VkPhysicalDeviceProperties props;
VkPhysicalDeviceMemoryProperties mprops;
+ /* Queues */
+ uint32_t qfs[3];
+ int num_qfs;
+
/* Debug callback */
VkDebugUtilsMessengerEXT debug_ctx;
list = av_realloc_array(list, sizeof(*list), ++count); \
if (!list) { \
err = AVERROR(ENOMEM); \
- goto end; \
+ goto fail; \
+ } \
+ list[count - 1] = av_strdup(val); \
+ if (!list[count - 1]) { \
+ err = AVERROR(ENOMEM); \
+ goto fail; \
} \
- list[count - 1] = val; \
} while(0)
static const struct {
EXT_EXTERNAL_FD_MEMORY = 1ULL << 2, /* VK_KHR_external_memory_fd */
EXT_EXTERNAL_FD_SEM = 1ULL << 3, /* VK_KHR_external_semaphore_fd */
- EXT_OPTIONAL = 1ULL << 62,
- EXT_REQUIRED = 1ULL << 63,
+ EXT_NO_FLAG = 1ULL << 63,
};
typedef struct VulkanOptExtension {
} VulkanOptExtension;
static const VulkanOptExtension optional_instance_exts[] = {
- /* For future use */
+ { VK_KHR_SURFACE_EXTENSION_NAME, EXT_NO_FLAG },
};
static const VulkanOptExtension optional_device_exts[] = {
return 0;
}
-static int check_extensions(AVHWDeviceContext *ctx, int dev,
+static int check_extensions(AVHWDeviceContext *ctx, int dev, AVDictionary *opts,
const char * const **dst, uint32_t *num, int debug)
{
const char *tstr;
const char *mod;
int optional_exts_num;
uint32_t sup_ext_count;
+ char *user_exts_str = NULL;
+ AVDictionaryEntry *user_exts;
VkExtensionProperties *sup_ext;
const VulkanOptExtension *optional_exts;
mod = "instance";
optional_exts = optional_instance_exts;
optional_exts_num = FF_ARRAY_ELEMS(optional_instance_exts);
+ user_exts = av_dict_get(opts, "instance_extensions", NULL, 0);
+ if (user_exts) {
+ user_exts_str = av_strdup(user_exts->value);
+ if (!user_exts_str) {
+ err = AVERROR(ENOMEM);
+ goto fail;
+ }
+ }
vkEnumerateInstanceExtensionProperties(NULL, &sup_ext_count, NULL);
sup_ext = av_malloc_array(sup_ext_count, sizeof(VkExtensionProperties));
if (!sup_ext)
mod = "device";
optional_exts = optional_device_exts;
optional_exts_num = FF_ARRAY_ELEMS(optional_device_exts);
+ user_exts = av_dict_get(opts, "device_extensions", NULL, 0);
+ if (user_exts) {
+ user_exts_str = av_strdup(user_exts->value);
+ if (!user_exts_str) {
+ err = AVERROR(ENOMEM);
+ goto fail;
+ }
+ }
vkEnumerateDeviceExtensionProperties(hwctx->phys_dev, NULL,
&sup_ext_count, NULL);
sup_ext = av_malloc_array(sup_ext_count, sizeof(VkExtensionProperties));
}
for (int i = 0; i < optional_exts_num; i++) {
- int req = optional_exts[i].flag & EXT_REQUIRED;
tstr = optional_exts[i].name;
-
found = 0;
for (int j = 0; j < sup_ext_count; j++) {
if (!strcmp(tstr, sup_ext[j].extensionName)) {
break;
}
}
- if (!found) {
- int lvl = req ? AV_LOG_ERROR : AV_LOG_VERBOSE;
- av_log(ctx, lvl, "Extension \"%s\" not found!\n", tstr);
- if (req) {
- err = AVERROR(EINVAL);
- goto end;
- }
+ if (!found)
