| File: | root/firefox-clang/media/ffvpx/libavutil/vulkan.c |
| Warning: | line 306, column 13 Access to field 'fence' results in a dereference of a null pointer (loaded from variable 'e') |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 1 | /* | |||
| 2 | * Copyright (c) Lynne | |||
| 3 | * | |||
| 4 | * This file is part of FFmpeg. | |||
| 5 | * | |||
| 6 | * FFmpeg is free software; you can redistribute it and/or | |||
| 7 | * modify it under the terms of the GNU Lesser General Public | |||
| 8 | * License as published by the Free Software Foundation; either | |||
| 9 | * version 2.1 of the License, or (at your option) any later version. | |||
| 10 | * | |||
| 11 | * FFmpeg is distributed in the hope that it will be useful, | |||
| 12 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |||
| 13 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |||
| 14 | * Lesser General Public License for more details. | |||
| 15 | * | |||
| 16 | * You should have received a copy of the GNU Lesser General Public | |||
| 17 | * License along with FFmpeg; if not, write to the Free Software | |||
| 18 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA | |||
| 19 | */ | |||
| 20 | ||||
| 21 | #include "config.h" | |||
| 22 | #include "avassert.h" | |||
| 23 | #include "mem.h" | |||
| 24 | ||||
| 25 | #include "vulkan.h" | |||
| 26 | #include "libavutil/vulkan_loader.h" | |||
| 27 | ||||
| 28 | #if CONFIG_SHADER_COMPRESSION0 | |||
| 29 | #include "libavutil/zlib_utils.h" | |||
| 30 | #endif | |||
| 31 | ||||
| 32 | const VkComponentMapping ff_comp_identity_map = { | |||
| 33 | .r = VK_COMPONENT_SWIZZLE_IDENTITY, | |||
| 34 | .g = VK_COMPONENT_SWIZZLE_IDENTITY, | |||
| 35 | .b = VK_COMPONENT_SWIZZLE_IDENTITY, | |||
| 36 | .a = VK_COMPONENT_SWIZZLE_IDENTITY, | |||
| 37 | }; | |||
| 38 | ||||
| 39 | /* Converts return values to strings */ | |||
| 40 | const char *ff_vk_ret2str(VkResult res) | |||
| 41 | { | |||
| 42 | #define CASE(VAL) case VAL: return #VAL | |||
| 43 | switch (res) { | |||
| 44 | CASE(VK_SUCCESS); | |||
| 45 | CASE(VK_NOT_READY); | |||
| 46 | CASE(VK_TIMEOUT); | |||
| 47 | CASE(VK_EVENT_SET); | |||
| 48 | CASE(VK_EVENT_RESET); | |||
| 49 | CASE(VK_INCOMPLETE); | |||
| 50 | CASE(VK_ERROR_OUT_OF_HOST_MEMORY); | |||
| 51 | CASE(VK_ERROR_OUT_OF_DEVICE_MEMORY); | |||
| 52 | CASE(VK_ERROR_INITIALIZATION_FAILED); | |||
| 53 | CASE(VK_ERROR_DEVICE_LOST); | |||
| 54 | CASE(VK_ERROR_MEMORY_MAP_FAILED); | |||
| 55 | CASE(VK_ERROR_LAYER_NOT_PRESENT); | |||
| 56 | CASE(VK_ERROR_EXTENSION_NOT_PRESENT); | |||
| 57 | CASE(VK_ERROR_FEATURE_NOT_PRESENT); | |||
| 58 | CASE(VK_ERROR_INCOMPATIBLE_DRIVER); | |||
| 59 | CASE(VK_ERROR_TOO_MANY_OBJECTS); | |||
| 60 | CASE(VK_ERROR_FORMAT_NOT_SUPPORTED); | |||
| 61 | CASE(VK_ERROR_FRAGMENTED_POOL); | |||
| 62 | CASE(VK_ERROR_UNKNOWN); | |||
| 63 | CASE(VK_ERROR_OUT_OF_POOL_MEMORY); | |||
| 64 | CASE(VK_ERROR_INVALID_EXTERNAL_HANDLE); | |||
| 65 | CASE(VK_ERROR_FRAGMENTATION); | |||
| 66 | CASE(VK_ERROR_INVALID_OPAQUE_CAPTURE_ADDRESS); | |||
| 67 | CASE(VK_PIPELINE_COMPILE_REQUIRED); | |||
