Adaptation for cuda11 and legacy gpu/cpu (#17143)

- compile gpu with cuda11
  - add support for sm35, sm50 for legacy GPUs
  - add sm86 for ampere GPU, like RTX 30xx, Tesla Ax
  - support cpu only support SSE42
Signed-off-by: NJi Bin <matrixji@live.com>

core/CMakeLists.txt

@@ -168,7 +168,7 @@ if (MILVUS_GPU_VERSION)
     message(STATUS "Building Milvus GPU version")
     add_compile_definitions("MILVUS_GPU_VERSION")
     enable_language(CUDA)
-    find_package(CUDA 10 REQUIRED)
+    find_package(CUDA 11 REQUIRED)
     set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} -Xcompiler -fPIC -std=c++11 -D_FORCE_INLINES --expt-extended-lambda")
 else ()
     message(STATUS "Building Milvus CPU version")

core/src/index/CMakeLists.txt

@@ -79,7 +79,7 @@ if (MILVUS_GPU_VERSION OR KNOWHERE_GPU_VERSION)
     message(STATUS "Building Knowhere GPU version")
     add_compile_definitions("MILVUS_GPU_VERSION")
     enable_language(CUDA)
-    find_package(CUDA 10 REQUIRED)
+    find_package(CUDA 11 REQUIRED)
     set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} -Xcompiler -fPIC -std=c++11 -D_FORCE_INLINES --expt-extended-lambda")
 else ()
     message(STATUS "Building Knowhere CPU version")

core/src/index/cmake/ThirdPartyPackagesCore.cmake

@@ -513,7 +513,7 @@ macro(build_faiss)
     set(FAISS_CONFIGURE_ARGS
             "--prefix=${FAISS_PREFIX}"
             "CFLAGS=${EP_C_FLAGS}"
-            "CXXFLAGS=${EP_CXX_FLAGS} -mf16c -O3"
+            "CXXFLAGS=${EP_CXX_FLAGS} -msse4.2 -O3"
             --without-python)
 
     if (FAISS_WITH_MKL)
@@ -535,7 +535,7 @@ macro(build_faiss)
     if (KNOWHERE_GPU_VERSION)
         set(FAISS_CONFIGURE_ARGS ${FAISS_CONFIGURE_ARGS}
                 "--with-cuda=${CUDA_TOOLKIT_ROOT_DIR}"
-                "--with-cuda-arch=-gencode=arch=compute_60,code=sm_60 -gencode=arch=compute_61,code=sm_61 -gencode=arch=compute_70,code=sm_70 -gencode=arch=compute_75,code=sm_75"
+                "--with-cuda-arch=-gencode=arch=compute_35,code=sm_35 -gencode=arch=compute_50,code=sm_50 -gencode=arch=compute_60,code=sm_60 -gencode=arch=compute_61,code=sm_61 -gencode=arch=compute_70,code=sm_70 -gencode=arch=compute_75,code=sm_75 -gencode=arch=compute_86,code=sm_86"
                 )
     else ()
         set(FAISS_CONFIGURE_ARGS ${FAISS_CONFIGURE_ARGS}

core/src/index/thirdparty/faiss/Makefile

@@ -45,11 +45,11 @@ libfaiss.$(SHAREDEXT): $(OBJ)
 
 # support avx
 %avx.o: %avx.cpp
-	$(CXX) $(CPPFLAGS) $(CXXFLAGS) $(CPUFLAGS) -mavx2 -c $< -o $@
+	$(CXX) $(CPPFLAGS) $(CXXFLAGS) $(CPUFLAGS) -mf16c -mavx2 -c $< -o $@
 
 # support avx512
 %avx512.o: %avx512.cpp
-	$(CXX) $(CPPFLAGS) $(CXXFLAGS) $(CPUFLAGS) -mavx512f -mavx512dq -mavx512bw -c $< -o $@
+	$(CXX) $(CPPFLAGS) $(CXXFLAGS) $(CPUFLAGS) -mf16c -mavx512f -mavx512dq -mavx512bw -c $< -o $@
 
 %.o: %.cu
 	$(NVCC) $(NVCCFLAGS) -c $< -o $@

core/src/index/thirdparty/faiss/gpu/utils/DeviceDefs.cuh

@@ -13,7 +13,7 @@
 namespace faiss { namespace gpu {
 
 #ifdef __CUDA_ARCH__
-#if __CUDA_ARCH__ <= 750
+#if __CUDA_ARCH__ <= 860
 constexpr int kWarpSize = 32;
 #else
 #error Unknown __CUDA_ARCH__; please define parameters for compute capability

core/src/index/thirdparty/faiss/utils/distances_simd_avx.cpp

@@ -188,16 +188,16 @@ float fvec_Linf_avx (const float* x, const float* y, size_t d) {
     return  _mm_cvtss_f32 (msum2);
 }
 