continue;
- }
- if (!req)
- p->extensions |= optional_exts[i].flag;
av_log(ctx, AV_LOG_VERBOSE, "Using %s extension \"%s\"\n", mod, tstr);
-
+ p->extensions |= optional_exts[i].flag;
ADD_VAL_TO_LIST(extension_names, extensions_found, tstr);
}
}
}
if (found) {
+ av_log(ctx, AV_LOG_VERBOSE, "Using %s extension \"%s\"\n", mod, tstr);
ADD_VAL_TO_LIST(extension_names, extensions_found, tstr);
} else {
av_log(ctx, AV_LOG_ERROR, "Debug extension \"%s\" not found!\n",
tstr);
err = AVERROR(EINVAL);
- goto end;
+ goto fail;
+ }
+ }
+
+ if (user_exts_str) {
+ char *save, *token = av_strtok(user_exts_str, "+", &save);
+ while (token) {
+ found = 0;
+ for (int j = 0; j < sup_ext_count; j++) {
+ if (!strcmp(token, sup_ext[j].extensionName)) {
+ found = 1;
+ break;
+ }
+ }
+ if (found) {
+ 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_WARNING, "%s extension \"%s\" not found, excluding.\n",
+ mod, token);
+ }
+ token = av_strtok(NULL, "+", &save);
}
}
*dst = extension_names;
*num = extensions_found;
-end:
+ av_free(user_exts_str);
+ av_free(sup_ext);
+ return 0;
+
+fail:
+ if (extension_names)
+ for (int i = 0; i < extensions_found; i++)
+ av_free((void *)extension_names[i]);
+ av_free(extension_names);
+ av_free(user_exts_str);
av_free(sup_ext);
return err;
}
};
/* Check for present/missing extensions */
- err = check_extensions(ctx, 0, &inst_props.ppEnabledExtensionNames,
+ err = check_extensions(ctx, 0, opts, &inst_props.ppEnabledExtensionNames,
&inst_props.enabledExtensionCount, debug_mode);
if (err < 0)
return err;
if (debug_mode) {
- static const char *layers[] = { "VK_LAYER_LUNARG_standard_validation" };
+ static const char *layers[] = { "VK_LAYER_KHRONOS_validation" };
inst_props.ppEnabledLayerNames = layers;
inst_props.enabledLayerCount = FF_ARRAY_ELEMS(layers);
}
/* Try to create the instance */
ret = vkCreateInstance(&inst_props, hwctx->alloc, &hwctx->inst);
- /* Free used memory */
- av_free((void *)inst_props.ppEnabledExtensionNames);
-
/* Check for errors */
if (ret != VK_SUCCESS) {
av_log(ctx, AV_LOG_ERROR, "Instance creation failure: %s\n",
vk_ret2str(ret));
+ for (int i = 0; i < inst_props.enabledExtensionCount; i++)
+ av_free((void *)inst_props.ppEnabledExtensionNames[i]);
+ av_free((void *)inst_props.ppEnabledExtensionNames);
return AVERROR_EXTERNAL;
}
hwctx->alloc, &p->debug_ctx);
}
+ hwctx->enabled_inst_extensions = inst_props.ppEnabledExtensionNames;
+ hwctx->nb_enabled_inst_extensions = inst_props.enabledExtensionCount;
+
return 0;
}
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,
}
vkDestroyInstance(hwctx->inst, hwctx->alloc);
+
+ for (int i = 0; i < hwctx->nb_enabled_inst_extensions; i++)
+ av_free((void *)hwctx->enabled_inst_extensions[i]);
+ av_free((void *)hwctx->enabled_inst_extensions);
+
