.cu | Bin 1914 -> 1847 bytes
 ...dp4a_ncdiv4hw4_1x1_32x64x32_32x64x32_id.cu | Bin 1908 -> 1841 bytes
 ...4a_ncdiv4hw4_1x1_32x64x32_32x64x32_relu.cu | Bin 1912 -> 1845 bytes
 ...ncdiv4hw4_1x1_64x128x32_64x32x32_hswish.cu | Bin 1915 -> 1848 bytes
 ...p4a_ncdiv4hw4_1x1_64x128x32_64x32x32_id.cu | Bin 1909 -> 1842 bytes
 ...a_ncdiv4hw4_1x1_64x128x32_64x32x32_relu.cu | Bin 1913 -> 1846 bytes
 ..._ncdiv4hw4_1x1_64x32x32_64x32x32_hswish.cu | Bin 1914 -> 1847 bytes
 ...dp4a_ncdiv4hw4_1x1_64x32x32_64x32x32_id.cu | Bin 1908 -> 1841 bytes
 ...4a_ncdiv4hw4_1x1_64x32x32_64x32x32_relu.cu | Bin 1912 -> 1845 bytes
 ..._ncdiv4hw4_1x1_64x64x32_64x32x32_hswish.cu | Bin 1914 -> 1847 bytes
 ...dp4a_ncdiv4hw4_1x1_64x64x32_64x32x32_id.cu | Bin 1908 -> 1841 bytes
 ...4a_ncdiv4hw4_1x1_64x64x32_64x32x32_relu.cu | Bin 1912 -> 1845 bytes
 ...p4a_ncdiv4hw4_32x128x32_32x64x32_hswish.cu | Bin 1914 -> 1847 bytes
 ...mm_dp4a_ncdiv4hw4_32x128x32_32x64x32_id.cu | Bin 1908 -> 1841 bytes
 ..._dp4a_ncdiv4hw4_32x128x32_32x64x32_relu.cu | Bin 1912 -> 1845 bytes
 ...dp4a_ncdiv4hw4_32x32x32_32x32x32_hswish.cu | Bin 1913 -> 1846 bytes
 ...emm_dp4a_ncdiv4hw4_32x32x32_32x32x32_id.cu | Bin 1907 -> 1840 bytes
 ...m_dp4a_ncdiv4hw4_32x32x32_32x32x32_relu.cu | Bin 1911 -> 1844 bytes
 ...dp4a_ncdiv4hw4_32x64x32_32x64x32_hswish.cu | Bin 1913 -> 1846 bytes
 ...emm_dp4a_ncdiv4hw4_32x64x32_32x64x32_id.cu | Bin 1907 -> 1840 bytes
 ...m_dp4a_ncdiv4hw4_32x64x32_32x64x32_relu.cu | Bin 1911 -> 1844 bytes
 ...p4a_ncdiv4hw4_64x128x32_64x32x32_hswish.cu | Bin 1914 -> 1847 bytes
 ...mm_dp4a_ncdiv4hw4_64x128x32_64x32x32_id.cu | Bin 1908 -> 1841 bytes
 ..._dp4a_ncdiv4hw4_64x128x32_64x32x32_relu.cu | Bin 1912 -> 1845 bytes
 ...dp4a_ncdiv4hw4_64x32x32_64x32x32_hswish.cu | Bin 1913 -> 1846 bytes
 ...emm_dp4a_ncdiv4hw4_64x32x32_64x32x32_id.cu | Bin 1907 -> 1840 bytes
 ...m_dp4a_ncdiv4hw4_64x32x32_64x32x32_relu.cu | Bin 1911 -> 1844 bytes
 ...dp4a_ncdiv4hw4_64x64x32_64x32x32_hswish.cu | Bin 1913 -> 1846 bytes
 ...emm_dp4a_ncdiv4hw4_64x64x32_64x32x32_id.cu | Bin 1907 -> 1840 bytes
 ...m_dp4a_ncdiv4hw4_64x64x32_64x32x32_relu.cu | Bin 1911 -> 1844 bytes
 ..._ncdiv32hw32_128x128x32_64x32x32_hswish.cu | Bin 1916 -> 1849 bytes
 ...4hw4_ncdiv32hw32_128x128x32_64x32x32_id.cu | Bin 1910 -> 1843 bytes
 ...w4_ncdiv32hw32_128x128x32_64x32x32_relu.cu | Bin 1914 -> 1847 bytes
 ...4_ncdiv32hw32_128x32x32_64x32x32_hswish.cu | Bin 1915 -> 1848 bytes
 ...v4hw4_ncdiv32hw32_128x32x32_64x32x32_id.cu | Bin 1909 -> 1842 bytes
 ...hw4_ncdiv32hw32_128x32x32_64x32x32_relu.cu | Bin 1913 -> 1846 bytes
 ...4_ncdiv32hw32_128x64x32_64x32x32_hswish.cu | Bin 1915 -> 1848 bytes
 ...v4hw4_ncdiv32hw32_128x64x32_64x32x32_id.cu | Bin 1909 -> 1842 bytes
 ...hw4_ncdiv32hw32_128x64x32_64x32x32_relu.cu | Bin 1913 -> 1846 bytes
 ...iv32hw32_1x1_128x128x32_64x32x32_hswish.cu | Bin 1917 -> 1850 bytes
 ..._ncdiv32hw32_1x1_128x128x32_64x32x32_id.cu | Bin 1911 -> 1844 bytes
 ...cdiv32hw32_1x1_128x128x32_64x32x32_relu.cu | Bin 1915 -> 1848 bytes
 ...div32hw32_1x1_128x32x32_64x32x32_hswish.cu | Bin 1916 -> 1849 bytes
 ...4_ncdiv32hw32_1x1_128x32x32_64x32x32_id.cu | Bin 1910 -> 1843 bytes
 ...ncdiv32hw32_1x1_128x32x32_64x32x32_relu.cu | Bin 1914 -> 1847 bytes
 ...div32hw32_1x1_128x64x32_64x32x32_hswish.cu | Bin 1916 -> 1849 bytes
 ...4_ncdiv32hw32_1x1_128x64x32_64x32x32_id.cu | Bin 1910 -> 1843 bytes
 ...ncdiv32hw32_1x1_128x64x32_64x32x32_relu.cu | Bin 1914 -> 1847 bytes
 ...div32hw32_1x1_32x128x32_32x64x32_hswish.cu | Bin 1916 -> 1849 bytes
 ...4_ncdiv32hw32_1x1_32x128x32_32x64x32_id.cu | Bin 1910 -> 1843 bytes
 ...ncdiv32hw32_1x1_32x128x32_32x64x32_relu.cu | Bin 1914 -> 1847 bytes
 ...cdiv32hw32_1x1_32x32x32_32x32x32_hswish.cu | Bin 1915 -> 1848 bytes
 ...w4_ncdiv32hw32_1x1_32x32x32_32x32x32_id.cu | Bin 1909 -> 1842 bytes
 ..._ncdiv32hw32_1x1_32x32x32_32x32x32_relu.cu | Bin 1913 -> 1846 bytes
 ...cdiv32hw32_1x1_32x64x32_32x64x32_hswish.cu | Bin 1915 -> 1848 bytes
 ...w4_ncdiv32hw32_1x1_32x64x32_32x64x32_id.cu | Bin 1909 -> 1842 bytes
 ..._ncdiv32hw32_1x1_32x64x32_32x64x32_relu.cu | Bin 1913 -> 1846 bytes
 ...div32hw32_1x1_64x128x32_64x32x32_hswish.cu | Bin 1916 -> 1849 bytes
 ...4_ncdiv32hw32_1x1_64x128x32_64x32x32_id.cu | Bin 1910 -> 1843 bytes
 ...ncdiv32hw32_1x1_64x128x32_64x32x32_relu.cu | Bin 1914 -> 1847 bytes
 ...cdiv32hw32_1x1_64x32x32_64x32x32_hswish.cu | Bin 1915 -> 1848 bytes
 ...w4_ncdiv32hw32_1x1_64x32x32_64x32x32_id.cu | Bin 1909 -> 1842 bytes
 ..._ncdiv32hw32_1x1_64x32x32_64x32x32_relu.cu | Bin 1913 -> 1846 bytes
 ...cdiv32hw32_1x1_64x64x32_64x32x32_hswish.cu | Bin 1915 -> 1848 bytes
 ...w4_ncdiv32hw32_1x1_64x64x32_64x32x32_id.cu | Bin 1909 -> 1842 bytes
 ..._ncdiv32hw32_1x1_64x64x32_64x32x32_relu.cu | Bin 1913 -> 1846 bytes
 ...4_ncdiv32hw32_32x128x32_32x64x32_hswish.cu | Bin 1915 -> 1848 bytes
 ...v4hw4_ncdiv32hw32_32x128x32_32x64x32_id.cu | Bin 1909 -> 1842 bytes
 ...hw4_ncdiv32hw32_32x128x32_32x64x32_relu.cu | Bin 1913 -> 1846 bytes
 ...w4_ncdiv32hw32_32x32x32_32x32x32_hswish.cu | Bin 1914 -> 1847 bytes
 ...iv4hw4_ncdiv32hw32_32x32x32_32x32x32_id.cu | Bin 1908 -> 1841 bytes
 ...4hw4_ncdiv32hw32_32x32x32_32x32x32_relu.cu | Bin 1912 -> 1845 bytes
 ...w4_ncdiv32hw32_32x64x32_32x64x32_hswish.cu | Bin 1914 -> 1847 bytes
 ...iv4hw4_ncdiv32hw32_32x64x32_32x64x32_id.cu | Bin 1908 -> 1841 bytes
 ...4hw4_ncdiv32hw32_32x64x32_32x64x32_relu.cu | Bin 1912 -> 1845 bytes
 ...4_ncdiv32hw32_64x128x32_64x32x32_hswish.cu | Bin 1915 -> 1848 bytes
 ...v4hw4_ncdiv32hw32_64x128x32_64x32x32_id.cu | Bin 1909 -> 1842 bytes
 ...hw4_ncdiv32hw32_64x128x32_64x32x32_relu.cu | Bin 1913 -> 1846 bytes
 ...w4_ncdiv32hw32_64x32x32_64x32x32_hswish.cu | Bin 1914 -> 1847 bytes
 ...iv4hw4_ncdiv32hw32_64x32x32_64x32x32_id.cu | Bin 1908 -> 1841 bytes
 ...4hw4_ncdiv32hw32_64x32x32_64x32x32_relu.cu | Bin 1912 -> 1845 bytes
 ...w4_ncdiv32hw32_64x64x32_64x32x32_hswish.cu | Bin 1914 -> 1847 bytes
 ...iv4hw4_ncdiv32hw32_64x64x32_64x32x32_id.cu | Bin 1908 -> 1841 bytes
 ...4hw4_ncdiv32hw32_64x64x32_64x32x32_relu.cu | Bin 1912 -> 1845 bytes
 ...div4hw4_nchw_128x128x32_64x32x32_hswish.cu | Bin 1891 -> 1824 bytes
 ...a_ncdiv4hw4_nchw_128x128x32_64x32x32_id.cu | Bin 1885 -> 1818 bytes
 ...ncdiv4hw4_nchw_128x128x32_64x32x32_relu.cu | Bin 1889 -> 1822 bytes
 ...cdiv4hw4_nchw_128x32x32_64x32x32_hswish.cu | Bin 1890 -> 1823 bytes
 ...4a_ncdiv4hw4_nchw_128x32x32_64x32x32_id.cu | Bin 1884 -> 1817 bytes
 ..._ncdiv4hw4_nchw_128x32x32_64x32x32_relu.cu | Bin 1888 -> 1821 bytes
 ...cdiv4hw4_nchw_128x64x32_64x32x32_hswish.cu | Bin 1890 -> 1823 bytes
 ...4a_ncdiv4hw4_nchw_128x64x32_64x32x32_id.cu | Bin 1884 -> 1817 bytes
 ..._ncdiv4hw4_nchw_128x64x32_64x32x32_relu.cu | Bin 1888 -> 1821 bytes
 ...div4hw4_nchw_16x128x16_16x128x16_hswish.cu | Bin 1890 -> 1823 bytes
 ...a_ncdiv4hw4_nchw_16x128x16_16x128x16_id.cu | Bin 1884 -> 1817 bytes
 ...ncdiv4hw4_nchw_16x128x16_16x128x16_relu.cu | Bin 1888 -> 1821 bytes
 ...a_ncdiv4hw4_nchw_16x64x8_16x64x8_hswish.cu | Bin 1886 -> 1819 bytes
 ..._dp4a_ncdiv4hw4_nchw_16x64x8_16x64x8_id.cu | Bin 1880 -> 1813 bytes
 ...p4a_ncdiv4hw4_nchw_16x64x8_16x64x8_relu.cu | Bin 1884 -> 1817 bytes
 ...hw4_nchw_1x1_128x128x32_64x32x32_hswish.cu | Bin 1892 -> 1825 bytes
 ...div4hw4_nchw_1x1_128x128x32_64x32x32_id.cu | Bin 1886 -> 1819 bytes
 ...v4hw4_nchw_1x1_128x128x32_64x32x32_relu.cu | Bin 1890 -> 1823 bytes
 ...4hw4_nchw_1x1_128x32x32_64x32x32_hswish.cu | Bin 1891 -> 1824 bytes
 ...cdiv4hw4_nchw_1x1_128x32x32_64x32x32_id.cu | Bin 1885 -> 1818 bytes
 ...iv4hw4_nchw_1x1_128x32x32_64x32x32_relu.cu | Bin 1889 -> 1822 bytes
 ...4hw4_nchw_1x1_128x64x32_64x32x32_hswish.cu | Bin 1891 -> 1824 bytes
 ...cdiv4hw4_nchw_1x1_128x64x32_64x32x32_id.cu | Bin 1885 -> 1818 bytes
 ...iv4hw4_nchw_1x1_128x64x32_64x32x32_relu.cu | Bin 1889 -> 1822 bytes
 ...hw4_nchw_1x1_16x128x16_16x128x16_hswish.cu | Bin 1891 -> 1824 bytes
 ...div4hw4_nchw_1x1_16x128x16_16x128x16_id.cu | Bin 1885 -> 1818 bytes
 ...v4hw4_nchw_1x1_16x128x16_16x128x16_relu.cu | Bin 1889 -> 1822 bytes
 ...div4hw4_nchw_1x1_16x64x8_16x64x8_hswish.cu | Bin 1887 -> 1820 bytes
 ...a_ncdiv4hw4_nchw_1x1_16x64x8_16x64x8_id.cu | Bin 1881 -> 1814 bytes
 ...ncdiv4hw4_nchw_1x1_16x64x8_16x64x8_relu.cu | Bin 1885 -> 1818 bytes
 ...4hw4_nchw_1x1_32x128x32_32x64x32_hswish.cu | Bin 1891 -> 1824 bytes
 ...cdiv4hw4_nchw_1x1_32x128x32_32x64x32_id.cu | Bin 1885 -> 1818 bytes
 ...iv4hw4_nchw_1x1_32x128x32_32x64x32_relu.cu | Bin 1889 -> 1822 bytes
 ...v4hw4_nchw_1x1_32x32x32_32x32x32_hswish.cu | Bin 1890 -> 1823 bytes
 ...ncdiv4hw4_nchw_1x1_32x32x32_32x32x32_id.cu | Bin 1884 -> 1817 bytes
 ...div4hw4_nchw_1x1_32x32x32_32x32x32_relu.cu | Bin 1888 -> 1821 bytes
 ...v4hw4_nchw_1x1_32x64x32_32x64x32_hswish.cu | Bin 1890 -> 1823 bytes
 ...ncdiv4hw4_nchw_1x1_32x64x32_32x64x32_id.cu | Bin 1884 -> 1817 bytes
 ...div4hw4_nchw_1x1_32x64x32_32x64x32_relu.cu | Bin 1888 -> 1821 bytes
 ...4hw4_nchw_1x1_64x128x32_64x32x32_hswish.cu | Bin 1891 -> 1824 bytes
 ...cdiv4hw4_nchw_1x1_64x128x32_64x32x32_id.cu | Bin 1885 -> 1818 bytes
 ...iv4hw4_nchw_1x1_64x128x32_64x32x32_relu.cu | Bin 1889 -> 1822 bytes
 ...v4hw4_nchw_1x1_64x32x32_64x32x32_hswish.cu | Bin 1890 -> 1823 bytes
 ...ncdiv4hw4_nchw_1x1_64x32x32_64x32x32_id.cu | Bin 1884 -> 1817 bytes
 ...div4hw4_nchw_1x1_64x32x32_64x32x32_relu.cu | Bin 1888 -> 1821 bytes
 ...v4hw4_nchw_1x1_64x64x32_64x32x32_hswish.cu | Bin 1890 -> 1823 bytes
 ...ncdiv4hw4_nchw_1x1_64x64x32_64x32x32_id.cu | Bin 1884 -> 1817 bytes
 ...div4hw4_nchw_1x1_64x64x32_64x32x32_relu.cu | Bin 1888 -> 1821 bytes
 ...cdiv4hw4_nchw_32x128x32_32x64x32_hswish.cu | Bin 1890 -> 1823 bytes
 ...4a_ncdiv4hw4_nchw_32x128x32_32x64x32_id.cu | Bin 1884 -> 1817 bytes
 ..._ncdiv4hw4_nchw_32x128x32_32x64x32_relu.cu | Bin 1888 -> 1821 bytes
 ...ncdiv4hw4_nchw_32x32x32_32x32x32_hswish.cu | Bin 1889 -> 1822 bytes
 ...p4a_ncdiv4hw4_nchw_32x32x32_32x32x32_id.cu | Bin 1883 -> 1816 bytes
 ...a_ncdiv4hw4_nchw_32x32x32_32x32x32_relu.cu | Bin 1887 -> 1820 bytes
 ...ncdiv4hw4_nchw_32x64x32_32x64x32_hswish.cu | Bin 1889 -> 1822 bytes
 ...p4a_ncdiv4hw4_nchw_32x64x32_32x64x32_id.cu | Bin 1883 -> 1816 bytes
 ...a_ncdiv4hw4_nchw_32x64x32_32x64x32_relu.cu | Bin 1887 -> 1820 bytes
 ...cdiv4hw4_nchw_64x128x32_64x32x32_hswish.cu | Bin 1890 -> 1823 bytes
 ...4a_ncdiv4hw4_nchw_64x128x32_64x32x32_id.cu | Bin 1884 -> 1817 bytes
 ..._ncdiv4hw4_nchw_64x128x32_64x32x32_relu.cu | Bin 1888 -> 1821 bytes
 ...ncdiv4hw4_nchw_64x32x32_64x32x32_hswish.cu | Bin 1889 -> 1822 bytes
 ...p4a_ncdiv4hw4_nchw_64x32x32_64x32x32_id.cu | Bin 1883 -> 1816 bytes
 ...a_ncdiv4hw4_nchw_64x32x32_64x32x32_relu.cu | Bin 1887 -> 1820 bytes
 ...ncdiv4hw4_nchw_64x64x32_64x32x32_hswish.cu | Bin 1889 -> 1822 bytes
 ...p4a_ncdiv4hw4_nchw_64x64x32_64x32x32_id.cu | Bin 1883 -> 1816 bytes
 ...a_ncdiv4hw4_nchw_64x64x32_64x32x32_relu.cu | Bin 1887 -> 1820 bytes
 ..._ncdiv32hw32_128x128x64_64x64x64_hswish.cu | Bin 1924 -> 1857 bytes
 ...imma_ncdiv32hw32_128x128x64_64x64x64_id.cu | Bin 1918 -> 1851 bytes
 ...ma_ncdiv32hw32_128x128x64_64x64x64_relu.cu | Bin 1922 -> 1855 bytes
 ..._ncdiv32hw32_128x256x64_64x64x64_hswish.cu | Bin 1924 -> 1857 bytes
 ...imma_ncdiv32hw32_128x256x64_64x64x64_id.cu | Bin 1918 -> 1851 bytes
 ...ma_ncdiv32hw32_128x256x64_64x64x64_relu.cu | Bin 1922 -> 1855 bytes
 ...a_ncdiv32hw32_128x64x64_64x32x64_hswish.cu | Bin 1923 -> 1856 bytes
 ..._imma_ncdiv32hw32_128x64x64_64x32x64_id.cu | Bin 1917 -> 1850 bytes
 ...mma_ncdiv32hw32_128x64x64_64x32x64_relu.cu | Bin 1921 -> 1854 bytes
 ...iv32hw32_1x1_128x128x64_64x64x64_hswish.cu | Bin 1925 -> 1858 bytes
 ..._ncdiv32hw32_1x1_128x128x64_64x64x64_id.cu | Bin 1919 -> 1852 bytes
 ...cdiv32hw32_1x1_128x128x64_64x64x64_relu.cu | Bin 1923 -> 1856 bytes
 ...iv32hw32_1x1_128x256x64_64x64x64_hswish.cu | Bin 1925 -> 1858 bytes
 ..._ncdiv32hw32_1x1_128x256x64_64x64x64_id.cu | Bin 1919 -> 1852 bytes
 ...cdiv32hw32_1x1_128x256x64_64x64x64_relu.cu | Bin 1923 -> 1856 bytes
 ...div32hw32_1x1_128x64x64_64x32x64_hswish.cu | Bin 1924 -> 1857 bytes
 ...a_ncdiv32hw32_1x1_128x64x64_64x32x64_id.cu | Bin 1918 -> 1851 bytes
 ...ncdiv32hw32_1x1_128x64x64_64x32x64_relu.cu | Bin 1922 -> 1855 bytes
 ...iv32hw32_1x1_256x128x64_64x64x64_hswish.cu | Bin 1925 -> 1858 bytes
 ..._ncdiv32hw32_1x1_256x128x64_64x64x64_id.cu | Bin 1919 -> 1852 bytes
 ...cdiv32hw32_1x1_256x128x64_64x64x64_relu.cu | Bin 1923 -> 1856 bytes
 ...cdiv32hw32_1x1_32x64x64_32x16x64_hswish.cu | Bin 1923 -> 1856 bytes
 ...ma_ncdiv32hw32_1x1_32x64x64_32x16x64_id.cu | Bin 1917 -> 1850 bytes
 ..._ncdiv32hw32_1x1_32x64x64_32x16x64_relu.cu | Bin 1921 -> 1854 bytes
 ...div32hw32_1x1_64x128x64_32x64x64_hswish.cu | Bin 1924 -> 1857 bytes
 ...a_ncdiv32hw32_1x1_64x128x64_32x64x64_id.cu | Bin 1918 -> 1851 bytes
 ...ncdiv32hw32_1x1_64x128x64_32x64x64_relu.cu | Bin 1922 -> 1855 bytes
 ...cdiv32hw32_1x1_64x64x64_32x32x64_hswish.cu | Bin 1923 -> 1856 bytes
 ...ma_ncdiv32hw32_1x1_64x64x64_32x32x64_id.cu | Bin 1917 -> 1850 bytes
 ..._ncdiv32hw32_1x1_64x64x64_32x32x64_relu.cu | Bin 1921 -> 1854 bytes
 ..._ncdiv32hw32_256x128x64_64x64x64_hswish.cu | Bin 1924 -> 1857 bytes
 ...imma_ncdiv32hw32_256x128x64_64x64x64_id.cu | Bin 1918 -> 1851 bytes
 ...ma_ncdiv32hw32_256x128x64_64x64x64_relu.cu | Bin 1922 -> 1855 bytes
 ...ma_ncdiv32hw32_32x64x64_32x16x64_hswish.cu | Bin 1922 -> 1855 bytes
 ...m_imma_ncdiv32hw32_32x64x64_32x16x64_id.cu | Bin 1916 -> 1849 bytes
 ...imma_ncdiv32hw32_32x64x64_32x16x64_relu.cu | Bin 1920 -> 1853 bytes
 ...a_ncdiv32hw32_64x128x64_32x64x64_hswish.cu | Bin 1923 -> 1856 bytes
 ..._imma_ncdiv32hw32_64x128x64_32x64x64_id.cu | Bin 1917 -> 1850 bytes
 ...mma_ncdiv32hw32_64x128x64_32x64x64_relu.cu | Bin 1921 -> 1854 bytes
 ...ma_ncdiv32hw32_64x64x64_32x32x64_hswish.cu | Bin 1922 -> 1855 bytes
 ...m_imma_ncdiv32hw32_64x64x64_32x32x64_id.cu | Bin 1916 -> 1849 bytes
 ...imma_ncdiv32hw32_64x64x64_32x32x64_relu.cu | Bin 1920 -> 1853 bytes
 ...32_ncdiv4hw4_128x128x64_64x64x64_hswish.cu | Bin 1923 -> 1856 bytes
 ...32hw32_ncdiv4hw4_128x128x64_64x64x64_id.cu | Bin 1917 -> 1850 bytes
 ...hw32_ncdiv4hw4_128x128x64_64x64x64_relu.cu | Bin 1921 -> 1854 bytes
 ...32_ncdiv4hw4_128x256x64_64x64x64_hswish.cu | Bin 1923 -> 1856 bytes
 ...32hw32_ncdiv4hw4_128x256x64_64x64x64_id.cu | Bin 1917 -> 1850 bytes
 ...hw32_ncdiv4hw4_128x256x64_64x64x64_relu.cu | Bin 1921 -> 1854 bytes
 ...w32_ncdiv4hw4_128x64x64_64x32x64_hswish.cu | Bin 1922 -> 1855 bytes
 ...v32hw32_ncdiv4hw4_128x64x64_64x32x64_id.cu | Bin 1916 -> 1849 bytes
 ...2hw32_ncdiv4hw4_128x64x64_64x32x64_relu.cu | Bin 1920 -> 1853 bytes
 ...cdiv4hw4_1x1_128x128x64_64x64x64_hswish.cu | Bin 1924 -> 1857 bytes
 ...32_ncdiv4hw4_1x1_128x128x64_64x64x64_id.cu | Bin 1918 -> 1851 bytes
 ..._ncdiv4hw4_1x1_128x128x64_64x64x64_relu.cu | Bin 1922 -> 1855 bytes
 ...cdiv4hw4_1x1_128x256x64_64x64x64_hswish.cu | Bin 1924 -> 1857 bytes
 ...32_ncdiv4hw4_1x1_128x256x64_64x64x64_id.cu | Bin 1918 -> 1851 bytes
 ..._ncdiv4hw4_1x1_128x256x64_64x64x64_relu.cu | Bin 1922 -> 1855 bytes
 ...ncdiv4hw4_1x1_128x64x64_64x32x64_hswish.cu | Bin 1923 -> 1856 bytes
 ...w32_ncdiv4hw4_1x1_128x64x64_64x32x64_id.cu | Bin 1917 -> 1850 bytes
 ...2_ncdiv4hw4_1x1_128x64x64_64x32x64_relu.cu | Bin 1921 -> 1854 bytes
 ...cdiv4hw4_1x1_256x128x64_64x64x64_hswish.cu | Bin 1924 -> 1857 bytes
 ...32_ncdiv4hw4_1x1_256x128x64_64x64x64_id.cu | Bin 1918 -> 1851 bytes
 ..._ncdiv4hw4_1x1_256x128x64_64x64x64_relu.cu | Bin 1922 -> 1855 bytes
 ..._ncdiv4hw4_1x1_32x64x64_16x32x64_hswish.cu | Bin 1922 -> 1855 bytes
 ...hw32_ncdiv4hw4_1x1_32x64x64_16x32x64_id.cu | Bin 1916 -> 1849 bytes
 ...32_ncdiv4hw4_1x1_32x64x64_16x32x64_relu.cu | Bin 1920 -> 1853 bytes
 ...ncdiv4hw4_1x1_64x128x64_32x64x64_hswish.cu | Bin 1923 -> 1856 bytes
 ...w32_ncdiv4hw4_1x1_64x128x64_32x64x64_id.cu | Bin 1917 -> 1850 bytes
 ...2_ncdiv4hw4_1x1_64x128x64_32x64x64_relu.cu | Bin 1921 -> 1854 bytes
 ..._ncdiv4hw4_1x1_64x64x64_32x32x64_hswish.cu | Bin 1922 -> 1855 bytes
 ...hw32_ncdiv4hw4_1x1_64x64x64_32x32x64_id.cu | Bin 1916 -> 1849 bytes
 ...32_ncdiv4hw4_1x1_64x64x64_32x32x64_relu.cu | Bin 1920 -> 1853 bytes
 ...32_ncdiv4hw4_256x128x64_64x64x64_hswish.cu | Bin 1923 -> 1856 bytes
 ...32hw32_ncdiv4hw4_256x128x64_64x64x64_id.cu | Bin 1917 -> 1850 bytes
 ...hw32_ncdiv4hw4_256x128x64_64x64x64_relu.cu | Bin 1921 -> 1854 bytes
 ...hw32_ncdiv4hw4_32x64x64_16x32x64_hswish.cu | Bin 1921 -> 1854 bytes
 ...iv32hw32_ncdiv4hw4_32x64x64_16x32x64_id.cu | Bin 1915 -> 1848 bytes
 ...32hw32_ncdiv4hw4_32x64x64_16x32x64_relu.cu | Bin 1919 -> 1852 bytes
 ...w32_ncdiv4hw4_64x128x64_32x64x64_hswish.cu | Bin 1922 -> 1855 bytes
 ...v32hw32_ncdiv4hw4_64x128x64_32x64x64_id.cu | Bin 1916 -> 1849 bytes
 ...2hw32_ncdiv4hw4_64x128x64_32x64x64_relu.cu | Bin 1920 -> 1853 bytes
 ...hw32_ncdiv4hw4_64x64x64_32x32x64_hswish.cu | Bin 1921 -> 1854 bytes
 ...iv32hw32_ncdiv4hw4_64x64x64_32x32x64_id.cu | Bin 1915 -> 1848 bytes
 ...32hw32_ncdiv4hw4_64x64x64_32x32x64_relu.cu | Bin 1919 -> 1852 bytes
 .../cuda/convolution/backward_data/algo.cpp   |  97 ++++---
 dnn/src/cuda/convolution/backward_data/algo.h |  51 +++-
 .../convolution/backward_data/chanwise.cpp    |  10 +-
 .../backward_data/chanwise_small.cpp          |  33 +--
 .../cutlass_deconvolution_wrapper.cu          | 100 +++++++
 .../cutlass_deconvolution_wrapper.cuh         |  44 +++
 .../backward_data/deconv_int8_helper.cu       |  76 +++++
 .../backward_data/deconv_int8_helper.cuh      |  27 ++
 .../convolution/backward_data/group_conv.cpp  |  46 +--
 .../implicit_gemm_int8_nchw4_dp4a.cpp         | 127 +++++++++
 ...v_int8_implicit_gemm_cutlass_wrapper.cuinl |  62 ++++
 ...m_dp4a_ncdiv4hw4_16x128x16_16x128x16_id.cu | Bin 0 -> 1837 bytes
 ...mm_dp4a_ncdiv4hw4_16x128x16_16x64x16_id.cu | Bin 0 -> 1836 bytes
 ..._gemm_dp4a_ncdiv4hw4_16x64x8_16x64x8_id.cu | Bin 0 -> 1833 bytes
 ...mm_dp4a_ncdiv4hw4_32x128x32_32x64x32_id.cu | Bin 0 -> 1837 bytes
 ...mm_dp4a_ncdiv4hw4_64x128x32_64x32x32_id.cu | Bin 0 -> 1837 bytes
 dnn/src/cuda/convolution/opr_impl.cpp         |  86 +++---
 dnn/src/cuda/convolution/opr_impl.h           |   1 +
 dnn/test/common/convolution.cpp               | 172 +++++------
 dnn/test/common/convolution.h                 |   4 +-
 dnn/test/cuda/convolution.cpp                 | 267 +++++++++++-------
 src/opr/test/dnn/convolution.cpp              |  45 +++
 296 files changed, 1024 insertions(+), 395 deletions(-)
 create mode 100644 dnn/src/cuda/convolution/backward_data/cutlass_deconvolution_wrapper.cu
 create mode 100644 dnn/src/cuda/convolution/backward_data/cutlass_deconvolution_wrapper.cuh
 create mode 100644 dnn/src/cuda/convolution/backward_data/deconv_int8_helper.cu
 create mode 100644 dnn/src/cuda/convolution/backward_data/deconv_int8_helper.cuh
 create mode 100644 dnn/src/cuda/convolution/backward_data/implicit_gemm_int8_nchw4_dp4a.cpp
 create mode 100644 dnn/src/cuda/convolution/backward_data/int8/deconv_int8_implicit_gemm_cutlass_wrapper.cuinl
 create mode 100644 dnn/src/cuda/convolution/backward_data/int8/kimpl/deconv_int8_implicit_gemm_dp4a_ncdiv4hw4_16x128x16_16x128x16_id.cu
 create mode 100644 dnn/src/cuda/convolution/backward_data/int8/kimpl/deconv_int8_implicit_gemm_dp4a_ncdiv4hw4_16x128x16_16x64x16_id.cu
 create mode 100644 dnn/src/cuda/convolution/backward_data/int8/kimpl/deconv_int8_implicit_gemm_dp4a_ncdiv4hw4_16x64x8_16x64x8_id.cu
 create mode 100644 dnn/src/cuda/convolution/backward_data/int8/kimpl/deconv_int8_implicit_gemm_dp4a_ncdiv4hw4_32x128x32_32x64x32_id.cu
 create mode 100644 dnn/src/cuda/convolution/backward_data/int8/kimpl/deconv_int8_implicit_gemm_dp4a_ncdiv4hw4_64x128x32_64x32x32_id.cu

