未验证 提交 c0b08078 编写于 作者: L Lynne

hwcontext_vulkan: use all enabled queues for transfers, make uploads async

This commit makes full use of the enabled queues to provide asynchronous
uploads of images (downloads remain synchronous).
For a pure uploading use cases, the performance gains can be significant.
上级 cdb949a0
......@@ -41,11 +41,23 @@
#define CHECK_CU(x) FF_CUDA_CHECK_DL(cuda_cu, cu, x)
#endif
typedef struct VulkanQueueCtx {
VkFence fence;
VkQueue queue;
int was_synchronous;
/* Buffer dependencies */
AVBufferRef **buf_deps;
int nb_buf_deps;
int buf_deps_alloc_size;
} VulkanQueueCtx;
typedef struct VulkanExecCtx {
VkCommandPool pool;
VkCommandBuffer buf;
VkQueue queue;
VkFence fence;
VkCommandBuffer *bufs;
VulkanQueueCtx *queues;
int nb_queues;
int cur_queue_idx;
} VulkanExecCtx;
typedef struct VulkanDevicePriv {
......@@ -60,8 +72,9 @@ typedef struct VulkanDevicePriv {
/* Debug callback */
VkDebugUtilsMessengerEXT debug_ctx;
/* Image uploading */
VulkanExecCtx cmd;
/* Image transfers */
VulkanExecCtx upload_ctx;
VulkanExecCtx download_ctx;
/* Extensions */
uint64_t extensions;
......@@ -89,6 +102,16 @@ typedef struct AVVkFrameInternal {
#endif
} AVVkFrameInternal;
#define GET_QUEUE_COUNT(hwctx, graph, comp, tx) ( \
graph ? hwctx->nb_graphics_queues : \
comp ? (hwctx->nb_comp_queues ? \
hwctx->nb_comp_queues : hwctx->nb_graphics_queues) : \
tx ? (hwctx->nb_tx_queues ? hwctx->nb_tx_queues : \
(hwctx->nb_comp_queues ? \
hwctx->nb_comp_queues : hwctx->nb_graphics_queues)) : \
0 \
)
#define VK_LOAD_PFN(inst, name) PFN_##name pfn_##name = (PFN_##name) \
vkGetInstanceProcAddr(inst, #name)
......@@ -709,7 +732,7 @@ fail:
}
static int create_exec_ctx(AVHWDeviceContext *ctx, VulkanExecCtx *cmd,
int queue_family_index)
int queue_family_index, int num_queues)
{
VkResult ret;
AVVulkanDeviceContext *hwctx = ctx->hwctx;
......@@ -722,21 +745,20 @@ static int create_exec_ctx(AVHWDeviceContext *ctx, VulkanExecCtx *cmd,
VkCommandBufferAllocateInfo cbuf_create = {
.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO,
.level = VK_COMMAND_BUFFER_LEVEL_PRIMARY,
.commandBufferCount = 1,
.commandBufferCount = num_queues,
};
VkFenceCreateInfo fence_spawn = {
.sType = VK_STRUCTURE_TYPE_FENCE_CREATE_INFO,
};
cmd->nb_queues = num_queues;
ret = vkCreateFence(hwctx->act_dev, &fence_spawn,
hwctx->alloc, &cmd->fence);
if (ret != VK_SUCCESS) {
av_log(ctx, AV_LOG_ERROR, "Failed to create frame fence: %s\n",
vk_ret2str(ret));
return AVERROR_EXTERNAL;
}
cmd->queues = av_mallocz(num_queues * sizeof(*cmd->queues));
if (!cmd->queues)
return AVERROR(ENOMEM);
cmd->bufs = av_mallocz(num_queues * sizeof(*cmd->bufs));
if (!cmd->bufs)
return AVERROR(ENOMEM);
/* Create command pool */
ret = vkCreateCommandPool(hwctx->act_dev, &cqueue_create,
hwctx->alloc, &cmd->pool);
if (ret != VK_SUCCESS) {
......@@ -747,15 +769,19 @@ static int create_exec_ctx(AVHWDeviceContext *ctx, VulkanExecCtx *cmd,
cbuf_create.commandPool = cmd->pool;
ret = vkAllocateCommandBuffers(hwctx->act_dev, &cbuf_create, &cmd->buf);
/* Allocate command buffer */