+ for (int i = 0; i < hwctx->nb_enabled_dev_extensions; i++)
+ av_free((void *)hwctx->enabled_dev_extensions[i]);
+ av_free((void *)hwctx->enabled_dev_extensions);
}
static int vulkan_device_create_internal(AVHWDeviceContext *ctx,
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 = {
if ((err = search_queue_families(ctx, &dev_info)))
goto end;
- if ((err = check_extensions(ctx, 1, &dev_info.ppEnabledExtensionNames,
- &dev_info.enabledExtensionCount, 0)))
+ if ((err = check_extensions(ctx, 1, opts, &dev_info.ppEnabledExtensionNames,
+ &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 *)dev_info.ppEnabledExtensionNames);
+ 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));
+ for (int i = 0; i < dev_info.enabledExtensionCount; i++)
+ av_free((void *)dev_info.ppEnabledExtensionNames[i]);
+ av_free((void *)dev_info.ppEnabledExtensionNames);
err = AVERROR_EXTERNAL;
goto end;
}
if (opt_d)
p->use_linear_images = strtol(opt_d->value, NULL, 10);
+ hwctx->enabled_dev_extensions = dev_info.ppEnabledExtensionNames;
+ hwctx->nb_enabled_dev_extensions = dev_info.enabledExtensionCount;
+
end:
return err;
}
AVVulkanDeviceContext *hwctx = ctx->hwctx;
VulkanDevicePriv *p = ctx->internal->priv;
+ /* Set device extension flags */
+ for (int i = 0; i < hwctx->nb_enabled_dev_extensions; i++) {
+ for (int j = 0; j < FF_ARRAY_ELEMS(optional_device_exts); j++) {
+ if (!strcmp(hwctx->enabled_dev_extensions[i],
+ optional_device_exts[j].name)) {
+ p->extensions |= optional_device_exts[j].flag;
+ break;
+ }
+ }
+ }
+
vkGetPhysicalDeviceQueueFamilyProperties(hwctx->phys_dev, &queue_num, NULL);
if (!queue_num) {
av_log(ctx, AV_LOG_ERROR, "Failed to get queues!\n");
#undef CHECK_QUEUE
+ p->qfs[p->num_qfs++] = hwctx->queue_family_index;
+ if ((hwctx->queue_family_tx_index != hwctx->queue_family_index) &&
+ (hwctx->queue_family_tx_index != hwctx->queue_family_comp_index))
+ p->qfs[p->num_qfs++] = hwctx->queue_family_tx_index;
+ if ((hwctx->queue_family_comp_index != hwctx->queue_family_index) &&
+ (hwctx->queue_family_comp_index != hwctx->queue_family_tx_index))
+ 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);
if (err)
return 0;
}
-static int prepare_frame(AVHWFramesContext *hwfc, AVVkFrame *frame)
+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;
AVVulkanDeviceContext *hwctx = ctx->hwctx;
- VulkanFramesPriv *s = hwfc->internal->priv;
const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
VkImageMemoryBarrier img_bar[AV_NUM_DATA_POINTERS] = { 0 };
VkSubmitInfo s_info = {
.sType = VK_STRUCTURE_TYPE_SUBMIT_INFO,
.commandBufferCount = 1,
- .pCommandBuffers = &s->cmd.buf,
+ .pCommandBuffers = &ectx->buf,
.pSignalSemaphores = frame->sem,
.signalSemaphoreCount = planes,
};
- ret = vkBeginCommandBuffer(s->cmd.buf, &cmd_start);