diff --git a/.gitattributes b/.gitattributes
index be56319f3..07945c657 100644
--- a/.gitattributes
+++ b/.gitattributes
@@ -5,6 +5,7 @@ dnn/src/cuda/conv_bias/int8_imma/kimpl/* binary
 dnn/src/cuda/batch_conv_bias/int8/kimpl/* binary
 dnn/src/cuda/matrix_mul/fp32_simt/kimpl/* binary
 dnn/src/cuda/sass/prebuilt/map_defs.cpp binary
+dnn/src/cuda/convolution/backward_data/int8/kimpl/* binary
 tools/mlir/mlir-tblgen filter=lfs diff=lfs merge=lfs -text
 *.caffemodel filter=lfs diff=lfs merge=lfs -text
 imperative/python/test/integration/data/*.mge filter=lfs diff=lfs merge=lfs -text
diff --git a/dnn/src/common/convolution.cpp b/dnn/src/common/convolution.cpp
index e2c7aa916..b23f4f83e 100644
--- a/dnn/src/common/convolution.cpp
+++ b/dnn/src/common/convolution.cpp
@@ -46,7 +46,7 @@ void make_canonized_filter_meta_nchw_nhwc(
         size_t src_ndim, const TensorLayout& filter, const Param& param,
         typename ConvolutionBase<Parameter>::CanonizedFilterMeta& ret) {
     megdnn_assert(param.format == Param::Format::NCHW ||
-                  param.format == Param::Format::NHWC );
+                  param.format == Param::Format::NHWC);
     auto img_ndim = src_ndim - 2;
     size_t flt_start, flt_spatial_start, ocpg_pos, icpg_pos;
     if (param.sparse == Param::Sparse::DENSE) {
@@ -320,8 +320,8 @@ void make_canonized_filter_meta_nchwxx(
                           img_ndim, filter.ndim);
             megdnn_assert((filter[filter.ndim - 1] == pack_size &&
                            filter[filter.ndim - 2] == pack_size) ||
-                           (filter[filter.ndim - 1] == 2 * pack_size &&
-                            filter[filter.ndim - 2] == 2 * pack_size),
+                                  (filter[filter.ndim - 1] == 2 * pack_size &&
+                                   filter[filter.ndim - 2] == 2 * pack_size),
                           "last 2 dim of filter must be %zu, but got %zu, %zu",
                           pack_size, filter[filter.ndim - 2],
                           filter[filter.ndim - 1]);
@@ -684,7 +684,8 @@ ConvolutionBase<Parameter>::deduce_layout_fwd(const TensorLayout& src,
         }
         if (param().format == Param::Format::NCHW44 ||
             param().format == Param::Format::NCHW44_DOT) {
-            //!support nchw44 filter change to 88 for int8 winogradf23_88 using MK8 mamtul
+            //! support nchw44 filter change to 88 for int8 winogradf23_88 using
+            //! MK8 mamtul
             megdnn_assert((src.ndim == 4 && filter.ndim == 5 &&
                            filter[filter.ndim - 1] == 4) ||
                                   (src.ndim == 5 &&
@@ -716,7 +717,7 @@ ConvolutionBase<Parameter>::deduce_layout_fwd(const TensorLayout& src,
                   "currently only convolution on 2D image is supported");
     auto cflt = make_canonized_filter_meta(src.ndim, filter);
     if (param().format == Param::Format::NCHW ||
-        param().format == Param::Format::NHWC ) {
+        param().format == Param::Format::NHWC) {
         size_t src_or_dst_c_pos = 0;
         size_t src_or_dst_spatial_start = 0;
         if (param().format == Param::Format::NCHW) {
@@ -790,7 +791,7 @@ ConvolutionBase<Parameter>::deduce_layout_fwd(const TensorLayout& src,
         dst[3] = infer_conv_shape(src[3], cflt.dilated_spatial[1],
                                   cflt.stride[1], cflt.padding[1]);
         dst[4] = 32;
-    } else if (param().format == Param::Format::NCHW88 ) {
+    } else if (param().format == Param::Format::NCHW88) {
         megdnn_assert(src.ndim == 5 || (src.ndim == 4 && src[1] <= 8),
                       "invalid src ndim for NCHW88, expected=5 or 4, got=%zu",
                       src.ndim);
@@ -1042,10 +1043,10 @@ void ConvolutionBackwardData::deduce_dtype(DType filter, DType diff,
     }
     megdnn_assert(param().compute_mode != Param::ComputeMode::FLOAT32
 #if !MEGDNN_DISABLE_FLOAT16
-                          || filter.enumv() == DTypeEnum::Float16
-                          || filter.enumv() == DTypeEnum::BFloat16
+                          || filter.enumv() == DTypeEnum::Float16 ||
+                          filter.enumv() == DTypeEnum::BFloat16
 #endif
-                          ,
+                  ,
                   "ComputeMode::FLOAT32 is only available for Float16/BFloat16 "
                   "input / output.");
 }
@@ -1096,6 +1097,24 @@ void ConvolutionBackwardData::deduce_layout(const TensorLayout& filter,
                     diff[i + src_or_dst_spatial_start], cflt.dilated_spatial[i],
                     cflt.stride[i], cflt.padding[i]);
         }
+    } else if (param().format == Param::Format::NCHW4) {
+        megdnn_assert(diff.ndim == 5,
+                      "valid diff ndim for NCHW4, expected=5, got=%zu",
+                      diff.ndim);
+        megdnn_assert(cflt.group == 1, "%s", errmsg().c_str());
+        megdnn_assert(cflt.ocpg * cflt.group == diff[1] * 4, "%s",
+                      errmsg().c_str());
+        grad.ndim = diff.ndim;
+        grad[0] = diff[0];
+        auto ic = cflt.icpg * cflt.group;
+        megdnn_assert(ic % 4 == 0);
+        grad[1] = ic / 4;
+        grad[2] = deduce(diff[2], cflt.dilated_spatial[0], cflt.stride[0],
+                         cflt.padding[0]);
+        grad[3] = deduce(diff[3], cflt.dilated_spatial[1], cflt.stride[1],
+                         cflt.padding[1]);
+        megdnn_assert(diff[4] == 4);
+        grad[4] = 4;
     } else {
         megdnn_assert(param().format == Param::Format::NHWCD4);
         megdnn_assert(diff.ndim == 5,
diff --git a/dnn/src/cuda/conv_bias/cutlass_convolution_wrapper.cu b/dnn/src/cuda/conv_bias/cutlass_convolution_wrapper.cu
index 667b5771d..61921a219 100644
--- a/dnn/src/cuda/conv_bias/cutlass_convolution_wrapper.cu
+++ b/dnn/src/cuda/conv_bias/cutlass_convolution_wrapper.cu
@@ -62,22 +62,21 @@ void megdnn::cuda::cutlass_wrapper::
                                          threadblock_k_>;                      \
         using WarpShape = cutlass::gemm::GemmShape<warp_m_, warp_n_, warp_k_>; \
         using InstructionShape = cutlass::gemm::GemmShape<8, 8, 16>;           \
-        using Convolution = cutlass::convolution::device::Convolution<         \
+        using Convolution = cutlass::conv::device::Convolution<                \
                 int8_t, cutlass::layout::TensorNCxHWx<32>, int8_t,             \
                 cutlass::layout::TensorCxRSKx<32>, ElementOutput,              \
                 cutlass::layout::TensorNCxHWx<32>, int32_t,                    \
                 cutlass::layout::TensorNCxHWx<32>, int32_t,                    \
-                cutlass::convolution::ConvType::kConvolution,                  \
+                cutlass::conv::ConvType::kConvolution,                         \
                 cutlass::arch::OpClassTensorOp, cutlass::arch::Sm75,           \
                 ThreadBlockShape, WarpShape, InstructionShape, EpilogueOp,     \
-                cutlass::convolution::threadblock::                            \
-                        ConvolutionNCxHWxThreadblockSwizzle<                   \
-                                cutlass::convolution::ConvType::kConvolution>, \
+                cutlass::conv::threadblock::                                   \
+                        ConvolutionFpropNCxHWxThreadblockSwizzle,              \
                 2, 16, 16, NeedLoadFromConstMem>;                              \
-        typename Convolution::ConvolutionParameter conv_param{                 \
-                param.n,  param.ci, param.co, param.hi, param.wi,              \
-                param.fh, param.fw, param.ho, param.wo, param.sh,              \
-                param.sw, param.ph, param.pw, 1,        1};                    \
+        typename Convolution::ConvolutionParameter conv_param(                 \
+                param.n, param.hi, param.wi, param.ci, param.co, param.fh,     \
+                param.fw, param.ho, param.wo, param.ph, param.pw, param.sh,    \
+                param.sw, 1, 1, cutlass::conv::Mode::kCrossCorrelation);       \
         return cutlass_convolution_wrapper<Convolution>(                       \
                 d_src, d_filter, d_bias, d_z, d_dst, workspace, conv_param,    \
                 epilogue, stream);                                             \
@@ -186,22 +185,21 @@ void megdnn::cuda::cutlass_wrapper::
                                          threadblock_k_>;                      \
         using WarpShape = cutlass::gemm::GemmShape<warp_m_, warp_n_, warp_k_>; \
         using InstructionShape = cutlass::gemm::GemmShape<8, 8, 16>;           \
-        using Convolution = cutlass::convolution::device::Convolution<         \
+        using Convolution = cutlass::conv::device::Convolution<                \
                 int8_t, cutlass::layout::TensorNCxHWx<32>, int8_t,             \
                 cutlass::layout::TensorCxRSKx<32>, ElementOutput,              \
                 cutlass::layout::TensorNCxHWx<4>, int32_t,                     \
                 cutlass::layout::TensorNCxHWx<4>, int32_t,                     \
-                cutlass::convolution::ConvType::kConvolution,                  \
+                cutlass::conv::ConvType::kConvolution,                         \
                 cutlass::arch::OpClassTensorOp, cutlass::arch::Sm75,           \
                 ThreadBlockShape, WarpShape, InstructionShape, EpilogueOp,     \
-                cutlass::convolution::threadblock::                            \
-                        ConvolutionNCxHWxThreadblockSwizzle<                   \
-                                cutlass::convolution::ConvType::kConvolution>, \
+                cutlass::conv::threadblock::                                   \
+                        ConvolutionFpropNCxHWxThreadblockSwizzle,              \
                 2, 16, 16, NeedLoadFromConstMem>;                              \
-        typename Convolution::ConvolutionParameter conv_param{                 \
-                param.n,  param.ci, param.co, param.hi, param.wi,              \
-                param.fh, param.fw, param.ho, param.wo, param.sh,              \
-                param.sw, param.ph, param.pw, 1,        1};                    \
+        typename Convolution::ConvolutionParameter conv_param(                 \
+                param.n, param.hi, param.wi, param.ci, param.co, param.fh,     \
+                param.fw, param.ho, param.wo, param.ph, param.pw, param.sh,    \
+                param.sw, 1, 1, cutlass::conv::Mode::kCrossCorrelation);       \
         return cutlass_convolution_wrapper<Convolution>(                       \
                 d_src, d_filter, d_bias, d_z, d_dst, workspace, conv_param,    \
                 epilogue, stream);                                             \
@@ -311,22 +309,21 @@ void megdnn::cuda::cutlass_wrapper::
                                          threadblock_k_>;                      \
         using WarpShape = cutlass::gemm::GemmShape<warp_m_, warp_n_, warp_k_>; \
         using InstructionShape = cutlass::gemm::GemmShape<1, 1, 4>;            \
-        using Convolution = cutlass::convolution::device::Convolution<         \
+        using Convolution = cutlass::conv::device::Convolution<                \
                 int8_t, cutlass::layout::TensorNCxHWx<4>, int8_t,              \
                 cutlass::layout::TensorCxRSKx<4>, ElementOutput,               \
                 cutlass::layout::TensorNCxHWx<4>, int32_t,                     \
                 cutlass::layout::TensorNCxHWx<4>, int32_t,                     \
-                cutlass::convolution::ConvType::kConvolution,                  \
+                cutlass::conv::ConvType::kConvolution,                         \
                 cutlass::arch::OpClassSimt, cutlass::arch::Sm61,               \
                 ThreadBlockShape, WarpShape, InstructionShape, EpilogueOp,     \
-                cutlass::convolution::threadblock::                            \
-                        ConvolutionNCxHWxThreadblockSwizzle<                   \
-                                cutlass::convolution::ConvType::kConvolution>, \
+                cutlass::conv::threadblock::                                   \
+                        ConvolutionFpropNCxHWxThreadblockSwizzle,              \
                 stage_, 4, aligned_, NeedLoadFromConstMem>;                    \
-        typename Convolution::ConvolutionParameter conv_param{                 \
-                param.n,  param.ci, param.co, param.hi, param.wi,              \
-                param.fh, param.fw, param.ho, param.wo, param.sh,              \
-                param.sw, param.ph, param.pw, 1,        1};                    \
+        typename Convolution::ConvolutionParameter conv_param(                 \
+                param.n, param.hi, param.wi, param.ci, param.co, param.fh,     \
+                param.fw, param.ho, param.wo, param.ph, param.pw, param.sh,    \
+                param.sw, 1, 1, cutlass::conv::Mode::kCrossCorrelation);       \
         return cutlass_convolution_wrapper<Convolution>(                       \
                 d_src, d_filter, d_bias, d_z, d_dst, workspace, conv_param,    \
                 epilogue, stream);                                             \
@@ -441,23 +438,22 @@ void megdnn::cuda::cutlass_wrapper::
                                          threadblock_k_>;                      \
         using WarpShape = cutlass::gemm::GemmShape<warp_m_, warp_n_, warp_k_>; \
         using InstructionShape = cutlass::gemm::GemmShape<1, 1, 4>;            \
-        using Convolution = cutlass::convolution::device::Convolution<         \
+        using Convolution = cutlass::conv::device::Convolution<                \
                 int8_t, cutlass::layout::TensorNCxHWx<4>, int8_t,              \
                 cutlass::layout::TensorCxRSKx<4>, ElementOutput,               \
                 cutlass::layout::TensorNCHW, float,                            \
                 cutlass::layout::TensorNCHW, int32_t,                          \
-                cutlass::convolution::ConvType::kConvolution,                  \
+                cutlass::conv::ConvType::kConvolution,                         \
                 cutlass::arch::OpClassSimt, cutlass::arch::Sm61,               \
                 ThreadBlockShape, WarpShape, InstructionShape, EpilogueOp,     \
-                cutlass::convolution::threadblock::                            \
-                        ConvolutionNCxHWxThreadblockSwizzle<                   \
-                                cutlass::convolution::ConvType::kConvolution>, \
+                cutlass::conv::threadblock::                                   \
+                        ConvolutionFpropNCxHWxThreadblockSwizzle,              \
                 stages_, 4, aligned_, NeedLoadFromConstMem,                    \
                 cutlass::arch::OpMultiplyAdd>;                                 \
-        typename Convolution::ConvolutionParameter conv_param{                 \
-                param.n,  param.ci, param.co, param.hi, param.wi,              \
-                param.fh, param.fw, param.ho, param.wo, param.sh,              \
-                param.sw, param.ph, param.pw, 1,        1};                    \
+        typename Convolution::ConvolutionParameter conv_param(                 \
+                param.n, param.hi, param.wi, param.ci, param.co, param.fh,     \
+                param.fw, param.ho, param.wo, param.ph, param.pw, param.sh,    \
+                param.sw, 1, 1, cutlass::conv::Mode::kCrossCorrelation);       \
         return cutlass_convolution_wrapper<Convolution>(                       \
                 d_src, d_filter, d_bias, d_z, d_dst, workspace, conv_param,    \
                 epilogue, stream);                                             \
@@ -572,36 +568,35 @@ void megdnn::cuda::cutlass_wrapper::
                                          threadblock_k_>;                      \
         using WarpShape = cutlass::gemm::GemmShape<warp_m_, warp_n_, warp_k_>; \
         using InstructionShape = cutlass::gemm::GemmShape<1, 1, 4>;            \
-        using Convolution = cutlass::convolution::device::Convolution<         \
+        using Convolution = cutlass::conv::device::Convolution<                \
                 int8_t, cutlass::layout::TensorNCxHWx<4>, int8_t,              \
                 cutlass::layout::TensorCxRSKx<4>, ElementOutput,               \
                 cutlass::layout::TensorNCxHWx<32>, int32_t,                    \
                 cutlass::layout::TensorNCxHWx<32>, int32_t,                    \
-                cutlass::convolution::ConvType::kConvolution,                  \
+                cutlass::conv::ConvType::kConvolution,                         \
                 cutlass::arch::OpClassSimt, cutlass::arch::Sm61,               \
                 ThreadBlockShape, WarpShape, InstructionShape, EpilogueOp,     \
-                cutlass::convolution::threadblock::                            \
-                        ConvolutionNCxHWxThreadblockSwizzle<                   \
-                                cutlass::convolution::ConvType::kConvolution>, \
+                cutlass::conv::threadblock::                                   \
+                        ConvolutionFpropNCxHWxThreadblockSwizzle,              \
                 stages_, 4, aligned_, NeedLoadFromConstMem>;                   \
-        typename Convolution::ConvolutionParameter conv_param{                 \
-                param.n,  param.ci, param.co, param.hi, param.wi,              \
-                param.fh, param.fw, param.ho, param.wo, param.sh,              \
-                param.sw, param.ph, param.pw, 1,        1};                    \
+        typename Convolution::ConvolutionParameter conv_param(                 \
+                param.n, param.hi, param.wi, param.ci, param.co, param.fh,     \
+                param.fw, param.ho, param.wo, param.ph, param.pw, param.sh,    \
+                param.sw, 1, 1, cutlass::conv::Mode::kCrossCorrelation);       \
         return cutlass_convolution_wrapper<Convolution>(                       \
                 d_src, d_filter, d_bias, d_z, d_dst, workspace, conv_param,    \
                 epilogue, stream);                                             \
     }
 #define DISPATCH_KERNEL                                                      \
-    DISPATCH_KERNEL_WITH_TILE_SHAPE(128, 128, 32, 64, 32, 32, 2, 16);           \
-    DISPATCH_KERNEL_WITH_TILE_SHAPE(128, 64, 32, 64, 32, 32, 2, 16);            \
-    DISPATCH_KERNEL_WITH_TILE_SHAPE(64, 128, 32, 64, 32, 32, 2, 16);            \
-    DISPATCH_KERNEL_WITH_TILE_SHAPE(128, 32, 32, 64, 32, 32, 2, 16);            \
-    DISPATCH_KERNEL_WITH_TILE_SHAPE(32, 128, 32, 32, 64, 32, 2, 16);            \
-    DISPATCH_KERNEL_WITH_TILE_SHAPE(64, 64, 32, 64, 32, 32, 2, 16);             \
-    DISPATCH_KERNEL_WITH_TILE_SHAPE(32, 64, 32, 32, 64, 32, 2, 16);             \
-    DISPATCH_KERNEL_WITH_TILE_SHAPE(64, 32, 32, 64, 32, 32, 2, 16);             \
-    DISPATCH_KERNEL_WITH_TILE_SHAPE(32, 32, 32, 32, 32, 32, 2, 16);             \
+    DISPATCH_KERNEL_WITH_TILE_SHAPE(128, 128, 32, 64, 32, 32, 2, 16);        \
+    DISPATCH_KERNEL_WITH_TILE_SHAPE(128, 64, 32, 64, 32, 32, 2, 16);         \
+    DISPATCH_KERNEL_WITH_TILE_SHAPE(64, 128, 32, 64, 32, 32, 2, 16);         \
+    DISPATCH_KERNEL_WITH_TILE_SHAPE(128, 32, 32, 64, 32, 32, 2, 16);         \
+    DISPATCH_KERNEL_WITH_TILE_SHAPE(32, 128, 32, 32, 64, 32, 2, 16);         \
+    DISPATCH_KERNEL_WITH_TILE_SHAPE(64, 64, 32, 64, 32, 32, 2, 16);          \
+    DISPATCH_KERNEL_WITH_TILE_SHAPE(32, 64, 32, 32, 64, 32, 2, 16);          \
+    DISPATCH_KERNEL_WITH_TILE_SHAPE(64, 32, 32, 64, 32, 32, 2, 16);          \
+    DISPATCH_KERNEL_WITH_TILE_SHAPE(32, 32, 32, 32, 32, 32, 2, 16);          \
     megdnn_assert(false,                                                     \
                   "unsupported threadblock shape (%dx%dx%d) and warp shape " \
                   "(%dx%dx%d)",                                              \
diff --git a/dnn/src/cuda/conv_bias/int8/conv_bias_int8_implicit_gemm_cutlass_wrapper.cuinl b/dnn/src/cuda/conv_bias/int8/conv_bias_int8_implicit_gemm_cutlass_wrapper.cuinl
index 31b9c40b9..cf20f616d 100644
--- a/dnn/src/cuda/conv_bias/int8/conv_bias_int8_implicit_gemm_cutlass_wrapper.cuinl
+++ b/dnn/src/cuda/conv_bias/int8/conv_bias_int8_implicit_gemm_cutlass_wrapper.cuinl
@@ -29,28 +29,30 @@ void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper(
         cudaStream_t stream) {
     typename Convolution::TensorRefSrc tensor_src{
             const_cast<typename Convolution::ElementSrc*>(d_src),
-            Convolution::LayoutSrc::packed({conv_param.n(), conv_param.hi(),
-                                            conv_param.wi(), conv_param.ci()})};
+            Convolution::LayoutSrc::packed(
+                    {conv_param.N, conv_param.H, conv_param.W, conv_param.C})};
     typename Convolution::TensorRefFilter tensor_filter{
             const_cast<typename Convolution::ElementFilter*>(d_filter),
-            Convolution::LayoutFilter::packed({conv_param.co(), conv_param.fh(),
-                                               conv_param.fw(),
-                                               conv_param.ci()})};
+            Convolution::LayoutFilter::packed(
+                    {conv_param.K, conv_param.R, conv_param.S, conv_param.C})};
     typename Convolution::TensorRefBias tensor_bias{
             const_cast<typename Convolution::ElementBias*>(d_bias),
-            Convolution::LayoutBias::packed({1, 1, 1, conv_param.co()})};
+            Convolution::LayoutBias::packed({1, 1, 1, conv_param.K})};
     typename Convolution::TensorRefDst tensor_z{
             const_cast<typename Convolution::ElementDst*>(d_z),
-            Convolution::LayoutDst::packed({conv_param.n(), conv_param.ho(),
-                                            conv_param.wo(), conv_param.co()})};
+            Convolution::LayoutDst::packed(
+                    {conv_param.N, conv_param.P, conv_param.Q, conv_param.K})};
     typename Convolution::TensorRefDst tensor_dst{
             d_dst,
-            Convolution::LayoutDst::packed({conv_param.n(), conv_param.ho(),
-                                            conv_param.wo(), conv_param.co()})};
-    typename Convolution::Arguments arguments{
-            conv_param,  tensor_src, tensor_filter,
-            tensor_bias, tensor_z,   tensor_dst.non_const_ref(),
-            epilogue};
+            Convolution::LayoutDst::packed(
+                    {conv_param.N, conv_param.P, conv_param.Q, conv_param.K})};
+    typename Convolution::Arguments arguments{conv_param,
+                                              tensor_src.non_const_ref(),
+                                              tensor_filter.non_const_ref(),
+                                              tensor_bias.non_const_ref(),
+                                              tensor_z.non_const_ref(),
+                                              tensor_dst.non_const_ref(),
+                                              epilogue};
     Convolution conv_op;
     cutlass_check(conv_op.initialize(arguments, workspace));
     cutlass_check(conv_op(stream));
diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_128x128x32_64x32x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_128x128x32_64x32x32_hswish.cu
index 4f0eebb74b659460ee9929c44833da05abb82de9..aede9980adb36d3caa3e4acf8383c11af39132d9 100644
GIT binary patch
delta 54
zcmey(w}Wqk9rI*Y=I+VYndeTfW|5M0D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-TZ)MHWL6oj}<Tg

delta 164
zcmdnN_nU8n9W#4=PH9PIe%@p|=59FqB=g*QsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!-oVnq1OP1$I}HE;

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_128x128x32_64x32x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_128x128x32_64x32x32_id.cu
index b92f7681e340f99238674782187829367a5dcbaf..57f38b13b383bcb954ea374c6bba84d36996fe6e 100644
GIT binary patch
delta 56
zcmey$w~23q1@mNk=I+TCndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhru
MXOY-^i)A(w0BDXCumAu6

delta 104
zcmdnQ_myvh1v7hoPH9PIe%@pY=59FqF!NkEJAp-Nay^TLu??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZu4}O*-QYuP$lpH

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_128x128x32_64x32x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_128x128x32_64x32x32_relu.cu
index e47ab05203d3163c749f99423c61ed013d76343e..a73b75fa3390f2925f59f580133918f0e5c8ee1d 100644
GIT binary patch
delta 76
zcmey#w~cRu4fAAY=I+T?ndh2YS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m)D$1<A<0C(~p9RL6T

delta 164
zcmdnS_mgjf4KsUwPH9PIe%@po=59FqIP=_ksJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!UdPhG1OO)mI{^Ry

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_128x32x32_64x32x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_128x32x32_64x32x32_hswish.cu
index 30eec1369454a5e7c09e70ece0ab6a1801f8019d..4121d19ad3adb33749522c1beba9fd274206eb25 100644
GIT binary patch
delta 54
zcmeyxx1Dc;E%Rg-=I+VYnCDKeVv&+{D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-F%;AHWL6nauq26

delta 164
zcmdna_ls|XEi-$5PH9PIe%@qT=59Fq1oPZ_sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!UeD6O1OO@uI|l#&

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_128x32x32_64x32x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_128x32x32_64x32x32_id.cu
index 1307f93751e6255a7c323df7416f8a6fa5914179..ebc98a406aea46f81ea813f9ec5276a82a52c20b 100644
GIT binary patch
delta 56
zcmeyuw~=pyIrC&Y=I+TCnCHq{S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhru
MW0BZ=lVvs&0B402sQ>@~

delta 104
zcmdnU_l0kRIWv2HPH9PIe%@qr=59Fq5c6C(JDx>qavh6=u??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZu2yj*-QYr`z7cA

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_128x32x32_64x32x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_128x32x32_64x32x32_relu.cu
index 1e095d923d257e8afea08c5da62d667ca76ce6f7..c70ca16048b586dce33d6ba0391ec3581d3ac7cc 100644
GIT binary patch
delta 76
zcmeytx0P>$HS=UA=I+T?nCF^XS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m)D%QBk@0CtTY7XSbN

delta 164
zcmdnW_k(YPH8XpDPH9PIe%@qj=59Fq81vkEsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!Udz(K1OOxeI{N?s

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_128x64x32_64x32x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_128x64x32_64x32x32_hswish.cu
index dbf1e900ff5a8301db7e1857f1ee03646ed0e757..9fbad0d37c0b533e6ce61591cab60f17ce2b17e1 100644
GIT binary patch
delta 54
zcmeyxx1Dc;E%Rg-=I+VYnCDKeVv&+{D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-F%;AHWL6nauq26

delta 164
zcmdna_ls|XEi-$5PH9PIe%@qT=59Fq1oPZ_sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!UeD6O1OO@uI|l#&

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_128x64x32_64x32x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_128x64x32_64x32x32_id.cu
index 1ef223f9670095bb94b07cf08483f3534dbe4395..9fd33cba3fb2c1f5d1cdf2a2ae6595ea400bea05 100644
GIT binary patch
delta 56
zcmeyuw~=pyIrC&Y=I+TCnCHq{S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhru
MW0BZ=lVvs&0B402sQ>@~

delta 104
zcmdnU_l0kRIWv2HPH9PIe%@qr=59Fq5c6C(JDx>qavh6=u??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZu2yj*-QYr`z7cA

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_128x64x32_64x32x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_128x64x32_64x32x32_relu.cu
index 65a5fcc71dae4599e09c4acae3af3f8ff49a642e..8d5cc67ace1e77eece1a9251b1e49c4c763c7da2 100644
GIT binary patch
delta 76
zcmeytx0P>$HS=UA=I+T?nCF^XS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m)D%QBk@0CtTY7XSbN

delta 164
zcmdnW_k(YPH8XpDPH9PIe%@qj=59Fq81vkEsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!Udz(K1OOxeI{N?s

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_16x128x16_16x128x16_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_16x128x16_16x128x16_hswish.cu
index 0c779fe15d9856d56b82073d75e6c284fdf267af..820c97570893a0729537a37742e4c802a5cb793f 100644
GIT binary patch
delta 54
zcmeyxx1Dc;9rI*Y=I+VYndeTfW|5M0D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-TZ)M783wIBNZ(G

delta 164
zcmdna_ls|X9W#4=PH9PIe%@p|=59FqB=g*QsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C>L{4#
RC|KwyloXY2-oVn%1OP49J5K-r

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_16x128x16_16x128x16_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_16x128x16_16x128x16_id.cu
index 85fdecd79fda6911989011249fdb9b5c2f050410..f3d1a46305aafd269b66388742a92fe20600b16e 100644
GIT binary patch
delta 56
zcmeyuw~=py1@mNk=I+TCndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhru
MXOY-^i)9uQ0B9Z+tpET3

delta 104
zcmdnU_l0kR1v7hoPH9PIe%@pY=59FqF!NkEJAp-Nay^TLu??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZu4}OSxf-CfF<n!

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_16x128x16_16x128x16_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_16x128x16_16x128x16_relu.cu
index c088e6f6ec533ac73e814d8f65b420a89b129b67..7f35a20cdebb6273a5dad8ac28779b1a0aa66674 100644
GIT binary patch
delta 76
zcmeytx0P>$4fAAY=I+T?ndh2YS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m)D$1;lv0C!p*8UO$Q

delta 164
zcmdnW_k(YP4KsUwPH9PIe%@po=59FqIP=_ksJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C>L{4#
RC|KwyloXY2UdPhT1OO+^J3{~f

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_16x64x8_16x64x8_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_16x64x8_16x64x8_hswish.cu
index 931f3716f967cee115a607de08e4be4fd838a63c..a952578d9a877f0c35c54d2c17f2999c40e53efd 100644
GIT binary patch
delta 56
zcmeyyx0!E)74u|A=I+Utndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhqz
M$Re@%4$CYi0BSH5y8r+H

delta 126
zcmdnY_l<9Z6*GH&PH9PIe%@p&=59FqDDzxJsJN9?Nk&m>VoFj@esZ>zm2-Yx8B9kb
pi-fTamjVzZmzLxt78hGtB_pIEYC|duQmw4Ak+j)up2;$c2>`m7FAo3!

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_16x64x8_16x64x8_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_16x64x8_16x64x8_id.cu
index 70949c6586f2b6197e02ffe55433a5c1c71dd010..da9f49855f4cb2c261e11ffe9f95f25e9b66893b 100644
GIT binary patch
delta 54
zcmeysx0Y{%3G-xY=I+U7ndeR}WRa3}D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-F%H@783w70Tkx|

delta 105
zcmZ3>_knMN2{U_sPH9PIe%@pg=59E9Kl5BTJBCHd$j`aLBfKIcqbM~oB`GIAIXk#K
nv#Kg5)rLy}2$D-nauSP+t*nys^U7ex?Pr!?wA(zHWfl_v4xuJs

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_16x64x8_16x64x8_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_16x64x8_16x64x8_relu.cu
index 0aaa3f7a6f536dcf73c202bb3883f0fb90978701..065b70429f09de343662103a570412b5d34afaa2 100644
GIT binary patch
delta 56
zcmeyuw~=py1@mNk=I+TCndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhru
MXOY-^i)9uQ0B9Z+tpET3

delta 104
zcmdnU_l0kR1v7hoPH9PIe%@pY=59FqF!NkEJAp-Nay^TLu??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZu4}OSxf-CfF<n!

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_128x128x32_64x32x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_128x128x32_64x32x32_hswish.cu
index dc198d3496c7eba44e21221812b70c01a615cc50..2b3a319b14a63ccf6f2b42b29b17a3efbf559070 100644
GIT binary patch
delta 54
zcmeyvx07#!9rI*Y=I+VYndeTfW|5M0D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-TZ)M4if-A`xP?)

delta 164
zcmdnV_lIwT9W#4=PH9PIe%@p|=59FqB=g*QsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEB-oVny1OP3MI`IGi

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_128x128x32_64x32x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_128x128x32_64x32x32_id.cu
index e20f26dc91a6755403d95deb83b3693f319c1070..187f7b96c1ea0f66929b21c6dd922f9f366318ef 100644
GIT binary patch
delta 56
zcmeyyx0!E)1@mNk=I+TCndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhru
MXOY-^i)9WI0BHUdvj6}9

delta 104
zcmdnY_l<9Z1v7hoPH9PIe%@pY=59FqF!NkEJAp-Nay^TLu??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZu4}OIZOb(ASLqv

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_128x128x32_64x32x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_128x128x32_64x32x32_relu.cu
index f9e8b712fb27f4280e4aeccd49ec825fa833fb09..712822ff89ae2601856de785874764d165661835 100644
GIT binary patch
delta 76
zcmeyxx1Dc;4fAAY=I+T?ndh2YS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m)D$1;Zr0C<WXAOHXW

delta 164
zcmdna_ls|X4KsUwPH9PIe%@po=59FqIP=_ksJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEBUdPhO1OO+6I^_TW

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_128x32x32_64x32x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_128x32x32_64x32x32_hswish.cu
index ade39f7a1bfc9744876dcd36e25d04f061081734..4d7d9337ba98f5abba28d5e0aa940abc89b39940 100644
GIT binary patch
delta 54
zcmey(w}WqkE%Rg-=I+VYnCDKeVv&+{D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-F%;A4if-9-W4nW

delta 164
zcmdnN_nU8nEi-$5PH9PIe%@qT=59Fq1oPZ_sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEBUeD6W1OO_EI_m%c

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_128x32x32_64x32x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_128x32x32_64x32x32_id.cu
index e07c6dcf2f49ddfe1cd0d46b81cc6b06d63e3974..aeafaf1593745fdb37dea67a22e8757861dc6659 100644
GIT binary patch
delta 56
zcmey$w~23qIrC&Y=I+TCnCHq{S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhru
MW0BZ=lVuJQ0B7|TtN;K2

delta 104
zcmdnQ_myvhIWv2HPH9PIe%@qr=59Fq5c6C(JDx>qavh6=u??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZu2yjIZOb$%O&do

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_128x32x32_64x32x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_128x32x32_64x32x32_relu.cu
index 27aa11b5f9a74e75ae125496911c3c90ae3560b8..9d41256d4d37292c4e6f517345d77e2875fa80f4 100644
GIT binary patch
delta 76
zcmey#w~cRuHS=UA=I+T?nCF^XS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m)D%QA-v0Cy!G8UO$Q

delta 164
zcmdnS_mgjfH8XpDPH9PIe%@qj=59Fq81vkEsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEBUdz(S1OOy}I^O^Q

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_128x64x32_64x32x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_128x64x32_64x32x32_hswish.cu
index 3e38529ba083a97153b3c79534bda2693148f22b..66a31a7f2d8f55de987260bb6cc9983aa3643812 100644
GIT binary patch
delta 54
zcmey(w}WqkE%Rg-=I+VYnCDKeVv&+{D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-F%;A4if-9-W4nW

delta 164
zcmdnN_nU8nEi-$5PH9PIe%@qT=59Fq1oPZ_sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEBUeD6W1OO_EI_m%c

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_128x64x32_64x32x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_128x64x32_64x32x32_id.cu
index 1b3145c4e6255e3096048c20ebf6adaf953124ed..ca0cf04c47b412b769b3a6e722af19e068dd36c9 100644
GIT binary patch
delta 56
zcmey$w~23qIrC&Y=I+TCnCHq{S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhru
MW0BZ=lVuJQ0B7|TtN;K2

delta 104
zcmdnQ_myvhIWv2HPH9PIe%@qr=59Fq5c6C(JDx>qavh6=u??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZu2yjIZOb$%O&do

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_128x64x32_64x32x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_128x64x32_64x32x32_relu.cu
index 962dd2ffbb662d3cc8f13baef0115437586a4bda..f871dae4ed6b97fd373ceec9706bdb3b2f6f7e4f 100644
GIT binary patch
delta 76
zcmey#w~cRuHS=UA=I+T?nCF^XS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m)D%QA-v0Cy!G8UO$Q

delta 164
zcmdnS_mgjfH8XpDPH9PIe%@qj=59Fq81vkEsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEBUdz(S1OOy}I^O^Q

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_16x128x16_16x128x16_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_16x128x16_16x128x16_hswish.cu
index 14226ecd3b901e9aac04d5444f25e7ac6114efdb..725fed7fb736ae1901dfb7a24144c5961c34cc4a 100644
GIT binary patch
delta 54
zcmey(w}Wqk9rI*Y=I+VYndeTfW|5M0D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-TZ)MHWL6oj}<Tg

delta 164
zcmdnN_nU8n9W#4=PH9PIe%@p|=59FqB=g*QsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C>L{4#
RC|Kwyq$TET-oVnq1OP4qJ1PJG

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_16x128x16_16x128x16_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_16x128x16_16x128x16_id.cu
index 449dee42ef976533a07d19f702196fbe9a19ec54..e7688245da55156e422aaedfb50952d420b8e8d9 100644
GIT binary patch
delta 56
zcmey$w~23q1@mNk=I+TCndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhru
MXOY-^i)A(w0BDXCumAu6

delta 104
zcmdnQ_myvh1v7hoPH9PIe%@pY=59FqF!NkEJAp-Nay^TLu??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZu4}O*-QYuP$lpH

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_16x128x16_16x128x16_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_16x128x16_16x128x16_relu.cu
index 0ed7466912ebf3b72a88f4555ecc03afeb3652b1..100c48612f366ecebd2cde077ce56dd738c25433 100644
GIT binary patch
delta 76
zcmey#w~cRu4fAAY=I+T?ndh2YS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m)D$1<A<0C(~p9RL6T

delta 164
zcmdnS_mgjf4KsUwPH9PIe%@po=59FqIP=_ksJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C>L{4#
RC|Kwyq$TETUdPhG1OO-aJ01W4

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_16x64x8_16x64x8_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_16x64x8_16x64x8_hswish.cu
index 4d3feb12372ce71dfdfda70aa3eb904cdff55ca2..aa5ecf003a0f165b7e44cdfb67e4f5484bb2432f 100644
GIT binary patch
delta 56
zcmey)w}o$m74u|A=I+Utndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhqz
M$Re@%4$EvN0BWEWz5oCK

delta 126
zcmdnO_nmKp6*GH&PH9PIe%@p&=59FqDDzxJsJN9?Nk&m>VoFj@esZ>zm2-Yx8B9kb
pi-fTamjVzZmzLxt78hGtB_pIEYC|duQmw4Ak+j)up2;$s2>`p4FA@L%

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_16x64x8_16x64x8_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_16x64x8_16x64x8_id.cu
index 1dbfbf60739a530a23363f3698e3baf110e52ffb..0a1c1d3a54d987b53c70a6a54f6b527dcc52336f 100644
GIT binary patch
delta 54
zcmey!w~lXv3G-xY=I+U7ndeR}WRa3}D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-F%H@HWL6dZ4~MN

delta 105
zcmZ3-_mOXd2{U_sPH9PIe%@pg=59E9Kl5BTJBCHd$j`aLBfKIcqbM~oB`GIAIXk#K
nv#Kg5)rLy}2$D-nauSP+t*nys^U7ex?Pr!?wA(zHWi}H44|67C

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_16x64x8_16x64x8_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_16x64x8_16x64x8_relu.cu
index d4fcfb061335b0cdcee7a2f03f504dc0d775fb4c..a61a466ae3d904d36cef24d4c1e79581e8a87e20 100644
GIT binary patch
delta 56
zcmey$w~23q1@mNk=I+TCndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhru
MXOY-^i)A(w0BDXCumAu6

delta 104
zcmdnQ_myvh1v7hoPH9PIe%@pY=59FqF!NkEJAp-Nay^TLu??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZu4}O*-QYuP$lpH

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_32x128x32_32x64x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_32x128x32_32x64x32_hswish.cu
index ab97934953fb585e1ced968b17c9fb94f7e96daa..8ac06afb8ad77d9d59313647f07e0efb6c09106f 100644
GIT binary patch
delta 54
zcmey(w}WqkE%Rg-=I+VYnCDKeVv&+{D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-F%;A4if-9-W4nW

delta 164
zcmdnN_nU8nEi-$5PH9PIe%@qT=59Fq1oPZ_sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEBUeD6W1OO_EI_m%c

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_32x128x32_32x64x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_32x128x32_32x64x32_id.cu
index 4d9abe294e0187ce1a4e4b147c797a7c0ad862ca..7237f2de387d49308dcb46bf6216258dbe39b04e 100644
GIT binary patch
delta 56
zcmey$w~23qIrC&Y=I+TCnCHq{S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhru
MW0BZ=lVuJQ0B7|TtN;K2

delta 104
zcmdnQ_myvhIWv2HPH9PIe%@qr=59Fq5c6C(JDx>qavh6=u??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZu2yjIZOb$%O&do

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_32x128x32_32x64x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_32x128x32_32x64x32_relu.cu
index b1efef3676cfe4ff3eeed41c42c7a4b3fcd86555..a672e749dfde740b3b714a0bb47aabac0799c649 100644
GIT binary patch
delta 76
zcmey#w~cRuHS=UA=I+T?nCF^XS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m)D%QA-v0Cy!G8UO$Q

delta 164
zcmdnS_mgjfH8XpDPH9PIe%@qj=59Fq81vkEsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEBUdz(S1OOy}I^O^Q

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_32x32x32_32x32x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_32x32x32_32x32x32_hswish.cu
index fb6b71da008280971ad7e89c8b6f30bd341704dc..27ad631e576d40591409cfc0818e2bca4791f970 100644
GIT binary patch
delta 76
zcmeyxx1Dc;4fAAY=I+T?ndh2YS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m)D$1;Zr0C<WXAOHXW

delta 164
zcmdna_ls|X4KsUwPH9PIe%@po=59FqIP=_ksJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEBUdPhO1OO+6I^_TW

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_32x32x32_32x32x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_32x32x32_32x32x32_id.cu
index c38538c325ddd0810767ae25c67d36ffaf3dc83f..ddc61a512ebb176a428938093384cbf45d70d02e 100644
GIT binary patch
delta 56
zcmeyuw~=py8S`XY=I+Vondiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhru
MWs%r?gJljA0A}nJr2qf`

delta 85
zcmdnU_l0kR88dr+PH9PIe%@p==59FqAoE-}JB~$aaxII5lns{x5G0qD<RlgsTUjON
T=as><9%PnawA(zDWeyVnC)^$n

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_32x32x32_32x32x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_32x32x32_32x32x32_relu.cu
index b91efded3aa1b20c3ee78cedc181d72516c11c22..c5aaec6f49ca9f34ce39c2b8f9330fe894d6a980 100644
GIT binary patch
delta 56
zcmeytx0P>$74u|A=I+Utndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhqz
M$Re@%4$B-S0BaBx!2kdN

delta 126
zcmdnW_k(YP6*GH&PH9PIe%@p&=59FqDDzxJsJN9?Nk&m>VoFj@esZ>zm2-Yx8B9kb
pi-fTamjVzZmzLxt78hGtB_pIEYC|duQmw4Ak+j)up2;$Y2>`s1FBJd)

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_32x64x32_32x64x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_32x64x32_32x64x32_hswish.cu
index c877492e5be9a96a815876f51ee53353b2744090..a5d437a23ef8b72c79d250c928789cd383b6d45f 100644
GIT binary patch
delta 76
zcmeyxx1Dc;4fAAY=I+T?ndh2YS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m)D$1;Zr0C<WXAOHXW

delta 164
zcmdna_ls|X4KsUwPH9PIe%@po=59FqIP=_ksJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEBUdPhO1OO+6I^_TW

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_32x64x32_32x64x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_32x64x32_32x64x32_id.cu
index 40e6cb5d360682e04b42b29f83385e5500fa0805..d16c179b229532ef3272773b63b12bd06ffe9dad 100644
GIT binary patch
delta 56
zcmeyuw~=py8S`XY=I+Vondiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhru
MWs%r?gJljA0A}nJr2qf`

delta 85
zcmdnU_l0kR88dr+PH9PIe%@p==59FqAoE-}JB~$aaxII5lns{x5G0qD<RlgsTUjON
T=as><9%PnawA(zDWeyVnC)^$n

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_32x64x32_32x64x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_32x64x32_32x64x32_relu.cu
index 19f337a174bbaca1445e471cce1360e7255d40f9..e5968cf52aac02e3ffb410b2ec61534de5bf7ec9 100644
GIT binary patch
delta 56
zcmeytx0P>$74u|A=I+Utndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhqz
M$Re@%4$B-S0BaBx!2kdN

delta 126
zcmdnW_k(YP6*GH&PH9PIe%@p&=59FqDDzxJsJN9?Nk&m>VoFj@esZ>zm2-Yx8B9kb
pi-fTamjVzZmzLxt78hGtB_pIEYC|duQmw4Ak+j)up2;$Y2>`s1FBJd)

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_64x128x32_64x32x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_64x128x32_64x32x32_hswish.cu
index 27f7d5b869a4aeb61e56eedc69709b6960ffe8ce..61faa3e522870de1b9d88cd06cb52d28e957b182 100644
GIT binary patch
delta 54
zcmey(w}WqkE%Rg-=I+VYnCDKeVv&+{D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-F%;A4if-9-W4nW

delta 164
zcmdnN_nU8nEi-$5PH9PIe%@qT=59Fq1oPZ_sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEBUeD6W1OO_EI_m%c

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_64x128x32_64x32x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_64x128x32_64x32x32_id.cu
index 1a8e17fd6c533b32b14cc3f08b48993a79a50f83..b5177e0106608398e3096328cee966c9234dd578 100644
GIT binary patch
delta 56
zcmey$w~23qIrC&Y=I+TCnCHq{S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhru
MW0BZ=lVuJQ0B7|TtN;K2

delta 104
zcmdnQ_myvhIWv2HPH9PIe%@qr=59Fq5c6C(JDx>qavh6=u??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZu2yjIZOb$%O&do

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_64x128x32_64x32x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_64x128x32_64x32x32_relu.cu
index 1a4e679f6cf7ca1a9961125da69ad497c2c7b6a3..6e89d753b72c441b1beecbc30b8a10e070294f56 100644
GIT binary patch
delta 76
zcmey#w~cRuHS=UA=I+T?nCF^XS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m)D%QA-v0Cy!G8UO$Q

delta 164
zcmdnS_mgjfH8XpDPH9PIe%@qj=59Fq81vkEsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEBUdz(S1OOy}I^O^Q

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_64x32x32_64x32x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_64x32x32_64x32x32_hswish.cu
index 252a8f779310bdc004f1c3095a0853b8d67b4fd0..79d9c0f8efe456cd55466f6af1360f22c06b8213 100644
GIT binary patch
delta 76
zcmeyxx1Dc;4fAAY=I+T?ndh2YS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m)D$1;Zr0C<WXAOHXW

delta 164
zcmdna_ls|X4KsUwPH9PIe%@po=59FqIP=_ksJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEBUdPhO1OO+6I^_TW