| 68 | CASE(VK_ERROR_SURFACE_LOST_KHR); | |||
| 69 | CASE(VK_ERROR_NATIVE_WINDOW_IN_USE_KHR); | |||
| 70 | CASE(VK_SUBOPTIMAL_KHR); | |||
| 71 | CASE(VK_ERROR_OUT_OF_DATE_KHR); | |||
| 72 | CASE(VK_ERROR_INCOMPATIBLE_DISPLAY_KHR); | |||
| 73 | CASE(VK_ERROR_VALIDATION_FAILED_EXT); | |||
| 74 | CASE(VK_ERROR_INVALID_SHADER_NV); | |||
| 75 | CASE(VK_ERROR_VIDEO_PICTURE_LAYOUT_NOT_SUPPORTED_KHR); | |||
| 76 | CASE(VK_ERROR_VIDEO_PROFILE_OPERATION_NOT_SUPPORTED_KHR); | |||
| 77 | CASE(VK_ERROR_VIDEO_PROFILE_FORMAT_NOT_SUPPORTED_KHR); | |||
| 78 | CASE(VK_ERROR_VIDEO_PROFILE_CODEC_NOT_SUPPORTED_KHR); | |||
| 79 | CASE(VK_ERROR_VIDEO_STD_VERSION_NOT_SUPPORTED_KHR); | |||
| 80 | CASE(VK_ERROR_INVALID_DRM_FORMAT_MODIFIER_PLANE_LAYOUT_EXT); | |||
| 81 | CASE(VK_ERROR_NOT_PERMITTED_KHR); | |||
| 82 | CASE(VK_ERROR_FULL_SCREEN_EXCLUSIVE_MODE_LOST_EXT); | |||
| 83 | CASE(VK_THREAD_IDLE_KHR); | |||
| 84 | CASE(VK_THREAD_DONE_KHR); | |||
| 85 | CASE(VK_OPERATION_DEFERRED_KHR); | |||
| 86 | CASE(VK_OPERATION_NOT_DEFERRED_KHR); | |||
| 87 | default: return "Unknown error"; | |||
| 88 | } | |||
| 89 | #undef CASE | |||
| 90 | } | |||
| 91 | ||||
| 92 | /* Malitia pura, Khronos */ | |||
| 93 | #define FN_MAP_TO(dst_t, dst_name, src_t, src_name) \ | |||
| 94 | dst_t ff_vk_map_ ##src_name## _to_ ##dst_name(src_t src) \ | |||
| 95 | { \ | |||
| 96 | dst_t dst = 0x0; \ | |||
| 97 | MAP_TO(VK_FORMAT_FEATURE_2_SAMPLED_IMAGE_BIT, \ | |||
| 98 | VK_IMAGE_USAGE_SAMPLED_BIT); \ | |||
| 99 | MAP_TO(VK_FORMAT_FEATURE_2_TRANSFER_SRC_BIT, \ | |||
| 100 | VK_IMAGE_USAGE_TRANSFER_SRC_BIT); \ | |||
| 101 | MAP_TO(VK_FORMAT_FEATURE_2_TRANSFER_DST_BIT, \ | |||
| 102 | VK_IMAGE_USAGE_TRANSFER_DST_BIT); \ | |||
| 103 | MAP_TO(VK_FORMAT_FEATURE_2_STORAGE_IMAGE_BIT, \ | |||
| 104 | VK_IMAGE_USAGE_STORAGE_BIT); \ | |||
| 105 | MAP_TO(VK_FORMAT_FEATURE_2_COLOR_ATTACHMENT_BIT, \ | |||
| 106 | VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT); \ | |||
| 107 | MAP_TO(VK_FORMAT_FEATURE_2_VIDEO_DECODE_OUTPUT_BIT_KHR, \ | |||
| 108 | VK_IMAGE_USAGE_VIDEO_DECODE_DST_BIT_KHR); \ | |||
| 109 | MAP_TO(VK_FORMAT_FEATURE_2_VIDEO_DECODE_DPB_BIT_KHR, \ | |||
| 110 | VK_IMAGE_USAGE_VIDEO_DECODE_DPB_BIT_KHR); \ | |||
| 111 | MAP_TO(VK_FORMAT_FEATURE_2_VIDEO_ENCODE_DPB_BIT_KHR, \ | |||
| 112 | VK_IMAGE_USAGE_VIDEO_ENCODE_DPB_BIT_KHR); \ | |||
| 113 | MAP_TO(VK_FORMAT_FEATURE_2_VIDEO_ENCODE_INPUT_BIT_KHR, \ | |||
| 114 | VK_IMAGE_USAGE_VIDEO_ENCODE_SRC_BIT_KHR); \ | |||
| 115 | MAP_TO(VK_FORMAT_FEATURE_2_HOST_IMAGE_TRANSFER_BIT_EXT, \ | |||
| 116 | VK_IMAGE_USAGE_HOST_TRANSFER_BIT_EXT); \ | |||
| 117 | return dst; \ | |||
| 118 | } | |||
| 119 | ||||
| 120 | #define MAP_TO(flag1, flag2) if (src & flag2) dst |= flag1; | |||
| 121 | FN_MAP_TO(VkFormatFeatureFlagBits2, feats, VkImageUsageFlags, usage) | |||
| 122 | #undef MAP_TO | |||
| 123 | #define MAP_TO(flag1, flag2) if (src & flag1) dst |= flag2; | |||
| 124 | FN_MAP_TO(VkImageUsageFlags, usage, VkFormatFeatureFlagBits2, feats) | |||
| 125 | #undef MAP_TO | |||
| 126 | #undef FN_MAP_TO | |||