ret = vkAllocateCommandBuffers(hwctx->act_dev, &cbuf_create, cmd->bufs);
if (ret != VK_SUCCESS) {
av_log(ctx, AV_LOG_ERROR, "Command buffer alloc failure: %s\n",
vk_ret2str(ret));
return AVERROR_EXTERNAL;
}
vkGetDeviceQueue(hwctx->act_dev, cqueue_create.queueFamilyIndex, 0,
&cmd->queue);
for (int i = 0; i < num_queues; i++) {
VulkanQueueCtx *q = &cmd->queues[i];
vkGetDeviceQueue(hwctx->act_dev, queue_family_index, i, &q->queue);
q->was_synchronous = 1;
}
return 0;
}
......@@ -764,12 +790,154 @@ static void free_exec_ctx(AVHWDeviceContext *ctx, VulkanExecCtx *cmd)
{
AVVulkanDeviceContext *hwctx = ctx->hwctx;
if (cmd->fence)
vkDestroyFence(hwctx->act_dev, cmd->fence, hwctx->alloc);
if (cmd->buf)
vkFreeCommandBuffers(hwctx->act_dev, cmd->pool, 1, &cmd->buf);
/* Make sure all queues have finished executing */
for (int i = 0; i < cmd->nb_queues; i++) {
VulkanQueueCtx *q = &cmd->queues[i];
if (q->fence && !q->was_synchronous) {
vkWaitForFences(hwctx->act_dev, 1, &q->fence, VK_TRUE, UINT64_MAX);
vkResetFences(hwctx->act_dev, 1, &q->fence);
}
/* Free the fence */
if (q->fence)
vkDestroyFence(hwctx->act_dev, q->fence, hwctx->alloc);
/* Free buffer dependencies */
for (int j = 0; j < q->nb_buf_deps; j++)
av_buffer_unref(&q->buf_deps[j]);
av_free(q->buf_deps);
}
if (cmd->bufs)
vkFreeCommandBuffers(hwctx->act_dev, cmd->pool, cmd->nb_queues, cmd->bufs);
if (cmd->pool)
vkDestroyCommandPool(hwctx->act_dev, cmd->pool, hwctx->alloc);
av_freep(&cmd->bufs);
av_freep(&cmd->queues);
}
static VkCommandBuffer get_buf_exec_ctx(AVHWDeviceContext *ctx, VulkanExecCtx *cmd)
{
return cmd->bufs[cmd->cur_queue_idx];
}
static void unref_exec_ctx_deps(AVHWDeviceContext *ctx, VulkanExecCtx *cmd)
{
VulkanQueueCtx *q = &cmd->queues[cmd->cur_queue_idx];
for (int j = 0; j < q->nb_buf_deps; j++)
av_buffer_unref(&q->buf_deps[j]);
q->nb_buf_deps = 0;
}
static int wait_start_exec_ctx(AVHWDeviceContext *ctx, VulkanExecCtx *cmd)
{
VkResult ret;
AVVulkanDeviceContext *hwctx = ctx->hwctx;
VulkanQueueCtx *q = &cmd->queues[cmd->cur_queue_idx];
VkCommandBufferBeginInfo cmd_start = {
.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO,
.flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT,
};
/* Create the fence and don't wait for it initially */
if (!q->fence) {
VkFenceCreateInfo fence_spawn = {
.sType = VK_STRUCTURE_TYPE_FENCE_CREATE_INFO,
};
ret = vkCreateFence(hwctx->act_dev, &fence_spawn, hwctx->alloc,
&q->fence);
if (ret != VK_SUCCESS) {
av_log(ctx, AV_LOG_ERROR, "Failed to queue frame fence: %s\n",
vk_ret2str(ret));
return AVERROR_EXTERNAL;
}
} else if (!q->was_synchronous) {
vkWaitForFences(hwctx->act_dev, 1, &q->fence, VK_TRUE, UINT64_MAX);
vkResetFences(hwctx->act_dev, 1, &q->fence);
}
/* Discard queue dependencies */
unref_exec_ctx_deps(ctx, cmd);
ret = vkBeginCommandBuffer(cmd->bufs[cmd->cur_queue_idx], &cmd_start);
if (ret != VK_SUCCESS) {
av_log(ctx, AV_LOG_ERROR, "Unable to init command buffer: %s\n",
vk_ret2str(ret));
return AVERROR_EXTERNAL;
}
return 0;
}
static int add_buf_dep_exec_ctx(AVHWDeviceContext *ctx, VulkanExecCtx *cmd,