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_64x32x32_64x32x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_64x32x32_64x32x32_id.cu
index f5d01d75ad737eb3fcf870a47f8b7b59ce3f8df4..cc5c002637aadf653eebbd5ed9abcae67eefbcdc 100644
GIT binary patch
delta 56
zcmeyuw~=py8S`XY=I+Vondiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhru
MWs%r?gJljA0A}nJr2qf`

delta 85
zcmdnU_l0kR88dr+PH9PIe%@p==59FqAoE-}JB~$aaxII5lns{x5G0qD<RlgsTUjON
T=as><9%PnawA(zDWeyVnC)^$n

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_64x32x32_64x32x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_64x32x32_64x32x32_relu.cu
index a29be07e1d171b1d9084f9cfcb570612c32daf55..3a71ccf0f44f0682269252d878ad3b5e49d6e395 100644
GIT binary patch
delta 56
zcmeytx0P>$74u|A=I+Utndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhqz
M$Re@%4$B-S0BaBx!2kdN

delta 126
zcmdnW_k(YP6*GH&PH9PIe%@p&=59FqDDzxJsJN9?Nk&m>VoFj@esZ>zm2-Yx8B9kb
pi-fTamjVzZmzLxt78hGtB_pIEYC|duQmw4Ak+j)up2;$Y2>`s1FBJd)

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_64x64x32_64x32x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_64x64x32_64x32x32_hswish.cu
index 07333503fbc32c9874ace0c5a14b6d84a72d1822..de8806d6b4ef0ccecabc863eee761817ee4a9c84 100644
GIT binary patch
delta 76
zcmeyxx1Dc;4fAAY=I+T?ndh2YS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m)D$1;Zr0C<WXAOHXW

delta 164
zcmdna_ls|X4KsUwPH9PIe%@po=59FqIP=_ksJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEBUdPhO1OO+6I^_TW

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_64x64x32_64x32x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_64x64x32_64x32x32_id.cu
index 41f6db050537c7733e5eb259ac9a698b24cbddff..b5e3a342357167bd407ea6b9ebd58850e8672905 100644
GIT binary patch
delta 56
zcmeyuw~=py8S`XY=I+Vondiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhru
MWs%r?gJljA0A}nJr2qf`

delta 85
zcmdnU_l0kR88dr+PH9PIe%@p==59FqAoE-}JB~$aaxII5lns{x5G0qD<RlgsTUjON
T=as><9%PnawA(zDWeyVnC)^$n

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_64x64x32_64x32x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_1x1_64x64x32_64x32x32_relu.cu
index 34896baf1e5431ae8432e785b9f2ab05a6ded26b..2243305e9e1ecedae9d9444425172f5e1c6dce6b 100644
GIT binary patch
delta 56
zcmeytx0P>$74u|A=I+Utndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhqz
M$Re@%4$B-S0BaBx!2kdN

delta 126
zcmdnW_k(YP6*GH&PH9PIe%@p&=59FqDDzxJsJN9?Nk&m>VoFj@esZ>zm2-Yx8B9kb
pi-fTamjVzZmzLxt78hGtB_pIEYC|duQmw4Ak+j)up2;$Y2>`s1FBJd)

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_32x128x32_32x64x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_32x128x32_32x64x32_hswish.cu
index 5f247f0ccff0ff84ae3568364c643aaf427ba20b..6110868bf0e723daa9b163b8eb617c2ed7b5c60e 100644
GIT binary patch
delta 54
zcmeyxx1Dc;E%Rg-=I+VYnCDKeVv&+{D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-F%;AHWL6nauq26

delta 164
zcmdna_ls|XEi-$5PH9PIe%@qT=59Fq1oPZ_sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!UeD6O1OO@uI|l#&

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_32x128x32_32x64x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_32x128x32_32x64x32_id.cu
index 81cdd18cde0f493047267827589d9f7f00d4164a..0e5c018e8cd57ee3cfeed4a3fafc59ea3ca5be1e 100644
GIT binary patch
delta 56
zcmeyuw~=pyIrC&Y=I+TCnCHq{S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhru
MW0BZ=lVvs&0B402sQ>@~

delta 104
zcmdnU_l0kRIWv2HPH9PIe%@qr=59Fq5c6C(JDx>qavh6=u??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZu2yj*-QYr`z7cA

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_32x128x32_32x64x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_32x128x32_32x64x32_relu.cu
index 0bce8bf5ed58afbb4cc6f8d79940056e08aaff3e..33d950ece601a4f1ad7a11ab92f8e4141ad57e06 100644
GIT binary patch
delta 76
zcmeytx0P>$HS=UA=I+T?nCF^XS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m)D%QBk@0CtTY7XSbN

delta 164
zcmdnW_k(YPH8XpDPH9PIe%@qj=59Fq81vkEsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!Udz(K1OOxeI{N?s

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_32x32x32_32x32x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_32x32x32_32x32x32_hswish.cu
index 1ffdb3c4872a2d8d7e3658362ffd560d3ae8313d..012b70b064cb8dfac7a73dd73efa3bd1e5ac81c5 100644
GIT binary patch
delta 76
zcmey#w~cRu4fAAY=I+T?ndh2YS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m)D$1<A<0C(~p9RL6T

delta 164
zcmdnS_mgjf4KsUwPH9PIe%@po=59FqIP=_ksJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!UdPhG1OO)mI{^Ry

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_32x32x32_32x32x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_32x32x32_32x32x32_id.cu
index bf0fdcff5ca0517944a7ea3e032b4e31b8cfce5a..aff240c47cdeeca03641591b8290b82c8af1cb91 100644
GIT binary patch
delta 56
zcmey&w}Eei8S`XY=I+Vondiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhru
MWs%r?gJm`o0A_p@q5uE@

delta 85
zcmdnM_nB{l88dr+PH9PIe%@p==59FqAoE-}JB~$aaxII5lns{x5G0qD<RlgsTUjON
T=as><9%PnawA(zDWi}H4Co&!k

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_32x32x32_32x32x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_32x32x32_32x32x32_relu.cu
index bb3a17ba68ac931d5f3256443598c30e5b49c5be..c101cd36ce39e2709ad49d82b79c41f9ecb8b253 100644
GIT binary patch
delta 56
zcmey)w}o$m74u|A=I+Utndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhqz
M$Re@%4$EvN0BWEWz5oCK

delta 126
zcmdnO_nmKp6*GH&PH9PIe%@p&=59FqDDzxJsJN9?Nk&m>VoFj@esZ>zm2-Yx8B9kb
pi-fTamjVzZmzLxt78hGtB_pIEYC|duQmw4Ak+j)up2;$s2>`p4FA@L%

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_32x64x32_32x64x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_32x64x32_32x64x32_hswish.cu
index 9df0b4f40ce5b18810253283b0238825a21b3f00..487a06fb97a0ca8024fb8b670d69ef2a24b96b96 100644
GIT binary patch
delta 76
zcmey#w~cRu4fAAY=I+T?ndh2YS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m)D$1<A<0C(~p9RL6T

delta 164
zcmdnS_mgjf4KsUwPH9PIe%@po=59FqIP=_ksJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!UdPhG1OO)mI{^Ry

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_32x64x32_32x64x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_32x64x32_32x64x32_id.cu
index c0da1d27367253dca4f67cc6aced74da2b8f7874..450fd723d8955ef91bb10a5cd3184517db4e4b22 100644
GIT binary patch
delta 56
zcmey&w}Eei8S`XY=I+Vondiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhru
MWs%r?gJm`o0A_p@q5uE@

delta 85
zcmdnM_nB{l88dr+PH9PIe%@p==59FqAoE-}JB~$aaxII5lns{x5G0qD<RlgsTUjON
T=as><9%PnawA(zDWi}H4Co&!k

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_32x64x32_32x64x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_32x64x32_32x64x32_relu.cu
index 309aa23941b716f9b538f189e1182e4ca3e4f532..efc249c2d471ecb0c46851b4860f199c56519fcb 100644
GIT binary patch
delta 56
zcmey)w}o$m74u|A=I+Utndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhqz
M$Re@%4$EvN0BWEWz5oCK

delta 126
zcmdnO_nmKp6*GH&PH9PIe%@p&=59FqDDzxJsJN9?Nk&m>VoFj@esZ>zm2-Yx8B9kb
pi-fTamjVzZmzLxt78hGtB_pIEYC|duQmw4Ak+j)up2;$s2>`p4FA@L%

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_64x128x32_64x32x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_64x128x32_64x32x32_hswish.cu
index 210c52e07b76f0c45b10369fdf7fcdf6a0de6785..8503592dfbd119060b78ad5dadbecf6c5145881d 100644
GIT binary patch
delta 54
zcmeyxx1Dc;E%Rg-=I+VYnCDKeVv&+{D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-F%;AHWL6nauq26

delta 164
zcmdna_ls|XEi-$5PH9PIe%@qT=59Fq1oPZ_sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!UeD6O1OO@uI|l#&

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_64x128x32_64x32x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_64x128x32_64x32x32_id.cu
index 4f4f35f8410cff630615d9bd99b9affcf1c6329a..2d9f42b02acf5bca792b89a544b8477a3e9d0b94 100644
GIT binary patch
delta 56
zcmeyuw~=pyIrC&Y=I+TCnCHq{S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhru
MW0BZ=lVvs&0B402sQ>@~

delta 104
zcmdnU_l0kRIWv2HPH9PIe%@qr=59Fq5c6C(JDx>qavh6=u??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZu2yj*-QYr`z7cA

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_64x128x32_64x32x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_64x128x32_64x32x32_relu.cu
index 57e4e210a81491bcfb5bfa75723faa590a1b13ab..d1cbc3ce80788e4c98e19a65ed59f6b37c04712d 100644
GIT binary patch
delta 76
zcmeytx0P>$HS=UA=I+T?nCF^XS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m)D%QBk@0CtTY7XSbN

delta 164
zcmdnW_k(YPH8XpDPH9PIe%@qj=59Fq81vkEsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!Udz(K1OOxeI{N?s

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_64x32x32_64x32x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_64x32x32_64x32x32_hswish.cu
index 224e8a63670d2bdc4e298efb6ce5ffd8bcc8ed58..5c473b18408de1b367d724d2b20e4c07019a636a 100644
GIT binary patch
delta 76
zcmey#w~cRu4fAAY=I+T?ndh2YS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m)D$1<A<0C(~p9RL6T

delta 164
zcmdnS_mgjf4KsUwPH9PIe%@po=59FqIP=_ksJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!UdPhG1OO)mI{^Ry

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_64x32x32_64x32x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_64x32x32_64x32x32_id.cu
index 557610101842ffb4c909c3513141d3cef1dd125e..31a15cc73ad7e93278a0448c8a53b576199d21fb 100644
GIT binary patch
delta 56
zcmey&w}Eei8S`XY=I+Vondiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhru
MWs%r?gJm`o0A_p@q5uE@

delta 85
zcmdnM_nB{l88dr+PH9PIe%@p==59FqAoE-}JB~$aaxII5lns{x5G0qD<RlgsTUjON
T=as><9%PnawA(zDWi}H4Co&!k

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_64x32x32_64x32x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_64x32x32_64x32x32_relu.cu
index 170f197105f9a19d953030df5143e806ef18b2e3..fc86fa43585df6bb7a54364bae7c93684b20ebe7 100644
GIT binary patch
delta 56
zcmey)w}o$m74u|A=I+Utndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhqz
M$Re@%4$EvN0BWEWz5oCK

delta 126
zcmdnO_nmKp6*GH&PH9PIe%@p&=59FqDDzxJsJN9?Nk&m>VoFj@esZ>zm2-Yx8B9kb
pi-fTamjVzZmzLxt78hGtB_pIEYC|duQmw4Ak+j)up2;$s2>`p4FA@L%

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_64x64x32_64x32x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_64x64x32_64x32x32_hswish.cu
index 03cbfbc5921f85b9674b867c2890b785fd8d104b..2f66a3b3dbc0b30bd4bca69291b11577b66885ee 100644
GIT binary patch
delta 76
zcmey#w~cRu4fAAY=I+T?ndh2YS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m)D$1<A<0C(~p9RL6T

delta 164
zcmdnS_mgjf4KsUwPH9PIe%@po=59FqIP=_ksJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!UdPhG1OO)mI{^Ry

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_64x64x32_64x32x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_64x64x32_64x32x32_id.cu
index 4698d99f6739fd8927e49591ad6ef8cfa3c1d4cd..7da1dd3ffc0ee98e78abe64a418463fd91989a28 100644
GIT binary patch
delta 56
zcmey&w}Eei8S`XY=I+Vondiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhru
MWs%r?gJm`o0A_p@q5uE@

delta 85
zcmdnM_nB{l88dr+PH9PIe%@p==59FqAoE-}JB~$aaxII5lns{x5G0qD<RlgsTUjON
T=as><9%PnawA(zDWi}H4Co&!k

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_64x64x32_64x32x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_64x64x32_64x32x32_relu.cu
index 38358f7ead2b030d7be39a69b27c36490d64ccf8..199c133d590acc9429bcddd3d835abaff3ab5fa0 100644
GIT binary patch
delta 56
zcmey)w}o$m74u|A=I+Utndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhqz
M$Re@%4$EvN0BWEWz5oCK

delta 126
zcmdnO_nmKp6*GH&PH9PIe%@p&=59FqDDzxJsJN9?Nk&m>VoFj@esZ>zm2-Yx8B9kb
pi-fTamjVzZmzLxt78hGtB_pIEYC|duQmw4Ak+j)up2;$s2>`p4FA@L%

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_128x128x32_64x32x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_128x128x32_64x32x32_hswish.cu
index b2c9c4628f4f1084d10517b099b89824b4beb5c9..b17a642029f69358ddc6da430132cf6e03f85e8d 100644
GIT binary patch
delta 54
zcmeyvx07#!J@aHY=I+TinCDKeVUdz`D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-TaVcHWL6ptQ9u^

delta 145
zcmdnV_lIwTJu`cLPH9PIe%@qz=59Fq6!Y91sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG92Bp2sqq2>|{S
BH<JJW

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_128x128x32_64x32x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_128x128x32_64x32x32_id.cu
index e758d0da7b78dc2c888d8b1bcc587eddc394e681..af4824b2cfc35f1ffa3b8759f7e5f100335b2682 100644
GIT binary patch
delta 56
zcmeyyx0!E)CG%tl=I+UtnCHq{S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhqz
Mz#_5vHp^@#0BM&Mw*UYD

delta 104
zcmdnY_l<9ZB{O?|PH9PIe%@qD=59Fq2=iPxJCQ|das!Kmu??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZu1P5*-QYwr6u?P

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_128x128x32_64x32x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_128x128x32_64x32x32_relu.cu
index f707002d6b15af94eb308bad30475306f8ae1e49..5d99e6fa90b570c6997bd3f6ebdd7924da0f1e5d 100644
GIT binary patch
delta 54
zcmeyxx1Dc;E%Rg-=I+VYnCDKeVv&+{D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-F%;AHWL6nauq26

delta 164
zcmdna_ls|XEi-$5PH9PIe%@qT=59Fq1oPZ_sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!UeD6O1OO@uI|l#&

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_128x32x32_64x32x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_128x32x32_64x32x32_hswish.cu
index d1c44bec4e93537c1c512eec4b703cbd8decf0ad..c854ef867270e133745a3c585ad251f1e1def372 100644
GIT binary patch
delta 54
zcmey(w}Wqk9rI*Y=I+VYndeTfW|5M0D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-TZ)MHWL6oj}<Tg

delta 164
zcmdnN_nU8n9W#4=PH9PIe%@p|=59FqB=g*QsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!-oVnq1OP1$I}HE;

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_128x32x32_64x32x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_128x32x32_64x32x32_id.cu
index 5158b527d76fc935e73de1de686870785afc514e..1eb4a3aacb3957b7fbba5076fb12c31f15e43f86 100644
GIT binary patch
delta 56
zcmey$w~23q1@mNk=I+TCndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhru
MXOY-^i)A(w0BDXCumAu6

delta 104
zcmdnQ_myvh1v7hoPH9PIe%@pY=59FqF!NkEJAp-Nay^TLu??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZu4}O*-QYuP$lpH

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_128x32x32_64x32x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_128x32x32_64x32x32_relu.cu
index 21c01d3675fc3eeec3f93b2bcb86fe02c95c25be..c56b0b4859b9cb6d2e40fd48576f40226c5ac2f6 100644
GIT binary patch
delta 76
zcmey#w~cRu4fAAY=I+T?ndh2YS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m)D$1<A<0C(~p9RL6T

delta 164
zcmdnS_mgjf4KsUwPH9PIe%@po=59FqIP=_ksJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!UdPhG1OO)mI{^Ry

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_128x64x32_64x32x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_128x64x32_64x32x32_hswish.cu
index b28f68247219e07bca34c94fcb166aacd470fab4..d07b72e9f4a2028acd951bd94239acff9ff305ba 100644
GIT binary patch
delta 54
zcmey(w}Wqk9rI*Y=I+VYndeTfW|5M0D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-TZ)MHWL6oj}<Tg

delta 164
zcmdnN_nU8n9W#4=PH9PIe%@p|=59FqB=g*QsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!-oVnq1OP1$I}HE;

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_128x64x32_64x32x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_128x64x32_64x32x32_id.cu
index f106af213e1889d6744bb87bfc84f268b48ee5bf..c9e2b4494452709d53d52433237c3be69b2f148b 100644
GIT binary patch
delta 56
zcmey$w~23q1@mNk=I+TCndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhru
MXOY-^i)A(w0BDXCumAu6

delta 104
zcmdnQ_myvh1v7hoPH9PIe%@pY=59FqF!NkEJAp-Nay^TLu??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZu4}O*-QYuP$lpH

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_128x64x32_64x32x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_128x64x32_64x32x32_relu.cu
index 7f45ef6209a6aa7d3a35034a79017ed3a4df73c7..18c8cea87bf706b8ba6ead7973931fb4303154ef 100644
GIT binary patch
delta 76
zcmey#w~cRu4fAAY=I+T?ndh2YS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m)D$1<A<0C(~p9RL6T

delta 164
zcmdnS_mgjf4KsUwPH9PIe%@po=59FqIP=_ksJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!UdPhG1OO)mI{^Ry

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_128x128x32_64x32x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_128x128x32_64x32x32_hswish.cu
index 130b04b863e3fa754e9fb4b24127f227e3c05b27..8f0745635ea3dd38ec6f5f122bcf7838c6d879cf 100644
GIT binary patch
delta 54
zcmey%w~KFsJ@aHY=I+TinCDKeVUdz`D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-TaVc4if-C7!^7I

delta 145
zcmdnR_m^*jJu`cLPH9PIe%@qz=59Fq6!Y91sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG92Bp2sqW2>|~#
BH<kbZ

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_128x128x32_64x32x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_128x128x32_64x32x32_id.cu
index 428024430f170a0ccf0fd834633ba37896839800..b7f76e686cee5fe475cda62cfde6adddcb111a1f 100644
GIT binary patch
delta 56
zcmey)w}o$mCG%tl=I+UtnCHq{S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhqz
Mz#_5vHp?6)0BQ#nx&QzG

delta 104
zcmdnO_nmKpB{O?|PH9PIe%@qD=59Fq2=iPxJCQ|das!Kmu??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZu1P5IZOb*btU@%

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_128x128x32_64x32x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_128x128x32_64x32x32_relu.cu
index 0a7d3c990753413a214bc256be4e40b13aebc819..ae4ce8654a9a30006d9eadd22699733533771535 100644
GIT binary patch
delta 54
zcmey(w}WqkE%Rg-=I+VYnCDKeVv&+{D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-F%;A4if-9-W4nW

delta 164
zcmdnN_nU8nEi-$5PH9PIe%@qT=59Fq1oPZ_sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEBUeD6W1OO_EI_m%c

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_128x32x32_64x32x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_128x32x32_64x32x32_hswish.cu
index 716a540b711bef86ac13934b1de12ba137433287..670b6403db995d43b3ef3adf60c28718c086edef 100644
GIT binary patch
delta 54
zcmeyvx07#!9rI*Y=I+VYndeTfW|5M0D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-TZ)M4if-A`xP?)

delta 164
zcmdnV_lIwT9W#4=PH9PIe%@p|=59FqB=g*QsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEB-oVny1OP3MI`IGi

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_128x32x32_64x32x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_128x32x32_64x32x32_id.cu
index 35b97bcda599aed2b2404d4968942c8f36d5f7b4..5a85818a9e8acffe680c433a7e49999aa77b882f 100644
GIT binary patch
delta 56
zcmeyyx0!E)1@mNk=I+TCndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhru
MXOY-^i)9WI0BHUdvj6}9

delta 104
zcmdnY_l<9Z1v7hoPH9PIe%@pY=59FqF!NkEJAp-Nay^TLu??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZu4}OIZOb(ASLqv

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_128x32x32_64x32x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_128x32x32_64x32x32_relu.cu
index 8ea93a1c4d9f0621c4d5fcb17b5e95defec5b051..18b743fc2477951dc2d4de2e5cc477fc84d43c93 100644
GIT binary patch
delta 76
zcmeyxx1Dc;4fAAY=I+T?ndh2YS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m)D$1;Zr0C<WXAOHXW

delta 164
zcmdna_ls|X4KsUwPH9PIe%@po=59FqIP=_ksJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEBUdPhO1OO+6I^_TW

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_128x64x32_64x32x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_128x64x32_64x32x32_hswish.cu
index 3681959837ff5df0c0583c4b36c43592099f1aab..0d5fdc72cabf6bcbb6ac96ed3b2b7a9615a3ba84 100644
GIT binary patch
delta 54
zcmeyvx07#!9rI*Y=I+VYndeTfW|5M0D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-TZ)M4if-A`xP?)

delta 164
zcmdnV_lIwT9W#4=PH9PIe%@p|=59FqB=g*QsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEB-oVny1OP3MI`IGi

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_128x64x32_64x32x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_128x64x32_64x32x32_id.cu
index ebeb82911b4e51fbc55541f23a10a9ab2fb6a9e5..b0f4bd6f7544fcdf756244446adbdd01e557f410 100644
GIT binary patch
delta 56
zcmeyyx0!E)1@mNk=I+TCndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhru
MXOY-^i)9WI0BHUdvj6}9

delta 104
zcmdnY_l<9Z1v7hoPH9PIe%@pY=59FqF!NkEJAp-Nay^TLu??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZu4}OIZOb(ASLqv

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_128x64x32_64x32x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_128x64x32_64x32x32_relu.cu
index d9b73fa2dbdf5da5653e58f66b25ce534fd27ae5..f2ee1cdca63f022ecf5c46edc7c24b9add26bb1a 100644
GIT binary patch
delta 76
zcmeyxx1Dc;4fAAY=I+T?ndh2YS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m)D$1;Zr0C<WXAOHXW

delta 164
zcmdna_ls|X4KsUwPH9PIe%@po=59FqIP=_ksJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEBUdPhO1OO+6I^_TW

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_32x128x32_32x64x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_32x128x32_32x64x32_hswish.cu
index 0cd9b194f6a68d2ccd1b833fc7b3925d87b2161a..157cd07b49bb6a3b0e19742df15f3834e18a29b9 100644
GIT binary patch
delta 54
zcmeyvx07#!9rI*Y=I+VYndeTfW|5M0D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-TZ)M4if-A`xP?)

delta 164
zcmdnV_lIwT9W#4=PH9PIe%@p|=59FqB=g*QsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEB-oVny1OP3MI`IGi

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_32x128x32_32x64x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_32x128x32_32x64x32_id.cu
index 4a89381a8c71bc9e076d8028880be7f2f0788428..eb93fa6cd7d753ab18df0d0526005ebacefc8eb3 100644
GIT binary patch
delta 56
zcmeyyx0!E)1@mNk=I+TCndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhru
MXOY-^i)9WI0BHUdvj6}9

delta 104
zcmdnY_l<9Z1v7hoPH9PIe%@pY=59FqF!NkEJAp-Nay^TLu??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZu4}OIZOb(ASLqv

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_32x128x32_32x64x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_32x128x32_32x64x32_relu.cu
index 0011b40b18136af0a689e5b900e523b826363b6f..3d47893181f6d516ee2cd52495718f2690f5856f 100644
GIT binary patch
delta 76
zcmeyxx1Dc;4fAAY=I+T?ndh2YS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m)D$1;Zr0C<WXAOHXW

delta 164
zcmdna_ls|X4KsUwPH9PIe%@po=59FqIP=_ksJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEBUdPhO1OO+6I^_TW

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_32x32x32_32x32x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_32x32x32_32x32x32_hswish.cu
index a60e3b181f2bc9f11f21596075924161e45cb4cf..99f335be139b5dffdd310147aa2137248528efac 100644
GIT binary patch
delta 54
zcmey(w}WqkE%Rg-=I+VYnCDKeVv&+{D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-F%;A4if-9-W4nW

delta 164
zcmdnN_nU8nEi-$5PH9PIe%@qT=59Fq1oPZ_sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEBUeD6W1OO_EI_m%c

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_32x32x32_32x32x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_32x32x32_32x32x32_id.cu
index 4bee34e863810b9e00e6e2eabbe6f433f11396a8..643f43caacae89bc1fd1565e500dd37a6c23e5f1 100644
GIT binary patch
delta 56
zcmey$w~23qIrC&Y=I+TCnCHq{S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhru
MW0BZ=lVuJQ0B7|TtN;K2

delta 104
zcmdnQ_myvhIWv2HPH9PIe%@qr=59Fq5c6C(JDx>qavh6=u??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZu2yjIZOb$%O&do

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_32x32x32_32x32x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_32x32x32_32x32x32_relu.cu
index 69ab4e102db4ca1372b6cc177fc6efc6d0107e78..18ba41d9d24e1f8c130b26ab47413336a0dcadfb 100644
GIT binary patch
delta 76
zcmey#w~cRuHS=UA=I+T?nCF^XS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m)D%QA-v0Cy!G8UO$Q

delta 164
zcmdnS_mgjfH8XpDPH9PIe%@qj=59Fq81vkEsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEBUdz(S1OOy}I^O^Q

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_32x64x32_32x64x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_32x64x32_32x64x32_hswish.cu
index eb4d164e9f163ad737e9d2aba221c5ca5afa2c67..b84d3f0bd87771e44975863c2644782d5d59130c 100644
GIT binary patch
delta 54
zcmey(w}WqkE%Rg-=I+VYnCDKeVv&+{D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-F%;A4if-9-W4nW

delta 164
zcmdnN_nU8nEi-$5PH9PIe%@qT=59Fq1oPZ_sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEBUeD6W1OO_EI_m%c

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_32x64x32_32x64x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_32x64x32_32x64x32_id.cu
index 576e237832cdb914ddea1b42c64d27ac69be9e68..83482b72c363f7fbe53d4345996a3140ac6410d4 100644
GIT binary patch
delta 56
zcmey$w~23qIrC&Y=I+TCnCHq{S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhru
MW0BZ=lVuJQ0B7|TtN;K2

delta 104
zcmdnQ_myvhIWv2HPH9PIe%@qr=59Fq5c6C(JDx>qavh6=u??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZu2yjIZOb$%O&do

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_32x64x32_32x64x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_32x64x32_32x64x32_relu.cu
index 32917d0042db7617b6fffa77bd88ab2f754230d7..f66b7b28f4bd9e3e801b18ecca24e2a508895264 100644
GIT binary patch
delta 76
zcmey#w~cRuHS=UA=I+T?nCF^XS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m)D%QA-v0Cy!G8UO$Q

delta 164
zcmdnS_mgjfH8XpDPH9PIe%@qj=59Fq81vkEsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEBUdz(S1OOy}I^O^Q

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_64x128x32_64x32x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_64x128x32_64x32x32_hswish.cu
index 367f6012d731606ffcd95b2905785dbdf4b7c483..6ba333db3b1351e0beca332bbbcdd32a267d75de 100644
GIT binary patch
delta 54
zcmeyvx07#!9rI*Y=I+VYndeTfW|5M0D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-TZ)M4if-A`xP?)

delta 164
zcmdnV_lIwT9W#4=PH9PIe%@p|=59FqB=g*QsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEB-oVny1OP3MI`IGi

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_64x128x32_64x32x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_64x128x32_64x32x32_id.cu
index e5ba977b18faf2a326b89835afe139284cb16cd9..9fd0f3e94b15f34fbcf08f3326100e023d13b9e8 100644
GIT binary patch
delta 56
zcmeyyx0!E)1@mNk=I+TCndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhru
MXOY-^i)9WI0BHUdvj6}9

delta 104
zcmdnY_l<9Z1v7hoPH9PIe%@pY=59FqF!NkEJAp-Nay^TLu??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZu4}OIZOb(ASLqv

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_64x128x32_64x32x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_64x128x32_64x32x32_relu.cu
index c832049c8becbf04d6a5767cbeb08aada92afceb..936502d62f0adb15311002e6fc0c719332ff80cc 100644
GIT binary patch
delta 76
zcmeyxx1Dc;4fAAY=I+T?ndh2YS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m)D$1;Zr0C<WXAOHXW

delta 164
zcmdna_ls|X4KsUwPH9PIe%@po=59FqIP=_ksJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEBUdPhO1OO+6I^_TW

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_64x32x32_64x32x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_64x32x32_64x32x32_hswish.cu
index 8a81f58713a332d785f5e84de8641d6a5c09a08d..f3f4977aa88ab56a0396040108e6fdd2f443058f 100644
GIT binary patch
delta 54
zcmey(w}WqkE%Rg-=I+VYnCDKeVv&+{D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-F%;A4if-9-W4nW

delta 164
zcmdnN_nU8nEi-$5PH9PIe%@qT=59Fq1oPZ_sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEBUeD6W1OO_EI_m%c

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_64x32x32_64x32x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_64x32x32_64x32x32_id.cu
index e391dd646f6040f53bb180d2bf83e4c005adc911..dfe25c5735625432cbb7d2c6e747eb56e565a1ad 100644
GIT binary patch
delta 56
zcmey$w~23qIrC&Y=I+TCnCHq{S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhru
MW0BZ=lVuJQ0B7|TtN;K2

delta 104
zcmdnQ_myvhIWv2HPH9PIe%@qr=59Fq5c6C(JDx>qavh6=u??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZu2yjIZOb$%O&do

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_64x32x32_64x32x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_64x32x32_64x32x32_relu.cu
index 27309b701e1460f4a00f2f1dd9233a24b05aa254..1d341e09de10ba09725a1f132f3baff3c400fcb9 100644
GIT binary patch
delta 76
zcmey#w~cRuHS=UA=I+T?nCF^XS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m)D%QA-v0Cy!G8UO$Q

delta 164
zcmdnS_mgjfH8XpDPH9PIe%@qj=59Fq81vkEsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEBUdz(S1OOy}I^O^Q

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_64x64x32_64x32x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_64x64x32_64x32x32_hswish.cu
index 026080c88204d95408fa592b3091cc96dc6ab669..ca51165edb82bd20c2dc98dc6241bbee3ed1ce25 100644
GIT binary patch
delta 54
zcmey(w}WqkE%Rg-=I+VYnCDKeVv&+{D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-F%;A4if-9-W4nW

delta 164
zcmdnN_nU8nEi-$5PH9PIe%@qT=59Fq1oPZ_sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEBUeD6W1OO_EI_m%c

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_64x64x32_64x32x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_64x64x32_64x32x32_id.cu
index 4deac474aac6e0e6de47e28f64325e71250240c0..eb0fcc2c3e423f10138e346c2036141f99f9cc03 100644
GIT binary patch
delta 56
zcmey$w~23qIrC&Y=I+TCnCHq{S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhru
MW0BZ=lVuJQ0B7|TtN;K2

delta 104
zcmdnQ_myvhIWv2HPH9PIe%@qr=59Fq5c6C(JDx>qavh6=u??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZu2yjIZOb$%O&do

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_64x64x32_64x32x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_1x1_64x64x32_64x32x32_relu.cu
index b0a42c147bbc697cb0d15d746c2a6c3b0b71fcf7..ebd3b131482246fc1dec4b924fe4046d457bc392 100644
GIT binary patch
delta 76
zcmey#w~cRuHS=UA=I+T?nCF^XS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m)D%QA-v0Cy!G8UO$Q

delta 164
zcmdnS_mgjfH8XpDPH9PIe%@qj=59Fq81vkEsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEBUdz(S1OOy}I^O^Q

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_32x128x32_32x64x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_32x128x32_32x64x32_hswish.cu
index 48036be9e0b7f96f188f1afa782f508b00914e18..c10268e670f8dec35d470e2edc488cfe3d9ec47b 100644
GIT binary patch
delta 54
zcmey(w}Wqk9rI*Y=I+VYndeTfW|5M0D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-TZ)MHWL6oj}<Tg

delta 164
zcmdnN_nU8n9W#4=PH9PIe%@p|=59FqB=g*QsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!-oVnq1OP1$I}HE;

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_32x128x32_32x64x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_32x128x32_32x64x32_id.cu
index 6d1a69d287cb30995e7139d4d6080ac0f842f18e..258c39ba946e20519cbdd60e3a7a066d009b51d4 100644
GIT binary patch
delta 56
zcmey$w~23q1@mNk=I+TCndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhru
MXOY-^i)A(w0BDXCumAu6

delta 104
zcmdnQ_myvh1v7hoPH9PIe%@pY=59FqF!NkEJAp-Nay^TLu??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZu4}O*-QYuP$lpH

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_32x128x32_32x64x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_32x128x32_32x64x32_relu.cu
index 91ce847d7c0a6f30c87ea54a3e76be8ff3fd3214..475b3585063f3cc3199f7eef182e3f9394097ce6 100644
GIT binary patch
delta 76
zcmey#w~cRu4fAAY=I+T?ndh2YS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m)D$1<A<0C(~p9RL6T

delta 164
zcmdnS_mgjf4KsUwPH9PIe%@po=59FqIP=_ksJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!UdPhG1OO)mI{^Ry

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_32x32x32_32x32x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_32x32x32_32x32x32_hswish.cu
index 6d4a698f7f9e66e88afa75556a565dcec5b86437..e178f8b208a506604de755e44e15d58b5aa7ecb2 100644
GIT binary patch
delta 54
zcmeyxx1Dc;E%Rg-=I+VYnCDKeVv&+{D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-F%;AHWL6nauq26

delta 164
zcmdna_ls|XEi-$5PH9PIe%@qT=59Fq1oPZ_sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!UeD6O1OO@uI|l#&

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_32x32x32_32x32x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_32x32x32_32x32x32_id.cu
index aeee09ac7e4bf9e66e7f8319788848958d48dd86..33ab33785e3550738d7c8ec4db87fb357b6e3e4b 100644
GIT binary patch
delta 56
zcmeyuw~=pyIrC&Y=I+TCnCHq{S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhru
MW0BZ=lVvs&0B402sQ>@~

delta 104
zcmdnU_l0kRIWv2HPH9PIe%@qr=59Fq5c6C(JDx>qavh6=u??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZu2yj*-QYr`z7cA

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_32x32x32_32x32x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_32x32x32_32x32x32_relu.cu
index d601e01df0257b3fb27ae757184487f0edb5abf3..9149da9ef00e7d427ab48318dd71894038d6f8b4 100644
GIT binary patch
delta 76
zcmeytx0P>$HS=UA=I+T?nCF^XS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m)D%QBk@0CtTY7XSbN

delta 164
zcmdnW_k(YPH8XpDPH9PIe%@qj=59Fq81vkEsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!Udz(K1OOxeI{N?s

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_32x64x32_32x64x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_32x64x32_32x64x32_hswish.cu
index fd235e0aab8e34cbb3112af60c674eaff31b41d5..447d3c9ed3155c4395623bd3dd079e048d2437f9 100644
GIT binary patch
delta 54
zcmeyxx1Dc;E%Rg-=I+VYnCDKeVv&+{D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-F%;AHWL6nauq26

delta 164
zcmdna_ls|XEi-$5PH9PIe%@qT=59Fq1oPZ_sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!UeD6O1OO@uI|l#&

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_32x64x32_32x64x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_32x64x32_32x64x32_id.cu
index ac180bd42d3ae8a8f9c97de1caae4d54199d02a2..aad75d8a98ce3c3fa6a8a838bf9a61fdfdb7d0b2 100644
GIT binary patch
delta 56
zcmeyuw~=pyIrC&Y=I+TCnCHq{S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhru
MW0BZ=lVvs&0B402sQ>@~

delta 104
zcmdnU_l0kRIWv2HPH9PIe%@qr=59Fq5c6C(JDx>qavh6=u??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZu2yj*-QYr`z7cA

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_32x64x32_32x64x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_32x64x32_32x64x32_relu.cu
index 352535b74739e956adb292b2200735c356386e8c..c70fcadfc2c0d2d54e468b372af196a0455d453c 100644
GIT binary patch
delta 76
zcmeytx0P>$HS=UA=I+T?nCF^XS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m)D%QBk@0CtTY7XSbN

delta 164
zcmdnW_k(YPH8XpDPH9PIe%@qj=59Fq81vkEsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!Udz(K1OOxeI{N?s

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_64x128x32_64x32x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_64x128x32_64x32x32_hswish.cu
index ccaca6e8ea2fc21d02487839ff0076b2b603ea2c..88d1ec1bc1bff88e6b6697783ac6b3658bdfb5b5 100644
GIT binary patch
delta 54
zcmey(w}Wqk9rI*Y=I+VYndeTfW|5M0D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-TZ)MHWL6oj}<Tg

delta 164
zcmdnN_nU8n9W#4=PH9PIe%@p|=59FqB=g*QsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!-oVnq1OP1$I}HE;

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_64x128x32_64x32x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_64x128x32_64x32x32_id.cu
index 35915b26724c1ec1b1f622f34c6359e9b8f1516d..b414833fb8ad26501e044ac0deec9ba6cc974b8f 100644
GIT binary patch
delta 56
zcmey$w~23q1@mNk=I+TCndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhru
MXOY-^i)A(w0BDXCumAu6

delta 104
zcmdnQ_myvh1v7hoPH9PIe%@pY=59FqF!NkEJAp-Nay^TLu??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZu4}O*-QYuP$lpH

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_64x128x32_64x32x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_64x128x32_64x32x32_relu.cu
index 1c4b308e8d2717e9c4177af54197f14eb6b939ad..aa3ee9cb39c9b5ae9d51f1c7f4486c2d1bede777 100644
GIT binary patch
delta 76
zcmey#w~cRu4fAAY=I+T?ndh2YS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m)D$1<A<0C(~p9RL6T

delta 164
zcmdnS_mgjf4KsUwPH9PIe%@po=59FqIP=_ksJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!UdPhG1OO)mI{^Ry

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_64x32x32_64x32x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_64x32x32_64x32x32_hswish.cu
index c97b5e918f1362e796388aafc53503ca016b112e..b8a5f97345a1cda9c83e7e0c0a823f020f896150 100644
GIT binary patch
delta 54
zcmeyxx1Dc;E%Rg-=I+VYnCDKeVv&+{D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-F%;AHWL6nauq26

delta 164
zcmdna_ls|XEi-$5PH9PIe%@qT=59Fq1oPZ_sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!UeD6O1OO@uI|l#&

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_64x32x32_64x32x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_64x32x32_64x32x32_id.cu
index 8327475ad764906b99f5c5499abb507f208ef675..45d4a9c13a083b99c864fac5f47315b8d7225717 100644
GIT binary patch
delta 56
zcmeyuw~=pyIrC&Y=I+TCnCHq{S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhru
MW0BZ=lVvs&0B402sQ>@~

delta 104
zcmdnU_l0kRIWv2HPH9PIe%@qr=59Fq5c6C(JDx>qavh6=u??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZu2yj*-QYr`z7cA

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_64x32x32_64x32x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_64x32x32_64x32x32_relu.cu
index 1b0ba002d56de04a7f64cc51c3be396caccb687a..736fa4a17e470136a4ee913a7262b56d14c40db6 100644
GIT binary patch
delta 76
zcmeytx0P>$HS=UA=I+T?nCF^XS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m)D%QBk@0CtTY7XSbN

delta 164
zcmdnW_k(YPH8XpDPH9PIe%@qj=59Fq81vkEsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!Udz(K1OOxeI{N?s

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_64x64x32_64x32x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_64x64x32_64x32x32_hswish.cu
index cabcf1ae65705db3bbe6701f06a589efed9ddbba..afb78f94fab35f6c3300588d42c3ff39e1cecd72 100644
GIT binary patch
delta 54
zcmeyxx1Dc;E%Rg-=I+VYnCDKeVv&+{D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-F%;AHWL6nauq26

delta 164
zcmdna_ls|XEi-$5PH9PIe%@qT=59Fq1oPZ_sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!UeD6O1OO@uI|l#&