| 127 | ||||
| 128 | static void load_enabled_qfs(FFVulkanContext *s) | |||
| 129 | { | |||
| 130 | s->nb_qfs = 0; | |||
| 131 | for (int i = 0; i < s->hwctx->nb_qf; i++) { | |||
| 132 | /* Skip duplicates */ | |||
| 133 | int skip = 0; | |||
| 134 | for (int j = 0; j < s->nb_qfs; j++) { | |||
| 135 | if (s->qfs[j] == s->hwctx->qf[i].idx) { | |||
| 136 | skip = 1; | |||
| 137 | break; | |||
| 138 | } | |||
| 139 | } | |||
| 140 | if (skip) | |||
| 141 | continue; | |||
| 142 | ||||
| 143 | s->qfs[s->nb_qfs++] = s->hwctx->qf[i].idx; | |||
| 144 | } | |||
| 145 | } | |||
| 146 | ||||
| 147 | int ff_vk_load_props(FFVulkanContext *s) | |||
| 148 | { | |||
| 149 | FFVulkanFunctions *vk = &s->vkfn; | |||
| 150 | ||||
| 151 | s->props = (VkPhysicalDeviceProperties2) { | |||
| 152 | .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROPERTIES_2, | |||
| 153 | }; | |||
| 154 | ||||
| 155 | FF_VK_STRUCT_EXT(s, &s->props, &s->props_11, FF_VK_EXT_NO_FLAG,do { if (((1ULL << 63) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 63))) { (&s->props_11) ->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES ; ff_vk_link_struct(&s->props, &s->props_11); } } while (0) | |||
| 156 | VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES)do { if (((1ULL << 63) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 63))) { (&s->props_11) ->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES ; ff_vk_link_struct(&s->props, &s->props_11); } } while (0); | |||
| 157 | FF_VK_STRUCT_EXT(s, &s->props, &s->driver_props, FF_VK_EXT_NO_FLAG,do { if (((1ULL << 63) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 63))) { (&s->driver_props )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES ; ff_vk_link_struct(&s->props, &s->driver_props ); } } while (0) | |||
| 158 | VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES)do { if (((1ULL << 63) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 63))) { (&s->driver_props )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES ; ff_vk_link_struct(&s->props, &s->driver_props ); } } while (0); | |||
| 159 | FF_VK_STRUCT_EXT(s, &s->props, &s->subgroup_props, FF_VK_EXT_NO_FLAG,do { if (((1ULL << 63) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 63))) { (&s->subgroup_props )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_PROPERTIES ; ff_vk_link_struct(&s->props, &s->subgroup_props ); } } while (0) | |||
| 160 | VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_PROPERTIES)do { if (((1ULL << 63) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 63))) { (&s->subgroup_props )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_PROPERTIES ; ff_vk_link_struct(&s->props, &s->subgroup_props ); } } while (0); | |||
| 161 | ||||
| 162 | FF_VK_STRUCT_EXT(s, &s->props, &s->push_desc_props, FF_VK_EXT_PUSH_DESCRIPTOR,do { if (((1ULL << 14) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 14))) { (&s->push_desc_props )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR ; ff_vk_link_struct(&s->props, &s->push_desc_props ); } } while (0) | |||