+ 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;
+ }
+
+ ret = vkBeginCommandBuffer(ectx->buf, &cmd_start);
if (ret != VK_SUCCESS)
return AVERROR_EXTERNAL;
for (int i = 0; i < planes; i++) {
img_bar[i].sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER;
img_bar[i].srcAccessMask = 0x0;
- img_bar[i].dstAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT;
+ img_bar[i].dstAccessMask = new_access;
img_bar[i].oldLayout = frame->layout[i];
- img_bar[i].newLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
+ 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;
frame->access[i] = img_bar[i].dstAccessMask;
}
- vkCmdPipelineBarrier(s->cmd.buf, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
+ 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(s->cmd.buf);
+ ret = vkEndCommandBuffer(ectx->buf);
if (ret != VK_SUCCESS)
return AVERROR_EXTERNAL;
- ret = vkQueueSubmit(s->cmd.queue, 1, &s_info, s->cmd.fence);
+ ret = vkQueueSubmit(ectx->queue, 1, &s_info, ectx->fence);
if (ret != VK_SUCCESS) {
return AVERROR_EXTERNAL;
} else {
- vkWaitForFences(hwctx->act_dev, 1, &s->cmd.fence, VK_TRUE, UINT64_MAX);
- vkResetFences(hwctx->act_dev, 1, &s->cmd.fence);
+ vkWaitForFences(hwctx->act_dev, 1, &ectx->fence, VK_TRUE, UINT64_MAX);
+ vkResetFences(hwctx->act_dev, 1, &ectx->fence);
}
return 0;
const int p_h = i > 0 ? AV_CEIL_RSHIFT(h, desc->log2_chroma_h) : h;
VkImageCreateInfo image_create_info = {
- .sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO,
- .pNext = create_pnext,
- .imageType = VK_IMAGE_TYPE_2D,
- .format = img_fmts[i],
- .extent.width = p_w,
- .extent.height = p_h,
- .extent.depth = 1,
- .mipLevels = 1,
- .arrayLayers = 1,
- .flags = VK_IMAGE_CREATE_ALIAS_BIT,
- .tiling = tiling,
- .initialLayout = VK_IMAGE_LAYOUT_UNDEFINED,
- .usage = usage,
- .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
- .samples = VK_SAMPLE_COUNT_1_BIT,
+ .sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO,
+ .pNext = create_pnext,
+ .imageType = VK_IMAGE_TYPE_2D,
+ .format = img_fmts[i],
+ .extent.width = p_w,
+ .extent.height = p_h,
+ .extent.depth = 1,
+ .mipLevels = 1,
+ .arrayLayers = 1,
+ .flags = VK_IMAGE_CREATE_ALIAS_BIT,
+ .tiling = tiling,
+ .initialLayout = VK_IMAGE_LAYOUT_UNDEFINED,
+ .usage = usage,
+ .samples = VK_SAMPLE_COUNT_1_BIT,
+ .pQueueFamilyIndices = p->qfs,
+ .queueFamilyIndexCount = p->num_qfs,
+ .sharingMode = p->num_qfs > 1 ? VK_SHARING_MODE_CONCURRENT :
+ VK_SHARING_MODE_EXCLUSIVE,
};
ret = vkCreateImage(hwctx->act_dev, &image_create_info,
if (err)
goto fail;
- err = prepare_frame(hwfc, f);
+ err = prepare_frame(hwfc, &p->cmd, f, PREP_MODE_WRITE);
if (err)
goto fail;
int err = 0;
VkResult ret;
AVVkFrame *f;
+ int bind_counts = 0;
AVHWDeviceContext *ctx = hwfc->device_ctx;
AVVulkanDeviceContext *hwctx = ctx->hwctx;