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_64x64x32_64x32x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_64x64x32_64x32x32_id.cu
index 2f111fd96779010b8e6353ed1d8c33679a68cb46..20cd85303ec85187244f6373a0945d4f22cbaed0 100644
GIT binary patch
delta 56
zcmeyuw~=pyIrC&Y=I+TCnCHq{S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhru
MW0BZ=lVvs&0B402sQ>@~

delta 104
zcmdnU_l0kRIWv2HPH9PIe%@qr=59Fq5c6C(JDx>qavh6=u??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZu2yj*-QYr`z7cA

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_64x64x32_64x32x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_ncdiv32hw32_64x64x32_64x32x32_relu.cu
index b49faba5573568eac9a748fce52fbe5f28522d9a..c4d2fc3c75c96a5dd1627b21c9918554dd44b4af 100644
GIT binary patch
delta 76
zcmeytx0P>$HS=UA=I+T?nCF^XS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m)D%QBk@0CtTY7XSbN

delta 164
zcmdnW_k(YPH8XpDPH9PIe%@qj=59Fq81vkEsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!Udz(K1OOxeI{N?s

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_128x128x32_64x32x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_128x128x32_64x32x32_hswish.cu
index 357366755eee27f49c787f1e952fe6cd70d805d8..889ded1a86dff674d17c3cfbd447ee2807e94ac7 100644
GIT binary patch
delta 54
zcmaFNw}5YhA@gJl=8nn7ndeT<W|5M0D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-F$&%G7|tP_!O%E

delta 164
zcmZ3$_n2>kAv1e^PH9PIe%@q5<_<V}C-dBTsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!Ucl181OW6?I+y?e

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_128x128x32_64x32x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_128x128x32_64x32x32_id.cu
index c447d319401b553ec8317177424a026ec544b9ee..c7e9abbad953f1f575a1924dc3878d317b8ecbd2 100644
GIT binary patch
delta 56
zcmcc1H;ZqBHuGd7=8nk+ndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhq@
MXOY-^ie)kr0973oD*ylh

delta 104
zcmbQmcb9L2HZyyEPH9PIe%@ql<_<V}GxJ<HJAg%Maz2ZMu??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZgV@!WF`QSuO(Xm

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_128x128x32_64x32x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_128x128x32_64x32x32_relu.cu
index 898466e16fced80c4d1ec1e5b5518962b4185a0c..e0e825fa8b73dc5aa696cbdcdd2684db4ee0dfa5 100644
GIT binary patch
delta 76
zcmaFJH;-?FKJ#QV=8nlnndh2YS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m&t$1<4-0A0Nur~m)}

delta 164
zcmbQo_mFRcJ~MlMPH9PIe%@q#<_<V}JM-LnsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!p2yO_1OV<yI*b4S

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_128x32x32_64x32x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_128x32x32_64x32x32_hswish.cu
index 83dd420481fb91f3ea88a77d51cc16a0897cfd57..52c3111ba1434105707f79c62cf1734396f7a7c5 100644
GIT binary patch
delta 54
zcmaFFH=l2V0rO;Y=8nn7nCDK;Vv&+{D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-F%*9G7|tO+Z3b#

delta 164
zcmbQw_lR$U0W*7kPH9PIe%@pQ<_<V}2lL!|sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!p3l<21OV|)I+6eY

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_128x32x32_64x32x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_128x32x32_64x32x32_id.cu
index 5dc680a7104601b625c5b7e5d604e438a327484f..5ba1e043c05741f0dded11a01637be0e42b6c787 100644
GIT binary patch
delta 56
zcmcb^H<NFJ7V~67=8nk+nCHq{S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhq@
MW0BZ=l4UXz08|teBme*a

delta 104
zcmbQqcZY9-7BhQ(PH9PIe%@p)<_<V}6Z2d++n+^havqC>u??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZgU&UWF`QQS|w8e

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_128x32x32_64x32x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_128x32x32_64x32x32_relu.cu
index bbf92f637050a627267e742b1b3e971673fbed1d..2c66f76a0b34c0b0855b2d18c422048208bab1eb 100644
GIT binary patch
delta 76
zcmaFBH<xdN9`j^V=8nlnnCF^XS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m&t%QBe>09;rdq5uE@

delta 164
zcmbQs_keGM9y5D>PH9PIe%@p~<_<V}8}r<HsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!p3Bm}1OV$qI)(rM

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_128x64x32_64x32x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_128x64x32_64x32x32_hswish.cu
index faa8ea1883358e766fb291f3597b24546e68c882..bec858a1a4c9d7b7d31b934c1d505b2db1f7c78a 100644
GIT binary patch
delta 54
zcmaFFH=l2V0rO;Y=8nn7nCDK;Vv&+{D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-F%*9G7|tO+Z3b#

delta 164
zcmbQw_lR$U0W*7kPH9PIe%@pQ<_<V}2lL!|sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!p3l<21OV|)I+6eY

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_128x64x32_64x32x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_128x64x32_64x32x32_id.cu
index 7e4273a7ab9825f6075a546ad6027b83545356f7..cd029a0c1e2e42d9e1d8dee24c7af8a0c58d20f1 100644
GIT binary patch
delta 56
zcmcb^H<NFJ7V~67=8nk+nCHq{S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhq@
MW0BZ=l4UXz08|teBme*a

delta 104
zcmbQqcZY9-7BhQ(PH9PIe%@p)<_<V}6Z2d++n+^havqC>u??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZgU&UWF`QQS|w8e

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_128x64x32_64x32x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_128x64x32_64x32x32_relu.cu
index 7fa35c727afb0317167a6879ad8e10142bcd16f8..38dc02c2da8fff42eedd3a066f3abb119d593012 100644
GIT binary patch
delta 76
zcmaFBH<xdN9`j^V=8nlnnCF^XS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m&t%QBe>09;rdq5uE@

delta 164
zcmbQs_keGM9y5D>PH9PIe%@p~<_<V}8}r<HsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!p3Bm}1OV$qI)(rM

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_16x128x16_16x128x16_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_16x128x16_16x128x16_hswish.cu
index 8865080e356c1386e72b302a6160bf6c88d97ebe..7c3d55322ab5080ab498433166524dec23a299cb 100644
GIT binary patch
delta 54
zcmaFFH=l2VA@gJl=8nn7ndeT<W|5M0D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-F$&%5)%L^j1;H<

delta 164
zcmbQw_lR$UAv1e^PH9PIe%@q5<_<V}C-dBTsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C>L{4#
RC|KwyloXY2Ucge%1OW9LI@$mL

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_16x128x16_16x128x16_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_16x128x16_16x128x16_id.cu
index 6a93ab6171cf7b4163c863df6d4bc59a588bcab7..af59cd30a84a09d6e122d2dc5be1dde2984e4f83 100644
GIT binary patch
delta 56
zcmcb^H<NFJHuGd7=8nk+ndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhq@
MXOY-^ie(ZL0936NC;$Ke

delta 104
zcmbQqcZY9-HZyyEPH9PIe%@ql<_<V}GxJ<HJAg%Maz2ZMu??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZgV@!Bqjim-z8W8

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_16x128x16_16x128x16_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_16x128x16_16x128x16_relu.cu
index 6e01aaaf1e058b2ca59ba90945c422d4a2a9c9cc..0b447b2be320d85ac7fd08cf6dfec1ee798def84 100644
GIT binary patch
delta 76
zcmaFBH<xdNKJ#QV=8nlnndh2YS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m&t$1;ft09_>=r2qf`

delta 164
zcmbQs_keGMJ~MlMPH9PIe%@q#<_<V}JM-LnsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C>L{4#
RC|KwyloXY2p2t$p1OV?5I?ez9

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_16x64x8_16x64x8_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_16x64x8_16x64x8_hswish.cu
index ec78da97a6961366625a51984ef9af3d5f844b3d..d8832567fce7b006537b39c02f3d2d51765ba284 100644
GIT binary patch
delta 56
zcmcb|H=A#RF7sp)=8nmSndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhqD
MWRci>hGh~H09L;hHUIzs

delta 126
zcmbQucaLv_E;D<6PH9PIe%@qV<_<V}EAw1MsJN9?Nk&m>VoFj@esZ>zm2-Yx8B9kZ
pi-fTamjVzZmzLxt78hGtB_pIEYC|duQmw4Ak+j)u?qr$71OSWqE{XsE

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_16x64x8_16x64x8_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_16x64x8_16x64x8_id.cu
index 27df4ce87072239f0b236d7559e98958152071ea..9c5f1ec04920b8dd128718b2b2d6a615fd71aaa6 100644
GIT binary patch
delta 54
zcmcb?H<fRLI`d>b=8nmGndeSUWRa3}D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-F%E?5)%L(Y7}As

delta 105
zcmbQrcY|+(Ix~BIPH9PIe%@qt<_<V}J@Z^R+lNKU$j`aLBfKIcqbM~oB`GIAIXk#K
nv#Kg5)rLy}2$D-nauSP+t*nys^U7ext!I{CwA<XwGKmQQ=A<Rd

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_16x64x8_16x64x8_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_16x64x8_16x64x8_relu.cu
index b08e9563ae022006755111314c7d9c336f2ae7e2..86d700c60936c168383573180f02ed98959a6411 100644
GIT binary patch
delta 56
zcmcb^H<NFJHuGd7=8nk+ndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhq@
MXOY-^ie(ZL0936NC;$Ke

delta 104
zcmbQqcZY9-HZyyEPH9PIe%@ql<_<V}GxJ<HJAg%Maz2ZMu??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZgV@!Bqjim-z8W8

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_128x128x32_64x32x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_128x128x32_64x32x32_hswish.cu
index 88d5dad60c40dd205afc9a5a36faec356e87b800..63e1b1388bd203c4d5b735bdf10da46c9f59a184 100644
GIT binary patch
delta 54
zcmaFDw~%jxA@gJl=8nn7ndeT<W|5M0D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-F$&%3KIY-WE8Fd

delta 164
zcmZ3;_k?eQAv1e^PH9PIe%@q5<_<V}C-dBTsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEBUcl1G1OW8YI(z^C

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_128x128x32_64x32x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_128x128x32_64x32x32_id.cu
index 021eee4646893c907859dd2eabcb918dd8559be6..5a83822c1ae74828503f04703c4a8381945875c9 100644
GIT binary patch
delta 56
zcmcb|H=A#RHuGd7=8nk+ndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhq@
MXOY-^ie(BD09B0@E&u=k

delta 104
zcmbQucaLv_HZyyEPH9PIe%@ql<_<V}GxJ<HJAg%Maz2ZMu??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZgV@!6ea+Ye<fZ3

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_128x128x32_64x32x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_128x128x32_64x32x32_relu.cu
index 540f13ec1b2dafca49f05cd7cfe47c6eafe11e09..ad0340f1f9c882d57e415fa703d0f9005ef00784 100644
GIT binary patch
delta 76
zcmaFFH=l2VKJ#QV=8nlnndh2YS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m&t$1;Tp0A5ucs{jB1

delta 164
zcmbQw_lR$UJ~MlMPH9PIe%@q#<_<V}JM-LnsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEBp2yP21OV>II&c60

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_128x32x32_64x32x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_128x32x32_64x32x32_hswish.cu
index d9fa27c5b76ad06afcf38290012703e3acb01e95..3bed635e63223000f2b6bc3613454f12482b8dba 100644
GIT binary patch
delta 54
zcmaFNw}5Yh0rO;Y=8nn7nCDK;Vv&+{D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-F%*93KIY+M--<3

delta 164
zcmZ3$_n2>k0W*7kPH9PIe%@pQ<_<V}2lL!|sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEBp3l<A1OV~QI(7g6

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_128x32x32_64x32x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_128x32x32_64x32x32_id.cu
index bd9b9db793a6809e3194b2d3ed8d863ab97b2580..69d3d131d6b569e82f86a1ce742ff0327752d62e 100644
GIT binary patch
delta 56
zcmcc1H;ZqB7V~67=8nk+nCHq{S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhq@
MW0BZ=l4S}L091q(CjbBd

delta 104
zcmbQmcb9L27BhQ(PH9PIe%@p)<_<V}6Z2d++n+^havqC>u??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZgU&U6ea+WDkW9`

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_128x32x32_64x32x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_128x32x32_64x32x32_relu.cu
index 3581ac6bc2dde32d33d4359e0802f30a2299a0de..316a3072ad0a6cc75392ec1c8d01b4ea262f5f2c 100644
GIT binary patch
delta 76
zcmaFJH;-?F9`j^V=8nlnnCF^XS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m&t%QA%t09^1Lr2qf`

delta 164
zcmbQo_mFRc9y5D>PH9PIe%@p~<_<V}8}r<HsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEBp3Bn61OV&AI%)s_

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_128x64x32_64x32x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_128x64x32_64x32x32_hswish.cu
index 2a13f48cec1f12ef1bd99191a3e769e8478a05eb..1070d213b771e4ef8151e45fd958dc837c6ccf9e 100644
GIT binary patch
delta 54
zcmaFNw}5Yh0rO;Y=8nn7nCDK;Vv&+{D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-F%*93KIY+M--<3

delta 164
zcmZ3$_n2>k0W*7kPH9PIe%@pQ<_<V}2lL!|sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEBp3l<A1OV~QI(7g6

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_128x64x32_64x32x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_128x64x32_64x32x32_id.cu
index 33a1b07f187ab90c46cb1fb50680b123df5e7a48..7b591fa4c1e4b4995f3fee52f4ad62a827e1e386 100644
GIT binary patch
delta 56
zcmcc1H;ZqB7V~67=8nk+nCHq{S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhq@
MW0BZ=l4S}L091q(CjbBd

delta 104
zcmbQmcb9L27BhQ(PH9PIe%@p)<_<V}6Z2d++n+^havqC>u??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZgU&U6ea+WDkW9`

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_128x64x32_64x32x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_128x64x32_64x32x32_relu.cu
index 2e86a099f9c6fb891474bacae88c7982df2aaf32..4c7bcb84a7d1f19ed94494f4be191d1ba705faaf 100644
GIT binary patch
delta 76
zcmaFJH;-?F9`j^V=8nlnnCF^XS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m&t%QA%t09^1Lr2qf`

delta 164
zcmbQo_mFRc9y5D>PH9PIe%@p~<_<V}8}r<HsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEBp3Bn61OV&AI%)s_

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_16x128x16_16x128x16_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_16x128x16_16x128x16_hswish.cu
index ddea6211e503043d806664aafd7d4bec23bfe154..080c322296e623dd8fff50771a54e1e564fd4b34 100644
GIT binary patch
delta 54
zcmaFNw}5YhA@gJl=8nn7ndeT<W|5M0D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-F$&%G7|tP_!O%E

delta 164
zcmZ3$_n2>kAv1e^PH9PIe%@q5<_<V}C-dBTsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C>L{4#
RC|Kwyq$TETUcl181OW9$I<)`*

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_16x128x16_16x128x16_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_16x128x16_16x128x16_id.cu
index 0aafd7fe2ea4d678f887ead8980b8490b7d4ec97..8b82f9a8d397c3541260c27f4e96d5855c065c3f 100644
GIT binary patch
delta 56
zcmcc1H;ZqBHuGd7=8nk+ndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhq@
MXOY-^ie)kr0973oD*ylh

delta 104
zcmbQmcb9L2HZyyEPH9PIe%@ql<_<V}GxJ<HJAg%Maz2ZMu??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZgV@!WF`QSuO(Xm

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_16x128x16_16x128x16_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_16x128x16_16x128x16_relu.cu
index 0809ddee14cae1c5c3514fbb1dbf6ae4872bc26d..cdcd6b855b1304df71909911e11fddf6002dc7a4 100644
GIT binary patch
delta 76
zcmaFJH;-?FKJ#QV=8nlnndh2YS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m&t$1<4-0A0Nur~m)}

delta 164
zcmbQo_mFRcJ~MlMPH9PIe%@q#<_<V}JM-LnsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C>L{4#
RC|Kwyq$TETp2yO_1OV?mI;j8v

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_16x64x8_16x64x8_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_16x64x8_16x64x8_hswish.cu
index 56c22fc6cd1723306d249cd07ab7b40021474cb9..5c8ee5495beec2484f0171ff6b415a97701a9c1b 100644
GIT binary patch
delta 56
zcmcc5H-~S7F7sp)=8nmSndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhqD
MWRci>hGjAn09P*+IRF3v

delta 126
zcmbQkcb{*AE;D<6PH9PIe%@qV<_<V}EAw1MsJN9?Nk&m>VoFj@esZ>zm2-Yx8B9kZ
pi-fTamjVzZmzLxt78hGtB_pIEYC|duQmw4Ak+j)u?qr$F1OSZnE{y;H

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_16x64x8_16x64x8_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_16x64x8_16x64x8_id.cu
index 68a8a13bbeaf52764aff28e76e7d90f0a8d1462a..10bb5915ec1f1380b725249e1ba3eebe5c3bd209 100644
GIT binary patch
delta 54
zcmcb~H;r$DI`d>b=8nmGndeSUWRa3}D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-F%E?G7|tE))Zv`

delta 105
zcmbQncav{}Ix~BIPH9PIe%@qt<_<V}J@Z^R+lNKU$j`aLBfKIcqbM~oB`GIAIXk#K
nv#Kg5)rLy}2$D-nauSP+t*nys^U7ext!I{CwA<XwGMNbg=XNE|

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_16x64x8_16x64x8_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_16x64x8_16x64x8_relu.cu
index 8ca682f4f14d504a6604b93dc886c0146fe2f4bf..6a46eeac662fe5801f5eabb45071ea1b70441c18 100644
GIT binary patch
delta 56
zcmcc1H;ZqBHuGd7=8nk+ndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhq@
MXOY-^ie)kr0973oD*ylh

delta 104
zcmbQmcb9L2HZyyEPH9PIe%@ql<_<V}GxJ<HJAg%Maz2ZMu??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZgV@!WF`QSuO(Xm

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_32x128x32_32x64x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_32x128x32_32x64x32_hswish.cu
index b6923f4a59b272f9bbbc03069c646efcffa8e30d..90534a18eb74039f17612d06e97d1aa44d02c622 100644
GIT binary patch
delta 54
zcmaFNw}5Yh0rO;Y=8nn7nCDK;Vv&+{D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-F%*93KIY+M--<3

delta 164
zcmZ3$_n2>k0W*7kPH9PIe%@pQ<_<V}2lL!|sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEBp3l<A1OV~QI(7g6

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_32x128x32_32x64x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_32x128x32_32x64x32_id.cu
index 75acf999d7c056d31356a9ad5ba7bb686efb3163..9eb74ad125871338d5c2d4706c1add1a1b10b69f 100644
GIT binary patch
delta 56
zcmcc1H;ZqB7V~67=8nk+nCHq{S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhq@
MW0BZ=l4S}L091q(CjbBd

delta 104
zcmbQmcb9L27BhQ(PH9PIe%@p)<_<V}6Z2d++n+^havqC>u??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZgU&U6ea+WDkW9`

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_32x128x32_32x64x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_32x128x32_32x64x32_relu.cu
index 7248da5afefb5ebf42e73a20e54b21c78dfb797e..d470b22e271cba3f43ed56cd171ce3af597e22b1 100644
GIT binary patch
delta 76
zcmaFJH;-?F9`j^V=8nlnnCF^XS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m&t%QA%t09^1Lr2qf`

delta 164
zcmbQo_mFRc9y5D>PH9PIe%@p~<_<V}8}r<HsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEBp3Bn61OV&AI%)s_

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_32x32x32_32x32x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_32x32x32_32x32x32_hswish.cu
index b8c1e2789777ffabaeb72505fb9e4e42a3a68ea1..6b394e11c3aae570534d4d99ff89cf70ce39d09c 100644
GIT binary patch
delta 76
zcmaFFH=l2VKJ#QV=8nlnndh2YS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m&t$1;Tp0A5ucs{jB1

delta 164
zcmbQw_lR$UJ~MlMPH9PIe%@q#<_<V}JM-LnsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEBp2yP21OV>II&c60

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_32x32x32_32x32x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_32x32x32_32x32x32_id.cu
index 393b5a6cb0b4fd2e8043b3e0db6b54c218a37ada..652cea8833ded5dc4bcacad07c2d830bbfca0601 100644
GIT binary patch
delta 56
zcmcb^H<NFJCi7$i=8nnxndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhq@
MWs%r?f@KO508@JvAOHXW

delta 85
zcmbQqcZY9-CNq0}PH9PIe%@qF<_<V}BlBE1+mA(RaxRO6lns{x5G0qD<RlgsTUjON
T=as><Ze*5VwA<XuGKC2M1>zlQ

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_32x32x32_32x32x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_32x32x32_32x32x32_relu.cu
index c1ae6410e2759c092448213b16dac1026b5e29e1..5fc015576518cedcf963926db5ad7fdae603f4d7 100644
GIT binary patch
delta 56
zcmaFBH<xdNF7sp)=8nmSndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhqD
MWRci>hGhy909T(CJOBUy

delta 126
zcmbQs_keGME;D<6PH9PIe%@qV<_<V}EAw1MsJN9?Nk&m>VoFj@esZ>zm2-Yx8B9kZ
pi-fTamjVzZmzLxt78hGtB_pIEYC|duQmw4Ak+j)u?qr$51OSckE|35K

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_32x64x32_32x64x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_32x64x32_32x64x32_hswish.cu
index 2d327f97b792edbe47f677b0679a243d4b243238..9d743b45237a7202e57611f453de8ec734547fd2 100644
GIT binary patch
delta 76
zcmaFFH=l2VKJ#QV=8nlnndh2YS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m&t$1;Tp0A5ucs{jB1

delta 164
zcmbQw_lR$UJ~MlMPH9PIe%@q#<_<V}JM-LnsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEBp2yP21OV>II&c60

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_32x64x32_32x64x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_32x64x32_32x64x32_id.cu
index 33332c776f52be397055bfa28a50fff5bcbb093a..0921c3e9d7b0617d53287a068fe2e8459cacc53e 100644
GIT binary patch
delta 56
zcmcb^H<NFJCi7$i=8nnxndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhq@
MWs%r?f@KO508@JvAOHXW

delta 85
zcmbQqcZY9-CNq0}PH9PIe%@qF<_<V}BlBE1+mA(RaxRO6lns{x5G0qD<RlgsTUjON
T=as><Ze*5VwA<XuGKC2M1>zlQ

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_32x64x32_32x64x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_32x64x32_32x64x32_relu.cu
index 37369d3e85fd0620e08286d455a84fd97ac53878..a9c8dd404f0a772e503a9ed0af71f20f6ef0ffef 100644
GIT binary patch
delta 56
zcmaFBH<xdNF7sp)=8nmSndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhqD
MWRci>hGhy909T(CJOBUy

delta 126
zcmbQs_keGME;D<6PH9PIe%@qV<_<V}EAw1MsJN9?Nk&m>VoFj@esZ>zm2-Yx8B9kZ
pi-fTamjVzZmzLxt78hGtB_pIEYC|duQmw4Ak+j)u?qr$51OSckE|35K

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_64x128x32_64x32x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_64x128x32_64x32x32_hswish.cu
index 5ffd599ed326bee539aacadfd5cc040f63ade96f..23443ee28b2aca363b3a2cb677f5931ecddd4049 100644
GIT binary patch
delta 54
zcmaFNw}5Yh0rO;Y=8nn7nCDK;Vv&+{D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-F%*93KIY+M--<3

delta 164
zcmZ3$_n2>k0W*7kPH9PIe%@pQ<_<V}2lL!|sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEBp3l<A1OV~QI(7g6

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_64x128x32_64x32x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_64x128x32_64x32x32_id.cu
index 2fea2ac0f2524087e331043efe55da14da82ef1f..ac202d09355181577e4e7731d7202864fe72b857 100644
GIT binary patch
delta 56
zcmcc1H;ZqB7V~67=8nk+nCHq{S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhq@
MW0BZ=l4S}L091q(CjbBd

delta 104
zcmbQmcb9L27BhQ(PH9PIe%@p)<_<V}6Z2d++n+^havqC>u??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZgU&U6ea+WDkW9`

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_64x128x32_64x32x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_64x128x32_64x32x32_relu.cu
index 883b755d05451315eeec59e21d2d6cbd68e3a4c8..ce911fc9c925a663e0fc220ca368207abd3c3c89 100644
GIT binary patch
delta 76
zcmaFJH;-?F9`j^V=8nlnnCF^XS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m&t%QA%t09^1Lr2qf`

delta 164
zcmbQo_mFRc9y5D>PH9PIe%@p~<_<V}8}r<HsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEBp3Bn61OV&AI%)s_

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_64x32x32_64x32x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_64x32x32_64x32x32_hswish.cu
index 3cfe915fd6b12860cd3002d8de5258d8e01d87dd..1d2b47d20deabb03984dd2cf321c68511abb2452 100644
GIT binary patch
delta 76
zcmaFFH=l2VKJ#QV=8nlnndh2YS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m&t$1;Tp0A5ucs{jB1

delta 164
zcmbQw_lR$UJ~MlMPH9PIe%@q#<_<V}JM-LnsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEBp2yP21OV>II&c60

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_64x32x32_64x32x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_64x32x32_64x32x32_id.cu
index 6b711dc10e92661add3e0da67329bc5b73a544bb..d38622ac29bf2b62a9e81eb4e89c3ef3e7f75ce7 100644
GIT binary patch
delta 56
zcmcb^H<NFJCi7$i=8nnxndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhq@
MWs%r?f@KO508@JvAOHXW

delta 85
zcmbQqcZY9-CNq0}PH9PIe%@qF<_<V}BlBE1+mA(RaxRO6lns{x5G0qD<RlgsTUjON
T=as><Ze*5VwA<XuGKC2M1>zlQ

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_64x32x32_64x32x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_64x32x32_64x32x32_relu.cu
index c63406461b48c156fdcde77b06f66692ddcf2f2a..3845f7faf1ac6d6061e9269afbdc88ceb2afbf5c 100644
GIT binary patch
delta 56
zcmaFBH<xdNF7sp)=8nmSndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhqD
MWRci>hGhy909T(CJOBUy

delta 126
zcmbQs_keGME;D<6PH9PIe%@qV<_<V}EAw1MsJN9?Nk&m>VoFj@esZ>zm2-Yx8B9kZ
pi-fTamjVzZmzLxt78hGtB_pIEYC|duQmw4Ak+j)u?qr$51OSckE|35K

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_64x64x32_64x32x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_64x64x32_64x32x32_hswish.cu
index e1a4fac14886a77e8b3f3a79e213e63b5423a90f..2948cbbc43f2fc6a355d78e3ba9c5001b6f4852d 100644
GIT binary patch
delta 76
zcmaFFH=l2VKJ#QV=8nlnndh2YS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m&t$1;Tp0A5ucs{jB1

delta 164
zcmbQw_lR$UJ~MlMPH9PIe%@q#<_<V}JM-LnsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD5NEBp2yP21OV>II&c60

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_64x64x32_64x32x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_64x64x32_64x32x32_id.cu
index 8b2c0d5fdaa053272df6d35e97c63d824a30b33e..0f75f986896adee4307127b049b7e05dec306361 100644
GIT binary patch
delta 56
zcmcb^H<NFJCi7$i=8nnxndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhq@
MWs%r?f@KO508@JvAOHXW

delta 85
zcmbQqcZY9-CNq0}PH9PIe%@qF<_<V}BlBE1+mA(RaxRO6lns{x5G0qD<RlgsTUjON
T=as><Ze*5VwA<XuGKC2M1>zlQ

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_64x64x32_64x32x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_1x1_64x64x32_64x32x32_relu.cu
index eeb3c5c2510bae23df61ae56f3854146b548027f..4aa41a48f0fd98809c25864e2a452ddce8e5ae5e 100644
GIT binary patch
delta 56
zcmaFBH<xdNF7sp)=8nmSndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhqD
MWRci>hGhy909T(CJOBUy

delta 126
zcmbQs_keGME;D<6PH9PIe%@qV<_<V}EAw1MsJN9?Nk&m>VoFj@esZ>zm2-Yx8B9kZ
pi-fTamjVzZmzLxt78hGtB_pIEYC|duQmw4Ak+j)u?qr$51OSckE|35K

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_32x128x32_32x64x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_32x128x32_32x64x32_hswish.cu
index b04889b88617c7be2ab85ea843c26a6d02ab5c48..3a21872acb2ed3e758151676446575c484c3d2fc 100644
GIT binary patch
delta 54
zcmaFFH=l2V0rO;Y=8nn7nCDK;Vv&+{D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-F%*9G7|tO+Z3b#

delta 164
zcmbQw_lR$U0W*7kPH9PIe%@pQ<_<V}2lL!|sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!p3l<21OV|)I+6eY

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_32x128x32_32x64x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_32x128x32_32x64x32_id.cu
index 03d817af706a7be29991a33e27d49c846824237f..86e0bfb5cafa82e199e4328192925a63b99ae23a 100644
GIT binary patch
delta 56
zcmcb^H<NFJ7V~67=8nk+nCHq{S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhq@
MW0BZ=l4UXz08|teBme*a

delta 104
zcmbQqcZY9-7BhQ(PH9PIe%@p)<_<V}6Z2d++n+^havqC>u??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZgU&UWF`QQS|w8e

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_32x128x32_32x64x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_32x128x32_32x64x32_relu.cu
index 86009742c295cec0a4370cbfab0d1c2aca9e298d..265d639cb9d0c8b7c02e353b9795296845253c6e 100644
GIT binary patch
delta 76
zcmaFBH<xdN9`j^V=8nlnnCF^XS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m&t%QBe>09;rdq5uE@

delta 164
zcmbQs_keGM9y5D>PH9PIe%@p~<_<V}8}r<HsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!p3Bm}1OV$qI)(rM

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_32x32x32_32x32x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_32x32x32_32x32x32_hswish.cu
index 22a9bfb1ff672040957668ab385c8acc9a79cb5d..da540ae8f550ce2fb03bdb667ce12c3b066bbcdd 100644
GIT binary patch
delta 76
zcmaFJH;-?FKJ#QV=8nlnndh2YS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m&t$1<4-0A0Nur~m)}

delta 164
zcmbQo_mFRcJ~MlMPH9PIe%@q#<_<V}JM-LnsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!p2yO_1OV<yI*b4S

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_32x32x32_32x32x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_32x32x32_32x32x32_id.cu
index e1e200110916364e9550cbbeeb29693a47b06aa5..74c608a870bd0abf399e571950bce55aba828dcc 100644
GIT binary patch
delta 56
zcmcc3H-m43Ci7$i=8nnxndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhq@
MWs%r?f@Lxj08<MU9RL6T

delta 85
zcmbQicbjj6CNq0}PH9PIe%@qF<_<V}BlBE1+mA(RaxRO6lns{x5G0qD<RlgsTUjON
T=as><Ze*5VwA<XuGMNbg1vnjN

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_32x32x32_32x32x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_32x32x32_32x32x32_relu.cu
index d9b963e009a5753eaa32900910d6eb0e759f53f1..c8e8d91562d5b5c9f151e3692eb22283120e431e 100644
GIT binary patch
delta 56
zcmcc5H-~S7F7sp)=8nmSndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhqD
MWRci>hGjAn09P*+IRF3v

delta 126
zcmbQkcb{*AE;D<6PH9PIe%@qV<_<V}EAw1MsJN9?Nk&m>VoFj@esZ>zm2-Yx8B9kZ
pi-fTamjVzZmzLxt78hGtB_pIEYC|duQmw4Ak+j)u?qr$F1OSZnE{y;H

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_32x64x32_32x64x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_32x64x32_32x64x32_hswish.cu
index 4ad86a3f8c88e7712ab9c06b036545c1f79b2975..0dca550deaf741a6a7cb666d1327fcca124f953b 100644
GIT binary patch
delta 76
zcmaFJH;-?FKJ#QV=8nlnndh2YS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m&t$1<4-0A0Nur~m)}

delta 164
zcmbQo_mFRcJ~MlMPH9PIe%@q#<_<V}JM-LnsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!p2yO_1OV<yI*b4S

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_32x64x32_32x64x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_32x64x32_32x64x32_id.cu
index 6f7648c3f4f377b8a18b18f6d75acb77e0fc8cd4..989d55927abec35f3348258c613fdc03c3099e77 100644
GIT binary patch
delta 56
zcmcc3H-m43Ci7$i=8nnxndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhq@
MWs%r?f@Lxj08<MU9RL6T

delta 85
zcmbQicbjj6CNq0}PH9PIe%@qF<_<V}BlBE1+mA(RaxRO6lns{x5G0qD<RlgsTUjON
T=as><Ze*5VwA<XuGMNbg1vnjN

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_32x64x32_32x64x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_32x64x32_32x64x32_relu.cu
index f3f8802bc868137627c790a6efe2b1a2d1b383b5..684e5777d5fb8be485c47eebcec73a80c5317147 100644
GIT binary patch
delta 56
zcmcc5H-~S7F7sp)=8nmSndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhqD
MWRci>hGjAn09P*+IRF3v

delta 126
zcmbQkcb{*AE;D<6PH9PIe%@qV<_<V}EAw1MsJN9?Nk&m>VoFj@esZ>zm2-Yx8B9kZ
pi-fTamjVzZmzLxt78hGtB_pIEYC|duQmw4Ak+j)u?qr$F1OSZnE{y;H

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_64x128x32_64x32x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_64x128x32_64x32x32_hswish.cu
index 497591db5b63aadf94eb1a9ac915682b9d774fb6..0d47d1cd480d7a2e78f524bbf7df2770493fa94d 100644
GIT binary patch
delta 54
zcmaFFH=l2V0rO;Y=8nn7nCDK;Vv&+{D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-F%*9G7|tO+Z3b#

delta 164
zcmbQw_lR$U0W*7kPH9PIe%@pQ<_<V}2lL!|sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!p3l<21OV|)I+6eY

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_64x128x32_64x32x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_64x128x32_64x32x32_id.cu
index b06de5af17231728ee878556a4f15ee8cc4b240d..d8927fbbc2b0285986787dec124e29f49f11f6ad 100644
GIT binary patch
delta 56
zcmcb^H<NFJ7V~67=8nk+nCHq{S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhq@
MW0BZ=l4UXz08|teBme*a

delta 104
zcmbQqcZY9-7BhQ(PH9PIe%@p)<_<V}6Z2d++n+^havqC>u??335G0qD<RlgsTUjON
e=aoUVT3I;*IU$t=sa96mASP7YZgU&UWF`QQS|w8e

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_64x128x32_64x32x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_64x128x32_64x32x32_relu.cu
index 28aa9131c904076a517a1c862a6a0c52819cf0ea..5115a05b3baf9e26405d932b6397096e2f478675 100644
GIT binary patch
delta 76
zcmaFBH<xdN9`j^V=8nlnnCF^XS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m&t%QBe>09;rdq5uE@

delta 164
zcmbQs_keGM9y5D>PH9PIe%@p~<_<V}8}r<HsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!p3Bm}1OV$qI)(rM

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_64x32x32_64x32x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_64x32x32_64x32x32_hswish.cu
index 87e5aae00b259c8f4d3020ceb0d58e89b638a449..d448a3e6a7effee301451ac74969181217d4a501 100644
GIT binary patch
delta 76
zcmaFJH;-?FKJ#QV=8nlnndh2YS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m&t$1<4-0A0Nur~m)}

delta 164
zcmbQo_mFRcJ~MlMPH9PIe%@q#<_<V}JM-LnsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!p2yO_1OV<yI*b4S

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_64x32x32_64x32x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_64x32x32_64x32x32_id.cu
index 0f5d5d874d9904abdb604a79578f7438f55c2c2e..8e71d8caaadc77bd34b4a5857883e38844e5962e 100644
GIT binary patch
delta 56
zcmcc3H-m43Ci7$i=8nnxndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhq@
MWs%r?f@Lxj08<MU9RL6T

delta 85
zcmbQicbjj6CNq0}PH9PIe%@qF<_<V}BlBE1+mA(RaxRO6lns{x5G0qD<RlgsTUjON
T=as><Ze*5VwA<XuGMNbg1vnjN

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_64x32x32_64x32x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_64x32x32_64x32x32_relu.cu
index daadee35de348c3587239fbf8aa733d05a9f2275..d09c617156da4a39b2e23e34a2d9176978bb079a 100644
GIT binary patch
delta 56
zcmcc5H-~S7F7sp)=8nmSndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhqD
MWRci>hGjAn09P*+IRF3v

delta 126
zcmbQkcb{*AE;D<6PH9PIe%@qV<_<V}EAw1MsJN9?Nk&m>VoFj@esZ>zm2-Yx8B9kZ
pi-fTamjVzZmzLxt78hGtB_pIEYC|duQmw4Ak+j)u?qr$F1OSZnE{y;H

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_64x64x32_64x32x32_hswish.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_64x64x32_64x32x32_hswish.cu
index 6f19370ee918a046116e252562f90a5e81b9ced7..7aad8acb93a4fb23d9f715dd0ad62ba962dd2eae 100644
GIT binary patch
delta 76
zcmaFJH;-?FKJ#QV=8nlnndh2YS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rnxl
YS9pY1gdmgzmuFU0<)m&t$1<4-0A0Nur~m)}

delta 164
zcmbQo_mFRcJ~MlMPH9PIe%@q#<_<V}JM-LnsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_r`!
RC>WaQD3la!p2yO_1OV<yI*b4S

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_64x64x32_64x32x32_id.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_64x64x32_64x32x32_id.cu
index ca6f680bb626f62be1ba9d4d54fff100519e2d96..3774df1a549847c35e83ad3ba6333e7b53dca583 100644
GIT binary patch
delta 56
zcmcc3H-m43Ci7$i=8nnxndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhq@
MWs%r?f@Lxj08<MU9RL6T

delta 85
zcmbQicbjj6CNq0}PH9PIe%@qF<_<V}BlBE1+mA(RaxRO6lns{x5G0qD<RlgsTUjON
T=as><Ze*5VwA<XuGMNbg1vnjN

diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_64x64x32_64x32x32_relu.cu b/dnn/src/cuda/conv_bias/int8/kimpl/conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nchw_64x64x32_64x32x32_relu.cu
index b7eccd8c2cf6f67803179279bdceebd932c9e1b4..e23ea6f5cd0f41e18089445884d46084681910d4 100644
GIT binary patch
delta 56
zcmcc5H-~S7F7sp)=8nmSndiz|S(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW<rhqD
MWRci>hGjAn09P*+IRF3v

delta 126
zcmbQkcb{*AE;D<6PH9PIe%@qV<_<V}EAw1MsJN9?Nk&m>VoFj@esZ>zm2-Yx8B9kZ
pi-fTamjVzZmzLxt78hGtB_pIEYC|duQmw4Ak+j)u?qr$F1OSZnE{y;H

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_128x128x64_64x64x64_hswish.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_128x128x64_64x64x64_hswish.cu
index 42029b8aa35ec24d06f41307a886e3615d1a3e5c..b58174157fb3e7ac712eaee2cbabb5af6b32ecc0 100644
GIT binary patch
delta 60
zcmZqSKghSiiFvXobNA%)EE1DHGk3~cS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW
Q<rhqz%ObJ)1<M>J0I^mV3;+NC

delta 85
zcmX@e*TTQSiJ3h=r?ez9KX0-Vb2prQmU$tZoyQ_IxtB#k%7#k;2$D-nauSP+t*nys
T^U7da&oWCe+HGFSGKUEO8Ok0M

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_128x128x64_64x64x64_id.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_128x128x64_64x64x64_id.cu
index 9de29a4ae162050ba5779033d11a5890187fd739..7a0612f839a74299146e75f1d4595fe4e8be856c 100644
GIT binary patch
delta 54
zcmeyzx0`Q+4fAAY=I+T?nHNs3Ws#C~D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-Ta7U4if-Bu@yN0

delta 145
zcmdnZ_m6Lb4KsUwPH9PIe%@po=59FqIP=0BsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG92Bp3gFe2>|_|
BH<SPX