| 163 | VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR)do { if (((1ULL << 14) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 14))) { (&s->push_desc_props )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR ; ff_vk_link_struct(&s->props, &s->push_desc_props ); } } while (0); | |||
| 164 | FF_VK_STRUCT_EXT(s, &s->props, &s->hprops, FF_VK_EXT_EXTERNAL_HOST_MEMORY,do { if (((1ULL << 4) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 4))) { (&s->hprops)-> sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT ; ff_vk_link_struct(&s->props, &s->hprops); } } while (0) | |||
| 165 | VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT)do { if (((1ULL << 4) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 4))) { (&s->hprops)-> sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT ; ff_vk_link_struct(&s->props, &s->hprops); } } while (0); | |||
| 166 | FF_VK_STRUCT_EXT(s, &s->props, &s->coop_matrix_props, FF_VK_EXT_COOP_MATRIX,do { if (((1ULL << 11) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 11))) { (&s->coop_matrix_props )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COOPERATIVE_MATRIX_PROPERTIES_KHR ; ff_vk_link_struct(&s->props, &s->coop_matrix_props ); } } while (0) | |||
| 167 | VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COOPERATIVE_MATRIX_PROPERTIES_KHR)do { if (((1ULL << 11) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 11))) { (&s->coop_matrix_props )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COOPERATIVE_MATRIX_PROPERTIES_KHR ; ff_vk_link_struct(&s->props, &s->coop_matrix_props ); } } while (0); | |||
| 168 | FF_VK_STRUCT_EXT(s, &s->props, &s->optical_flow_props, FF_VK_EXT_OPTICAL_FLOW,do { if (((1ULL << 12) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 12))) { (&s->optical_flow_props )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_OPTICAL_FLOW_PROPERTIES_NV ; ff_vk_link_struct(&s->props, &s->optical_flow_props ); } } while (0) | |||
| 169 | VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_OPTICAL_FLOW_PROPERTIES_NV)do { if (((1ULL << 12) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 12))) { (&s->optical_flow_props )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_OPTICAL_FLOW_PROPERTIES_NV ; ff_vk_link_struct(&s->props, &s->optical_flow_props ); } } while (0); | |||
| 170 | FF_VK_STRUCT_EXT(s, &s->props, &s->host_image_props, FF_VK_EXT_HOST_IMAGE_COPY,do { if (((1ULL << 18) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 18))) { (&s->host_image_props )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_IMAGE_COPY_PROPERTIES_EXT ; ff_vk_link_struct(&s->props, &s->host_image_props ); } } while (0) | |||
| 171 | VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_IMAGE_COPY_PROPERTIES_EXT)do { if (((1ULL << 18) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 18))) { (&s->host_image_props )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_IMAGE_COPY_PROPERTIES_EXT ; ff_vk_link_struct(&s->props, &s->host_image_props ); } } while (0); | |||
| 172 | ||||
| 173 | #ifdef VK_EXT_shader_long_vector1 | |||