AVBufferRef * const *deps, int nb_deps)
{
AVBufferRef **dst;
VulkanQueueCtx *q = &cmd->queues[cmd->cur_queue_idx];
if (!deps || !nb_deps)
return 0;
dst = av_fast_realloc(q->buf_deps, &q->buf_deps_alloc_size,
(q->nb_buf_deps + nb_deps) * sizeof(*dst));
if (!dst)
goto err;
q->buf_deps = dst;
for (int i = 0; i < nb_deps; i++) {
q->buf_deps[q->nb_buf_deps] = av_buffer_ref(deps[i]);
if (!q->buf_deps[q->nb_buf_deps])
goto err;
q->nb_buf_deps++;
}
return 0;
err:
unref_exec_ctx_deps(ctx, cmd);
return AVERROR(ENOMEM);
}
static int submit_exec_ctx(AVHWDeviceContext *ctx, VulkanExecCtx *cmd,
VkSubmitInfo *s_info, int synchronous)
{
VkResult ret;
VulkanQueueCtx *q = &cmd->queues[cmd->cur_queue_idx];
ret = vkEndCommandBuffer(cmd->bufs[cmd->cur_queue_idx]);
if (ret != VK_SUCCESS) {
av_log(ctx, AV_LOG_ERROR, "Unable to finish command buffer: %s\n",
vk_ret2str(ret));
unref_exec_ctx_deps(ctx, cmd);
return AVERROR_EXTERNAL;
}
s_info->pCommandBuffers = &cmd->bufs[cmd->cur_queue_idx];
s_info->commandBufferCount = 1;
ret = vkQueueSubmit(q->queue, 1, s_info, q->fence);
if (ret != VK_SUCCESS) {
unref_exec_ctx_deps(ctx, cmd);
return AVERROR_EXTERNAL;
}
q->was_synchronous = synchronous;
if (synchronous) {
AVVulkanDeviceContext *hwctx = ctx->hwctx;
vkWaitForFences(hwctx->act_dev, 1, &q->fence, VK_TRUE, UINT64_MAX);
vkResetFences(hwctx->act_dev, 1, &q->fence);
unref_exec_ctx_deps(ctx, cmd);
} else { /* Rotate queues */
cmd->cur_queue_idx = (cmd->cur_queue_idx + 1) % cmd->nb_queues;
}
return 0;
}
static void vulkan_device_free(AVHWDeviceContext *ctx)
......@@ -937,7 +1105,8 @@ if (n >= queue_num) {
p->qfs[p->num_qfs++] = hwctx->queue_family_comp_index;
/* Create exec context - if there's something invalid this will error out */
err = create_exec_ctx(ctx, &p->cmd, hwctx->queue_family_tx_index);
err = create_exec_ctx(ctx, &p->cmd, hwctx->queue_family_tx_index,
GET_QUEUE_COUNT(hwctx, 0, 0, 1));
if (err)
return err;
......@@ -1257,26 +1426,17 @@ enum PrepMode {
static int prepare_frame(AVHWFramesContext *hwfc, VulkanExecCtx *ectx,
AVVkFrame *frame, enum PrepMode pmode)
{
VkResult ret;
int err;
uint32_t dst_qf;
VkImageLayout new_layout;
VkAccessFlags new_access;
AVHWDeviceContext *ctx = hwfc->device_ctx;
AVVulkanDeviceContext *hwctx = ctx->hwctx;
const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
VkImageMemoryBarrier img_bar[AV_NUM_DATA_POINTERS] = { 0 };
VkCommandBufferBeginInfo cmd_start = {
.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO,
.flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT,
};
VkSubmitInfo s_info = {
.sType = VK_STRUCTURE_TYPE_SUBMIT_INFO,
.commandBufferCount = 1,
.pCommandBuffers = &ectx->buf,
.pSignalSemaphores = frame->sem,
.signalSemaphoreCount = planes,
};
......@@ -1306,9 +1466,8 @@ static int prepare_frame(AVHWFramesContext *hwfc, VulkanExecCtx *ectx,
break;
}
ret = vkBeginCommandBuffer(ectx->buf, &cmd_start);
if (ret != VK_SUCCESS)
return AVERROR_EXTERNAL;
if ((err = wait_start_exec_ctx(ctx, ectx)))
return err;
/* Change the image layout to something more optimal for writes.
