CUexternalMemory ext_mem[AV_NUM_DATA_POINTERS];
CUmipmappedArray cu_mma[AV_NUM_DATA_POINTERS];
CUarray cu_array[AV_NUM_DATA_POINTERS];
- CUexternalSemaphore cu_sem;
+ CUexternalSemaphore cu_sem[AV_NUM_DATA_POINTERS];
#endif
} AVVkFrameInternal;
av_log(ctx, AV_LOG_VERBOSE, "Using %s extension \"%s\"\n", mod, tstr);
ADD_VAL_TO_LIST(extension_names, extensions_found, token);
} else {
- av_log(ctx, AV_LOG_ERROR, "%s extension \"%s\" not found!\n",
+ av_log(ctx, AV_LOG_WARNING, "%s extension \"%s\" not found, excluding.\n",
mod, token);
- err = AVERROR(EINVAL);
- goto fail;
}
token = av_strtok(NULL, "+", &save);
}
static int search_queue_families(AVHWDeviceContext *ctx, VkDeviceCreateInfo *cd)
{
uint32_t num;
+ float *weights;
VkQueueFamilyProperties *qs = NULL;
AVVulkanDeviceContext *hwctx = ctx->hwctx;
int graph_index = -1, comp_index = -1, tx_index = -1;
(i != comp_index), tx_index)
#undef SEARCH_FLAGS
-#define QF_FLAGS(flags) \
- ((flags) & VK_QUEUE_GRAPHICS_BIT ) ? "(graphics) " : "", \
- ((flags) & VK_QUEUE_COMPUTE_BIT ) ? "(compute) " : "", \
- ((flags) & VK_QUEUE_TRANSFER_BIT ) ? "(transfer) " : "", \
- ((flags) & VK_QUEUE_SPARSE_BINDING_BIT) ? "(sparse) " : ""
-
- av_log(ctx, AV_LOG_VERBOSE, "Using queue family %i for graphics, "
- "flags: %s%s%s%s\n", graph_index, QF_FLAGS(qs[graph_index].queueFlags));
-
+#define ADD_QUEUE(fidx, graph, comp, tx) \
+ av_log(ctx, AV_LOG_VERBOSE, "Using queue family %i (total queues: %i) for %s%s%s\n", \
+ fidx, qs[fidx].queueCount, graph ? "graphics " : "", \
+ comp ? "compute " : "", tx ? "transfers " : ""); \
+ av_log(ctx, AV_LOG_VERBOSE, " QF %i flags: %s%s%s%s\n", fidx, \
+ ((qs[fidx].queueFlags) & VK_QUEUE_GRAPHICS_BIT) ? "(graphics) " : "", \
+ ((qs[fidx].queueFlags) & VK_QUEUE_COMPUTE_BIT) ? "(compute) " : "", \
+ ((qs[fidx].queueFlags) & VK_QUEUE_TRANSFER_BIT) ? "(transfers) " : "", \
+ ((qs[fidx].queueFlags) & VK_QUEUE_SPARSE_BINDING_BIT) ? "(sparse) " : ""); \
+ pc[cd->queueCreateInfoCount].queueFamilyIndex = fidx; \
+ pc[cd->queueCreateInfoCount].queueCount = qs[fidx].queueCount; \
+ weights = av_malloc(qs[fidx].queueCount * sizeof(float)); \
+ pc[cd->queueCreateInfoCount].pQueuePriorities = weights; \
+ if (!weights) \
+ goto fail; \
+ for (int i = 0; i < qs[fidx].queueCount; i++) \
+ weights[i] = 1.0f; \
+ cd->queueCreateInfoCount++;
+
+ ADD_QUEUE(graph_index, 1, comp_index < 0, tx_index < 0 && comp_index < 0)
hwctx->queue_family_index = graph_index;
- hwctx->queue_family_tx_index = graph_index;
hwctx->queue_family_comp_index = graph_index;
-
- pc[cd->queueCreateInfoCount++].queueFamilyIndex = graph_index;
+ hwctx->queue_family_tx_index = graph_index;
if (comp_index != -1) {
- av_log(ctx, AV_LOG_VERBOSE, "Using queue family %i for compute, "
- "flags: %s%s%s%s\n", comp_index, QF_FLAGS(qs[comp_index].queueFlags));
- hwctx->queue_family_tx_index = comp_index;
- hwctx->queue_family_comp_index = comp_index;