VulkanDevicePriv *p = ctx->internal->priv;
- const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
const AVPixFmtDescriptor *fmt_desc = av_pix_fmt_desc_get(hwfc->sw_format);
const int has_modifiers = p->extensions & EXT_DRM_MODIFIER_FLAGS;
- VkSubresourceLayout plane_data[AV_NUM_DATA_POINTERS];
- VkBindImageMemoryInfo bind_info[AV_NUM_DATA_POINTERS];
+ VkSubresourceLayout plane_data[AV_NUM_DATA_POINTERS] = { 0 };
+ 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;
VK_LOAD_PFN(hwctx->inst, vkGetMemoryFdPropertiesKHR);
for (int i = 0; i < desc->nb_layers; i++) {
- if (desc->layers[i].nb_planes > 1) {
- av_log(ctx, AV_LOG_ERROR, "Cannot import DMABUFS with more than 1 "
- "plane per layer!\n");
- return AVERROR(EINVAL);
- }
-
if (drm_to_vulkan_fmt(desc->layers[i].format) == VK_FORMAT_UNDEFINED) {
av_log(ctx, AV_LOG_ERROR, "Unsupported DMABUF layer format %#08x!\n",
desc->layers[i].format);
VkImportMemoryFdInfoKHR idesc = {
.sType = VK_STRUCTURE_TYPE_IMPORT_MEMORY_FD_INFO_KHR,
.handleType = htype,
- .fd = desc->objects[i].fd,
+ .fd = dup(desc->objects[i].fd),
};
ret = pfn_vkGetMemoryFdPropertiesKHR(hwctx->act_dev, htype,
- desc->objects[i].fd, &fdmp);
+ idesc.fd, &fdmp);
if (ret != VK_SUCCESS) {
av_log(hwfc, AV_LOG_ERROR, "Failed to get FD properties: %s\n",
vk_ret2str(ret));
err = AVERROR_EXTERNAL;
+ close(idesc.fd);
goto fail;
}
req.memoryTypeBits = fdmp.memoryTypeBits;
- err = alloc_mem(ctx, &req, 0x0, &idesc, &f->flags, &f->mem[i]);
- if (err)
+ err = alloc_mem(ctx, &req, VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT,
+ &idesc, &f->flags, &f->mem[i]);
+ if (err) {
+ close(idesc.fd);
return err;
+ }
f->size[i] = desc->objects[i].size;
}
VK_IMAGE_TILING_LINEAR : VK_IMAGE_TILING_OPTIMAL;
for (int i = 0; i < desc->nb_layers; i++) {
+ const int planes = desc->layers[i].nb_planes;
+ const int signal_p = has_modifiers && (planes > 1);
+
VkImageDrmFormatModifierExplicitCreateInfoEXT drm_info = {
.sType = VK_STRUCTURE_TYPE_IMAGE_DRM_FORMAT_MODIFIER_EXPLICIT_CREATE_INFO_EXT,
.drmFormatModifier = desc->objects[0].format_modifier,
- .drmFormatModifierPlaneCount = desc->layers[i].nb_planes,
+ .drmFormatModifierPlaneCount = planes,
.pPlaneLayouts = (const VkSubresourceLayout *)&plane_data,
};
const int p_h = i > 0 ? AV_CEIL_RSHIFT(hwfc->height, fmt_desc->log2_chroma_h) : hwfc->height;
VkImageCreateInfo image_create_info = {
- .sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO,
- .pNext = &einfo,
- .imageType = VK_IMAGE_TYPE_2D,
- .format = drm_to_vulkan_fmt(desc->layers[i].format),
- .extent.width = p_w,
- .extent.height = p_h,
- .extent.depth = 1,
- .mipLevels = 1,
- .arrayLayers = 1,
- .flags = VK_IMAGE_CREATE_ALIAS_BIT,
- .tiling = f->tiling,
- .initialLayout = VK_IMAGE_LAYOUT_UNDEFINED, /* specs say so */