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_128x128x64_64x64x64_relu.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_128x128x64_64x64x64_relu.cu
index a91bd76d23da062adce39f7a92d860848062ea73..cedd1f1b9de836e5e091d226955f14cf06d6e598 100644
GIT binary patch
delta 58
zcmZqT-_N(ffqAk!bNA%4EE1DHGIvgHWRa3}D=5k@@N=&42(JjqC`wICNy^Dj&JHfm
Otg6aM-TaJY4if-<85U9i

delta 124
zcmdnb*TlcUftfu&r?ez9KX0-Fb2prQnt36dox>vK?dM$K5nd6JQIwjPl9ZF5oE==A
ySyh#jYQv=f1j(f(If=!^R#wURd1X-JtgM`YoRG?bR4c1&5ECkHw|Ozk93}uO5H2ME

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_128x256x64_64x64x64_hswish.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_128x256x64_64x64x64_hswish.cu
index a156009f986b64346cea5af43ed9d5c0a037d224..d68997d0c397a0585230ccc969758bf9c5e528bd 100644
GIT binary patch
delta 60
zcmZqSKghSiiFvXobNA%)EE1DHGk3~cS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW
Q<rhqz%ObJ)1<M>J0I^mV3;+NC

delta 85
zcmX@e*TTQSiJ3h=r?ez9KX0-Vb2prQmU$tZoyQ_IxtB#k%7#k;2$D-nauSP+t*nys
T^U7da&oWCe+HGFSGKUEO8Ok0M

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_128x256x64_64x64x64_id.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_128x256x64_64x64x64_id.cu
index 645fb0f6e2992c4eaa575aed551dea21b174d1e4..d707e5f7f32a0d2d962a38b3f2ea3c1c579e02da 100644
GIT binary patch
delta 54
zcmeyzx0`Q+4fAAY=I+T?nHNs3Ws#C~D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-Ta7U4if-Bu@yN0

delta 145
zcmdnZ_m6Lb4KsUwPH9PIe%@po=59FqIP=0BsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG92Bp3gFe2>|_|
BH<SPX

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_128x256x64_64x64x64_relu.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_128x256x64_64x64x64_relu.cu
index b6451a72c50b4a82822f6c63ae761b89014cf5bd..0eb96352314fe976fa8b8076d434ef7aae69dc21 100644
GIT binary patch
delta 58
zcmZqT-_N(ffqAk!bNA%4EE1DHGIvgHWRa3}D=5k@@N=&42(JjqC`wICNy^Dj&JHfm
Otg6aM-TaJY4if-<85U9i

delta 124
zcmdnb*TlcUftfu&r?ez9KX0-Fb2prQnt36dox>vK?dM$K5nd6JQIwjPl9ZF5oE==A
ySyh#jYQv=f1j(f(If=!^R#wURd1X-JtgM`YoRG?bR4c1&5ECkHw|Ozk93}uO5H2ME

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_128x64x64_64x32x64_hswish.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_128x64x64_64x32x64_hswish.cu
index 0d5dc8c483833d735e88914be72604d60ef3beef..3dcf049746066b4fa6ddde9b8f4d143d15e6e23c 100644
GIT binary patch
delta 60
zcmZqXKft%ak$JKQbNA$PEE1DHF?Y&aS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW
Q<rhqz!y>WyIm;X-0I&@g1ONa4

delta 85
zcmX@W*UZ1ck(oU|r?ez9KX0-lb2prQhIt{Joy#IMxraqU%7#k;2$D-nauSP+t*nys
T^U7da&oE0c+HGFKGKUEO7w{en

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_128x64x64_64x32x64_id.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_128x64x64_64x32x64_id.cu
index dd1e9c2eadba5541d2c9d858f450efa3b7c9bf53..e9dc07cc839982d8f9ead4464f1e60d345222162 100644
GIT binary patch
delta 54
zcmey%w~KFsHS=UA=I+T?m={j2VUdz`D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-TaVc4if-Aloc`n

delta 145
zcmdnR_m^*jH8XpDPH9PIe%@qj=59Fq81up$sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG92Bp2sqW2>|-~
BH;w=R

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_128x64x64_64x32x64_relu.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_128x64x64_64x32x64_relu.cu
index 96f54e5dda5c78a5d273f4a75036f543849272ad..c5653626b65ac71e758e95daa6e443686b60a0cd 100644
GIT binary patch
delta 58
zcmZqV-^aJXo_VqxbNA#kEE1DHFn3OFV3Cq_D=5k@@N=&42(JjqC`wICNy^Dj&JHfm
Otg6aM-Tahg4if--z7|RV

delta 124
zcmdnT*T}!Yo|!#Ar?ez9KX0-<b2prQig_WNoy{WU?dM$K5nd6JQIwjPl9ZF5oE==A
ySyh#jYQv=f1j(f(If=!^R#wURd1X-JtgM`YoRG?bR4c1&5ECkHw|NoE93}uLe=Z&X

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_1x1_128x128x64_64x64x64_hswish.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_1x1_128x128x64_64x64x64_hswish.cu
index cdef6ce72ffb0be3c4347887d41f173d50878707..f51466f51b5e29f5f777549346b1021ca2531fa7 100644
GIT binary patch
delta 60
zcmZqWKg74eiFvXobNA%)EE1DHGk3~cS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW
Q<rhqz%ObJ)1<PC}0I|*&4*&oF

delta 85
zcmX@a*UG=aiJ3h=r?ez9KX0-Vb2prQmU$tZoyQ_IxtB#k%7#k;2$D-nauSP+t*nys
T^U7da&oWCe+HGFSGM5Pe8gw2P

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_1x1_128x128x64_64x64x64_id.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_1x1_128x128x64_64x64x64_id.cu
index b576c7969aa95a47455cbd095fc1b25ef5ab1879..7ae41fd2c808da298c662d3c04ae2f9d0f9df5cf 100644
GIT binary patch
delta 54
zcmey*w})?o4fAAY=I+T?nHNs3Ws#C~D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-Ta7UE)xJi9ThwP

delta 145
zcmdnP_n&Wr4KsUwPH9PIe%@po=59FqIP=0BsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG92Bp3gFu2>|}W
BH<tha

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_1x1_128x128x64_64x64x64_relu.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_1x1_128x128x64_64x64x64_relu.cu
index 7d6d5092c0feaafa20d0f0c9d3d690ebf5cedade..7e28c5bd84d500dbdec6e091be9bb01788947ac3 100644
GIT binary patch
delta 58
zcmZqXKft%afqAk!bNA%4EE1DHGIvgHWRa3}D=5k@@N=&42(JjqC`wICNy^Dj&JHfm
Otg6aM-TaJYE)xKKjTTh^

delta 124
zcmX@W*UZ1cftfu&r?ez9KX0-Fb2prQnt36dox>vK?dM$K5nd6JQIwjPl9ZF5oE==A
ySyh#jYQv=f1j(f(If=!^R#wURd1X-JtgM`YoRG?bR4c1&5ECkHw|OzkTqXc42rei9

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_1x1_128x256x64_64x64x64_hswish.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_1x1_128x256x64_64x64x64_hswish.cu
index b504399aa390894531aaaf32ea4af8d806c00176..bc75f547787bbdd022490a7a9c14a0862c1a59e3 100644
GIT binary patch
delta 60
zcmZqWKg74eiFvXobNA%)EE1DHGk3~cS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW
Q<rhqz%ObJ)1<PC}0I|*&4*&oF

delta 85
zcmX@a*UG=aiJ3h=r?ez9KX0-Vb2prQmU$tZoyQ_IxtB#k%7#k;2$D-nauSP+t*nys
T^U7da&oWCe+HGFSGM5Pe8gw2P

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_1x1_128x256x64_64x64x64_id.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_1x1_128x256x64_64x64x64_id.cu
index e1f9eaa29d169b9ed1d5f69a06540d0522314500..caeaba014fa86f49c0bf31e497af737066df17d2 100644
GIT binary patch
delta 54
zcmey*w})?o4fAAY=I+T?nHNs3Ws#C~D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-Ta7UE)xJi9ThwP

delta 145
zcmdnP_n&Wr4KsUwPH9PIe%@po=59FqIP=0BsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG92Bp3gFu2>|}W
BH<tha

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_1x1_128x256x64_64x64x64_relu.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_1x1_128x256x64_64x64x64_relu.cu
index 03f90194dcd64d861f26eccbc8efca84d4fa6bd6..af6c2a3c01119c3fd4c0c4863286b198ab705a8b 100644
GIT binary patch
delta 58
zcmZqXKft%afqAk!bNA%4EE1DHGIvgHWRa3}D=5k@@N=&42(JjqC`wICNy^Dj&JHfm
Otg6aM-TaJYE)xKKjTTh^

delta 124
zcmX@W*UZ1cftfu&r?ez9KX0-Fb2prQnt36dox>vK?dM$K5nd6JQIwjPl9ZF5oE==A
ySyh#jYQv=f1j(f(If=!^R#wURd1X-JtgM`YoRG?bR4c1&5ECkHw|OzkTqXc42rei9

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_1x1_128x64x64_64x32x64_hswish.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_1x1_128x64x64_64x32x64_hswish.cu
index 006f790a8868ae36f142ab78f1b25a9631950d77..32294cd5ae4b52c358d04bb58aa2a0fbcc40f68c 100644
GIT binary patch
delta 60
zcmZqSKghSik$JKQbNA$PEE1DHF?Y&aS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW
Q<rhqz!y>WyIm=uo0I-D@2LJ#7

delta 85
zcmX@e*TTQSk(oU|r?ez9KX0-lb2prQhIt{Joy#IMxraqU%7#k;2$D-nauSP+t*nys
T^U7da&oE0c+HGFKGM5Pe7@8gq

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_1x1_128x64x64_64x32x64_id.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_1x1_128x64x64_64x32x64_id.cu
index 0906068e26f8fbecc26350bae57d32e5bfe7b59b..34f00363d3efd00dfba02b64420ece4462f8dd0a 100644
GIT binary patch
delta 54
zcmeyzx0`Q+HS=UA=I+T?m={j2VUdz`D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-TaVcE)xJh02MU=

delta 145
zcmdnZ_m6LbH8XpDPH9PIe%@qj=59Fq81up$sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG92Bp2sqm2>|>Y
BH<17U

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_1x1_128x64x64_64x32x64_relu.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_1x1_128x64x64_64x32x64_relu.cu
index adcbfbeaf89988dc7ea7a100dfb5fc5715ebdc95..eea42928ebdc61619cfe57c61b1becd3632b4fda 100644
GIT binary patch
delta 58
zcmZqT-_N(fo_VqxbNA#kEE1DHFn3OFV3Cq_D=5k@@N=&42(JjqC`wICNy^Dj&JHfm
Otg6aM-TahgE)xKJG8Rn$

delta 124
zcmdnb*TlcUo|!#Ar?ez9KX0-<b2prQig_WNoy{WU?dM$K5nd6JQIwjPl9ZF5oE==A
ySyh#jYQv=f1j(f(If=!^R#wURd1X-JtgM`YoRG?bR4c1&5ECkHw|NoETqXc1cP=3S

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_1x1_256x128x64_64x64x64_hswish.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_1x1_256x128x64_64x64x64_hswish.cu
index 31bac99f39cac45836f3f002d9ccea48ebf2fee1..7f14fddceb7c3f75dc2f061c63f6ad966985dc1b 100644
GIT binary patch
delta 60
zcmZqWKg74eiFvXobNA%)EE1DHGk3~cS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW
Q<rhqz%ObJ)1<PC}0I|*&4*&oF

delta 85
zcmX@a*UG=aiJ3h=r?ez9KX0-Vb2prQmU$tZoyQ_IxtB#k%7#k;2$D-nauSP+t*nys
T^U7da&oWCe+HGFSGM5Pe8gw2P

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_1x1_256x128x64_64x64x64_id.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_1x1_256x128x64_64x64x64_id.cu
index d35a2be8cf91a6fcf1c77ba0f64c7679895427ed..ddc4869dd9794b9c4f637b4d435a47be768d22db 100644
GIT binary patch
delta 54
zcmey*w})?o4fAAY=I+T?nHNs3Ws#C~D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-Ta7UE)xJi9ThwP

delta 145
zcmdnP_n&Wr4KsUwPH9PIe%@po=59FqIP=0BsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG92Bp3gFu2>|}W
BH<tha

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_1x1_256x128x64_64x64x64_relu.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_1x1_256x128x64_64x64x64_relu.cu
index 272c72492732bba037872d3ca23ad4bebf0470ed..b7d0aba981c370128d30bea2e33fa57d6e7440e5 100644
GIT binary patch
delta 58
zcmZqXKft%afqAk!bNA%4EE1DHGIvgHWRa3}D=5k@@N=&42(JjqC`wICNy^Dj&JHfm
Otg6aM-TaJYE)xKKjTTh^

delta 124
zcmX@W*UZ1cftfu&r?ez9KX0-Fb2prQnt36dox>vK?dM$K5nd6JQIwjPl9ZF5oE==A
ySyh#jYQv=f1j(f(If=!^R#wURd1X-JtgM`YoRG?bR4c1&5ECkHw|OzkTqXc42rei9

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_1x1_32x64x64_32x16x64_hswish.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_1x1_32x64x64_32x16x64_hswish.cu
index 7315cecb3d71f45cef14843279a035f30a9674df..dee9e38c334b279dc82d34f9e205c80f7c4b4b13 100644
GIT binary patch
delta 58
zcmZqXKft%afqAk!bNA%4EE1DHGIvgHWRa3}D=5k@@N=&42(JjqC`wICNy^Dj&JHfm
Otg6aM-TaJYE)xKKjTTh^

delta 124
zcmX@W*UZ1cftfu&r?ez9KX0-Fb2prQnt36dox>vK?dM$K5nd6JQIwjPl9ZF5oE==A
ySyh#jYQv=f1j(f(If=!^R#wURd1X-JtgM`YoRG?bR4c1&5ECkHw|OzkTqXc42rei9

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_1x1_32x64x64_32x16x64_id.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_1x1_32x64x64_32x16x64_id.cu
index 9d4b8466913c723a7bb5d5acc983666d78410fb8..33939b7bbf45d9db56aed932157a1ad2a1041200 100644
GIT binary patch
delta 54
zcmey%w~KFs74u|A=I+UtnHNs3W|5M0D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-TZ)ME)xJf;}tFd

delta 164
zcmdnR_m^*j6*GH&PH9PIe%@p&=59FqDD%R4sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_nYQ
ML2%mU4J=(u04VG_z5oCK

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_1x1_32x64x64_32x16x64_relu.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_1x1_32x64x64_32x16x64_relu.cu
index 6a87dc00d8be6f116c273c56c72b224c08fc4aba..8926a9d6a5e51c34a9a0c731246e020d644ee085 100644
GIT binary patch
delta 54
zcmZqV-^aJXj(M^xbNA%y%nK*ivq;Ih6%^$c_&Ha2gja-Q6s0DnB<18MX9t&OR#oMs
KZhpctmk9tTj1<cN

delta 124
zcmdnT*T}!Yj+s3_r?ez9KX0-fb2prQl6fJVoy8*M?dM$K5nd6JQIwjPl9ZF5oE==A
ySyh#jYQv=f1j(f(If=!^R#wURd1X-JtgM`YoRG?bR4c1&5ECkHw|ODUTqXb}<}Mll

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_1x1_64x128x64_32x64x64_hswish.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_1x1_64x128x64_32x64x64_hswish.cu
index f234109f582ea2bf7587705f1f8bd0023916a1fe..2e9eda959141057b62637babad17acce330323af 100644
GIT binary patch
delta 60
zcmZqSKghSik$JKQbNA$PEE1DHF?Y&aS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW
Q<rhqz!y>WyIm=uo0I-D@2LJ#7

delta 85
zcmX@e*TTQSk(oU|r?ez9KX0-lb2prQhIt{Joy#IMxraqU%7#k;2$D-nauSP+t*nys
T^U7da&oE0c+HGFKGM5Pe7@8gq

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_1x1_64x128x64_32x64x64_id.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_1x1_64x128x64_32x64x64_id.cu
index 8846c82f408c075ca63914b01910440626c5d5f8..7e9fc7a451bde2772000b113393e3b7550a613d2 100644
GIT binary patch
delta 54
zcmeyzx0`Q+HS=UA=I+T?m={j2VUdz`D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-TaVcE)xJh02MU=

delta 145
zcmdnZ_m6LbH8XpDPH9PIe%@qj=59Fq81up$sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG92Bp2sqm2>|>Y
BH<17U

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_1x1_64x128x64_32x64x64_relu.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_1x1_64x128x64_32x64x64_relu.cu
index 54a3079223f9258a56dd14805909553c2b5d8264..00ed690b3e81d74bfa6609a4ee7ad0d1efe42611 100644
GIT binary patch
delta 58
zcmZqT-_N(fo_VqxbNA#kEE1DHFn3OFV3Cq_D=5k@@N=&42(JjqC`wICNy^Dj&JHfm
Otg6aM-TahgE)xKJG8Rn$

delta 124
zcmdnb*TlcUo|!#Ar?ez9KX0-<b2prQig_WNoy{WU?dM$K5nd6JQIwjPl9ZF5oE==A
ySyh#jYQv=f1j(f(If=!^R#wURd1X-JtgM`YoRG?bR4c1&5ECkHw|NoETqXc1cP=3S

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_1x1_64x64x64_32x32x64_hswish.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_1x1_64x64x64_32x32x64_hswish.cu
index df3571a2da61c4c1ead1d9b17ddcbba8ef50c153..e644b16658730c8859c8d8d1fde6840938ca6985 100644
GIT binary patch
delta 58
zcmZqXKft%afqAk!bNA%4EE1DHGIvgHWRa3}D=5k@@N=&42(JjqC`wICNy^Dj&JHfm
Otg6aM-TaJYE)xKKjTTh^

delta 124
zcmX@W*UZ1cftfu&r?ez9KX0-Fb2prQnt36dox>vK?dM$K5nd6JQIwjPl9ZF5oE==A
ySyh#jYQv=f1j(f(If=!^R#wURd1X-JtgM`YoRG?bR4c1&5ECkHw|OzkTqXc42rei9

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_1x1_64x64x64_32x32x64_id.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_1x1_64x64x64_32x32x64_id.cu
index deba71bbab2ea1184566f223df2a10c331b39ec4..f870aed546abc00659c04e8f57ee2612b866ff2b 100644
GIT binary patch
delta 54
zcmey%w~KFs74u|A=I+UtnHNs3W|5M0D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-TZ)ME)xJf;}tFd

delta 164
zcmdnR_m^*j6*GH&PH9PIe%@p&=59FqDD%R4sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_nYQ
ML2%mU4J=(u04VG_z5oCK

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_1x1_64x64x64_32x32x64_relu.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_1x1_64x64x64_32x32x64_relu.cu
index 2392dc7bc24bdc1bf70dcbc5fa39e6a1c867d3ca..d6bebaf2ca1651c9f671eeb63962c0d019f03fb3 100644
GIT binary patch
delta 54
zcmZqV-^aJXj(M^xbNA%y%nK*ivq;Ih6%^$c_&Ha2gja-Q6s0DnB<18MX9t&OR#oMs
KZhpctmk9tTj1<cN

delta 124
zcmdnT*T}!Yj+s3_r?ez9KX0-fb2prQl6fJVoy8*M?dM$K5nd6JQIwjPl9ZF5oE==A
ySyh#jYQv=f1j(f(If=!^R#wURd1X-JtgM`YoRG?bR4c1&5ECkHw|ODUTqXb}<}Mll

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_256x128x64_64x64x64_hswish.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_256x128x64_64x64x64_hswish.cu
index fe0b58496728ae10d1c5d02cff8a9ffb8a69a1e9..33027d350eea1663a82c07d37dac5b8c9ad28b65 100644
GIT binary patch
delta 60
zcmZqSKghSiiFvXobNA%)EE1DHGk3~cS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW
Q<rhqz%ObJ)1<M>J0I^mV3;+NC

delta 85
zcmX@e*TTQSiJ3h=r?ez9KX0-Vb2prQmU$tZoyQ_IxtB#k%7#k;2$D-nauSP+t*nys
T^U7da&oWCe+HGFSGKUEO8Ok0M

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_256x128x64_64x64x64_id.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_256x128x64_64x64x64_id.cu
index 61ac1b497198efb99ce2771d88ba595df2447304..adca5a47d93382a872400fd07e881e064fdebb08 100644
GIT binary patch
delta 54
zcmeyzx0`Q+4fAAY=I+T?nHNs3Ws#C~D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-Ta7U4if-Bu@yN0

delta 145
zcmdnZ_m6Lb4KsUwPH9PIe%@po=59FqIP=0BsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG92Bp3gFe2>|_|
BH<SPX

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_256x128x64_64x64x64_relu.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_256x128x64_64x64x64_relu.cu
index 4ea0b892f8cde45bc10b77465df872d2761e76ac..9ca3b5417055f1fb12d056fe69ad9739799cef9e 100644
GIT binary patch
delta 58
zcmZqT-_N(ffqAk!bNA%4EE1DHGIvgHWRa3}D=5k@@N=&42(JjqC`wICNy^Dj&JHfm
Otg6aM-TaJY4if-<85U9i

delta 124
zcmdnb*TlcUftfu&r?ez9KX0-Fb2prQnt36dox>vK?dM$K5nd6JQIwjPl9ZF5oE==A
ySyh#jYQv=f1j(f(If=!^R#wURd1X-JtgM`YoRG?bR4c1&5ECkHw|Ozk93}uO5H2ME

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_32x64x64_32x16x64_hswish.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_32x64x64_32x16x64_hswish.cu
index ffef0da2f7f107049c119d92598ace9ca2540a38..89b8055296b51c7884d243f3b3b082ed7ba5f485 100644
GIT binary patch
delta 58
zcmZqT-_N(ffqAk!bNA%4EE1DHGIvgHWRa3}D=5k@@N=&42(JjqC`wICNy^Dj&JHfm
Otg6aM-TaJY4if-<85U9i

delta 124
zcmdnb*TlcUftfu&r?ez9KX0-Fb2prQnt36dox>vK?dM$K5nd6JQIwjPl9ZF5oE==A
ySyh#jYQv=f1j(f(If=!^R#wURd1X-JtgM`YoRG?bR4c1&5ECkHw|Ozk93}uO5H2ME

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_32x64x64_32x16x64_id.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_32x64x64_32x16x64_id.cu
index aa544e2d48749eddaea471fd7d0c8418f23035f0..30e53fc0927533476b3a78114a1ca266fa44160c 100644
GIT binary patch
delta 54
zcmeyvx07#!74u|A=I+UtnHNs3W|5M0D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-TZ)M4if-9cNHrD

delta 164
zcmdnV_lIwT6*GH&PH9PIe%@p&=59FqDD%R4sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_nYQ
ML2$|D4J@5Z04Ml5$p8QV

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_32x64x64_32x16x64_relu.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_32x64x64_32x16x64_relu.cu
index 762bf3e9f3fd79b5d45dfd223f55656808b6194b..465fe2653ca6b5d97fdfda8734e5d1a1219de4c7 100644
GIT binary patch
delta 54
zcmZqR-^;hbj(M^xbNA%y%nK*ivq;Ih6%^$c_&Ha2gja-Q6s0DnB<18MX9t&OR#oMs
KZhpcthY0{DAQZ>|

delta 124
zcmdnX*TBEQj+s3_r?ez9KX0-fb2prQl6fJVoy8*M?dM$K5nd6JQIwjPl9ZF5oE==A
ySyh#jYQv=f1j(f(If=!^R#wURd1X-JtgM`YoRG?bR4c1&5ECkHw|ODU93}uI?k*Pq

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_64x128x64_32x64x64_hswish.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_64x128x64_32x64x64_hswish.cu
index bad9a5ce96622a41c2cf80bb9fb7619714bc78c1..00fb0d97a57068e8274d698968b3131a1d61894d 100644
GIT binary patch
delta 60
zcmZqXKft%ak$JKQbNA$PEE1DHF?Y&aS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW
Q<rhqz!y>WyIm;X-0I&@g1ONa4

delta 85
zcmX@W*UZ1ck(oU|r?ez9KX0-lb2prQhIt{Joy#IMxraqU%7#k;2$D-nauSP+t*nys
T^U7da&oE0c+HGFKGKUEO7w{en

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_64x128x64_32x64x64_id.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_64x128x64_32x64x64_id.cu
index b08d7a953791729443132aedaa48999e3bd3be82..1b2402dee3697eaffccebc5028de925b41a37656 100644
GIT binary patch
delta 54
zcmey%w~KFsHS=UA=I+T?m={j2VUdz`D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-TaVc4if-Aloc`n

delta 145
zcmdnR_m^*jH8XpDPH9PIe%@qj=59Fq81up$sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG92Bp2sqW2>|-~
BH;w=R

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_64x128x64_32x64x64_relu.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_64x128x64_32x64x64_relu.cu
index 9173c90410bac96ef10a9ca19831ea473ced3cf1..647694ce46875c8dbc20fa000522d92fd20f0609 100644
GIT binary patch
delta 58
zcmZqV-^aJXo_VqxbNA#kEE1DHFn3OFV3Cq_D=5k@@N=&42(JjqC`wICNy^Dj&JHfm
Otg6aM-Tahg4if--z7|RV

delta 124
zcmdnT*T}!Yo|!#Ar?ez9KX0-<b2prQig_WNoy{WU?dM$K5nd6JQIwjPl9ZF5oE==A
ySyh#jYQv=f1j(f(If=!^R#wURd1X-JtgM`YoRG?bR4c1&5ECkHw|NoE93}uLe=Z&X

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_64x64x64_32x32x64_hswish.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_64x64x64_32x32x64_hswish.cu
index ceb383565200b238150600610bc88060b5edafa7..54cca4c8bf99dd90e0ebe7f2d283d60aa4a34f63 100644
GIT binary patch
delta 58
zcmZqT-_N(ffqAk!bNA%4EE1DHGIvgHWRa3}D=5k@@N=&42(JjqC`wICNy^Dj&JHfm
Otg6aM-TaJY4if-<85U9i

delta 124
zcmdnb*TlcUftfu&r?ez9KX0-Fb2prQnt36dox>vK?dM$K5nd6JQIwjPl9ZF5oE==A
ySyh#jYQv=f1j(f(If=!^R#wURd1X-JtgM`YoRG?bR4c1&5ECkHw|Ozk93}uO5H2ME

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_64x64x64_32x32x64_id.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_64x64x64_32x32x64_id.cu
index 7f1b55c0f98d4734a3eaf0ee1552e0f4ce11ebc6..48c0ed8f0c909cf0b44936077e3deebda38be882 100644
GIT binary patch
delta 54
zcmeyvx07#!74u|A=I+UtnHNs3W|5M0D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-TZ)M4if-9cNHrD

delta 164
zcmdnV_lIwT6*GH&PH9PIe%@p&=59FqDD%R4sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_nYQ
ML2$|D4J@5Z04Ml5$p8QV

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_64x64x64_32x32x64_relu.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_64x64x64_32x32x64_relu.cu
index c599f7acde577bdf6c1f80afebdbe40838f30682..bbfd2311dc9a29e8254e96f4e66eb3edaff6aad8 100644
GIT binary patch
delta 54
zcmZqR-^;hbj(M^xbNA%y%nK*ivq;Ih6%^$c_&Ha2gja-Q6s0DnB<18MX9t&OR#oMs
KZhpcthY0{DAQZ>|

delta 124
zcmdnX*TBEQj+s3_r?ez9KX0-fb2prQl6fJVoy8*M?dM$K5nd6JQIwjPl9ZF5oE==A
ySyh#jYQv=f1j(f(If=!^R#wURd1X-JtgM`YoRG?bR4c1&5ECkHw|ODU93}uI?k*Pq

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_128x128x64_64x64x64_hswish.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_128x128x64_64x64x64_hswish.cu
index e11f92ff8df3965d38121d3f17051e112d851d40..b9333b454200f5148045de5121bf3bba916bbb19 100644
GIT binary patch
delta 60
zcmZqXKft%ak$JKQbNA$PEE1DHF?Y&aS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW
Q<rhqz!y>WyIm;X-0I&@g1ONa4

delta 85
zcmX@W*UZ1ck(oU|r?ez9KX0-lb2prQhIt{Joy#IMxraqU%7#k;2$D-nauSP+t*nys
T^U7da&oE0c+HGFKGKUEO7w{en

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_128x128x64_64x64x64_id.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_128x128x64_64x64x64_id.cu
index bf638233d3e95f19637a082b05fc27dcdb385e96..0ed1e56ac59af1da4be7a89bf09dd9bb7c8cd120 100644
GIT binary patch
delta 54
zcmey%w~KFsHS=UA=I+T?m={j2VUdz`D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-TaVc4if-Aloc`n

delta 145
zcmdnR_m^*jH8XpDPH9PIe%@qj=59Fq81up$sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG92Bp2sqW2>|-~
BH;w=R

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_128x128x64_64x64x64_relu.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_128x128x64_64x64x64_relu.cu
index 5ef27dfd260d4f04676311b442ca52b132f3a29d..9ffd151c2f830452b1a3746d7872d574da734805 100644
GIT binary patch
delta 58
zcmZqV-^aJXo_VqxbNA#kEE1DHFn3OFV3Cq_D=5k@@N=&42(JjqC`wICNy^Dj&JHfm
Otg6aM-Tahg4if--z7|RV

delta 124
zcmdnT*T}!Yo|!#Ar?ez9KX0-<b2prQig_WNoy{WU?dM$K5nd6JQIwjPl9ZF5oE==A
ySyh#jYQv=f1j(f(If=!^R#wURd1X-JtgM`YoRG?bR4c1&5ECkHw|NoE93}uLe=Z&X

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_128x256x64_64x64x64_hswish.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_128x256x64_64x64x64_hswish.cu
index f46311f22deda0d9f8f4c555751febbcaaa9a562..5371a42a0c7a85bcd1bb559f1a9265af748c9334 100644
GIT binary patch
delta 60
zcmZqXKft%ak$JKQbNA$PEE1DHF?Y&aS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW
Q<rhqz!y>WyIm;X-0I&@g1ONa4

delta 85
zcmX@W*UZ1ck(oU|r?ez9KX0-lb2prQhIt{Joy#IMxraqU%7#k;2$D-nauSP+t*nys
T^U7da&oE0c+HGFKGKUEO7w{en

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_128x256x64_64x64x64_id.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_128x256x64_64x64x64_id.cu
index f9b97cd47327d336721a342f29e89a98b8869f4d..c152bd5e4ebd4ad6f77b4acfd6e45fa649a46a56 100644
GIT binary patch
delta 54
zcmey%w~KFsHS=UA=I+T?m={j2VUdz`D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-TaVc4if-Aloc`n

delta 145
zcmdnR_m^*jH8XpDPH9PIe%@qj=59Fq81up$sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG92Bp2sqW2>|-~
BH;w=R

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_128x256x64_64x64x64_relu.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_128x256x64_64x64x64_relu.cu
index 3ac0bd364f159da54e205f74bd38156d57adc41e..dfe97c7f59e81e37bc8dd4ac2d402fad4a46d3aa 100644
GIT binary patch
delta 58
zcmZqV-^aJXo_VqxbNA#kEE1DHFn3OFV3Cq_D=5k@@N=&42(JjqC`wICNy^Dj&JHfm
Otg6aM-Tahg4if--z7|RV

delta 124
zcmdnT*T}!Yo|!#Ar?ez9KX0-<b2prQig_WNoy{WU?dM$K5nd6JQIwjPl9ZF5oE==A
ySyh#jYQv=f1j(f(If=!^R#wURd1X-JtgM`YoRG?bR4c1&5ECkHw|NoE93}uLe=Z&X

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_128x64x64_64x32x64_hswish.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_128x64x64_64x32x64_hswish.cu
index c426dddf71e88169516ba2ae61ee62748fab4b3a..9db800d3a53b2468a3321c61b630fd00e49474c2 100644
GIT binary patch
delta 58
zcmZqT-_N(ffqAk!bNA%4EE1DHGIvgHWRa3}D=5k@@N=&42(JjqC`wICNy^Dj&JHfm
Otg6aM-TaJY4if-<85U9i

delta 124
zcmdnb*TlcUftfu&r?ez9KX0-Fb2prQnt36dox>vK?dM$K5nd6JQIwjPl9ZF5oE==A
ySyh#jYQv=f1j(f(If=!^R#wURd1X-JtgM`YoRG?bR4c1&5ECkHw|Ozk93}uO5H2ME

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_128x64x64_64x32x64_id.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_128x64x64_64x32x64_id.cu
index 9fe447f3bc5cdc66ae5068f5948b67f71386064b..4f0e90ad14d0fd176fdde27dc0f65458041773c7 100644
GIT binary patch
delta 54
zcmeyvx07#!74u|A=I+UtnHNs3W|5M0D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-TZ)M4if-9cNHrD

delta 164
zcmdnV_lIwT6*GH&PH9PIe%@p&=59FqDD%R4sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_nYQ
ML2$|D4J@5Z04Ml5$p8QV

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_128x64x64_64x32x64_relu.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_128x64x64_64x32x64_relu.cu
index 198293e2e5f9c383d75bd9efb258a50edf7d25ae..c664ccd86647945b89566d4881db9e6d3b04578a 100644
GIT binary patch
delta 54
zcmZqR-^;hbj(M^xbNA%y%nK*ivq;Ih6%^$c_&Ha2gja-Q6s0DnB<18MX9t&OR#oMs
KZhpcthY0{DAQZ>|

delta 124
zcmdnX*TBEQj+s3_r?ez9KX0-fb2prQl6fJVoy8*M?dM$K5nd6JQIwjPl9ZF5oE==A
ySyh#jYQv=f1j(f(If=!^R#wURd1X-JtgM`YoRG?bR4c1&5ECkHw|ODU93}uI?k*Pq

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_1x1_128x128x64_64x64x64_hswish.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_1x1_128x128x64_64x64x64_hswish.cu
index 1d4a974f24ef62589264e938ebf6209341f57c17..ea9cedcbeab4f5bab02abe016c3a6ec3936a7981 100644
GIT binary patch
delta 60
zcmZqSKghSik$JKQbNA$PEE1DHF?Y&aS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW
Q<rhqz!y>WyIm=uo0I-D@2LJ#7

delta 85
zcmX@e*TTQSk(oU|r?ez9KX0-lb2prQhIt{Joy#IMxraqU%7#k;2$D-nauSP+t*nys
T^U7da&oE0c+HGFKGM5Pe7@8gq

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_1x1_128x128x64_64x64x64_id.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_1x1_128x128x64_64x64x64_id.cu
index 069e96db4230c0ee9b4b0ac6cb33f58b603d7775..9652ea56abe4ced9b3f459ac955dc2e2685776dc 100644
GIT binary patch
delta 54
zcmeyzx0`Q+HS=UA=I+T?m={j2VUdz`D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-TaVcE)xJh02MU=

delta 145
zcmdnZ_m6LbH8XpDPH9PIe%@qj=59Fq81up$sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG92Bp2sqm2>|>Y
BH<17U

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_1x1_128x128x64_64x64x64_relu.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_1x1_128x128x64_64x64x64_relu.cu
index 80a947c22da4cd66991b3ece5122ecb7077089a4..1beff87fff88b55b5597e5513bc8e2bfbe3f5845 100644
GIT binary patch
delta 58
zcmZqT-_N(fo_VqxbNA#kEE1DHFn3OFV3Cq_D=5k@@N=&42(JjqC`wICNy^Dj&JHfm
Otg6aM-TahgE)xKJG8Rn$

delta 124
zcmdnb*TlcUo|!#Ar?ez9KX0-<b2prQig_WNoy{WU?dM$K5nd6JQIwjPl9ZF5oE==A
ySyh#jYQv=f1j(f(If=!^R#wURd1X-JtgM`YoRG?bR4c1&5ECkHw|NoETqXc1cP=3S

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_1x1_128x256x64_64x64x64_hswish.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_1x1_128x256x64_64x64x64_hswish.cu
index 5fe0f19c1f92f7e50c037772dcd2332a25e71d31..876e0dae7e90c3fddc8575b6c9f128bcf3ec9250 100644
GIT binary patch
delta 60
zcmZqSKghSik$JKQbNA$PEE1DHF?Y&aS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW
Q<rhqz!y>WyIm=uo0I-D@2LJ#7

delta 85
zcmX@e*TTQSk(oU|r?ez9KX0-lb2prQhIt{Joy#IMxraqU%7#k;2$D-nauSP+t*nys
T^U7da&oE0c+HGFKGM5Pe7@8gq

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_1x1_128x256x64_64x64x64_id.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_1x1_128x256x64_64x64x64_id.cu
index ca80dda287c9f82ab8714da8f4c50493a4e27953..e86c52382450c53f3580d1fbae9c0b7d55520b06 100644
GIT binary patch
delta 54
zcmeyzx0`Q+HS=UA=I+T?m={j2VUdz`D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-TaVcE)xJh02MU=

delta 145
zcmdnZ_m6LbH8XpDPH9PIe%@qj=59Fq81up$sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG92Bp2sqm2>|>Y
BH<17U

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_1x1_128x256x64_64x64x64_relu.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_1x1_128x256x64_64x64x64_relu.cu
index 812cdc8124b4ae6df3ade810bdd81473d1cfd6ce..36b3b3c888506817a045c41c4bfd37a2548ca522 100644
GIT binary patch
delta 58
zcmZqT-_N(fo_VqxbNA#kEE1DHFn3OFV3Cq_D=5k@@N=&42(JjqC`wICNy^Dj&JHfm
Otg6aM-TahgE)xKJG8Rn$

delta 124
zcmdnb*TlcUo|!#Ar?ez9KX0-<b2prQig_WNoy{WU?dM$K5nd6JQIwjPl9ZF5oE==A
ySyh#jYQv=f1j(f(If=!^R#wURd1X-JtgM`YoRG?bR4c1&5ECkHw|NoETqXc1cP=3S

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_1x1_128x64x64_64x32x64_hswish.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_1x1_128x64x64_64x32x64_hswish.cu
index 0b0da908e235ad22e5185606a2b5ca3c8eb6d73f..4e115ae846048645a4f58411159b09d345e1504d 100644
GIT binary patch
delta 58
zcmZqXKft%afqAk!bNA%4EE1DHGIvgHWRa3}D=5k@@N=&42(JjqC`wICNy^Dj&JHfm
Otg6aM-TaJYE)xKKjTTh^

delta 124
zcmX@W*UZ1cftfu&r?ez9KX0-Fb2prQnt36dox>vK?dM$K5nd6JQIwjPl9ZF5oE==A
ySyh#jYQv=f1j(f(If=!^R#wURd1X-JtgM`YoRG?bR4c1&5ECkHw|OzkTqXc42rei9

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_1x1_128x64x64_64x32x64_id.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_1x1_128x64x64_64x32x64_id.cu
index 89608b5bb926679024612b54664113a32342319d..a7829ddcbd2528dbd0e5f3fca9bc57b00546798b 100644
GIT binary patch
delta 54
zcmey%w~KFs74u|A=I+UtnHNs3W|5M0D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-TZ)ME)xJf;}tFd

delta 164
zcmdnR_m^*j6*GH&PH9PIe%@p&=59FqDD%R4sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_nYQ
ML2%mU4J=(u04VG_z5oCK

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_1x1_128x64x64_64x32x64_relu.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_1x1_128x64x64_64x32x64_relu.cu
index 80aac1fcaccd6edd6950694a98200d426f993b2b..e9cd6b18679c6b59db6e29511a01e5dc174ffc65 100644
GIT binary patch
delta 54
zcmZqV-^aJXj(M^xbNA%y%nK*ivq;Ih6%^$c_&Ha2gja-Q6s0DnB<18MX9t&OR#oMs
KZhpctmk9tTj1<cN

delta 124
zcmdnT*T}!Yj+s3_r?ez9KX0-fb2prQl6fJVoy8*M?dM$K5nd6JQIwjPl9ZF5oE==A
ySyh#jYQv=f1j(f(If=!^R#wURd1X-JtgM`YoRG?bR4c1&5ECkHw|ODUTqXb}<}Mll

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_1x1_256x128x64_64x64x64_hswish.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_1x1_256x128x64_64x64x64_hswish.cu
index 338226f8164bf67916a24a709ffaaaf55f377e42..b98a9376bfc5a06310cd78e4a42446498e677601 100644
GIT binary patch
delta 60
zcmZqSKghSik$JKQbNA$PEE1DHF?Y&aS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW
Q<rhqz!y>WyIm=uo0I-D@2LJ#7

delta 85
zcmX@e*TTQSk(oU|r?ez9KX0-lb2prQhIt{Joy#IMxraqU%7#k;2$D-nauSP+t*nys
T^U7da&oE0c+HGFKGM5Pe7@8gq