- pc[cd->queueCreateInfoCount++].queueFamilyIndex = comp_index;
+ ADD_QUEUE(comp_index, 0, 1, tx_index < 0)
+ hwctx->queue_family_tx_index = comp_index;
+ hwctx->queue_family_comp_index = comp_index;
}
if (tx_index != -1) {
- av_log(ctx, AV_LOG_VERBOSE, "Using queue family %i for transfers, "
- "flags: %s%s%s%s\n", tx_index, QF_FLAGS(qs[tx_index].queueFlags));
- hwctx->queue_family_tx_index = tx_index;
- pc[cd->queueCreateInfoCount++].queueFamilyIndex = tx_index;
+ ADD_QUEUE(tx_index, 0, 0, 1)
+ hwctx->queue_family_tx_index = tx_index;
}
-#undef QF_FLAGS
-
+#undef ADD_QUEUE
av_free(qs);
return 0;
+
+fail:
+ av_freep(&pc[0].pQueuePriorities);
+ av_freep(&pc[1].pQueuePriorities);
+ av_freep(&pc[2].pQueuePriorities);
+ av_free(qs);
+
+ return AVERROR(ENOMEM);
}
static int create_exec_ctx(AVHWDeviceContext *ctx, VulkanExecCtx *cmd,
VulkanDevicePriv *p = ctx->internal->priv;
AVVulkanDeviceContext *hwctx = ctx->hwctx;
VkDeviceQueueCreateInfo queue_create_info[3] = {
- { .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO,
- .pQueuePriorities = (float []){ 1.0f },
- .queueCount = 1, },
- { .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO,
- .pQueuePriorities = (float []){ 1.0f },
- .queueCount = 1, },
- { .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO,
- .pQueuePriorities = (float []){ 1.0f },
- .queueCount = 1, },
+ { .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO, },
+ { .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO, },
+ { .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO, },
};
VkDeviceCreateInfo dev_info = {
goto end;
if ((err = check_extensions(ctx, 1, opts, &dev_info.ppEnabledExtensionNames,
- &dev_info.enabledExtensionCount, 0)))
+ &dev_info.enabledExtensionCount, 0))) {
+ av_free((void *)queue_create_info[0].pQueuePriorities);
+ av_free((void *)queue_create_info[1].pQueuePriorities);
+ av_free((void *)queue_create_info[2].pQueuePriorities);
goto end;
+ }
ret = vkCreateDevice(hwctx->phys_dev, &dev_info, hwctx->alloc,
&hwctx->act_dev);
+ av_free((void *)queue_create_info[0].pQueuePriorities);
+ av_free((void *)queue_create_info[1].pQueuePriorities);
+ av_free((void *)queue_create_info[2].pQueuePriorities);
+
if (ret != VK_SUCCESS) {
av_log(ctx, AV_LOG_ERROR, "Device creation failure: %s\n",
vk_ret2str(ret));
AVCUDADeviceContextInternal *cu_internal = cuda_dev->internal;
CudaFunctions *cu = cu_internal->cuda_dl;
- if (internal->cu_sem)
- CHECK_CU(cu->cuDestroyExternalSemaphore(internal->cu_sem));
-
for (int i = 0; i < planes; i++) {
+ if (internal->cu_sem[i])
+ CHECK_CU(cu->cuDestroyExternalSemaphore(internal->cu_sem[i]));
if (internal->cu_mma[i])
CHECK_CU(cu->cuMipmappedArrayDestroy(internal->cu_mma[i]));
if (internal->ext_mem[i])
for (int i = 0; i < planes; i++) {
vkDestroyImage(hwctx->act_dev, f->img[i], hwctx->alloc);
vkFreeMemory(hwctx->act_dev, f->mem[i], hwctx->alloc);