* This also signals the newly created semaphore, making it usable
......@@ -1330,23 +1489,12 @@ static int prepare_frame(AVHWFramesContext *hwfc, VulkanExecCtx *ectx,
frame->access[i] = img_bar[i].dstAccessMask;
}
vkCmdPipelineBarrier(ectx->buf, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
VK_PIPELINE_STAGE_TRANSFER_BIT, 0,
0, NULL, 0, NULL, planes, img_bar);
ret = vkEndCommandBuffer(ectx->buf);
if (ret != VK_SUCCESS)
return AVERROR_EXTERNAL;
ret = vkQueueSubmit(ectx->queue, 1, &s_info, ectx->fence);
if (ret != VK_SUCCESS) {
return AVERROR_EXTERNAL;
} else {
vkWaitForFences(hwctx->act_dev, 1, &ectx->fence, VK_TRUE, UINT64_MAX);
vkResetFences(hwctx->act_dev, 1, &ectx->fence);
}
vkCmdPipelineBarrier(get_buf_exec_ctx(ctx, ectx),
VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
VK_PIPELINE_STAGE_TRANSFER_BIT,
0, 0, NULL, 0, NULL, planes, img_bar);
return 0;
return submit_exec_ctx(ctx, ectx, &s_info, 0);
}
static int create_frame(AVHWFramesContext *hwfc, AVVkFrame **frame,
......@@ -1559,7 +1707,8 @@ static int vulkan_frames_init(AVHWFramesContext *hwfc)
hwctx->usage |= DEFAULT_USAGE_FLAGS;
err = create_exec_ctx(hwfc->device_ctx, &fp->cmd,
dev_hwctx->queue_family_tx_index);
dev_hwctx->queue_family_tx_index,
GET_QUEUE_COUNT(dev_hwctx, 0, 0, 1));
if (err)
return err;
......@@ -2620,12 +2769,12 @@ static int unmap_buffers(AVHWDeviceContext *ctx, AVBufferRef **bufs,
return err;
}
static int transfer_image_buf(AVHWDeviceContext *ctx, AVVkFrame *frame,
static int transfer_image_buf(AVHWDeviceContext *ctx, const AVFrame *f,
AVBufferRef **bufs, const int *buf_stride, int w,
int h, enum AVPixelFormat pix_fmt, int to_buf)
{
VkResult ret;
AVVulkanDeviceContext *hwctx = ctx->hwctx;
int err;
AVVkFrame *frame = (AVVkFrame *)f->data[0];
VulkanDevicePriv *s = ctx->internal->priv;
int bar_num = 0;
......@@ -2634,17 +2783,11 @@ static int transfer_image_buf(AVHWDeviceContext *ctx, AVVkFrame *frame,
const int planes = av_pix_fmt_count_planes(pix_fmt);
const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(pix_fmt);
VkCommandBufferBeginInfo cmd_start = {
.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO,
.flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT,
};
VkImageMemoryBarrier img_bar[AV_NUM_DATA_POINTERS] = { 0 };
VkCommandBuffer cmd_buf = get_buf_exec_ctx(ctx, &s->cmd);
VkSubmitInfo s_info = {
.sType = VK_STRUCTURE_TYPE_SUBMIT_INFO,
.commandBufferCount = 1,
.pCommandBuffers = &s->cmd.buf,
.pSignalSemaphores = frame->sem,
.pWaitSemaphores = frame->sem,
.pWaitDstStageMask = sem_wait_dst,
......@@ -2652,12 +2795,8 @@ static int transfer_image_buf(AVHWDeviceContext *ctx, AVVkFrame *frame,
.waitSemaphoreCount = planes,
};
ret = vkBeginCommandBuffer(s->cmd.buf, &cmd_start);
if (ret != VK_SUCCESS) {
av_log(ctx, AV_LOG_ERROR, "Unable to init command buffer: %s\n",
vk_ret2str(ret));
return AVERROR_EXTERNAL;
}
if ((err = wait_start_exec_ctx(ctx, &s->cmd)))
return err;
/* Change the image layout to something more optimal for transfers */
for (int i = 0; i < planes; i++) {
......@@ -2691,7 +2830,7 @@ static int transfer_image_buf(AVHWDeviceContext *ctx, AVVkFrame *frame,
}
if (bar_num)
vkCmdPipelineBarrier(s->cmd.buf, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