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_1x1_256x128x64_64x64x64_id.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_1x1_256x128x64_64x64x64_id.cu
index bde7fed6b1ff0937607d55bcdbd8728171d48cc4..f8ca88770f28f272da6900f7ade7c4c2820f8a6b 100644
GIT binary patch
delta 54
zcmeyzx0`Q+HS=UA=I+T?m={j2VUdz`D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-TaVcE)xJh02MU=

delta 145
zcmdnZ_m6LbH8XpDPH9PIe%@qj=59Fq81up$sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG92Bp2sqm2>|>Y
BH<17U

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_1x1_256x128x64_64x64x64_relu.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_1x1_256x128x64_64x64x64_relu.cu
index 5cc092b65c3a41ba1f4f1f3fb97eb88b397fd44e..88420f6da7f0695f243518c094b1b4108294d0c8 100644
GIT binary patch
delta 58
zcmZqT-_N(fo_VqxbNA#kEE1DHFn3OFV3Cq_D=5k@@N=&42(JjqC`wICNy^Dj&JHfm
Otg6aM-TahgE)xKJG8Rn$

delta 124
zcmdnb*TlcUo|!#Ar?ez9KX0-<b2prQig_WNoy{WU?dM$K5nd6JQIwjPl9ZF5oE==A
ySyh#jYQv=f1j(f(If=!^R#wURd1X-JtgM`YoRG?bR4c1&5ECkHw|NoETqXc1cP=3S

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_1x1_32x64x64_16x32x64_hswish.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_1x1_32x64x64_16x32x64_hswish.cu
index b49cdc6be5ee137b764147452395fc5246d08226..34234df52624f1ce5fa0b2dd640e77b6aa359e1b 100644
GIT binary patch
delta 58
zcmZqT-_N(fo_VqxbNA#kEE1DHFn3OFV3Cq_D=5k@@N=&42(JjqC`wICNy^Dj&JHfm
Otg6aM-TahgE)xKJG8Rn$

delta 124
zcmdnb*TlcUo|!#Ar?ez9KX0-<b2prQig_WNoy{WU?dM$K5nd6JQIwjPl9ZF5oE==A
ySyh#jYQv=f1j(f(If=!^R#wURd1X-JtgM`YoRG?bR4c1&5ECkHw|NoETqXc1cP=3S

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_1x1_32x64x64_16x32x64_id.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_1x1_32x64x64_16x32x64_id.cu
index 6957270fbc22a31e00543ef6ed17e920eff1b6a8..c69170cf5180c53f2c4ee7f8535c52520dfbfc0e 100644
GIT binary patch
delta 54
zcmeyvx07#!CG%tl=I+Utm={j2Vv&+{D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-F%;AE)xJe#uX<3

delta 164
zcmdnV_lIwTB{O?|PH9PIe%@qD=59Fq2=l^vsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_nYQ
ML2%mU^(<XX043)-xBvhE

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_1x1_32x64x64_16x32x64_relu.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_1x1_32x64x64_16x32x64_relu.cu
index 28960bc48cdb995ceaa4be6248359addb367e25d..ed347534cc59619209511ab13db1fe38b74c2704 100644
GIT binary patch
delta 54
zcmZqR-^;hbmU*%ZbNA$H%nK*iu}I0f6%^$c_&Ha2gja-Q6s0DnB<18MX9t&OR#oMs
KZhp)%mk9tSZxqA;

delta 145
zcmdnX*TBEQmYF?2r?ez9KX0-vb2prQf_Y&MRNTs{B%>%bF(oM{KRMgV$~ix;3@YR2
zT;UO35rU8nF3+s0%1O21QUHSF(vqCS;$kbSWP}cg_K?bgR4c1&B$MnmFJPI=1OU2r
BHy!{0

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_1x1_64x128x64_32x64x64_hswish.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_1x1_64x128x64_32x64x64_hswish.cu
index 99ff27cd0c4e2602dee58c53d6a8ec8ce0139248..8a9c6cdc3c8cc75a4465e230c336507e93895bac 100644
GIT binary patch
delta 58
zcmZqXKft%afqAk!bNA%4EE1DHGIvgHWRa3}D=5k@@N=&42(JjqC`wICNy^Dj&JHfm
Otg6aM-TaJYE)xKKjTTh^

delta 124
zcmX@W*UZ1cftfu&r?ez9KX0-Fb2prQnt36dox>vK?dM$K5nd6JQIwjPl9ZF5oE==A
ySyh#jYQv=f1j(f(If=!^R#wURd1X-JtgM`YoRG?bR4c1&5ECkHw|OzkTqXc42rei9

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_1x1_64x128x64_32x64x64_id.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_1x1_64x128x64_32x64x64_id.cu
index c913c159c6ef972b80ade9bf46476ec6028c08b0..1d081f4c9ec7fd7968ab50022799734d6e75bf6f 100644
GIT binary patch
delta 54
zcmey%w~KFs74u|A=I+UtnHNs3W|5M0D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-TZ)ME)xJf;}tFd

delta 164
zcmdnR_m^*j6*GH&PH9PIe%@p&=59FqDD%R4sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_nYQ
ML2%mU4J=(u04VG_z5oCK

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_1x1_64x128x64_32x64x64_relu.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_1x1_64x128x64_32x64x64_relu.cu
index 5144b3bcbcbaa55a6d08d60acb92f61bffcda14c..997cda702712ff8c03d93ecb245577a2fa64e712 100644
GIT binary patch
delta 54
zcmZqV-^aJXj(M^xbNA%y%nK*ivq;Ih6%^$c_&Ha2gja-Q6s0DnB<18MX9t&OR#oMs
KZhpctmk9tTj1<cN

delta 124
zcmdnT*T}!Yj+s3_r?ez9KX0-fb2prQl6fJVoy8*M?dM$K5nd6JQIwjPl9ZF5oE==A
ySyh#jYQv=f1j(f(If=!^R#wURd1X-JtgM`YoRG?bR4c1&5ECkHw|ODUTqXb}<}Mll

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_1x1_64x64x64_32x32x64_hswish.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_1x1_64x64x64_32x32x64_hswish.cu
index ef7e3b9de85be365046298f8a05750a26e325e72..b27c8248b8f251c0beddd6822aa04e3fc5e3db80 100644
GIT binary patch
delta 58
zcmZqT-_N(fo_VqxbNA#kEE1DHFn3OFV3Cq_D=5k@@N=&42(JjqC`wICNy^Dj&JHfm
Otg6aM-TahgE)xKJG8Rn$

delta 124
zcmdnb*TlcUo|!#Ar?ez9KX0-<b2prQig_WNoy{WU?dM$K5nd6JQIwjPl9ZF5oE==A
ySyh#jYQv=f1j(f(If=!^R#wURd1X-JtgM`YoRG?bR4c1&5ECkHw|NoETqXc1cP=3S

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_1x1_64x64x64_32x32x64_id.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_1x1_64x64x64_32x32x64_id.cu
index adb02359854a66feaeeda63ae1dfe13c4081b394..07d0fcd2fdadaf9a91161ebc00a017b69e7b82b0 100644
GIT binary patch
delta 54
zcmeyvx07#!CG%tl=I+Utm={j2Vv&+{D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-F%;AE)xJe#uX<3

delta 164
zcmdnV_lIwTB{O?|PH9PIe%@qD=59Fq2=l^vsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_nYQ
ML2%mU^(<XX043)-xBvhE

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_1x1_64x64x64_32x32x64_relu.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_1x1_64x64x64_32x32x64_relu.cu
index b3df1baf675f1eb03a7f5de7fa28df496279c315..94167d789b2053681a2cde2050a656cad846efff 100644
GIT binary patch
delta 54
zcmZqR-^;hbmU*%ZbNA$H%nK*iu}I0f6%^$c_&Ha2gja-Q6s0DnB<18MX9t&OR#oMs
KZhp)%mk9tSZxqA;

delta 145
zcmdnX*TBEQmYF?2r?ez9KX0-vb2prQf_Y&MRNTs{B%>%bF(oM{KRMgV$~ix;3@YR2
zT;UO35rU8nF3+s0%1O21QUHSF(vqCS;$kbSWP}cg_K?bgR4c1&B$MnmFJPI=1OU2r
BHy!{0

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_256x128x64_64x64x64_hswish.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_256x128x64_64x64x64_hswish.cu
index a0a2a8bee1689b7b418bdc6234b429dac2db6975..c27f507cd539c175ab6f4545b23724bad87bd79c 100644
GIT binary patch
delta 60
zcmZqXKft%ak$JKQbNA$PEE1DHF?Y&aS(RiIr6#5%<>V)4TUj~h=auE>l$K=X=eZRW
Q<rhqz!y>WyIm;X-0I&@g1ONa4

delta 85
zcmX@W*UZ1ck(oU|r?ez9KX0-lb2prQhIt{Joy#IMxraqU%7#k;2$D-nauSP+t*nys
T^U7da&oE0c+HGFKGKUEO7w{en

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_256x128x64_64x64x64_id.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_256x128x64_64x64x64_id.cu
index 89645723cd8f980f20255bda375f4e4e5540674c..44c2437d36be664bf784a40d1848be97c4a7808d 100644
GIT binary patch
delta 54
zcmey%w~KFsHS=UA=I+T?m={j2VUdz`D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-TaVc4if-Aloc`n

delta 145
zcmdnR_m^*jH8XpDPH9PIe%@qj=59Fq81up$sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG92Bp2sqW2>|-~
BH;w=R

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_256x128x64_64x64x64_relu.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_256x128x64_64x64x64_relu.cu
index 513ff2695ac3b18f7c6cb0880eee43999971493f..bd2802ee022c9a31c94267e9901daab3dd69e522 100644
GIT binary patch
delta 58
zcmZqV-^aJXo_VqxbNA#kEE1DHFn3OFV3Cq_D=5k@@N=&42(JjqC`wICNy^Dj&JHfm
Otg6aM-Tahg4if--z7|RV

delta 124
zcmdnT*T}!Yo|!#Ar?ez9KX0-<b2prQig_WNoy{WU?dM$K5nd6JQIwjPl9ZF5oE==A
ySyh#jYQv=f1j(f(If=!^R#wURd1X-JtgM`YoRG?bR4c1&5ECkHw|NoE93}uLe=Z&X

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_32x64x64_16x32x64_hswish.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_32x64x64_16x32x64_hswish.cu
index c9e96d49978360eb544ea2f40ac508315800ec32..ce9e3b039362bd44812361ca3a49d9f36fee895e 100644
GIT binary patch
delta 58
zcmZqV-^aJXo_VqxbNA#kEE1DHFn3OFV3Cq_D=5k@@N=&42(JjqC`wICNy^Dj&JHfm
Otg6aM-Tahg4if--z7|RV

delta 124
zcmdnT*T}!Yo|!#Ar?ez9KX0-<b2prQig_WNoy{WU?dM$K5nd6JQIwjPl9ZF5oE==A
ySyh#jYQv=f1j(f(If=!^R#wURd1X-JtgM`YoRG?bR4c1&5ECkHw|NoE93}uLe=Z&X

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_32x64x64_16x32x64_id.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_32x64x64_16x32x64_id.cu
index 20b9cddc891a1bc261ad5cc3910615fde50ec3dd..ff5f6d01857fca20bd7da3b68b235962b84ec08f 100644
GIT binary patch
delta 54
zcmey(w}WqkCG%tl=I+Utm={j2Vv&+{D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-F%;A4if-8S`{P!

delta 164
zcmdnN_nU8nB{O?|PH9PIe%@qD=59Fq2=l^vsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_nYQ
ML2$|D^(>uC03`D|!vFvP

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_32x64x64_16x32x64_relu.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_32x64x64_16x32x64_relu.cu
index 41cd9ae5656ec339616b4163c7cf0f4a69f36b37..a0786d6c80027e4cc820c1d97466a65148ffae7a 100644
GIT binary patch
delta 54
zcmey*w})?oE%Rg-=I+VYm={j2W08_|D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-Tatk4if-C&J{oa

delta 145
zcmdnP_n&WrEi-$5PH9PIe%@qT=59Fq1oOfisJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG92BUcfSk2><|9
BH<|zd

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_64x128x64_32x64x64_hswish.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_64x128x64_32x64x64_hswish.cu
index 3151e3dc38a1dc57985f575c35008ab43c303bb3..964f42196a6ad67aebb6eb78262a17dd43de33ff 100644
GIT binary patch
delta 58
zcmZqT-_N(ffqAk!bNA%4EE1DHGIvgHWRa3}D=5k@@N=&42(JjqC`wICNy^Dj&JHfm
Otg6aM-TaJY4if-<85U9i

delta 124
zcmdnb*TlcUftfu&r?ez9KX0-Fb2prQnt36dox>vK?dM$K5nd6JQIwjPl9ZF5oE==A
ySyh#jYQv=f1j(f(If=!^R#wURd1X-JtgM`YoRG?bR4c1&5ECkHw|Ozk93}uO5H2ME

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_64x128x64_32x64x64_id.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_64x128x64_32x64x64_id.cu
index c20fe3d06c6a28e861f168a35f98de11e7f743b0..c6dc75cc639d8876bdfee6007ae20d036bda5741 100644
GIT binary patch
delta 54
zcmeyvx07#!74u|A=I+UtnHNs3W|5M0D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-TZ)M4if-9cNHrD

delta 164
zcmdnV_lIwT6*GH&PH9PIe%@p&=59FqDD%R4sJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_nYQ
ML2$|D4J@5Z04Ml5$p8QV

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_64x128x64_32x64x64_relu.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_64x128x64_32x64x64_relu.cu
index dddbed50f612d50841807e11adcaa5f2a21a4b8d..e8c765cae8c3a43949fceeeeddb0593f7e0188aa 100644
GIT binary patch
delta 54
zcmZqR-^;hbj(M^xbNA%y%nK*ivq;Ih6%^$c_&Ha2gja-Q6s0DnB<18MX9t&OR#oMs
KZhpcthY0{DAQZ>|

delta 124
zcmdnX*TBEQj+s3_r?ez9KX0-fb2prQl6fJVoy8*M?dM$K5nd6JQIwjPl9ZF5oE==A
ySyh#jYQv=f1j(f(If=!^R#wURd1X-JtgM`YoRG?bR4c1&5ECkHw|ODU93}uI?k*Pq

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_64x64x64_32x32x64_hswish.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_64x64x64_32x32x64_hswish.cu
index 6028fba165cc043a786b248a69ad88ba773c9ddd..e0cbba9b09ab6fd111d747636da26fa5f11a9cb4 100644
GIT binary patch
delta 58
zcmZqV-^aJXo_VqxbNA#kEE1DHFn3OFV3Cq_D=5k@@N=&42(JjqC`wICNy^Dj&JHfm
Otg6aM-Tahg4if--z7|RV

delta 124
zcmdnT*T}!Yo|!#Ar?ez9KX0-<b2prQig_WNoy{WU?dM$K5nd6JQIwjPl9ZF5oE==A
ySyh#jYQv=f1j(f(If=!^R#wURd1X-JtgM`YoRG?bR4c1&5ECkHw|NoE93}uLe=Z&X

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_64x64x64_32x32x64_id.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_64x64x64_32x32x64_id.cu
index 0ea06e92e5d98051ef2238bbb2b03113a92216cf..0ff0f65f18bbfe9939c623d29b7ed5a36bdfc783 100644
GIT binary patch
delta 54
zcmey(w}WqkCG%tl=I+Utm={j2Vv&+{D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-F%;A4if-8S`{P!

delta 164
zcmdnN_nU8nB{O?|PH9PIe%@qD=59Fq2=l^vsJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG93oD1h}C=_nYQ
ML2$|D^(>uC03`D|!vFvP

diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_64x64x64_32x32x64_relu.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/conv_bias_int8_implicit_gemm_imma_ncdiv32hw32_ncdiv4hw4_64x64x64_32x32x64_relu.cu
index 688064da3aee8198af4f42d9d05da5e753446aba..822b04790310f355042ea30ce25496956e94eba4 100644
GIT binary patch
delta 54
zcmey*w})?oE%Rg-=I+VYm={j2W08_|D=5k@@N=&42(JjqC`wICNy^Dj&JHfmtg6aM
K-Tatk4if-C&J{oa

delta 145
zcmdnP_n&WrEi-$5PH9PIe%@qT=59Fq1oOfisJN9?Nk&m>VoFj@esZ>zm2-Yx8C1s4
zxxyp7A_O5DT%K7~m6K}2r2quUr6oCu#l==u$p{?~?ID!~sa96mNG92BUcfSk2><|9
BH<|zd

diff --git a/dnn/src/cuda/convolution/backward_data/algo.cpp b/dnn/src/cuda/convolution/backward_data/algo.cpp
index dcdcd5896..c643db1d2 100644
--- a/dnn/src/cuda/convolution/backward_data/algo.cpp
+++ b/dnn/src/cuda/convolution/backward_data/algo.cpp
@@ -6,7 +6,8 @@
  *
  * Unless required by applicable law or agreed to in writing,
  * software distributed under the License is distributed on an
- * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
  */
 
 #include "./algo.h"
@@ -20,32 +21,38 @@ ConvolutionBackwardDataImpl::AlgoPack::AlgoPack() {
     non_cudnn_algos.push_back(&chanwise_small);
     non_cudnn_algos.push_back(&matmul);
 
-    all_algos.push_back(&chanwise); // prefer chanwise
-    all_algos.push_back(&chanwise_small); // prefer small chanwise
+    all_algos.push_back(&chanwise);        // prefer chanwise
+    all_algos.push_back(&chanwise_small);  // prefer small chanwise
 
     fill_cudnn_algos();
-    for (auto &&i: cudnn) {
+    for (auto&& i : cudnn) {
         all_algos.push_back(&i);
     }
     all_algos.push_back(&matmul);
 
+    fill_int8_dp4a_algos();
+    for (auto&& algo : int8_nchw4_dotprod) {
+        all_algos.push_back(&algo);
+        int8_algos.push_back(&algo);
+    }
+
     all_algos.reserve(all_algos.size() * 2);
 
     // add gconv algos by AlgoGroupConvGeneral
     auto all_algos_data = all_algos.data();
     size_t group_algo_start = 2;
-    for (size_t i = group_algo_start; i < all_algos.size(); ++ i) {
+    for (size_t i = group_algo_start; i < all_algos.size(); ++i) {
         gconv.push_back({all_algos[i]});
     }
-    for (size_t i = group_algo_start; i < all_algos.size(); ++ i) {
+    for (size_t i = group_algo_start; i < all_algos.size(); ++i) {
         algo2gconv[all_algos[i]] = &gconv[i - group_algo_start];
     }
-    for (auto &&i: gconv) {
+    for (auto&& i : gconv) {
         all_algos.push_back(&i);
     }
     megdnn_assert(all_algos_data == all_algos.data());
 
-    non_cudnn_algos.push_back(all_algos.rbegin()[0]);   // group matmul
+    non_cudnn_algos.push_back(all_algos.rbegin()[0]);  // group matmul
     all_algos.push_back(&bfloat16);
     bfloat16_algos.push_back(&bfloat16);
 
@@ -59,63 +66,55 @@ MEGDNN_DEF_GET_ALGO_FROM_DESC(ConvolutionBackwardDataImpl)
 ConvolutionBackwardDataImpl::AlgoCUDNN*
 ConvolutionBackwardDataImpl::AlgoPack::cudnn_from_enum(
         cudnnConvolutionBwdDataAlgo_t algo) {
-    for (auto &&i: cudnn) {
+    for (auto&& i : cudnn) {
         if (i.cudnn_enum() == algo)
             return &i;
     }
-    megdnn_throw(megdnn_mangle(ssprintf(
-                    "can not find cudnn bwd_data algorithm %d",
-                    static_cast<int>(algo))));
+    megdnn_throw(
+            megdnn_mangle(ssprintf("can not find cudnn bwd_data algorithm %d",
+                                   static_cast<int>(algo))));
 }
 
 ConvolutionBackwardDataImpl::AlgoPack ConvolutionBackwardDataImpl::sm_algo_pack;
 
 ConvolutionBackwardDataImpl::AlgoBase::SizeArgs::SizeArgs(
-        ConvolutionBackwardDataImpl *o,
-        const TensorLayout &filter, const TensorLayout &diff,
-        const TensorLayout &grad):
-    SizeArgs(o, filter, o->check_layout_fwd(grad, filter, diff), diff, grad)
-{
-}
+        ConvolutionBackwardDataImpl* o, const TensorLayout& filter,
+        const TensorLayout& diff, const TensorLayout& grad)
+        : SizeArgs(o, filter, o->check_layout_fwd(grad, filter, diff), diff,
+                   grad) {}
 
 ConvolutionBackwardDataImpl::AlgoBase::SizeArgs::SizeArgs(
-        ConvolutionBackwardDataImpl *o, const TensorLayout& filter,
-        const CanonizedFilterMeta &filter_meta, const TensorLayout &diff,
-        const TensorLayout &grad):
-    handle{concrete_handle(o->handle())},
-    filter_meta{filter_meta},
-    diff_layout{&diff},
-    grad_layout{&grad},
-    filter_layout{&filter},
-    opr{o}
-{
-}
+        ConvolutionBackwardDataImpl* o, const TensorLayout& filter,
+        const CanonizedFilterMeta& filter_meta, const TensorLayout& diff,
+        const TensorLayout& grad)
+        : handle{concrete_handle(o->handle())},
+          filter_meta{filter_meta},
+          diff_layout{&diff},
+          grad_layout{&grad},
+          filter_layout{&filter},
+          opr{o} {}
 
 ConvolutionBackwardDataImpl::AlgoBase::ExecArgs::ExecArgs(
-        ConvolutionBackwardDataImpl *opr,
-        _megdnn_tensor_in filter,
-        _megdnn_tensor_in diff,
-        _megdnn_tensor_out grad,
-        _megdnn_workspace workspace):
-    SizeArgs(opr, filter.layout, diff.layout, grad.layout),
-    filter_tensor{&filter}, diff_tensor{&diff}, grad_tensor{&grad},
-    workspace{workspace}
-{
-}
+        ConvolutionBackwardDataImpl* opr, _megdnn_tensor_in filter,
+        _megdnn_tensor_in diff, _megdnn_tensor_out grad,
+        _megdnn_workspace workspace)
+        : SizeArgs(opr, filter.layout, diff.layout, grad.layout),
+          filter_tensor{&filter},
+          diff_tensor{&diff},
+          grad_tensor{&grad},
+          workspace{workspace} {}
 
 std::string ConvolutionBackwardDataImpl::AlgoBase::SizeArgs::to_string() const {
-    auto &&fm = filter_meta;
+    auto&& fm = filter_meta;
     MEGDNN_MARK_USED_VAR(fm);
     return megdnn_mangle(ssprintf(
-                "filter=%u{%u,%u,%u,%u}, diff=%s, grad=%s, "
-                "pad=%ux%u, stride=%ux%u, dilate=%ux%u, xcorr=%d, dtype=%s,%s",
-                fm.group, fm.ocpg, fm.icpg, fm.spatial[0], fm.spatial[1],
-                diff_layout->to_string().c_str(),
-                grad_layout->to_string().c_str(),
-                fm.padding[0], fm.padding[1], fm.stride[0], fm.stride[1],
-                fm.dilation[0], fm.dilation[1],
-                !fm.should_flip,
-                diff_layout->dtype.name(), grad_layout->dtype.name()));
+            "filter=%u{%u,%u,%u,%u}, diff=%s, grad=%s, "
+            "pad=%ux%u, stride=%ux%u, dilate=%ux%u, xcorr=%d, dtype=%s,%s",
+            fm.group, fm.ocpg, fm.icpg, fm.spatial[0], fm.spatial[1],
+            diff_layout->to_string().c_str(), grad_layout->to_string().c_str(),
+            fm.padding[0], fm.padding[1], fm.stride[0], fm.stride[1],
+            fm.dilation[0], fm.dilation[1], !fm.should_flip,
+            diff_layout->dtype.name(), grad_layout->dtype.name()));
 }
 
 // vim: syntax=cpp.doxygen
diff --git a/dnn/src/cuda/convolution/backward_data/algo.h b/dnn/src/cuda/convolution/backward_data/algo.h
index e3cd9f94d..0f2a1e47d 100644
--- a/dnn/src/cuda/convolution/backward_data/algo.h
+++ b/dnn/src/cuda/convolution/backward_data/algo.h
@@ -6,7 +6,8 @@
  *
  * Unless required by applicable law or agreed to in writing,
  * software distributed under the License is distributed on an
- * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
  */
 
 #pragma once
@@ -38,6 +39,7 @@ public:
         CUDA_CHANWISE_SMALL,
         CUDA_BFLOAT16,
         CUDA_GROUP_CONV_GENERAL,
+        CUDA_IMPLICIT_GEMM_NCHW4_DOTPROD_INT8
     };
     using Mapper = std::unordered_map<AlgorithmDesc, AlgoBase*>;
 
@@ -240,9 +242,53 @@ public:
     }
 };
 
+class ConvolutionBackwardDataImpl::AlgoInt8NCHW4DotProdImplicitGemm final
+        : public AlgoBase {
+public:
+    struct AlgoParam {
+        int threadblock_m;
+        int threadblock_n;
+        int threadblock_k;
+        int warp_m;
+        int warp_n;
+        int warp_k;
+        int stage;
+        std::string to_string() {
+            /// default algorithm
+            if (threadblock_m == 128 && threadblock_n == 128 &&
+                threadblock_k == 32 && warp_m == 32 && warp_n == 64 &&
+                warp_k == 32 && stage == 2) {
+                return "";
+            }
+            return ssprintf("_%dX%dX%d_%dX%dX%d_%dstage", threadblock_m,
+                            threadblock_n, threadblock_k, warp_m, warp_n,
+                            warp_k, stage);
+        }
+    };
+    AlgoInt8NCHW4DotProdImplicitGemm(AlgoParam algo_param)
+            : m_algo_param{algo_param},
+              m_name{ssprintf("INT8_NCHW4_DOTPROD_IMPLICIT_GEMM%s",
+                              m_algo_param.to_string().c_str())} {}
+    bool is_available(const SizeArgs& args) const override;
+    size_t get_workspace_in_bytes(const SizeArgs& args) const override;
+    void exec(const ExecArgs& args) const override;
+    const char* name() const override { return m_name.c_str(); }
+    AlgoAttribute attribute() const override {
+        return AlgoAttribute::REPRODUCIBLE;
+    }
+    MEGDNN_DECL_ALGO_TYPE(CUDA_IMPLICIT_GEMM_NCHW4_DOTPROD_INT8)
+private:
+    WorkspaceBundle get_workspace_bundle(dt_byte* raw_ptr,
+                                         const SizeArgs& args) const;
+    AlgoParam m_algo_param;
+    std::string m_name;
+};
+
 class ConvolutionBackwardDataImpl::AlgoPack : NonCopyableObj {
     // defined in cudnn.cpp
     void fill_cudnn_algos();
+    // defined in implicit_gemm_int8_nchw4_dp4a.cpp
+    void fill_int8_dp4a_algos();
 
     AlgoBase::Mapper m_all_algos_map;
 
@@ -256,12 +302,13 @@ public:
     std::vector<AlgoGroupConvGeneral> gconv;
     std::unordered_map<AlgoBase*, AlgoGroupConvGeneral*> algo2gconv;
     AlgoBFloat16 bfloat16;
+    std::vector<AlgoInt8NCHW4DotProdImplicitGemm> int8_nchw4_dotprod;
 
     std::vector<AlgoBase*>
             //! all algorithms
             all_algos,
             //! non-cudnn algos, used for heuristic if cudnn is not supported
-            non_cudnn_algos, bfloat16_algos;
+            non_cudnn_algos, bfloat16_algos, int8_algos;
 
     AlgoCUDNN* cudnn_from_enum(cudnnConvolutionBwdDataAlgo_t algo);
 
diff --git a/dnn/src/cuda/convolution/backward_data/chanwise.cpp b/dnn/src/cuda/convolution/backward_data/chanwise.cpp
index f6272fca7..89562f4d0 100644
--- a/dnn/src/cuda/convolution/backward_data/chanwise.cpp
+++ b/dnn/src/cuda/convolution/backward_data/chanwise.cpp
@@ -6,7 +6,8 @@
  *
  * Unless required by applicable law or agreed to in writing,
  * software distributed under the License is distributed on an
- * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
  */
 
 #include "./algo.h"
@@ -19,8 +20,10 @@ using namespace convolution;
 
 bool ConvolutionBackwardDataImpl::AlgoChanwise::is_available(
         const SizeArgs& args) const {
-    if (args.diff_layout->dtype == args.filter_layout->dtype &&
-        args.diff_layout->dtype == dtype::BFloat16()) {
+    if ((args.diff_layout->dtype == args.filter_layout->dtype &&
+         args.diff_layout->dtype == dtype::BFloat16()) ||
+        (args.diff_layout->dtype == args.filter_layout->dtype &&
+         args.diff_layout->dtype == dtype::QuantizedS8())) {
         return false;
     }
     auto&& fm = args.filter_meta;
@@ -74,4 +77,3 @@ void ConvolutionBackwardDataImpl::AlgoChanwise::exec(
 }
 
 // vim: syntax=cpp.doxygen
-
diff --git a/dnn/src/cuda/convolution/backward_data/chanwise_small.cpp b/dnn/src/cuda/convolution/backward_data/chanwise_small.cpp
index cc5b12e20..bb8ba182d 100644
--- a/dnn/src/cuda/convolution/backward_data/chanwise_small.cpp
+++ b/dnn/src/cuda/convolution/backward_data/chanwise_small.cpp
@@ -6,7 +6,8 @@
  *
  * Unless required by applicable law or agreed to in writing,
  * software distributed under the License is distributed on an
- * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
  */
 
 #include "src/cuda/convolution/backward_data/algo.h"
@@ -28,9 +29,11 @@ inline bool is_available_small(const chanwise::Param& param) {
 }  // anonymous namespace
 
 bool ConvolutionBackwardDataImpl::AlgoChanwiseSmall::is_available(
-        const SizeArgs &args) const {
-    if (args.diff_layout->dtype == args.filter_layout->dtype &&
-        args.diff_layout->dtype == dtype::BFloat16()) {
+        const SizeArgs& args) const {
+    if ((args.diff_layout->dtype == args.filter_layout->dtype &&
+         args.diff_layout->dtype == dtype::BFloat16()) ||
+        (args.diff_layout->dtype == args.filter_layout->dtype &&
+         args.diff_layout->dtype == dtype::QuantizedS8())) {
         return false;
     }
 #if CUDA_VERSION < 9000
@@ -38,30 +41,29 @@ bool ConvolutionBackwardDataImpl::AlgoChanwiseSmall::is_available(
         return false;
 #endif
     auto kparam = chanwise::Param::from_fwd_args(args.as_fwd_args());
-    auto &&fm = args.filter_meta;
+    auto&& fm = args.filter_meta;
     return args.filter_meta.format == Param::Format::NCHW &&
-        args.diff_layout->dtype.category() == DTypeCategory::FLOAT &&
+           args.diff_layout->dtype.category() == DTypeCategory::FLOAT &&
            args.opr->param().compute_mode == Param::ComputeMode::DEFAULT &&
-        fm.spatial_ndim == 2 && fm.icpg == 1 &&
-        fm.dilation[0] == 1 && fm.dilation[1] == 1 &&
-        !fm.should_flip && is_available_small(kparam);
+           fm.spatial_ndim == 2 && fm.icpg == 1 && fm.dilation[0] == 1 &&
+           fm.dilation[1] == 1 && !fm.should_flip && is_available_small(kparam);
 }
 
 size_t ConvolutionBackwardDataImpl::AlgoChanwiseSmall::get_workspace_in_bytes(
-        const SizeArgs &) const {
+        const SizeArgs&) const {
     return 0;
 }
 
 void ConvolutionBackwardDataImpl::AlgoChanwiseSmall::exec(
-        const ExecArgs &args) const {
+        const ExecArgs& args) const {
     auto kparam = chanwise::Param::from_fwd_args(args.as_fwd_args());
     auto stream = cuda_stream(args.handle);
     switch (args.grad_layout->dtype.enumv()) {
         case DTypeEnum::Float32:
-            return chanwise::run_bwd_data_small(args.grad_tensor->ptr<float>(),
-                                     args.diff_tensor->ptr<float>(),
-                                     args.filter_tensor->ptr<float>(), kparam,
-                                     stream);
+            return chanwise::run_bwd_data_small(
+                    args.grad_tensor->ptr<float>(),
+                    args.diff_tensor->ptr<float>(),
+                    args.filter_tensor->ptr<float>(), kparam, stream);
 #if CUDA_VERSION >= 9000
         case DTypeEnum::Float16:
             return chanwise::run_bwd_data_small(
@@ -77,4 +79,3 @@ void ConvolutionBackwardDataImpl::AlgoChanwiseSmall::exec(
 }
 
 // vim: syntax=cpp.doxygen
-
diff --git a/dnn/src/cuda/convolution/backward_data/cutlass_deconvolution_wrapper.cu b/dnn/src/cuda/convolution/backward_data/cutlass_deconvolution_wrapper.cu
new file mode 100644
index 000000000..1d054ae4f
--- /dev/null
+++ b/dnn/src/cuda/convolution/backward_data/cutlass_deconvolution_wrapper.cu
@@ -0,0 +1,100 @@
+/**
+ * \file src/cuda/convolution/backward_data/cutlass_deconvolution_wrapper.cu
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
+ */
+// ignore warning of cutlass
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wunused-parameter"
+#pragma GCC diagnostic ignored "-Wstrict-aliasing"
+
+#if !MEGDNN_TEGRA_X1
+#include "cutlass/convolution/device/convolution.h"
+#endif
+#include "src/common/opr_param_defs_enumv.cuh"
+#include "src/cuda/convolution/backward_data/cutlass_deconvolution_wrapper.cuh"
+#pragma GCC diagnostic pop
+
+using namespace megdnn;
+using namespace cuda;
+using namespace cutlass_wrapper;
+
+/* ================ cutlass kernel wrapper for nchw4 layout ================= */
+#if MEGDNN_TEGRA_X1
+void megdnn::cuda::cutlass_wrapper::do_deconv_int8_implicit_gemm_dp4a_ncdiv4hw4(
+        const int8_t* /* d_src */, const int8_t* /* d_filter */,
+        int8_t* /* d_dst */, int* /* workspace */,
+        const convolution::ConvParam& /* param */, float /* alpha */,
+        const GemmCoord& /* threadblock_shape */,
+        const GemmCoord& /* warp_shape */, int /* stages */,
+        cudaStream_t /* stream */) {}
+#else
+void megdnn::cuda::cutlass_wrapper::do_deconv_int8_implicit_gemm_dp4a_ncdiv4hw4(
+        const int8_t* d_src, const int8_t* d_filter, int8_t* d_dst,
+        int* workspace, const convolution::ConvParam& param, float alpha,
+        const GemmCoord& threadblock_shape, const GemmCoord& warp_shape,
+        int stages, cudaStream_t stream) {
+#define DISPATCH_KERNEL_WITH_TILE_SHAPE(threadblock_m_, threadblock_n_,        \
+                                        threadblock_k_, warp_m_, warp_n_,      \
+                                        warp_k_, stage_, aligned_)             \
+    if (threadblock_shape.m() == threadblock_m_ &&                             \
+        threadblock_shape.n() == threadblock_n_ &&                             \
+        threadblock_shape.k() == threadblock_k_ &&                             \
+        warp_shape.m() == warp_m_ && warp_shape.n() == warp_n_ &&              \
+        warp_shape.k() == warp_k_ && stages == stage_) {                       \
+        using ThreadBlockShape =                                               \
+                cutlass::gemm::GemmShape<threadblock_m_, threadblock_n_,       \
+                                         threadblock_k_>;                      \
+        using WarpShape = cutlass::gemm::GemmShape<warp_m_, warp_n_, warp_k_>; \
+        using InstructionShape = cutlass::gemm::GemmShape<1, 1, 4>;            \
+        using Deconvolution = cutlass::conv::device::Deconvolution<            \
+                int8_t, cutlass::layout::TensorNCxHWx<4>, int8_t,              \
+                cutlass::layout::TensorKxRSCx<4>, ElementOutput,               \
+                cutlass::layout::TensorNCxHWx<4>, int32_t,                     \
+                cutlass::layout::TensorNCxHWx<4>, int32_t,                     \
+                cutlass::arch::OpClassSimt, cutlass::arch::Sm61,               \
+                ThreadBlockShape, WarpShape, InstructionShape, EpilogueOp,     \
+                cutlass::conv::threadblock::                                   \
+                        ConvolutionDgradNCxHWxThreadblockSwizzle,              \
+                stage_, 4, aligned_>;                                          \
+        typename Deconvolution::ConvolutionParameter conv_param(               \
+                param.n, param.hi, param.wi, param.ci, param.co, param.fh,     \
+                param.fw, param.ho, param.wo, param.ph, param.pw, param.sh,    \
+                param.sw, 1, 1, cutlass::conv::Mode::kCrossCorrelation);       \
+        return cutlass_deconvolution_wrapper<Deconvolution>(                   \
+                d_src, d_filter, nullptr, nullptr, d_dst, workspace,           \
+                conv_param, epilogue, stream);                                 \
+    }
+#define DISPATCH_KERNEL                                                      \
+    DISPATCH_KERNEL_WITH_TILE_SHAPE(16, 64, 8, 16, 64, 8, 2, 4);             \
+    DISPATCH_KERNEL_WITH_TILE_SHAPE(16, 128, 16, 16, 64, 16, 2, 4);          \
+    DISPATCH_KERNEL_WITH_TILE_SHAPE(16, 128, 16, 16, 128, 16, 1, 8);         \
+    DISPATCH_KERNEL_WITH_TILE_SHAPE(32, 128, 32, 32, 64, 32, 2, 16);         \
+    DISPATCH_KERNEL_WITH_TILE_SHAPE(64, 128, 32, 64, 32, 32, 2, 16);         \
+    megdnn_assert(false,                                                     \
+                  "unsupported threadblock shape (%dx%dx%d) and warp shape " \
+                  "(%dx%dx%d)",                                              \
+                  threadblock_shape.m(), threadblock_shape.n(),              \
+                  threadblock_shape.k(), warp_shape.m(), warp_shape.n(),     \
+                  warp_shape.k());
+    using ElementOutput = int8_t;
+    using ElementAccumulator = int32_t;
+    using ElementBias = int32_t;
+    using ElementCompute = float;
+    using EpilogueOp = cutlass::epilogue::thread::BiasAddLinearCombinationClamp<
+            ElementOutput, 4, ElementAccumulator, ElementBias, ElementCompute>;
+    typename EpilogueOp::Params epilogue{alpha, 0, 0};
+    DISPATCH_KERNEL;
+
+#undef DISPATCH_KERNEL_WITH_TILE_SHAPE
+#undef DISPATCH_KERNEL
+}
+#endif
+
+// vim: syntax=cuda.doxygen
diff --git a/dnn/src/cuda/convolution/backward_data/cutlass_deconvolution_wrapper.cuh b/dnn/src/cuda/convolution/backward_data/cutlass_deconvolution_wrapper.cuh
new file mode 100644
index 000000000..35961673d
--- /dev/null
+++ b/dnn/src/cuda/convolution/backward_data/cutlass_deconvolution_wrapper.cuh
@@ -0,0 +1,44 @@
+/**
+ * \file src/cuda/convolution/backward_data/cutlass_deconvolution_wrapper.cuh
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
+ */
+#pragma once
+#include "cutlass/gemm/gemm.h"
+#include "src/cuda/convolution_helper/parameter.cuh"
+#include "src/cuda/utils.cuh"
+
+namespace megdnn {
+namespace cuda {
+namespace cutlass_wrapper {
+
+using GemmCoord = cutlass::gemm::GemmCoord;
+
+template <typename Convolution>
+void cutlass_deconvolution_wrapper(
+        const typename Convolution::ElementSrc* d_src,
+        const typename Convolution::ElementFilter* d_filter,
+        const typename Convolution::ElementBias* d_bias,
+        const typename Convolution::ElementDst* d_z,
+        typename Convolution::ElementDst* d_dst, int* workspace,
+        typename Convolution::ConvolutionParameter const& conv_param,
+        typename Convolution::EpilogueOutputOp::Params const& epilogue,
+        cudaStream_t stream);
+
+void do_deconv_int8_implicit_gemm_dp4a_ncdiv4hw4(
+        const int8_t* d_src, const int8_t* d_filter, int8_t* d_dst,
+        int* workspace, const convolution::ConvParam& param, float alpha,
+        const GemmCoord& threadblock_shape, const GemmCoord& warp_shape,
+        int stages, cudaStream_t stream);
+
+}  // namespace cutlass_wrapper
+}  // namespace cuda
+}  // namespace megdnn
+
+// vim: syntax=cuda.doxygen
diff --git a/dnn/src/cuda/convolution/backward_data/deconv_int8_helper.cu b/dnn/src/cuda/convolution/backward_data/deconv_int8_helper.cu
new file mode 100644
index 000000000..f3d284c39
--- /dev/null
+++ b/dnn/src/cuda/convolution/backward_data/deconv_int8_helper.cu
@@ -0,0 +1,76 @@
+/**
+ * \file src/cuda/convolution/backward_data/deconv_int8_helper.cu
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
+ */
+
+#include "src/cuda/convolution/backward_data/deconv_int8_helper.cuh"
+
+using namespace megdnn;
+using namespace cuda;
+using namespace deconv;
+
+#define BLOCKSIZE_X 16
+#define BLOCKSIZE_Y 16
+
+namespace {
+
+//
+__global__ void reorder_filter_nc4hw4_to_n4hwc4_kernel(
+        int8_t* __restrict__ dst, const int8_t* __restrict__ src, uint32_t OC,
+        uint32_t IC, uint32_t FHFW) {
+    const int32_t ocb = blockIdx.z;
+    const int32_t icb = blockIdx.y * BLOCKSIZE_X + threadIdx.y;
+    const int32_t fhfw = blockIdx.x * BLOCKSIZE_Y + threadIdx.x;
+
+    if (fhfw < FHFW && icb < IC / 4) {
+        int src0 = *reinterpret_cast<const int*>(
+                src + (ocb * 4 + 0) * IC * FHFW + (icb * FHFW + fhfw) * 4);
+        int src1 = *reinterpret_cast<const int*>(
+                src + (ocb * 4 + 1) * IC * FHFW + (icb * FHFW + fhfw) * 4);
+        int src2 = *reinterpret_cast<const int*>(
+                src + (ocb * 4 + 2) * IC * FHFW + (icb * FHFW + fhfw) * 4);
+        int src3 = *reinterpret_cast<const int*>(
+                src + (ocb * 4 + 3) * IC * FHFW + (icb * FHFW + fhfw) * 4);
+        // transpose 4x4
+        int dst01_lo = __byte_perm(src0, src1, 0x5140);
+        int dst01_hi = __byte_perm(src0, src1, 0x7362);
+        int dst23_lo = __byte_perm(src2, src3, 0x5140);
+        int dst23_hi = __byte_perm(src2, src3, 0x7362);
+        int dst0 = __byte_perm(dst01_lo, dst23_lo, 0x5410);
+        int dst1 = __byte_perm(dst01_lo, dst23_lo, 0x7632);
+        int dst2 = __byte_perm(dst01_hi, dst23_hi, 0x5410);
+        int dst3 = __byte_perm(dst01_hi, dst23_hi, 0x7632);
+
+        *reinterpret_cast<int*>(
+                dst + (ocb * FHFW * IC + fhfw * IC + icb * 4 + 0) * 4) = dst0;
+        *reinterpret_cast<int*>(
+                dst + (ocb * FHFW * IC + fhfw * IC + icb * 4 + 1) * 4) = dst1;
+        *reinterpret_cast<int*>(
+                dst + (ocb * FHFW * IC + fhfw * IC + icb * 4 + 2) * 4) = dst2;
+        *reinterpret_cast<int*>(
+                dst + (ocb * FHFW * IC + fhfw * IC + icb * 4 + 3) * 4) = dst3;
+    }
+}
+
+}  // namespace
+
+void megdnn::cuda::deconv::reorder_filter_nc4hw4_to_n4hwc4(
+        int8_t* dst, const int8_t* src, uint32_t OC, uint32_t IC, uint32_t FH,
+        uint32_t FW, cudaStream_t stream) {
+    dim3 threads(BLOCKSIZE_X, BLOCKSIZE_Y, 1);
+    dim3 blocks(DIVUP(FH * FW, BLOCKSIZE_X), DIVUP(IC / 4, BLOCKSIZE_Y),
+                OC / 4);
+
+    reorder_filter_nc4hw4_to_n4hwc4_kernel<<<blocks, threads, 0, stream>>>(
+            dst, src, OC, IC, FH * FW);
+    after_kernel_launch();
+}
+
+// vim: syntax=cuda.doxygen
diff --git a/dnn/src/cuda/convolution/backward_data/deconv_int8_helper.cuh b/dnn/src/cuda/convolution/backward_data/deconv_int8_helper.cuh
new file mode 100644
index 000000000..f50b3c36d
--- /dev/null
+++ b/dnn/src/cuda/convolution/backward_data/deconv_int8_helper.cuh
@@ -0,0 +1,27 @@
+/**
+ * \file src/cuda/convolution/backward_data/deconv_int8_helper.cuh
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
+ */
+#pragma once
+#include "src/cuda/utils.cuh"
+
+namespace megdnn {
+namespace cuda {
+namespace deconv {
+
+void reorder_filter_nc4hw4_to_n4hwc4(int8_t* dst, const int8_t* src,
+                                     uint32_t OC, uint32_t IC, uint32_t FH,
+                                     uint32_t FW, cudaStream_t stream);
+
+}  // namespace deconv
+}  // namespace cuda
+}  // namespace megdnn
+
+// vim: syntax=cuda.doxygen
diff --git a/dnn/src/cuda/convolution/backward_data/group_conv.cpp b/dnn/src/cuda/convolution/backward_data/group_conv.cpp
index e4df3ac8d..c2e3b9f45 100644
--- a/dnn/src/cuda/convolution/backward_data/group_conv.cpp
+++ b/dnn/src/cuda/convolution/backward_data/group_conv.cpp
@@ -6,7 +6,8 @@
  *
  * Unless required by applicable law or agreed to in writing,
  * software distributed under the License is distributed on an
- * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
  */
 