+ vkDestroySemaphore(hwctx->act_dev, f->sem[i], hwctx->alloc);
}
- vkDestroySemaphore(hwctx->act_dev, f->sem, hwctx->alloc);
-
av_free(f);
}
enum PrepMode {
PREP_MODE_WRITE,
PREP_MODE_RO_SHADER,
+ PREP_MODE_EXTERNAL_EXPORT,
};
static int prepare_frame(AVHWFramesContext *hwfc, VulkanExecCtx *ectx,
AVVkFrame *frame, enum PrepMode pmode)
{
VkResult ret;
+ uint32_t dst_qf;
VkImageLayout new_layout;
VkAccessFlags new_access;
AVHWDeviceContext *ctx = hwfc->device_ctx;
.commandBufferCount = 1,
.pCommandBuffers = &ectx->buf,
- .pSignalSemaphores = &frame->sem,
- .signalSemaphoreCount = 1,
+ .pSignalSemaphores = frame->sem,
+ .signalSemaphoreCount = planes,
};
+ VkPipelineStageFlagBits wait_st[AV_NUM_DATA_POINTERS];
+ for (int i = 0; i < planes; i++)
+ wait_st[i] = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT;
+
switch (pmode) {
case PREP_MODE_WRITE:
new_layout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
new_access = VK_ACCESS_TRANSFER_WRITE_BIT;
+ dst_qf = VK_QUEUE_FAMILY_IGNORED;
break;
case PREP_MODE_RO_SHADER:
new_layout = VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL;
new_access = VK_ACCESS_TRANSFER_READ_BIT;
+ dst_qf = VK_QUEUE_FAMILY_IGNORED;
+ break;
+ case PREP_MODE_EXTERNAL_EXPORT:
+ new_layout = VK_IMAGE_LAYOUT_GENERAL;
+ new_access = VK_ACCESS_MEMORY_READ_BIT | VK_ACCESS_MEMORY_WRITE_BIT;
+ dst_qf = VK_QUEUE_FAMILY_EXTERNAL_KHR;
+ s_info.pWaitSemaphores = frame->sem;
+ s_info.pWaitDstStageMask = wait_st;
+ s_info.waitSemaphoreCount = planes;
break;
}
img_bar[i].oldLayout = frame->layout[i];
img_bar[i].newLayout = new_layout;
img_bar[i].srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
- img_bar[i].dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
+ img_bar[i].dstQueueFamilyIndex = dst_qf;
img_bar[i].image = frame->img[i];
img_bar[i].subresourceRange.levelCount = 1;
img_bar[i].subresourceRange.layerCount = 1;
goto fail;
}
+ /* Create semaphore */
+ ret = vkCreateSemaphore(hwctx->act_dev, &sem_spawn,
+ hwctx->alloc, &f->sem[i]);
+ if (ret != VK_SUCCESS) {
+ av_log(hwctx, AV_LOG_ERROR, "Failed to create semaphore: %s\n",
+ vk_ret2str(ret));
+ return AVERROR_EXTERNAL;
+ }
+
f->layout[i] = image_create_info.initialLayout;
f->access[i] = 0x0;
}
- /* Create semaphore */
- ret = vkCreateSemaphore(hwctx->act_dev, &sem_spawn,
- hwctx->alloc, &f->sem);
- if (ret != VK_SUCCESS) {
- av_log(hwctx, AV_LOG_ERROR, "Failed to create semaphore: %s\n",
- vk_ret2str(ret));
- return AVERROR_EXTERNAL;
- }
-
f->flags = 0x0;
f->tiling = tiling;
for (int i = 0; i < planes; i++) {
vkDestroyImage(hwctx->act_dev, map->frame->img[i], hwctx->alloc);
vkFreeMemory(hwctx->act_dev, map->frame->mem[i], hwctx->alloc);
+ vkDestroySemaphore(hwctx->act_dev, map->frame->sem[i], hwctx->alloc);
}
- vkDestroySemaphore(hwctx->act_dev, map->frame->sem, hwctx->alloc);
-
av_freep(&map->frame);
}
VkBindImageMemoryInfo bind_info[AV_NUM_DATA_POINTERS] = { 0 };