- .usage = DEFAULT_USAGE_FLAGS,
- .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
- .samples = VK_SAMPLE_COUNT_1_BIT,
+ .sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO,
+ .pNext = &einfo,
+ .imageType = VK_IMAGE_TYPE_2D,
+ .format = drm_to_vulkan_fmt(desc->layers[i].format),
+ .extent.width = p_w,
+ .extent.height = p_h,
+ .extent.depth = 1,
+ .mipLevels = 1,
+ .arrayLayers = 1,
+ .flags = VK_IMAGE_CREATE_ALIAS_BIT,
+ .tiling = f->tiling,
+ .initialLayout = VK_IMAGE_LAYOUT_UNDEFINED, /* specs say so */
+ .usage = DEFAULT_USAGE_FLAGS,
+ .samples = VK_SAMPLE_COUNT_1_BIT,
+ .pQueueFamilyIndices = p->qfs,
+ .queueFamilyIndexCount = p->num_qfs,
+ .sharingMode = p->num_qfs > 1 ? VK_SHARING_MODE_CONCURRENT :
+ VK_SHARING_MODE_EXCLUSIVE,
};
- for (int j = 0; j < desc->layers[i].nb_planes; j++) {
+ for (int j = 0; j < planes; j++) {
plane_data[j].offset = desc->layers[i].planes[j].offset;
plane_data[j].rowPitch = desc->layers[i].planes[j].pitch;
plane_data[j].size = 0; /* The specs say so for all 3 */
/* 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
- * leave the semaphore unsignalled and enjoy the validation spam. */
+ * just signal the semaphore we created. */
f->layout[i] = image_create_info.initialLayout;
f->access[i] = 0x0;
- /* TODO: Fix to support more than 1 plane per layer */
- bind_info[i].sType = VK_STRUCTURE_TYPE_BIND_IMAGE_MEMORY_INFO;
- bind_info[i].pNext = NULL;
- bind_info[i].image = f->img[i];
- bind_info[i].memory = f->mem[desc->layers[i].planes[0].object_index];
- bind_info[i].memoryOffset = desc->layers[i].planes[0].offset;
+ for (int j = 0; j < planes; j++) {
+ VkImageAspectFlagBits aspect = j == 0 ? VK_IMAGE_ASPECT_MEMORY_PLANE_0_BIT_EXT :
+ j == 1 ? VK_IMAGE_ASPECT_MEMORY_PLANE_1_BIT_EXT :
+ VK_IMAGE_ASPECT_MEMORY_PLANE_2_BIT_EXT;
+
+ plane_info[bind_counts].sType = VK_STRUCTURE_TYPE_BIND_IMAGE_PLANE_MEMORY_INFO;
+ plane_info[bind_counts].planeAspect = aspect;
+
+ bind_info[bind_counts].sType = VK_STRUCTURE_TYPE_BIND_IMAGE_MEMORY_INFO;
+ bind_info[bind_counts].pNext = signal_p ? &plane_info[bind_counts] : NULL;
+ bind_info[bind_counts].image = f->img[i];
+ bind_info[bind_counts].memory = f->mem[desc->layers[i].planes[j].object_index];
+ bind_info[bind_counts].memoryOffset = desc->layers[i].planes[j].offset;
+ bind_counts++;
+ }
}
/* Bind the allocated memory to the images */
- ret = vkBindImageMemory2(hwctx->act_dev, planes, bind_info);
+ ret = vkBindImageMemory2(hwctx->act_dev, bind_counts, bind_info);
if (ret != VK_SUCCESS) {
av_log(ctx, AV_LOG_ERROR, "Failed to bind memory: %s\n",
vk_ret2str(ret));
return AVERROR_EXTERNAL;
}
+ /* NOTE: This is completely uneccesary and unneeded once we can import
+ * semaphores from DRM. Otherwise we have to activate the semaphores.