 #include "./algo.h"
@@ -16,8 +17,8 @@ using namespace cuda;
 using namespace convolution;
 
 void ConvolutionBackwardDataImpl::AlgoGroupConvGeneral::modify_size_args(
-        ConvolutionBackwardDataImpl::AlgoBase::SizeArgs &args,
-        TensorLayout &diff_pg, TensorLayout &grad_pg) {
+        ConvolutionBackwardDataImpl::AlgoBase::SizeArgs& args,
+        TensorLayout& diff_pg, TensorLayout& grad_pg) {
     diff_pg = *args.diff_layout;
     grad_pg = *args.grad_layout;
     auto nr_grp = args.filter_meta.group;
@@ -29,17 +30,18 @@ void ConvolutionBackwardDataImpl::AlgoGroupConvGeneral::modify_size_args(
 }
 
 ConvolutionBackwardDataImpl::AlgoGroupConvGeneral::AlgoGroupConvGeneral(
-        AlgoBase *impl):
-    m_impl{impl}
-{
+        AlgoBase* impl)
+        : m_impl{impl} {
     m_name = "group_conv:";
     m_name += impl->name();
 }
 
 bool ConvolutionBackwardDataImpl::AlgoGroupConvGeneral::is_available(
-        const SizeArgs &args) const {
-    if (args.diff_layout->dtype == args.filter_layout->dtype &&
-        args.diff_layout->dtype == dtype::BFloat16()) {
+        const SizeArgs& args) const {
+    if ((args.diff_layout->dtype == args.filter_layout->dtype &&
+         args.diff_layout->dtype == dtype::BFloat16()) ||
+        (args.diff_layout->dtype == args.filter_layout->dtype &&
+         args.diff_layout->dtype == dtype::QuantizedS8())) {
         return false;
     }
     auto sub_args = args;
@@ -48,8 +50,9 @@ bool ConvolutionBackwardDataImpl::AlgoGroupConvGeneral::is_available(
     return m_impl->is_available(sub_args);
 }
 
-size_t ConvolutionBackwardDataImpl::AlgoGroupConvGeneral::
-get_workspace_in_bytes(const SizeArgs &args) const {
+size_t
+ConvolutionBackwardDataImpl::AlgoGroupConvGeneral::get_workspace_in_bytes(
+        const SizeArgs& args) const {
     auto sub_args = args;
     TensorLayout diff_pg, grad_pg;
     modify_size_args(sub_args, diff_pg, grad_pg);
@@ -57,24 +60,24 @@ get_workspace_in_bytes(const SizeArgs &args) const {
 }
 
 void ConvolutionBackwardDataImpl::AlgoGroupConvGeneral::exec(
-        const ExecArgs &args) const {
+        const ExecArgs& args) const {
     auto sub_args = args;
     TensorND tflt{*args.filter_tensor}, tdiff{*args.diff_tensor},
-             tgrad{*args.grad_tensor};
+            tgrad{*args.grad_tensor};
     modify_size_args(sub_args, tdiff.layout, tgrad.layout);
     sub_args.filter_tensor = &tflt;
     sub_args.diff_tensor = &tdiff;
     sub_args.grad_tensor = &tgrad;
     auto grp = args.filter_meta.group;
 
-    auto &&fm = args.filter_meta;
-    auto strd_flt = (fm.icpg * fm.ocpg *
-            fm.spatial[0] * fm.spatial[1] * tflt.layout.dtype.size()),
-         strd_diff = (
-                 tdiff.layout.stride[1] * fm.ocpg * tdiff.layout.dtype.size()),
-         strd_grad = (
-                 tgrad.layout.stride[1] * fm.icpg * tgrad.layout.dtype.size());
-    for (uint32_t g = 0; g < grp; ++ g) {
+    auto&& fm = args.filter_meta;
+    auto strd_flt = (fm.icpg * fm.ocpg * fm.spatial[0] * fm.spatial[1] *
+                     tflt.layout.dtype.size()),
+         strd_diff =
+                 (tdiff.layout.stride[1] * fm.ocpg * tdiff.layout.dtype.size()),
+         strd_grad =
+                 (tgrad.layout.stride[1] * fm.icpg * tgrad.layout.dtype.size());
+    for (uint32_t g = 0; g < grp; ++g) {
         m_impl->exec(sub_args);
         incr_voidp(tflt.raw_ptr, strd_flt);
         incr_voidp(tdiff.raw_ptr, strd_diff);
@@ -83,4 +86,3 @@ void ConvolutionBackwardDataImpl::AlgoGroupConvGeneral::exec(
 }
 
 // vim: syntax=cpp.doxygen
-
diff --git a/dnn/src/cuda/convolution/backward_data/implicit_gemm_int8_nchw4_dp4a.cpp b/dnn/src/cuda/convolution/backward_data/implicit_gemm_int8_nchw4_dp4a.cpp
new file mode 100644
index 000000000..9cb8e647e
--- /dev/null
+++ b/dnn/src/cuda/convolution/backward_data/implicit_gemm_int8_nchw4_dp4a.cpp
@@ -0,0 +1,127 @@
+/**
+ * \file dnn/src/cuda/conv_bias/implicit_gemm_int8_nchw4_dp4a.cpp
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
+ */
+
+#include "./algo.h"
+#include "src/cuda/utils.h"
+#include "src/cuda/convolution_helper/parameter.cuh"
+#include "src/cuda/convolution/backward_data/cutlass_deconvolution_wrapper.cuh"
+#include "src/cuda/convolution/backward_data/deconv_int8_helper.cuh"
+
+using namespace megdnn;
+using namespace cuda;
+
+bool ConvolutionBackwardDataImpl::AlgoInt8NCHW4DotProdImplicitGemm::
+        is_available(const SizeArgs& args) const {
+    auto&& fm = args.filter_meta;
+    if (fm.format != Param::Format::NCHW4)
+        return false;
+
+    bool available = true;
+
+    auto src_dtype = args.diff_layout->dtype,
+         filter_dtype = args.filter_layout->dtype,
+         dst_dtype = args.grad_layout->dtype;
+
+    available &= (src_dtype.enumv() == DTypeEnum::QuantizedS8 &&
+                  filter_dtype.enumv() == DTypeEnum::QuantizedS8 &&
+                  dst_dtype.enumv() == DTypeEnum::QuantizedS8);
+    // TODO support group deconv int8
+    available &= (fm.group == 1);
+    // mode must be cross correlation
+    available &= !fm.should_flip;
+    // mode must be 2D
+    available &= fm.spatial_ndim == 2;
+    // TODO: support dialtion
+    available &= (fm.dilation[0] == 1 && fm.dilation[1] == 1);
+    // FIXME: too large filter size is not supported now
+    available &= fm.spatial[0] * fm.spatial[1] <= 64;
+    // only support sm_61 or later, platform should have fast native int8
+    // support
+    available &= is_compute_capability_required(6, 1);
+
+    return available;
+}
+
+WorkspaceBundle ConvolutionBackwardDataImpl::AlgoInt8NCHW4DotProdImplicitGemm::
+        get_workspace_bundle(dt_byte* raw_ptr, const SizeArgs& args) const {
+    size_t ws_filter = args.filter_layout->span().dist_byte();
+    return WorkspaceBundle{raw_ptr, {ws_filter}};
+}
+
+size_t ConvolutionBackwardDataImpl::AlgoInt8NCHW4DotProdImplicitGemm::
+        get_workspace_in_bytes(const SizeArgs& args) const {
+    return get_workspace_bundle(nullptr, args).total_size_in_bytes();
+}
+
+void ConvolutionBackwardDataImpl::AlgoInt8NCHW4DotProdImplicitGemm::exec(
+        const ExecArgs& args) const {
+    auto&& fm = args.filter_meta;
+    size_t n = args.diff_layout->operator[](0),
+           co = args.diff_layout->operator[](1) * 4,
+           ho = args.diff_layout->operator[](2),
+           wo = args.diff_layout->operator[](3);
+    size_t ci = args.grad_layout->operator[](1) * 4,
+           hi = args.grad_layout->operator[](2),
+           wi = args.grad_layout->operator[](3);
+    size_t fh = fm.spatial[0], fw = fm.spatial[1];
+    size_t sh = fm.stride[0], sw = fm.stride[1];
+    size_t ph = fm.padding[0], pw = fm.padding[1];
+
+    auto&& stream = cuda_stream(args.opr->handle());
+
+    int8_t* filter_ptr = nullptr;
+    // TODO: weight preprocess
+    {
+        filter_ptr = reinterpret_cast<int8_t*>(args.workspace.raw_ptr);
+        // reformat filter from nc4hw4 to n4hwc4
+        megdnn::cuda::deconv::reorder_filter_nc4hw4_to_n4hwc4(
+                filter_ptr, args.filter_tensor->compatible_ptr<int8_t>(), co,
+                ci, fh, fw, stream);
+    }
+    convolution::ConvParam kern_param;
+    kern_param.n = n, kern_param.co = co, kern_param.ci = ci,
+    kern_param.hi = hi, kern_param.wi = wi, kern_param.ho = ho,
+    kern_param.wo = wo, kern_param.ph = ph, kern_param.pw = pw,
+    kern_param.sh = sh, kern_param.sw = sw, kern_param.fh = fh,
+    kern_param.fw = fw;
+
+    float diff_scale =
+                  args.diff_layout->dtype.param<dtype::QuantizedS8>().scale,
+          filter_scale =
+                  args.filter_layout->dtype.param<dtype::QuantizedS8>().scale,
+          grad_scale =
+                  args.grad_layout->dtype.param<dtype::QuantizedS8>().scale;
+    float alpha = diff_scale * filter_scale / grad_scale;
+    cutlass_wrapper::do_deconv_int8_implicit_gemm_dp4a_ncdiv4hw4(
+            args.diff_tensor->compatible_ptr<int8_t>(), filter_ptr,
+            args.grad_tensor->compatible_ptr<int8_t>(), nullptr, kern_param,
+            alpha,
+            cutlass_wrapper::GemmCoord{m_algo_param.threadblock_m,
+                                       m_algo_param.threadblock_n,
+                                       m_algo_param.threadblock_k},
+            cutlass_wrapper::GemmCoord{m_algo_param.warp_m, m_algo_param.warp_n,
+                                       m_algo_param.warp_k},
+            m_algo_param.stage, stream);
+
+    after_kernel_launch();
+}
+
+void ConvolutionBackwardDataImpl::AlgoPack::fill_int8_dp4a_algos() {
+    using AlgoParam = AlgoInt8NCHW4DotProdImplicitGemm::AlgoParam;
+    int8_nchw4_dotprod.emplace_back(AlgoParam{16, 64, 8, 16, 64, 8, 2});
+    int8_nchw4_dotprod.emplace_back(AlgoParam{16, 128, 16, 16, 64, 16, 2});
+    int8_nchw4_dotprod.emplace_back(AlgoParam{16, 128, 16, 16, 128, 16, 1});
+    int8_nchw4_dotprod.emplace_back(AlgoParam{32, 128, 32, 32, 64, 32, 2});
+    int8_nchw4_dotprod.emplace_back(AlgoParam{64, 128, 32, 64, 32, 32, 2});
+}
+
+// vim: syntax=cpp.doxygen
diff --git a/dnn/src/cuda/convolution/backward_data/int8/deconv_int8_implicit_gemm_cutlass_wrapper.cuinl b/dnn/src/cuda/convolution/backward_data/int8/deconv_int8_implicit_gemm_cutlass_wrapper.cuinl
new file mode 100644
index 000000000..f5382f919
--- /dev/null
+++ b/dnn/src/cuda/convolution/backward_data/int8/deconv_int8_implicit_gemm_cutlass_wrapper.cuinl
@@ -0,0 +1,62 @@
+/**
+ * \file
+ * dnn/src/cuda/conv_bias/int8/conv_bias_int8_implicit_gemm_cutlass_wrapper.cuinl
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
+ */
+#include "cutlass/convolution/device/convolution.h"
+#include "src/cuda/convolution/backward_data/cutlass_deconvolution_wrapper.cuh"
+
+using namespace megdnn;
+using namespace cuda;
+using namespace cutlass_wrapper;
+
+template <typename Deconvolution>
+void megdnn::cuda::cutlass_wrapper::cutlass_deconvolution_wrapper(
+        const typename Deconvolution::ElementSrc* d_src,
+        const typename Deconvolution::ElementFilter* d_filter,
+        const typename Deconvolution::ElementBias* d_bias,
+        const typename Deconvolution::ElementDst* d_z,
+        typename Deconvolution::ElementDst* d_dst, int* workspace,
+        typename Deconvolution::ConvolutionParameter const& conv_param,
+        typename Deconvolution::EpilogueOutputOp::Params const& epilogue,
+        cudaStream_t stream) {
+    typename Deconvolution::TensorRefSrc tensor_src{
+            const_cast<typename Deconvolution::ElementSrc*>(d_src),
+            Deconvolution::LayoutSrc::packed(
+                    {conv_param.N, conv_param.P, conv_param.Q, conv_param.K})};
+    typename Deconvolution::TensorRefFilter tensor_filter{
+            const_cast<typename Deconvolution::ElementFilter*>(d_filter),
+            Deconvolution::LayoutFilter::packed(
+                    {conv_param.K, conv_param.R, conv_param.S, conv_param.C})};
+    typename Deconvolution::TensorRefBias tensor_bias{
+            const_cast<typename Deconvolution::ElementBias*>(d_bias),
+            Deconvolution::LayoutBias::packed({1, 1, 1, conv_param.K})};
+    typename Deconvolution::TensorRefDst tensor_z{
+            const_cast<typename Deconvolution::ElementDst*>(d_z),
+            Deconvolution::LayoutDst::packed(
+                    {conv_param.N, conv_param.H, conv_param.W, conv_param.C})};
+    typename Deconvolution::TensorRefDst tensor_dst{
+            d_dst,
+            Deconvolution::LayoutDst::packed(
+                    {conv_param.N, conv_param.H, conv_param.W, conv_param.C})};
+    typename Deconvolution::Arguments arguments{conv_param,
+                                                tensor_src.non_const_ref(),
+                                                tensor_filter.non_const_ref(),
+                                                tensor_bias.non_const_ref(),
+                                                tensor_z.non_const_ref(),
+                                                tensor_dst.non_const_ref(),
+                                                epilogue};
+    Deconvolution deconv_op;
+    cutlass_check(deconv_op.initialize(arguments, workspace));
+    cutlass_check(deconv_op(stream));
+    after_kernel_launch();
+}
+
+// vim: syntax=cuda.doxygen
diff --git a/dnn/src/cuda/convolution/backward_data/int8/kimpl/deconv_int8_implicit_gemm_dp4a_ncdiv4hw4_16x128x16_16x128x16_id.cu b/dnn/src/cuda/convolution/backward_data/int8/kimpl/deconv_int8_implicit_gemm_dp4a_ncdiv4hw4_16x128x16_16x128x16_id.cu
new file mode 100644
index 0000000000000000000000000000000000000000..296c397a9ee3dae4705d104ee50e0e22a1f7757f
GIT binary patch
literal 1837
zcmbVN*-jfV6n)QExIwB~6o_amgiMMANt#w=sTd^sHgardV#VXJyd;FL@3m*i1c-p<
z!EtP#yPb1y_QVW*xEfDJ*VpOYcyfE5{ypmS`zYgz8{#-a({=cfQqM?A3$;wAg4lGy
zjY>tXr9IT^01%lf3`Z+slu#KeW{7$xiM5?QZAg|AG#L&N69nE_Cn(b1&Uf9F#u(}z
z+<WCMX9t=Xl5@vR_YD~9jG*p-NQe}I0f<mkdd5-L8rl!3^g}9z^iCA2KP7YlfwPnm
zNBTm!lRo1y*RYRi3MyQh@jQn@#I(|tAzE{DNWD<9+v)h=3SE+Q;ay@V`d$|RV;K)I
zzT?Uk=6d+>^ZsG*?d)_5^Fv5TZ#V4s!)-Er4jWmw8}rQ7-P~}(E@VL$$((53hKB4g
zp1>Nf21my|boAu}mg6SWdt&r|0e>kV#nVvk*YV)8hc@*YYaxrw^Bdg`msh74yO<Ql
z7eMzpW0yj4VunRN6^aChLrHQy=%5!qkzWVSdI+{(zof8lHZxfe*Q7p*il`*h4kQQ#
zW5$<)a*Vg0l}>9XjW(i>li;g7DyiR`#!$;Eptlv#OK0(MB}$Bmp>vFHG~{e;B62`K
zxl8h6sBT0(RYMOo{PxgO#(HSOLxpcFmnA59m<DlSJlyoxC^LkWQY<~jC&@}YK1yDh
z3WWC%hz(RypEc&|q^`VlLd!LfoDk>3d;)qLhcN_Ov@8TeInS6<7>6bg>sGUF4$A(n
z*sObbutnb4r$#c6AGDcUYp!4lp{-{ak0sB!3QfO{SPB!S$|>mWO~{h}0G`Dm@Lqfn
zen5xG{9Y^wEI=PyxL5M55G*LOsSs#<AFT?ru$s`Cz)*l)##;LLQ_uFYS|7vLGK~|z
y<zbhhMwQ-a?`|~4(Vnf1)pMd{Drh5dFbVV2A(%_B`Fy(Te?=F%v&R(^v(7&QhjRx2

literal 0
HcmV?d00001

diff --git a/dnn/src/cuda/convolution/backward_data/int8/kimpl/deconv_int8_implicit_gemm_dp4a_ncdiv4hw4_16x128x16_16x64x16_id.cu b/dnn/src/cuda/convolution/backward_data/int8/kimpl/deconv_int8_implicit_gemm_dp4a_ncdiv4hw4_16x128x16_16x64x16_id.cu
new file mode 100644
index 0000000000000000000000000000000000000000..577303465bc238d2b0524ebeba2b7847ed8b3f36
GIT binary patch
literal 1836
zcmb7F*-jfV6n)QExIwB~6o{ZGLMBC`NlL4-R16Y*8#%T!vEuPqUJ}CB_u8{$0z^O_
z9LM&#+d21UPt4G#>&fN#<|e(LT;5%z|BgESKFYY_hB(g9bRB-A)H9OOLM_v&AU0ia
zqf(J;X%F=}07Rw=!_kTuB~*rr8KT}vVr^$n8<OP&U5-YG2?Fn|6BKE0=ezDoV+?f<
z9=!6Fvja^G$+_dE`wonCMo@P^Bt#0q07NJ%J>#fr4ef_i`XQA<dM66ipAx!&z*)+O
zBYmOVX`k_!YuLv$1r;vMc%DNcVp{3S5Usg6q+Tf5?R0!_g|5iD@GdbFeXk3Ev5bcp
z-*aUPb2EDU_3$`6IX~ON{1g(>+YS5kc$bV`!p7F^#=LNKKR2APAF`l}WKJ}1Lqm2L
zU&0!%hDU=QIyyduWzdBBK#YDDaB#AP`>lWzPeZleh6A#PHszUUA&boOTiuS9SEm@e
zm=nf7fb9#$u7u*mjEa0J6bTMTlH_{WL9cuwy$+o95Ny94r?77}Gg%PVWIm39s1(x<
zB<KWV#+QO}jJKYZPHQKPG@_5w;Hx|;nctkoP|GV|w-wQAXYp~RNsNi1bBu2_<ZNvs
zazH=3OY#9!H=>@4p@$lJd+0f1J+$GW!Z()75|lhlgSIdpZQ5&`8Nx~>mLB7iWF?-S
zB(F>zM-EPU2vky^HRkK2uDx_Z%QcXk5a+{q0(u;WDFj=zECfS2&zMpehaL~>R<CXj
z%J#0<tXp}wMc(<BMlz5ebeUUgu3!kEt!EfdB+t1DJ-?4w3InFfDd_E8$ddm6p2ZpP
zQG5`7K!?HnQ7i~7K%ZK;H}b3yEGV<75NLcKtqQZSn$ViSP=H;=TKf1~kM^=!pTpNO
zjRU{sVV9vsmELLZZZ*cyo~@15bE0J`Xd`hj2=mk-m`kwvdbaCtMHjlW#}yN^&VM^A
Ba?=0+

literal 0
HcmV?d00001

diff --git a/dnn/src/cuda/convolution/backward_data/int8/kimpl/deconv_int8_implicit_gemm_dp4a_ncdiv4hw4_16x64x8_16x64x8_id.cu b/dnn/src/cuda/convolution/backward_data/int8/kimpl/deconv_int8_implicit_gemm_dp4a_ncdiv4hw4_16x64x8_16x64x8_id.cu
new file mode 100644
index 0000000000000000000000000000000000000000..a22525cdcc33716dbbe00d571c7dc3ebc4429738
GIT binary patch
literal 1833
zcmbVNTTk0C6n^io@QO5TC{V!=LZ(B4w(F*GnFtbli(K1HtvZh7*uC)UckEo60u``6
zr0$8obNjw?x+kXS)6Hl+yuD2xM&tX-<nK|Z*Fz~++z`tdnymegggQnNnyW=J5u`}w
z+^9rkS{8?T?ExZHx#4I@j1nqE`4mxSB`J!|o;D=S2pSIthzSDs3M(kmu9)q5S2|;;
zd+_L#D>ysQ#E^_zZo2Ql6xIl84~T?JAsK)OMWtgLbqhm#zLcIXC6~^MT=gb|&LMG@
zFk(qhD0|jpJk;vXp-e%AO;Vm^(1=i0vNS|%ZVssvN_IOP=R=`uvd*224MpGU24F0M
z5#tB0irm}|RzDwC{gaFHEzA!gp}gI&->dt0@DesG>~749P!BW13A>UxoyRkxc^m4h
z!*~qO@Tz}w8llsZ2%R;FJ`$t<1NKV+A&&Z1zm1O2QG_<V8EGN&)bTsrPL)?yjBO|g
z<13*0lCf){I5C4fn+QdGz=0&0?sw2DpTMn!U=f1jx8nrP&0#8YVw=K;K@F8$+KG6P
zV9fYJP>%7|wK8ZOq|rr)apq%{rjquJHKba;0%}_sy=DuSD?wsR44q+or=jMeh(!kI
zXYb<d6uKKw&y^6NhTRA~XDmV+9x8m}xoknn!^8^<<H2UOhN&T}lwlb$T#1+B=}Gd+
z)p6k9BtoE)+Sa(Qi@I^r3N6<_a!joAy9wyA999q<(IOWNWjtj{VeALoKeq;TV<>04
zVzZv({uX%`UmD3keK2Bnt+|2?gto3>Jd!-)$`AZLVhL=QswS_ucOgsuJ$M=xz(?_3
z_#W-I@<*{CFb{od;ohjTLNKq)rbD3deYDKYywHT!1o{T-a@NwvpL(yC-TLgm7D?Fl
ztq8juHL7%0JA0=w4*rVTS-mG(uDmv43tKQtEP}lRhp*?m{!?_WJ9}I)G41>V9N}_A

literal 0
HcmV?d00001

diff --git a/dnn/src/cuda/convolution/backward_data/int8/kimpl/deconv_int8_implicit_gemm_dp4a_ncdiv4hw4_32x128x32_32x64x32_id.cu b/dnn/src/cuda/convolution/backward_data/int8/kimpl/deconv_int8_implicit_gemm_dp4a_ncdiv4hw4_32x128x32_32x64x32_id.cu
new file mode 100644
index 0000000000000000000000000000000000000000..400b5db240a6868af8ef39dcf6faa217eef16692
GIT binary patch
literal 1837
zcmb7FTTk0C6n^io@QO5TC{R%lLZ(B4w(F*GnFtbli(K1HtvZh7*e&JP@7TFC1uCEq
zse9t@+`jLe?ui-tax<QcZf}!^@#Ow8`FGUm^-#(cH^g#=rfdHrp^lM+=4zQt1u2pR
zH!2aCmc^l7dw@t)Za7*Iql8LPK10-5Ns6MgrwvInf+oWuVuHZE!U~GCE9Sf2mChLI
z9y~hb3eFBRF(l)bo9+iNg*AfO10o?)NCqH6QRx^*-NMkGFQw;8$)&R*SG_5r3rL(L
zj9AhW%AWNY54HMpC{s{jlayx}G$NFhtPIhbn?ve^lHE?n`B3PZtaE2$L(z}A0T|0*
z#Q1@$A~(0gr(cgxgOiK%EzD0Lp}gI&KTr4Z@HK2y*xi^{p&sUj6LuwYx`^jQ^ET91
zhw%iS;nm={AEBfES%i-JO{$N?=nnx;PquKs6>#FH@AkX!BM2LzO?$>#$UJrYPPen=
zl@((fYQp#m*uG@!S}0D;Fwdq!5g%|UNu~!K^u{O9Yav*K;CS3m;M^Q$GAFjFd=vyx
zDW;u>*9pdqF9qcoZ(S>c)<GI+gcxT&R%t4k-&jMc<tt#fmC;+aaJkYX#>CJ$#&;TO
zE{a%WfPV2V&Q77b0rgT05o+j-&`ZW5wBez`H=fHDlsrtmwlE%U#%q)s!b&BU5yO>u
zC7z!puUvuDPJs5)2!Tv$W8=TB>c&Ycv|IzrF|p3?C*a3&SVM3`%Um#&@sufru^)2(
z+#1%6p`7oE&U%&yTkKtYYcvA|!kF2$<_fkD+Pa4ESn`Z3KlJ;EC9q+tqP*Zfge(Q%
z!LzUgK8yFt_h`SFKZ^y0dFXQs_fDM^f_Z5+9RiW>qg8Geg(kEn&^KV0vz9>q*1Nsz
z)>r?vOv1)*McC!2k)^ZR**lGK@K@B%>P69V<+Tx8*o0YP5iBM+d_UjyzoK*9+2e|d
GS?50}6mtFm

literal 0
HcmV?d00001

diff --git a/dnn/src/cuda/convolution/backward_data/int8/kimpl/deconv_int8_implicit_gemm_dp4a_ncdiv4hw4_64x128x32_64x32x32_id.cu b/dnn/src/cuda/convolution/backward_data/int8/kimpl/deconv_int8_implicit_gemm_dp4a_ncdiv4hw4_64x128x32_64x32x32_id.cu
new file mode 100644
index 0000000000000000000000000000000000000000..c149a8e1b1bd7db1f72f2d06afee60c913ef14b8
GIT binary patch
literal 1837
zcmb7FTTk0C6n^io@QO5TC{R%lLZ(B4w(F*GnFtbli(K1HtvZh7*e&JP@7TFC1uCEq
zse9t@+`jLe?ui-tax<QcZf}!^@#Ow8`FGUm^-#(cH^g#=rfdHrp^lM+=4zQt1u2pR
zH!2aCmc^l7dw@t)Za7*Iql8LPK10-5Ns6MgrwvInf+oWuVuHZE!U~GCE9Sf2mChLI
z9y~hb3eFBRF(l)bo9+iNg*AfO10o?)NCqH6QRx^*-NMkGFQw;8$)&R*SG_5r3rL(L
zj9AhW%AWNY54HMpC{s{jlayx}G$NFhtPIhbn?ve^lHE?n`B3PZtaE2$L(z}A0T|0*
z#Q1@$A~(0gr(cgxgOiK%EzD0Lp}gI&KTr4Z@HK2y*xi^{p&sUj6LuwYx`^jQ^ET91
zhw%iS;nm>uBtl31vj`pcn^Yf((H{aH_qTAr6>#FH@AkX!BM2LzO?$>#$UJrYPPen=
zl@((fYQp#m*uG@!S}0D;Fwdq!5g%|UNu~!K^u{O9Yav*K;CS3m;M^Q$GAFjFd=vyx
zDW;u>*9pdqF9qcoZ(S>c)<GI+gcxT&R%t4k-&jMc<tt#fmC;+aaJkYX#>CJ$#&;TO
zE{a%WfPV2V&Q77b0rgT05o+j-&`ZW5wBez`H=fHDlsrtmwlE%U#%q)s!b&BU5yO>u
zC7z!puUvszPJs5)2!Tv$W8=TB>c&Ycv|IzrF|p3?C*a3&SVM3`%Um#&@sufru^)2(
z+#1%6p`7oE&U%&yTkKtYYcvA|!kF2$<_fkD+Pa4ESn`Z3KlJ;EC9q+tqP*Zfge(Q%
z!LzUgK8yFt_h`SFKZ^y0dFXQs_fDM^f_Z5+9RiW>qg8Geg(kEn&^KV0vz9>q*1Nsz
z)>r?vOv1)*McC!2k)^ZR**lGK@K@B%>P69V<+Tx8*o0YP5iBM+d_UjyzoK*9+2e|d
GS?51NfO7)?

literal 0
HcmV?d00001

diff --git a/dnn/src/cuda/convolution/opr_impl.cpp b/dnn/src/cuda/convolution/opr_impl.cpp
index badbde22d..a72d9c8f5 100644
--- a/dnn/src/cuda/convolution/opr_impl.cpp
+++ b/dnn/src/cuda/convolution/opr_impl.cpp
@@ -6,7 +6,8 @@
  *
  * Unless required by applicable law or agreed to in writing,
  * software distributed under the License is distributed on an
- * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
  */
 
 #include "src/cuda/convolution/opr_impl.h"
@@ -25,8 +26,9 @@ using namespace convolution;
 
 #define TO_STRING2(v) #v
 #define TO_STRING(v) TO_STRING2(v)
-#define CUDNN_VERSION_STR TO_STRING(CUDNN_MAJOR) "." \
-    TO_STRING(CUDNN_MINOR) "." TO_STRING(CUDNN_PATCHLEVEL)
+#define CUDNN_VERSION_STR  \
+    TO_STRING(CUDNN_MAJOR) \
+    "." TO_STRING(CUDNN_MINOR) "." TO_STRING(CUDNN_PATCHLEVEL)
 
 /* ============== ConvolutionForwardImpl ============== */
 ConvolutionForwardImpl::Algorithm*
@@ -72,25 +74,24 @@ void ConvolutionForwardImpl::exec(_megdnn_tensor_in src,
 }
 
 const char* ConvolutionForwardImpl::get_algorithm_set_name() const {
-    return "CUDA CONVOLUTION_FORWARD" ;
+    return "CUDA CONVOLUTION_FORWARD";
 }
 
 /* ============== ConvolutionBackwardDataImpl ============== */
 
 void ConvolutionBackwardDataImpl::exec(_megdnn_tensor_in filter,
-        _megdnn_tensor_in diff,
-        _megdnn_tensor_out grad,
-        _megdnn_workspace workspace) {
+                                       _megdnn_tensor_in diff,
+                                       _megdnn_tensor_out grad,
+                                       _megdnn_workspace workspace) {
     AlgoBase::ExecArgs args(this, filter, diff, grad, workspace);
-    auto algo = get_algorithm(this, filter.layout,
-                              diff.layout, grad.layout);
+    auto algo = get_algorithm(this, filter.layout, diff.layout, grad.layout);
     algo->check_workspace(args, workspace).exec(args);
 }
 
-std::vector<ConvolutionBackwardDataImpl::Algorithm *>
-ConvolutionBackwardDataImpl::get_all_algorithms(const TensorLayout &filter,
-        const TensorLayout &diff,
-        const TensorLayout &grad) {
+std::vector<ConvolutionBackwardDataImpl::Algorithm*>
+ConvolutionBackwardDataImpl::get_all_algorithms(const TensorLayout& filter,
+                                                const TensorLayout& diff,
+                                                const TensorLayout& grad) {
     return megdnn::get_all_algorithms<ConvolutionBackwardDataImpl>(
             {this, filter, diff, grad});
 }
@@ -106,10 +107,10 @@ ConvolutionBackwardDataImpl::get_algorithm_heuristic(
 }
 
 ConvolutionBackwardDataImpl::Algorithm*
-ConvolutionBackwardDataImpl::get_algorithm_heuristic(const TensorLayout& filter,
-        const CanonizedFilterMeta& filter_meta, const TensorLayout& diff,
-        const TensorLayout& grad, size_t workspace_limit_in_bytes,
-        bool reproducible) {
+ConvolutionBackwardDataImpl::get_algorithm_heuristic(
+        const TensorLayout& filter, const CanonizedFilterMeta& filter_meta,
+        const TensorLayout& diff, const TensorLayout& grad,
+        size_t workspace_limit_in_bytes, bool reproducible) {
     AlgoBase::SizeArgs args(this, filter, filter_meta, diff, grad);
 
     if (args.filter_meta.group > 1 &&
@@ -119,6 +120,19 @@ ConvolutionBackwardDataImpl::get_algorithm_heuristic(const TensorLayout& filter,
         return &sm_algo_pack.chanwise;
     }
 
+    if (args.filter_layout->dtype.enumv() ==
+        DTypeTrait<dtype::QuantizedS8>::enumv) {
+        if (reproducible) {
+            return megdnn::get_reproducible_algo<ConvolutionBackwardDataImpl>(
+                    sm_algo_pack.int8_algos, args, workspace_limit_in_bytes,
+                    "cuda conv bwd_data");
+        } else {
+            return megdnn::get_usable_algo<ConvolutionBackwardDataImpl>(
+                    sm_algo_pack.int8_algos, args, workspace_limit_in_bytes,
+                    "cuda conv bwd_data");
+        }
+    }
+
     auto get_cudnn_algo =
             [this, &args, workspace_limit_in_bytes,
              reproducible]() -> ConvolutionBackwardDataImpl::AlgoBase* {
@@ -206,12 +220,11 @@ ConvolutionBackwardDataImpl::get_algorithm_heuristic(const TensorLayout& filter,
 }
 
 size_t ConvolutionBackwardDataImpl::get_workspace_in_bytes(
-        const TensorLayout &filter,
-        const TensorLayout &diff,
-        const TensorLayout &grad) {
+        const TensorLayout& filter, const TensorLayout& diff,
+        const TensorLayout& grad) {
     AlgoBase::SizeArgs args(this, filter, diff, grad);
-    return get_algorithm(this, filter, args.filter_meta, diff, grad)->
-        get_workspace_in_bytes(args);
+    return get_algorithm(this, filter, args.filter_meta, diff, grad)
+            ->get_workspace_in_bytes(args);
 }
 
 const char* ConvolutionBackwardDataImpl::get_algorithm_set_name() const {
@@ -221,19 +234,19 @@ const char* ConvolutionBackwardDataImpl::get_algorithm_set_name() const {
 /* ============== ConvolutionBackwardFilterImpl ============== */
 
 void ConvolutionBackwardFilterImpl::exec(_megdnn_tensor_in src,
-        _megdnn_tensor_in diff,
-        _megdnn_tensor_out grad,
-        _megdnn_workspace workspace) {
+                                         _megdnn_tensor_in diff,
+                                         _megdnn_tensor_out grad,
+                                         _megdnn_workspace workspace) {
     AlgoBase::ExecArgs args(this, src, diff, grad, workspace);
-    auto algo = get_algorithm(this, src.layout, diff.layout,
-            grad.layout, args.grad_filter_meta);
+    auto algo = get_algorithm(this, src.layout, diff.layout, grad.layout,
+                              args.grad_filter_meta);
     algo->check_workspace(args, workspace).exec(args);
 }
 