-const __m256i lookup = _mm256_setr_epi8(
-        /* 0 */ 0, /* 1 */ 1, /* 2 */ 1, /* 3 */ 2,
-        /* 4 */ 1, /* 5 */ 2, /* 6 */ 2, /* 7 */ 3,
-        /* 8 */ 1, /* 9 */ 2, /* a */ 2, /* b */ 3,
-        /* c */ 2, /* d */ 3, /* e */ 3, /* f */ 4,
-
-        /* 0 */ 0, /* 1 */ 1, /* 2 */ 1, /* 3 */ 2,
-        /* 4 */ 1, /* 5 */ 2, /* 6 */ 2, /* 7 */ 3,
-        /* 8 */ 1, /* 9 */ 2, /* a */ 2, /* b */ 3,
-        /* c */ 2, /* d */ 3, /* e */ 3, /* f */ 4
+#define DECLARE_LOOKUP \
+const __m256i lookup = _mm256_setr_epi8( \
+        /* 0 */ 0, /* 1 */ 1, /* 2 */ 1, /* 3 */ 2, \
+        /* 4 */ 1, /* 5 */ 2, /* 6 */ 2, /* 7 */ 3, \
+        /* 8 */ 1, /* 9 */ 2, /* a */ 2, /* b */ 3, \
+        /* c */ 2, /* d */ 3, /* e */ 3, /* f */ 4, \
+        /* 0 */ 0, /* 1 */ 1, /* 2 */ 1, /* 3 */ 2, \
+        /* 4 */ 1, /* 5 */ 2, /* 6 */ 2, /* 7 */ 3, \
+        /* 8 */ 1, /* 9 */ 2, /* a */ 2, /* b */ 3, \
+        /* c */ 2, /* d */ 3, /* e */ 3, /* f */ 4 \
 );
 
 int popcnt_AVX2_lookup(const uint8_t* data, const size_t n) {
@@ -218,6 +218,7 @@ int popcnt_AVX2_lookup(const uint8_t* data, const size_t n) {
         i += 32; \
     }
 
+    DECLARE_LOOKUP
     while (i + 8*32 <= n) {
         __m256i local = _mm256_setzero_si256();
         ITER ITER ITER ITER
@@ -271,6 +272,7 @@ int xor_popcnt_AVX2_lookup(const uint8_t* data1, const uint8_t* data2, const siz
         i += 32; \
     }
 
+    DECLARE_LOOKUP
     while (i + 8*32 <= n) {
         __m256i local = _mm256_setzero_si256();
         ITER ITER ITER ITER
@@ -324,6 +326,7 @@ int or_popcnt_AVX2_lookup(const uint8_t* data1, const uint8_t* data2, const size
         i += 32; \
     }
 
+    DECLARE_LOOKUP
     while (i + 8*32 <= n) {
         __m256i local = _mm256_setzero_si256();
         ITER ITER ITER ITER
@@ -377,6 +380,7 @@ int and_popcnt_AVX2_lookup(const uint8_t* data1, const uint8_t* data2, const siz
         i += 32; \
     }
 
+    DECLARE_LOOKUP
     while (i + 8*32 <= n) {
         __m256i local = _mm256_setzero_si256();
         ITER ITER ITER ITER

core/src/index/unittest/faiss_ori/CMakeLists.txt

@@ -47,4 +47,4 @@ if (KNOWHERE_GPU_VERSION)
 
     install(TARGETS test_gpu DESTINATION unittest)
 
-endif ()
\ No newline at end of file
+endif ()

docker/build_env/gpu/centos7/Dockerfile

@@ -9,7 +9,7 @@
 # is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express
 # or implied. See the License for the specific language governing permissions and limitations under the License.
 
-FROM nvidia/cuda:10.1-devel-centos7
+FROM nvidia/cuda:11.6.2-devel-centos7
 
 ENV NVIDIA_DRIVER_CAPABILITIES compute,utility
 

docker/deploy/gpu/centos7/Dockerfile

@@ -10,8 +10,9 @@
 # or implied. See the License for the specific language governing permissions and limitations under the License.
 
 FROM hectormolinero/tini:v18 AS tini
+FROM nvidia/cuda:11.6.2-runtime-centos7 AS runtime
 
-FROM nvidia/cuda:10.1-devel-centos7
+FROM nvidia/cuda:11.6.2-base-centos7
 
 ENV NVIDIA_DRIVER_CAPABILITIES compute,utility
 
@@ -24,6 +25,10 @@ ENV LD_LIBRARY_PATH="$LD_LIBRARY_PATH:/var/lib/milvus/lib"
 
 COPY --from=tini /usr/bin/tini /tini
 
+# copy cublas library from runtime
+COPY --from=runtime /usr/local/cuda/targets/x86_64-linux/lib/libcublasLt.so.11 /var/lib/milvus/lib/libcublasLt.so.11
+COPY --from=runtime /usr/local/cuda/targets/x86_64-linux/lib/libcublas.so.11 /var/lib/milvus/lib/libcublas.so.11
+
 ENTRYPOINT ["/tini", "--"]
 
 WORKDIR /var/lib/milvus