+ * We're reusing the exec context that's also used for uploads/downloads. */
+ err = prepare_frame(hwfc, &p->cmd, f, PREP_MODE_RO_SHADER);
+ if (err)
+ goto fail;
+
*frame = f;
return 0;
fail:
- for (int i = 0; i < planes; i++) {
+ for (int i = 0; i < desc->nb_layers; 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);
}
+ for (int i = 0; i < desc->nb_objects; i++)
+ vkFreeMemory(hwctx->act_dev, f->mem[i], hwctx->alloc);
av_free(f);
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;
drm_desc->layers[i].planes[0].object_index = FFMIN(i, drm_desc->nb_objects - 1);
- if (f->tiling != VK_IMAGE_TILING_OPTIMAL)
+ if (f->tiling == VK_IMAGE_TILING_OPTIMAL)
continue;
vkGetImageSubresourceLayout(hwctx->act_dev, f->img[i], &sub, &layout);
VkResult ret;
AVVulkanDeviceContext *hwctx = ctx->hwctx;
VulkanDevicePriv *s = ctx->internal->priv;
+
+ int bar_num = 0;
VkPipelineStageFlagBits sem_wait_dst[AV_NUM_DATA_POINTERS];
const int planes = av_pix_fmt_count_planes(pix_fmt);
/* Change the image layout to something more optimal for transfers */
for (int i = 0; i < planes; i++) {
- img_bar[i].sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER;
- img_bar[i].srcAccessMask = 0x0;
- img_bar[i].dstAccessMask = to_buf ? VK_ACCESS_TRANSFER_READ_BIT :
+ VkImageLayout new_layout = to_buf ? VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL :
+ VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
+ VkAccessFlags new_access = to_buf ? VK_ACCESS_TRANSFER_READ_BIT :
VK_ACCESS_TRANSFER_WRITE_BIT;
- img_bar[i].oldLayout = frame->layout[i];
- img_bar[i].newLayout = to_buf ? VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL :
- VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
- img_bar[i].srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
- img_bar[i].dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
- img_bar[i].image = frame->img[i];
- img_bar[i].subresourceRange.levelCount = 1;
- img_bar[i].subresourceRange.layerCount = 1;
- img_bar[i].subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
sem_wait_dst[i] = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT;
- frame->layout[i] = img_bar[i].newLayout;
- frame->access[i] = img_bar[i].dstAccessMask;
+ /* If the layout matches and we have read access skip the barrier */
+ if ((frame->layout[i] == new_layout) && (frame->access[i] & new_access))
+ continue;
+
+ img_bar[bar_num].sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER;
+ img_bar[bar_num].srcAccessMask = 0x0;
+ img_bar[bar_num].dstAccessMask = new_access;
+ img_bar[bar_num].oldLayout = frame->layout[i];
+ img_bar[bar_num].newLayout = new_layout;
+ img_bar[bar_num].srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
+ img_bar[bar_num].dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
+ img_bar[bar_num].image = frame->img[i];
+ img_bar[bar_num].subresourceRange.levelCount = 1;
+ img_bar[bar_num].subresourceRange.layerCount = 1;
+ img_bar[bar_num].subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
+
+ frame->layout[i] = img_bar[bar_num].newLayout;
+ frame->access[i] = img_bar[bar_num].dstAccessMask;
+
+ bar_num++;
}
- vkCmdPipelineBarrier(s->cmd.buf, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
- VK_PIPELINE_STAGE_TRANSFER_BIT, 0,
- 0, NULL, 0, NULL, planes, img_bar);
+ if (bar_num)
+ vkCmdPipelineBarrier(s->cmd.buf, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
+ VK_PIPELINE_STAGE_TRANSFER_BIT, 0,
+ 0, NULL, 0, NULL, bar_num, img_bar);
/* Schedule a copy for each plane */
for (int i = 0; i < planes; i++) {
int h = src->height;
int p_height = i > 0 ? AV_CEIL_RSHIFT(h, log2_chroma) : h;
- tmp.linesize[i] = src->linesize[i];
+ tmp.linesize[i] = FFABS(src->linesize[i]);
err = create_buf(dev_ctx, &buf[i], p_height,
&tmp.linesize[i], VK_BUFFER_USAGE_TRANSFER_SRC_BIT,
VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT, NULL, NULL);
int h = dst->height;
int p_height = i > 0 ? AV_CEIL_RSHIFT(h, log2_chroma) : h;
- tmp.linesize[i] = dst->linesize[i];
+ tmp.linesize[i] = FFABS(dst->linesize[i]);
err = create_buf(dev_ctx, &buf[i], p_height,
&tmp.linesize[i], VK_BUFFER_USAGE_TRANSFER_DST_BIT,
VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT, NULL, NULL);
projects / ffmpeg / blobdiff