-std::vector<ConvolutionBackwardFilterImpl::Algorithm *>
-ConvolutionBackwardFilterImpl::get_all_algorithms(const TensorLayout &src,
-        const TensorLayout &diff,
-        const TensorLayout &grad) {
+std::vector<ConvolutionBackwardFilterImpl::Algorithm*>
+ConvolutionBackwardFilterImpl::get_all_algorithms(const TensorLayout& src,
+                                                  const TensorLayout& diff,
+                                                  const TensorLayout& grad) {
     return megdnn::get_all_algorithms<ConvolutionBackwardFilterImpl>(
             {this, src, diff, grad});
 }
@@ -269,7 +282,7 @@ ConvolutionBackwardFilterImpl::get_algorithm_heuristic(
         CUDNNBwdFilterDescs desc;
         args.init_desc(desc);
 
-        //disable, segfault in megbrain, need further investigate.
+        // disable, segfault in megbrain, need further investigate.
 #if 0
         auto is_heuristic_success =
                 convolution::PerformanceModelBackwardFilter::
@@ -358,12 +371,11 @@ ConvolutionBackwardFilterImpl::get_algorithm_heuristic(
 }
 
 size_t ConvolutionBackwardFilterImpl::get_workspace_in_bytes(
-        const TensorLayout &src,
-        const TensorLayout &diff,
-        const TensorLayout &grad) {
+        const TensorLayout& src, const TensorLayout& diff,
+        const TensorLayout& grad) {
     AlgoBase::SizeArgs args(this, src, diff, grad);
-    return get_algorithm(this, src, diff, grad, args.grad_filter_meta)->
-        get_workspace_in_bytes(args);
+    return get_algorithm(this, src, diff, grad, args.grad_filter_meta)
+            ->get_workspace_in_bytes(args);
 }
 
 const char* ConvolutionBackwardFilterImpl::get_algorithm_set_name() const {
diff --git a/dnn/src/cuda/convolution/opr_impl.h b/dnn/src/cuda/convolution/opr_impl.h
index 1ca8db092..7f9dbeaa7 100644
--- a/dnn/src/cuda/convolution/opr_impl.h
+++ b/dnn/src/cuda/convolution/opr_impl.h
@@ -105,6 +105,7 @@ public:
     class AlgoChanwiseSmall;
     class AlgoGroupConvGeneral;
     class AlgoBFloat16;
+    class AlgoInt8NCHW4DotProdImplicitGemm;
 
     class AlgoPack;
 
diff --git a/dnn/test/common/convolution.cpp b/dnn/test/common/convolution.cpp
index 5caebcf86..e3aea119e 100644
--- a/dnn/test/common/convolution.cpp
+++ b/dnn/test/common/convolution.cpp
@@ -6,7 +6,8 @@
  *
  * Unless required by applicable law or agreed to in writing,
  * software distributed under the License is distributed on an
- * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
  */
 
 #include "test/common/checker.h"
@@ -44,14 +45,12 @@ std::vector<TestArg> convolution::get_args_common() {
         param::Convolution param;
 
         param.mode = param::Convolution::Mode::CONVOLUTION;
-        args.emplace_back(param,
-                TensorShape{5, 2, i, i+1},
-                TensorShape{3, 2, 3, 4});
+        args.emplace_back(param, TensorShape{5, 2, i, i + 1},
+                          TensorShape{3, 2, 3, 4});
 
         param.mode = param::Convolution::Mode::CROSS_CORRELATION;
-        args.emplace_back(param,
-                TensorShape{5, 2, i, i+1},
-                TensorShape{3, 2, 3, 4});
+        args.emplace_back(param, TensorShape{5, 2, i, i + 1},
+                          TensorShape{3, 2, 3, 4});
     }
 
     return args;
@@ -65,14 +64,12 @@ std::vector<TestArg> convolution::get_args_padding() {
         param.pad_w = 2;
 
         param.mode = param::Convolution::Mode::CONVOLUTION;
-        args.emplace_back(param,
-                TensorShape{5, 2, i, i+1},
-                TensorShape{3, 2, 3, 4});
+        args.emplace_back(param, TensorShape{5, 2, i, i + 1},
+                          TensorShape{3, 2, 3, 4});
 
         param.mode = param::Convolution::Mode::CROSS_CORRELATION;
-        args.emplace_back(param,
-                TensorShape{5, 2, i, i+1},
-                TensorShape{3, 2, 3, 4});
+        args.emplace_back(param, TensorShape{5, 2, i, i + 1},
+                          TensorShape{3, 2, 3, 4});
     }
 
     return args;
@@ -84,14 +81,12 @@ std::vector<TestArg> convolution::get_args_large_channel() {
         param::Convolution param;
 
         param.mode = param::Convolution::Mode::CONVOLUTION;
-        args.emplace_back(param,
-                TensorShape{2, 20, i, i+1},
-                TensorShape{30, 20, 3, 4});
+        args.emplace_back(param, TensorShape{2, 20, i, i + 1},
+                          TensorShape{30, 20, 3, 4});
 
         param.mode = param::Convolution::Mode::CROSS_CORRELATION;
-        args.emplace_back(param,
-                TensorShape{2, 20, i, i+1},
-                TensorShape{30, 20, 3, 4});
+        args.emplace_back(param, TensorShape{2, 20, i, i + 1},
+                          TensorShape{30, 20, 3, 4});
     }
     for (size_t i = 16; i < 24; ++i) {
         param::Convolution param;
@@ -99,14 +94,12 @@ std::vector<TestArg> convolution::get_args_large_channel() {
         param.pad_w = 2;
 
         param.mode = param::Convolution::Mode::CONVOLUTION;
-        args.emplace_back(param,
-                TensorShape{2, 20, i, i+1},
-                TensorShape{30, 20, 3, 4});
+        args.emplace_back(param, TensorShape{2, 20, i, i + 1},
+                          TensorShape{30, 20, 3, 4});
 
         param.mode = param::Convolution::Mode::CROSS_CORRELATION;
-        args.emplace_back(param,
-                TensorShape{2, 20, i, i+1},
-                TensorShape{30, 20, 3, 4});
+        args.emplace_back(param, TensorShape{2, 20, i, i + 1},
+                          TensorShape{30, 20, 3, 4});
     }
 
     return args;
@@ -118,14 +111,12 @@ std::vector<TestArg> convolution::get_args_1x1() {
         param::Convolution param;
 
         param.mode = param::Convolution::Mode::CONVOLUTION;
-        args.emplace_back(param,
-                TensorShape{2, 20, i, i+1},
-                TensorShape{30, 20, 1, 1});
+        args.emplace_back(param, TensorShape{2, 20, i, i + 1},
+                          TensorShape{30, 20, 1, 1});
 
         param.mode = param::Convolution::Mode::CROSS_CORRELATION;
-        args.emplace_back(param,
-                TensorShape{2, 20, i, i+1},
-                TensorShape{30, 20, 1, 1});
+        args.emplace_back(param, TensorShape{2, 20, i, i + 1},
+                          TensorShape{30, 20, 1, 1});
     }
 
     return args;
@@ -137,14 +128,12 @@ std::vector<TestArg> convolution::get_args_large_filter() {
         param::Convolution param;
 
         param.mode = param::Convolution::Mode::CONVOLUTION;
-        args.emplace_back(param,
-                TensorShape{2, 2, i, i+1},
-                TensorShape{3, 2, 7, 8});
+        args.emplace_back(param, TensorShape{2, 2, i, i + 1},
+                          TensorShape{3, 2, 7, 8});
 
         param.mode = param::Convolution::Mode::CROSS_CORRELATION;
-        args.emplace_back(param,
-                TensorShape{2, 2, i, i+1},
-                TensorShape{3, 2, 7, 8});
+        args.emplace_back(param, TensorShape{2, 2, i, i + 1},
+                          TensorShape{3, 2, 7, 8});
     }
 
     return args;
@@ -181,9 +170,8 @@ std::vector<TestArg> convolution::get_args_4x4() {
     for (size_t oh = 1; oh < 20; ++oh) {
         param::Convolution param;
         param.mode = param::Convolution::Mode::CROSS_CORRELATION;
-        args.emplace_back(param,
-                TensorShape{4, 3, oh+3, oh+4},
-                TensorShape{2, 3, 4, 4});
+        args.emplace_back(param, TensorShape{4, 3, oh + 3, oh + 4},
+                          TensorShape{2, 3, 4, 4});
     }
 
     return args;
@@ -289,26 +277,22 @@ std::vector<TestArg> convolution::get_args_fallback_non_templated_impl() {
 std::vector<TestArg> convolution::get_args_cudnn_5_1_failures() {
     std::vector<TestArg> args;
     args.emplace_back(
-            param::Convolution{
-                param::Convolution::Mode::CROSS_CORRELATION, 0, 4, 1, 2},
-            TensorShape{5, 3, 25, 20},
-            TensorShape{10, 3, 7, 4}
-    );
+            param::Convolution{param::Convolution::Mode::CROSS_CORRELATION, 0,
+                               4, 1, 2},
+            TensorShape{5, 3, 25, 20}, TensorShape{10, 3, 7, 4});
 
     return args;
 }
 
 std::vector<TestArg> convolution::get_args_x86_winograd_algorithm() {
     std::vector<TestArg> args;
-    for (size_t ic_size: {8, 16})
-    {
+    for (size_t ic_size : {8, 16}) {
         param::Convolution param;
         param.mode = param::Convolution::Mode::CROSS_CORRELATION;
         param.stride_h = param.stride_w = 1;
         param.pad_h = param.pad_w = 0;
-        args.emplace_back(param,
-                TensorShape{2, ic_size, 102, 102},
-                TensorShape{8, ic_size, 3, 3});
+        args.emplace_back(param, TensorShape{2, ic_size, 102, 102},
+                          TensorShape{8, ic_size, 3, 3});
     }
 
     return args;
@@ -317,18 +301,15 @@ std::vector<TestArg> convolution::get_args_x86_winograd_algorithm() {
 std::vector<TestArg> convolution::get_args_BRAIN_481() {
     std::vector<TestArg> args;
     {
-        param::Convolution param{param::Convolution::Mode::CROSS_CORRELATION,
-            0, 1, 1, 2};
-        args.emplace_back(param,
-                TensorShape{4, 4, 14, 13},
-                TensorShape{3, 4, 8, 13});
-        for (size_t margin = 0; margin < 5; ++margin)
-        {
-            param::Convolution param{param::Convolution::Mode::CROSS_CORRELATION,
-                1, 1, 2, 2};
-            args.emplace_back(param,
-                    TensorShape{4, 4, 14, 13},
-                    TensorShape{3, 4, 16-margin, 15-margin});
+        param::Convolution param{param::Convolution::Mode::CROSS_CORRELATION, 0,
+                                 1, 1, 2};
+        args.emplace_back(param, TensorShape{4, 4, 14, 13},
+                          TensorShape{3, 4, 8, 13});
+        for (size_t margin = 0; margin < 5; ++margin) {
+            param::Convolution param{
+                    param::Convolution::Mode::CROSS_CORRELATION, 1, 1, 2, 2};
+            args.emplace_back(param, TensorShape{4, 4, 14, 13},
+                              TensorShape{3, 4, 16 - margin, 15 - margin});
         }
     }
 
@@ -337,7 +318,7 @@ std::vector<TestArg> convolution::get_args_BRAIN_481() {
 
 std::vector<TestArg> convolution::get_args() {
     std::vector<TestArg> all_args, args;
-#define ADD_ARGS(NAME) \
+#define ADD_ARGS(NAME)        \
     args = get_args_##NAME(); \
     all_args.insert(all_args.end(), args.begin(), args.end());
     ADD_ARGS(common)
@@ -356,12 +337,12 @@ std::vector<TestArg> convolution::get_args() {
     ADD_ARGS(BRAIN_481)
 #undef ADD_ARGS
 
-   return all_args;
+    return all_args;
 }
 
 std::vector<TestArg> convolution::get_args_cuda_conv_bwd_data() {
     std::vector<TestArg> all_args, args;
-#define ADD_ARGS(NAME) \
+#define ADD_ARGS(NAME)        \
     args = get_args_##NAME(); \
     all_args.insert(all_args.end(), args.begin(), args.end());
     ADD_ARGS(common)
@@ -378,19 +359,19 @@ std::vector<TestArg> convolution::get_args_cuda_conv_bwd_data() {
     ADD_ARGS(x86_winograd_algorithm)
 #undef ADD_ARGS
 
-   return all_args;
+    return all_args;
 }
 
 std::vector<TestArg> convolution::get_args_cudnn_7_5_failures() {
     std::vector<TestArg> all_args, args;
-#define ADD_ARGS(NAME) \
+#define ADD_ARGS(NAME)        \
     args = get_args_##NAME(); \
     all_args.insert(all_args.end(), args.begin(), args.end());
     ADD_ARGS(cudnn_5_1_failures)
     ADD_ARGS(BRAIN_481)
 #undef ADD_ARGS
 
-   return all_args;
+    return all_args;
 }
 std::vector<TestArg> convolution::get_chanwise_args() {
     std::vector<TestArg> args;
@@ -421,12 +402,9 @@ std::vector<TestArg> convolution::get_dilated_args() {
     param::Convolution param;
     param.pad_h = param.pad_w = 2;
     param.dilate_h = param.dilate_w = 2;
-    size_t n = 1, ic = 15, ih = 128, iw = 128,
-           fh = 3, fw = 3,
-           oc = 17;
-    args.emplace_back(param,
-            TensorShape{n, ic, ih, iw},
-            TensorShape{oc, ic, fh, fw});
+    size_t n = 1, ic = 15, ih = 128, iw = 128, fh = 3, fw = 3, oc = 17;
+    args.emplace_back(param, TensorShape{n, ic, ih, iw},
+                      TensorShape{oc, ic, fh, fw});
     // exhaustive search
     // clang-format off
     for (size_t n: {2})
@@ -451,13 +429,44 @@ std::vector<TestArg> convolution::get_dilated_args() {
     return args;
 }
 
-void convolution::test_conv_config_combinations(int k_size,
-                                                Handle* handle, bool test_int8,
-                                                bool test_backward,
-                                                bool is_cuda,
-                                                ConvEPSGetter eps_getter,
-                                                bool use_io16xc32) {
-    Checker<Convolution> checker(handle);
+std::vector<TestArg> convolution::get_args_int8_nchw4_conv_bwd_data() {
+    std::vector<TestArg> args;
+    param::Convolution cur_param;
+
+    // clang-format off
+    for (auto mode : {param::ConvBias::Mode::CROSS_CORRELATION}) {
+    for (size_t b : {64, 16}) {
+    for (size_t ic : {16, 32}) {
+    for (size_t oc : {16, 32}) {
+    for (size_t h : {8}) {
+    for (size_t w : {8, 11}) {
+    for (size_t kernel_size : {3, 4, 5, 7}) {
+    for (int p : {0, static_cast<int>(kernel_size / 2)}) {
+    for (size_t s : {2}) {
+        if (kernel_size >= 7) {
+            b = std::min(b, 32_z);
+        }
+        size_t f = kernel_size;
+        cur_param.mode = mode;
+
+        cur_param.format = param::ConvBias::Format::NCHW4;
+        cur_param.sparse = param::ConvBias::Sparse::DENSE;
+        cur_param.pad_h = cur_param.pad_w = p;
+        cur_param.stride_h = cur_param.stride_w = s;
+
+        //! bias channel
+        args.emplace_back(cur_param, TensorShape{b, ic / 4, h, w, 4},
+                          TensorShape{oc, ic / 4, f, f, 4});
+    } } } } } } } } }
+    // clang-format on
+
+    return args;
+}
+
+void convolution::test_conv_config_combinations(
+        int k_size, Handle* handle, bool test_int8, bool test_backward,
+        bool is_cuda, ConvEPSGetter eps_getter, bool use_io16xc32) {
+Checker<Convolution> checker(handle);
     std::unique_ptr<Checker<ConvolutionBackwardData>> checker_bwd_data_ptr;
     std::unique_ptr<Checker<ConvolutionBackwardFilter>> checker_bwd_filter_ptr;
     if (test_backward) {
@@ -657,7 +666,6 @@ void convolution::test_conv_config_combinations(int k_size,
             }
         }
     }
-
 }
 
 // vim: syntax=cpp.doxygen
diff --git a/dnn/test/common/convolution.h b/dnn/test/common/convolution.h
index 2d9a338d6..99b3a8f97 100644
--- a/dnn/test/common/convolution.h
+++ b/dnn/test/common/convolution.h
@@ -6,7 +6,8 @@
  *
  * Unless required by applicable law or agreed to in writing,
  * software distributed under the License is distributed on an
- * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
  */
 #pragma once
 
@@ -47,6 +48,7 @@ std::vector<TestArg> get_args_cudnn_7_5_failures();
 std::vector<TestArg> get_1x1_args();
 std::vector<TestArg> get_dilated_args();
 std::vector<TestArg> get_chanwise_args();
+std::vector<TestArg> get_args_int8_nchw4_conv_bwd_data();
 
 //! \param stage 0 for fwd, 1 for bwd data, 2 for bwd filter
 using ConvEPSGetter =
diff --git a/dnn/test/cuda/convolution.cpp b/dnn/test/cuda/convolution.cpp
index 8dd7f4b8c..5d6a2a6f9 100644
--- a/dnn/test/cuda/convolution.cpp
+++ b/dnn/test/cuda/convolution.cpp
@@ -6,7 +6,8 @@
  *
  * Unless required by applicable law or agreed to in writing,
  * software distributed under the License is distributed on an
- * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
  */
 #include "megdnn/dtype.h"
 #include "megdnn/oprs.h"
@@ -18,7 +19,6 @@
 #include "test/common/convolution.h"
 #include "test/common/rng.h"
 #include "test/cuda/benchmark.h"
-
 #include "src/cuda/utils.h"
 
 #define V1(x) #x
@@ -29,8 +29,7 @@
 namespace megdnn {
 namespace test {
 
-TEST_F(CUDA, CONVOLUTION_8X8X32)
-{
+TEST_F(CUDA, CONVOLUTION_8X8X32) {
     if (!cuda::is_compute_capability_required(6, 1)) {
         printf("Skip CUDA.CONVOLUTION_8X8X32 test as current device"
                "doesn't support\n");
@@ -41,66 +40,63 @@ TEST_F(CUDA, CONVOLUTION_8X8X32)
     std::vector<TestArg> args;
     {
         auto v = get_args();
-        for (auto &&a: v) {
+        for (auto&& a : v) {
             args.push_back(std::move(a));
         }
     }
     {
         auto v = get_dilated_args();
-        for (auto &&a: v) {
+        for (auto&& a : v) {
             args.push_back(std::move(a));
         }
     }
     {
         auto v = get_chanwise_args();
-        for (auto &&a: v) {
+        for (auto&& a : v) {
             args.push_back(std::move(a));
         }
     }
     Checker<ConvolutionForward> checker(handle_cuda());
     UniformIntRNG rng(-4, 4);
-    for (auto arg: args) {
+    for (auto arg : args) {
         arg.param.format = param::Convolution::Format::NHWC;
         arg.src = cvt_src_or_dst_nchw2nhwc(arg.src);
         arg.filter = cvt_filter_nchw2nhwc(arg.filter);
-        checker.set_dtype(0, dtype::Int8()).
-            set_dtype(1, dtype::Int8()).
-            set_dtype(2, dtype::Int32()).
-            set_param(arg.param).
-            set_rng(0, &rng).
-            set_rng(1, &rng).
-            execs({arg.src, arg.filter, {}});
+        checker.set_dtype(0, dtype::Int8())
+                .set_dtype(1, dtype::Int8())
+                .set_dtype(2, dtype::Int32())
+                .set_param(arg.param)
+                .set_rng(0, &rng)
+                .set_rng(1, &rng)
+                .execs({arg.src, arg.filter, {}});
     }
 }
 
-TEST_F(CUDA, CONVOLUTION_FORWARD)
-{
+TEST_F(CUDA, CONVOLUTION_FORWARD) {
     using namespace convolution;
     std::vector<TestArg> args = get_args();
     Checker<ConvolutionForward> checker(handle_cuda());
     NormalRNG default_rng;
-    for (auto &&arg: args) {
+    for (auto&& arg : args) {
         float scale =
                 1.0f / sqrt(arg.filter[1] * arg.filter[2] * arg.filter[3]);
         UniformFloatRNG rng(scale, 2 * scale);
-        checker.
-            set_dtype(0, dtype::Float32()).
-            set_dtype(1, dtype::Float32()).
-            set_dtype(2, dtype::Float32()).
-            set_rng(0, &default_rng).
-            set_rng(1, &default_rng).
-            set_epsilon(1e-3).
-            set_param(arg.param).
-            execs({arg.src, arg.filter, {}});
-        checker.
-            set_dtype(0, dtype::Float16()).
-            set_dtype(1, dtype::Float16()).
-            set_dtype(2, dtype::Float16()).
-            set_rng(0, &rng).
-            set_rng(1, &rng).
-            set_epsilon(1e-1).
-            set_param(arg.param).
-            execs({arg.src, arg.filter, {}});
+        checker.set_dtype(0, dtype::Float32())
+                .set_dtype(1, dtype::Float32())
+                .set_dtype(2, dtype::Float32())
+                .set_rng(0, &default_rng)
+                .set_rng(1, &default_rng)
+                .set_epsilon(1e-3)
+                .set_param(arg.param)
+                .execs({arg.src, arg.filter, {}});
+        checker.set_dtype(0, dtype::Float16())
+                .set_dtype(1, dtype::Float16())
+                .set_dtype(2, dtype::Float16())
+                .set_rng(0, &rng)
+                .set_rng(1, &rng)
+                .set_epsilon(1e-1)
+                .set_param(arg.param)
+                .execs({arg.src, arg.filter, {}});
         arg.param.compute_mode = param::Convolution::ComputeMode::FLOAT32;
         checker.set_dtype(0, dtype::Float16())
                 .set_dtype(1, dtype::Float16())
@@ -152,51 +148,49 @@ TEST_F(CUDA, CONV_FORWARD_MATMUL_NCHW4) {
     checker.exec({{8, 64, 12, 12, 4}, {256, 64, 3, 3, 4}, {}});
 }
 
-TEST_F(CUDA, CONVOLUTION_1X1_FORWARD)
-{
+TEST_F(CUDA, CONVOLUTION_1X1_FORWARD) {
     using namespace convolution;
     std::vector<TestArg> args = get_1x1_args();
     Checker<ConvolutionForward> checker(handle_cuda());
     NormalRNG default_rng;
-    for (auto &&arg: args) {
-        float scale = 1.0f / sqrt(arg.filter[1] * arg.filter[2] * arg.filter[3]);
+    for (auto&& arg : args) {
+        float scale =
+                1.0f / sqrt(arg.filter[1] * arg.filter[2] * arg.filter[3]);
         UniformFloatRNG rng(scale, 2 * scale);
-        checker.
-            set_dtype(0, dtype::Float32()).
-            set_dtype(1, dtype::Float32()).
-            set_rng(0, &default_rng).
-            set_rng(1, &default_rng).
-            set_epsilon(1e-3).
-            set_param(arg.param).
-            execs({arg.src, arg.filter, {}});
+        checker.set_dtype(0, dtype::Float32())
+                .set_dtype(1, dtype::Float32())
+                .set_rng(0, &default_rng)
+                .set_rng(1, &default_rng)
+                .set_epsilon(1e-3)
+                .set_param(arg.param)
+                .execs({arg.src, arg.filter, {}});
     }
 }
 
-TEST_F(CUDA, BENCHMARK_CONVOLUTION_1X1_FORWARD)
-{
+TEST_F(CUDA, BENCHMARK_CONVOLUTION_1X1_FORWARD) {
     using namespace convolution;
     std::vector<TestArg> args = get_1x1_args();
     Benchmarker<ConvolutionForward> marker(handle_cuda());
     NormalRNG default_rng;
-    for (auto &&arg: args) {
-        float scale = 1.0f / sqrt(arg.filter[1] * arg.filter[2] * arg.filter[3]);
+    for (auto&& arg : args) {
+        float scale =
+                1.0f / sqrt(arg.filter[1] * arg.filter[2] * arg.filter[3]);
         UniformFloatRNG rng(scale, 2 * scale);
-        marker.set_dtype(0, dtype::Float32()).
-            set_dtype(1, dtype::Float32()).
-            set_rng(0, &default_rng).
-            set_rng(1, &default_rng).
-            set_param(arg.param).
-            execs({arg.src, arg.filter, {}});
+        marker.set_dtype(0, dtype::Float32())
+                .set_dtype(1, dtype::Float32())
+                .set_rng(0, &default_rng)
+                .set_rng(1, &default_rng)
+                .set_param(arg.param)
+                .execs({arg.src, arg.filter, {}});
     }
 }
 
-TEST_F(CUDA, CONVOLUTION_BACKWARD_DATA)
-{
+TEST_F(CUDA, CONVOLUTION_BACKWARD_DATA) {
     using namespace convolution;
     std::vector<TestArg> args = get_args_cuda_conv_bwd_data();
     Checker<ConvolutionBackwardData> checker(handle_cuda());
     NormalRNG default_rng;
-    for (auto &&arg: args) {
+    for (auto&& arg : args) {
         float scale =
                 64.f / sqrt(arg.filter[0] * arg.filter[2] * arg.filter[3]);
         UniformFloatRNG rng(scale, 2 * scale);
@@ -243,8 +237,7 @@ TEST_F(CUDA, CONVOLUTION_BACKWARD_DATA)
     }
 }
 
-TEST_F(CUDA, CONVOLUTION_BACKWARD_DATA_MATMUL)
-{
+TEST_F(CUDA, CONVOLUTION_BACKWARD_DATA_MATMUL) {
     using namespace convolution;
     std::vector<TestArg> args = get_args_cuda_conv_bwd_data();
     Checker<ConvolutionBackwardData> checker(handle_cuda());
@@ -252,7 +245,7 @@ TEST_F(CUDA, CONVOLUTION_BACKWARD_DATA_MATMUL)
     checker.set_before_exec_callback(AlgoChecker<ConvolutionBackwardData>(
             ExecutionPolicyAlgoName{"MATMUL", {{"CUBLAS", {}}}}));
     NormalRNG default_rng;
-    for (auto &&arg: args) {
+    for (auto&& arg : args) {
         float scale =
                 64.f / sqrt(arg.filter[0] * arg.filter[2] * arg.filter[3]);
         UniformFloatRNG rng(scale, 2 * scale);
@@ -273,9 +266,39 @@ TEST_F(CUDA, CONVOLUTION_BACKWARD_DATA_MATMUL)
     }
 }
 
+TEST_F(CUDA, CONVOLUTION_BACKWARD_DATA_INT8_DP4A) {
+    if (!cuda::is_compute_capability_required(6, 1)) {
+        printf("Skip CUDA.CONVOLUTION_BACKWARD_DATA_INT8_DP4A test as current "
+               "device doesn't support\n");
+        return;
+    }
+
+    using namespace convolution;
+    std::vector<TestArg> args = get_args_int8_nchw4_conv_bwd_data();
+    Checker<ConvolutionBackwardData> checker(handle_cuda());
+
+    checker.set_before_exec_callback(AlgoChecker<ConvolutionBackwardData>(
+            "INT8_NCHW4_DOTPROD_IMPLICIT_GEMM"));
+
+    checker.set_epsilon(1 + 1e-3).set_max_avg_error(1e-1);
 
-TEST_F(CUDA, CONVOLUTION_BACKWARD_DATA_FAILED_CUDNN7_5)
-{
+    for (auto&& arg : args) {
+        UniformIntRNG rng(-3, 3);
+        auto src = TensorLayout(arg.src, dtype::QuantizedS8{1.2f});
+        auto filter = TensorLayout(arg.filter, dtype::QuantizedS8{1.3f});
+        TensorLayout dst;
+        dst.dtype = dtype::QuantizedS8{1.2f};
+        {
+            auto opr = handle_cuda()->create_operator<Convolution>();
+            opr->param() = arg.param;
+            opr->deduce_layout(src, filter, dst);
+        }
+        checker.set_rng(0, &rng).set_rng(1, &rng).set_param(arg.param).exec(
+                TensorLayoutArray{filter, dst, src});
+    }
+}
+
+TEST_F(CUDA, CONVOLUTION_BACKWARD_DATA_FAILED_CUDNN7_5) {
     // BRAIN-481 failed on architectures 7.0, remove the following if statement,
     // when cudnn fixed the problem.
     if (cuda::is_compute_capability_required(7, 0))
@@ -284,8 +307,9 @@ TEST_F(CUDA, CONVOLUTION_BACKWARD_DATA_FAILED_CUDNN7_5)
     std::vector<TestArg> args = get_args_cudnn_7_5_failures();
     Checker<ConvolutionBackwardData> checker(handle_cuda());
     NormalRNG default_rng;
-    for (auto &&arg: args) {
-        float scale = 128.f / sqrt(arg.filter[0] * arg.filter[2] * arg.filter[3]);
+    for (auto&& arg : args) {
+        float scale =
+                128.f / sqrt(arg.filter[0] * arg.filter[2] * arg.filter[3]);
         scale = std::max(scale, 1.f);
         UniformFloatRNG rng(scale, 2 * scale);
         auto src = TensorLayout(arg.src, dtype::Float32());
@@ -297,19 +321,17 @@ TEST_F(CUDA, CONVOLUTION_BACKWARD_DATA_FAILED_CUDNN7_5)
             opr->deduce_layout(src, filter, dst);
         }
         src.dtype = dst.dtype = filter.dtype = dtype::Float32();
-        checker.
-            set_rng(0, &default_rng).
-            set_rng(1, &default_rng).
-            set_epsilon(1e-3).
-            set_param(arg.param).
-            exec(TensorLayoutArray{filter, dst, src});
+        checker.set_rng(0, &default_rng)
+                .set_rng(1, &default_rng)
+                .set_epsilon(1e-3)
+                .set_param(arg.param)
+                .exec(TensorLayoutArray{filter, dst, src});
         src.dtype = dst.dtype = filter.dtype = dtype::Float16();
-        checker.
-            set_rng(0, &rng).
-            set_rng(1, &rng).
-            set_epsilon(1e-1).
-            set_param(arg.param).
-            exec(TensorLayoutArray{filter, dst, src});
+        checker.set_rng(0, &rng)
+                .set_rng(1, &rng)
+                .set_epsilon(1e-1)
+                .set_param(arg.param)
+                .exec(TensorLayoutArray{filter, dst, src});
         arg.param.compute_mode = param::Convolution::ComputeMode::FLOAT32;
         checker.set_rng(0, &rng)
                 .set_rng(1, &rng)
@@ -319,13 +341,12 @@ TEST_F(CUDA, CONVOLUTION_BACKWARD_DATA_FAILED_CUDNN7_5)
     }
 }
 
-TEST_F(CUDA, CONVOLUTION_BACKWARD_FILTER)
-{
+TEST_F(CUDA, CONVOLUTION_BACKWARD_FILTER) {
     using namespace convolution;
     std::vector<TestArg> args = get_args();
     Checker<ConvolutionBackwardFilter> checker(handle_cuda());
     bool f16_checked = false;
-    for (auto &&arg: args) {
+    for (auto&& arg : args) {
         auto src = TensorLayout(arg.src, dtype::Float32());
         auto filter = TensorLayout(arg.filter, dtype::Float32());
         TensorLayout dst;
@@ -337,12 +358,11 @@ TEST_F(CUDA, CONVOLUTION_BACKWARD_FILTER)
         float scale = 1.0f / sqrt(dst[2] * dst[3]);
         UniformFloatRNG rng(scale, 2 * scale);
         src.dtype = dst.dtype = filter.dtype = dtype::Float32();
-        checker.
-            set_rng(0, &rng).
-            set_rng(1, &rng).
-            set_epsilon(1e-3).
-            set_param(arg.param).
-            exec(TensorLayoutArray{src, dst, filter});
+        checker.set_rng(0, &rng)
+                .set_rng(1, &rng)
+                .set_epsilon(1e-3)
+                .set_param(arg.param)
+                .exec(TensorLayoutArray{src, dst, filter});
 
         // reduce on large f16 array may introduce significant error
         if (dst.total_nr_elems() >= 1000 && f16_checked)
@@ -350,12 +370,11 @@ TEST_F(CUDA, CONVOLUTION_BACKWARD_FILTER)
 
         f16_checked = true;
         src.dtype = dst.dtype = filter.dtype = dtype::Float16();
-        checker.
-            set_rng(0, &rng).
-            set_rng(1, &rng).
-            set_epsilon(1e-1).
-            set_param(arg.param).
-            exec(TensorLayoutArray{src, dst, filter});
+        checker.set_rng(0, &rng)
+                .set_rng(1, &rng)
+                .set_epsilon(1e-1)
+                .set_param(arg.param)
+                .exec(TensorLayoutArray{src, dst, filter});
         arg.param.compute_mode = param::Convolution::ComputeMode::FLOAT32;
         checker.set_rng(0, &rng)
                 .set_rng(1, &rng)
@@ -377,14 +396,13 @@ TEST_F(CUDA, CONVOLUTION_BACKWARD_FILTER)
     }
 }
 
-TEST_F(CUDA, CONVOLUTION_BACKWARD_FILTER_MATMUL)
-{
+TEST_F(CUDA, CONVOLUTION_BACKWARD_FILTER_MATMUL) {
     using namespace convolution;
     std::vector<TestArg> args = get_args();
     Checker<ConvolutionBackwardFilter> checker(handle_cuda());
     checker.set_before_exec_callback(AlgoChecker<ConvolutionBackwardFilter>(
             ExecutionPolicyAlgoName{"MATMUL", {{"CUBLAS", {}}}}));
-    for (auto &&arg: args) {
+    for (auto&& arg : args) {
         auto src = TensorLayout(arg.src, dtype::Float32());
         auto filter = TensorLayout(arg.filter, dtype::Float32());
         TensorLayout dst;
@@ -396,17 +414,16 @@ TEST_F(CUDA, CONVOLUTION_BACKWARD_FILTER_MATMUL)
         float scale = 1.0f / sqrt(dst[2] * dst[3]);
         UniformFloatRNG rng(scale, 2 * scale);
         src.dtype = dst.dtype = filter.dtype = dtype::Float32();
-        checker.
-            set_rng(0, &rng).
-            set_rng(1, &rng).
-            set_epsilon(1e-3).
-            set_param(arg.param).
-            exec(TensorLayoutArray{src, dst, filter});
+        checker.set_rng(0, &rng)
+                .set_rng(1, &rng)
+                .set_epsilon(1e-3)
+                .set_param(arg.param)
+                .exec(TensorLayoutArray{src, dst, filter});
     }
 }
 
 TEST_F(CUDA, CONV_CONFIG_COMBINATIONS) {
-    auto eps_getter = [](bool f16, int stage, const char *name) -> float {
+    auto eps_getter = [](bool f16, int stage, const char* name) -> float {
         if (f16) {
             return stage == 2 ? 0.5 : 0.2;
         }
@@ -687,6 +704,46 @@ TEST_F(CUDA, BENCHMARK_CONVOLUTION_BWD_DATA_BF16) {
     run(32, 64, 64, 56, 56, 1, 1, 0);
 }
 
+TEST_F(CUDA, BENCHMARK_CONVOLUTION_BWD_DATA_INT8_DP4A) {
+    CUBenchmarker<ConvolutionBackwardData> bench{handle_cuda()};
+    std::unique_ptr<OprProxy<ConvolutionBackwardData>> proxy{
+            new OprProxy<ConvolutionBackwardData>{true}};
+    size_t RUNS = 10;
+    bench.set_proxy(proxy).set_times(RUNS);
+
+    auto run = [&](size_t N, size_t OC, size_t IC, size_t IH, size_t IW,
+                   size_t FH, size_t SH, size_t PH) {
+        bench.set_dtype(0, dtype::QuantizedS8{1.0f})
+                .set_dtype(1, dtype::QuantizedS8{1.0f})
+                .set_dtype(2, dtype::QuantizedS8{1.0f});
+        param::Convolution param;
+        param.format = param::Convolution::Format::NCHW4;
+        param.stride_h = param.stride_w = SH;
+        param.pad_h = param.pad_w = PH;
+        param.compute_mode = param::Convolution::ComputeMode::DEFAULT;
+        bench.set_param(param);
+        bench.proxy()->target_execution_policy = {};
+        TensorLayout src{{N, IC / 4, IH, IW, 4}, dtype::QuantizedS8{1.0f}},
+                filter{{OC, IC / 4, FH, FH, 4}, dtype::QuantizedS8{1.0f}};
+        TensorLayout dst;
+        dst.dtype = dtype::QuantizedS8{1.0f};
+        {
+            auto&& opr = handle_cuda()->create_operator<Convolution>();
+            opr->param() = param;
+            opr->deduce_layout(src, filter, dst);
+        }
+        auto used = bench.execl({filter, dst, src}) / RUNS;
+        float flo = 2.0 * N * OC * IC * dst[2] * dst[3] * FH * FH;
+        printf("inp=%s, kern=%s, dst=%s ", src.to_string().c_str(),
+               filter.to_string().c_str(), dst.to_string().c_str());
+        printf("time_fp32=%.2fms, flops=%.3fTFLOPS\n", used,
+               (flo / (used * 1e9)));
+    };
+    run(64, 32, 32, 92, 180, 4, 2, 2);
+    run(64, 32, 32, 46, 80, 4, 2, 2);
+    run(16, 16, 16, 92, 180, 4, 2, 2);
+    run(16, 16, 16, 46, 80, 4, 2, 2);
+}
 
 TEST_F(CUDA, CONVOLUTION_BWD_FILTER_BENCHMARK) {
     CUBenchmarker<ConvolutionBackwardFilter> bench{handle_cuda()};
diff --git a/src/opr/test/dnn/convolution.cpp b/src/opr/test/dnn/convolution.cpp
index 9e779f271..899726049 100644
--- a/src/opr/test/dnn/convolution.cpp
+++ b/src/opr/test/dnn/convolution.cpp
@@ -598,6 +598,51 @@ TEST(TestOprDNN, Deconvolution) {
         run({TensorShape{4, 6, 7, 2}, {2, 3, 4, 8, 13}}, opt);
 }
 
+TEST(TestOprDNN, DeconvolutionExePolicy_QuantizedS8) {
+    REQUIRE_GPU(1);
+    auto cn = CompNode::load("gpu0");
+    cn.activate();
+    REQUIRE_CUDA_COMPUTE_CAPABILITY(6, 1);
+
+    Param param;
+    using Policy = opr::ConvolutionBackwardData::ExecutionPolicy;
+    using S = Policy::Strategy;
+
+#if MGB_ENABLE_FASTRUN
+    for (auto strategy : {S::PROFILE, S::HEURISTIC, S::PROFILE_REPRODUCIBLE,
+                          S::PROFILE_HEURISTIC}) {
+#else
+    for (auto strategy : {S : HEURISTIC, S::PROFILE_HEURISTIC}) {
+#endif
+        auto graph = ComputingGraph::make();
+        HostTensorGenerator<> gen;
+
+        auto mkvar = [&](const char* name, const TensorShape& shp,
+                         const DType& dtype) {
+            return opr::TypeCvt::make(
+                    opr::Host2DeviceCopy::make(*graph, gen(shp)).rename(name),
+                    dtype);
+        };
+
+        auto x = mkvar("x", {16, 4, 50, 50, 4}, dtype::QuantizedS8(1.2f));
+        auto w = mkvar("w", {16, 4, 4, 4, 4}, dtype::QuantizedS8(1.3f));
+
+        param.format = Param::Format::NCHW4;
+        param.pad_h = param.pad_w = 2;
+        param.stride_h = param.stride_w = 2;
+
+        Policy policy;
+        policy.strategy = strategy;
+
+        auto deconv = opr::ConvolutionBackwardData::make_deconv(
+                x, w, param, policy,
+                OperatorNodeConfig{dtype::QuantizedS8(1.2f)});
+        HostTensorND host_y;
+        auto func = graph->compile({make_callback_copy(deconv, host_y)});
+        func->execute();
+    }
+}
+
 TEST(TestOprDNN, ConvolutionBackwardFilter) {
     using Checker = AutoOprChecker<3, 1>;
 
-- 
GitLab