VkBindImagePlaneMemoryInfo plane_info[AV_NUM_DATA_POINTERS] = { 0 };
VkExternalMemoryHandleTypeFlagBits htype = VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT;
- VkSemaphoreCreateInfo sem_spawn = {
- .sType = VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO,
- };
VK_LOAD_PFN(hwctx->inst, vkGetMemoryFdPropertiesKHR);
.handleTypes = htype,
};
+ VkSemaphoreCreateInfo sem_spawn = {
+ .sType = VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO,
+ };
+
const int p_w = i > 0 ? AV_CEIL_RSHIFT(hwfc->width, fmt_desc->log2_chroma_w) : hwfc->width;
const int p_h = i > 0 ? AV_CEIL_RSHIFT(hwfc->height, fmt_desc->log2_chroma_h) : hwfc->height;
goto fail;
}
+ ret = vkCreateSemaphore(hwctx->act_dev, &sem_spawn,
+ hwctx->alloc, &f->sem[i]);
+ if (ret != VK_SUCCESS) {
+ av_log(hwctx, AV_LOG_ERROR, "Failed to create semaphore: %s\n",
+ vk_ret2str(ret));
+ return AVERROR_EXTERNAL;
+ }
+
+ /* We'd import a semaphore onto the one we created using
+ * vkImportSemaphoreFdKHR but unfortunately neither DRM nor VAAPI
+ * offer us anything we could import and sync with, so instead
+ * just signal the semaphore we created. */
+
f->layout[i] = image_create_info.initialLayout;
f->access[i] = 0x0;
}
}
- ret = vkCreateSemaphore(hwctx->act_dev, &sem_spawn,
- hwctx->alloc, &f->sem);
- if (ret != VK_SUCCESS) {
- av_log(hwctx, AV_LOG_ERROR, "Failed to create semaphore: %s\n",
- vk_ret2str(ret));
- return AVERROR_EXTERNAL;
- }
-
- /* We'd import a semaphore onto the one we created using
- * vkImportSemaphoreFdKHR but unfortunately neither DRM nor VAAPI
- * offer us anything we could import and sync with, so instead
- * just signal the semaphore we created. */
-
/* Bind the allocated memory to the images */
ret = vkBindImageMemory2(hwctx->act_dev, bind_counts, bind_info);
if (ret != VK_SUCCESS) {
return 0;
fail:
- for (int i = 0; i < desc->nb_layers; i++)
+ for (int i = 0; i < desc->nb_layers; i++) {
vkDestroyImage(hwctx->act_dev, f->img[i], hwctx->alloc);
+ vkDestroySemaphore(hwctx->act_dev, f->sem[i], hwctx->alloc);
+ }
for (int i = 0; i < desc->nb_objects; i++)
vkFreeMemory(hwctx->act_dev, f->mem[i], hwctx->alloc);
- vkDestroySemaphore(hwctx->act_dev, f->sem, hwctx->alloc);
av_free(f);
dst_int = dst_f->internal;
if (!dst_int || !dst_int->cuda_fc_ref) {
- VkSemaphoreGetFdInfoKHR sem_export = {
- .sType = VK_STRUCTURE_TYPE_SEMAPHORE_GET_FD_INFO_KHR,
- .semaphore = dst_f->sem,
- .handleType = VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT,
- };
- CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC ext_sem_desc = {
- .type = CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD,
- };
-
if (!dst_f->internal)
dst_f->internal = dst_int = av_mallocz(sizeof(*dst_f->internal));
.memory = dst_f->mem[i],
.handleType = VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR,
};
+ VkSemaphoreGetFdInfoKHR sem_export = {
+ .sType = VK_STRUCTURE_TYPE_SEMAPHORE_GET_FD_INFO_KHR,