vkCmdPipelineBarrier(cmd_buf, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
VK_PIPELINE_STAGE_TRANSFER_BIT, 0,
0, NULL, 0, NULL, bar_num, img_bar);
......@@ -2715,33 +2854,33 @@ static int transfer_image_buf(AVHWDeviceContext *ctx, AVVkFrame *frame,
};
if (to_buf)
vkCmdCopyImageToBuffer(s->cmd.buf, frame->img[i], frame->layout[i],
vkCmdCopyImageToBuffer(cmd_buf, frame->img[i], frame->layout[i],
vkbuf->buf, 1, &buf_reg);
else
vkCmdCopyBufferToImage(s->cmd.buf, vkbuf->buf, frame->img[i],
vkCmdCopyBufferToImage(cmd_buf, vkbuf->buf, frame->img[i],
frame->layout[i], 1, &buf_reg);
}
ret = vkEndCommandBuffer(s->cmd.buf);
if (ret != VK_SUCCESS) {
av_log(ctx, AV_LOG_ERROR, "Unable to finish command buffer: %s\n",
vk_ret2str(ret));
return AVERROR_EXTERNAL;
}
/* Wait for the download/upload to finish if uploading, otherwise the
* semaphore will take care of synchronization when uploading */
ret = vkQueueSubmit(s->cmd.queue, 1, &s_info, s->cmd.fence);
if (ret != VK_SUCCESS) {
av_log(ctx, AV_LOG_ERROR, "Unable to submit command buffer: %s\n",
vk_ret2str(ret));
return AVERROR_EXTERNAL;
/* When uploading, do this asynchronously if the source is refcounted by
* keeping the buffers as a submission dependency.
* The hwcontext is guaranteed to not be freed until all frames are freed
* in the frames_unint function.
* When downloading to buffer, do this synchronously and wait for the
* queue submission to finish executing */
if (!to_buf) {
int ref;
for (ref = 0; ref < AV_NUM_DATA_POINTERS; ref++) {
if (!f->buf[ref])
break;
if ((err = add_buf_dep_exec_ctx(hwfc, &s->cmd, &f->buf[ref], 1)))
return err;
}
if (ref && (err = add_buf_dep_exec_ctx(hwfc, &s->cmd, bufs, planes)))
return err;
return submit_exec_ctx(hwfc, &s->cmd, &s_info, !ref);
} else {
vkWaitForFences(hwctx->act_dev, 1, &s->cmd.fence, VK_TRUE, UINT64_MAX);
vkResetFences(hwctx->act_dev, 1, &s->cmd.fence);
return submit_exec_ctx(hwfc, &s->cmd, &s_info, 1);
}
return 0;
}
/* Technically we can use VK_EXT_external_memory_host to upload and download,
......@@ -2778,11 +2917,11 @@ static int vulkan_transfer_data_from_mem(AVHWFramesContext *hwfc, AVFrame *dst,
err = vulkan_map_frame_to_mem(hwfc, map, dst, AV_HWFRAME_MAP_WRITE);
if (err)
goto end;
return err;
err = av_frame_copy(map, src);
av_frame_free(&map);
goto end;
return err;
}
/* Create buffers */
......@@ -2809,7 +2948,7 @@ static int vulkan_transfer_data_from_mem(AVHWFramesContext *hwfc, AVFrame *dst,
goto end;
/* Copy buffers to image */
err = transfer_image_buf(dev_ctx, f, bufs, tmp.linesize,
err = transfer_image_buf(dev_ctx, dst, bufs, tmp.linesize,
src->width, src->height, src->format, 0);
end:
......@@ -2949,10 +3088,12 @@ static int vulkan_transfer_data_to_mem(AVHWFramesContext *hwfc, AVFrame *dst,
err = create_buf(dev_ctx, &bufs[i], p_height,
&tmp.linesize[i], VK_BUFFER_USAGE_TRANSFER_DST_BIT,
VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT, NULL, NULL);
if (err)
goto end;
}
/* Copy image to buffer */
if ((err = transfer_image_buf(dev_ctx, f, bufs, tmp.linesize,
if ((err = transfer_image_buf(dev_ctx, src, bufs, tmp.linesize,
dst->width, dst->height, dst->format, 1)))
goto end;
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册