+ .semaphore = dst_f->sem[i],
+ .handleType = VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT,
+ };
+ CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC ext_sem_desc = {
+ .type = CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD,
+ };
ret = pfn_vkGetMemoryFdKHR(hwctx->act_dev, &export_info,
&ext_desc.handle.fd);
err = AVERROR_EXTERNAL;
goto fail;
}
- }
- ret = pfn_vkGetSemaphoreFdKHR(hwctx->act_dev, &sem_export,
- &ext_sem_desc.handle.fd);
- if (ret != VK_SUCCESS) {
- av_log(ctx, AV_LOG_ERROR, "Failed to export semaphore: %s\n",
- vk_ret2str(ret));
- err = AVERROR_EXTERNAL;
- goto fail;
- }
+ ret = pfn_vkGetSemaphoreFdKHR(hwctx->act_dev, &sem_export,
+ &ext_sem_desc.handle.fd);
+ if (ret != VK_SUCCESS) {
+ av_log(ctx, AV_LOG_ERROR, "Failed to export semaphore: %s\n",
+ vk_ret2str(ret));
+ err = AVERROR_EXTERNAL;
+ goto fail;
+ }
- ret = CHECK_CU(cu->cuImportExternalSemaphore(&dst_int->cu_sem,
- &ext_sem_desc));
- if (ret < 0) {
- err = AVERROR_EXTERNAL;
- goto fail;
+ ret = CHECK_CU(cu->cuImportExternalSemaphore(&dst_int->cu_sem[i],
+ &ext_sem_desc));
+ if (ret < 0) {
+ err = AVERROR_EXTERNAL;
+ goto fail;
+ }
}
}
AVCUDADeviceContext *cuda_dev = cuda_cu->hwctx;
AVCUDADeviceContextInternal *cu_internal = cuda_dev->internal;
CudaFunctions *cu = cu_internal->cuda_dl;
- CUDA_EXTERNAL_SEMAPHORE_WAIT_PARAMS s_w_par = { 0 };
- CUDA_EXTERNAL_SEMAPHORE_SIGNAL_PARAMS s_s_par = { 0 };
+ CUDA_EXTERNAL_SEMAPHORE_WAIT_PARAMS s_w_par[AV_NUM_DATA_POINTERS] = { 0 };
+ CUDA_EXTERNAL_SEMAPHORE_SIGNAL_PARAMS s_s_par[AV_NUM_DATA_POINTERS] = { 0 };
ret = CHECK_CU(cu->cuCtxPushCurrent(cuda_dev->cuda_ctx));
if (ret < 0) {
}
dst_int = dst_f->internal;
- ret = CHECK_CU(cu->cuWaitExternalSemaphoresAsync(&dst_int->cu_sem, &s_w_par,
- 1, cuda_dev->stream));
+ ret = CHECK_CU(cu->cuWaitExternalSemaphoresAsync(dst_int->cu_sem, s_w_par,
+ planes, cuda_dev->stream));
if (ret < 0) {
err = AVERROR_EXTERNAL;
goto fail;
}
}
- ret = CHECK_CU(cu->cuSignalExternalSemaphoresAsync(&dst_int->cu_sem, &s_s_par,
- 1, cuda_dev->stream));
+ ret = CHECK_CU(cu->cuSignalExternalSemaphoresAsync(dst_int->cu_sem, s_s_par,
+ planes, cuda_dev->stream));
if (ret < 0) {
err = AVERROR_EXTERNAL;
goto fail;
if (!drm_desc)
return AVERROR(ENOMEM);
+ err = prepare_frame(hwfc, &p->cmd, f, PREP_MODE_EXTERNAL_EXPORT);
+ if (err < 0)
+ goto end;
+
err = ff_hwframe_map_create(src->hw_frames_ctx, dst, src, &vulkan_unmap_to_drm, drm_desc);
if (err < 0)
goto end;
.sType = VK_STRUCTURE_TYPE_SUBMIT_INFO,
.commandBufferCount = 1,
.pCommandBuffers = &s->cmd.buf,
- .pSignalSemaphores = &frame->sem,
- .pWaitSemaphores = &frame->sem,
+ .pSignalSemaphores = frame->sem,
+ .pWaitSemaphores = frame->sem,
.pWaitDstStageMask = sem_wait_dst,
- .signalSemaphoreCount = 1,
- .waitSemaphoreCount = 1,
+ .signalSemaphoreCount = planes,
+ .waitSemaphoreCount = planes,
};
ret = vkBeginCommandBuffer(s->cmd.buf, &cmd_start);