| 174 | FF_VK_STRUCT_EXT(s, &s->props, &s->long_vector_props, FF_VK_EXT_LONG_VECTOR,do { if (((1ULL << 22) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 22))) { (&s->long_vector_props )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_LONG_VECTOR_PROPERTIES_EXT ; ff_vk_link_struct(&s->props, &s->long_vector_props ); } } while (0) | |||
| 175 | VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_LONG_VECTOR_PROPERTIES_EXT)do { if (((1ULL << 22) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 22))) { (&s->long_vector_props )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_LONG_VECTOR_PROPERTIES_EXT ; ff_vk_link_struct(&s->props, &s->long_vector_props ); } } while (0); | |||
| 176 | #endif | |||
| 177 | ||||
| 178 | s->feats = (VkPhysicalDeviceFeatures2) { | |||
| 179 | .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2, | |||
| 180 | }; | |||
| 181 | ||||
| 182 | FF_VK_STRUCT_EXT(s, &s->feats, &s->feats_12, FF_VK_EXT_NO_FLAG,do { if (((1ULL << 63) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 63))) { (&s->feats_12) ->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES ; ff_vk_link_struct(&s->feats, &s->feats_12); } } while (0) | |||
| 183 | VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES)do { if (((1ULL << 63) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 63))) { (&s->feats_12) ->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES ; ff_vk_link_struct(&s->feats, &s->feats_12); } } while (0); | |||
| 184 | FF_VK_STRUCT_EXT(s, &s->feats, &s->atomic_float_feats, FF_VK_EXT_ATOMIC_FLOAT,do { if (((1ULL << 10) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 10))) { (&s->atomic_float_feats )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_FLOAT_FEATURES_EXT ; ff_vk_link_struct(&s->feats, &s->atomic_float_feats ); } } while (0) | |||
| 185 | VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_FLOAT_FEATURES_EXT)do { if (((1ULL << 10) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 10))) { (&s->atomic_float_feats )->sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_FLOAT_FEATURES_EXT ; ff_vk_link_struct(&s->feats, &s->atomic_float_feats ); } } while (0); | |||
| 186 | ||||
| 187 | /* Try allocating 1024 layouts */ | |||
| 188 | s->host_image_copy_layouts = av_malloc(sizeof(*s->host_image_copy_layouts)*1024); | |||
| 189 | s->host_image_props.pCopySrcLayouts = s->host_image_copy_layouts; | |||
| 190 | s->host_image_props.copySrcLayoutCount = 512; | |||
| 191 | s->host_image_props.pCopyDstLayouts = s->host_image_copy_layouts + 512; | |||
| 192 | s->host_image_props.copyDstLayoutCount = 512; | |||
| 193 | ||||
| 194 | vk->GetPhysicalDeviceProperties2(s->hwctx->phys_dev, &s->props); | |||
| 195 | ||||
| 196 | /* Check if we had enough memory for all layouts */ | |||
| 197 | if (s->host_image_props.copySrcLayoutCount == 512 || | |||
| 198 | s->host_image_props.copyDstLayoutCount == 512) { | |||
| 199 | VkImageLayout *new_array; | |||
| 200 | size_t new_size; | |||
| 201 | s->host_image_props.pCopySrcLayouts = | |||
| 202 | s->host_image_props.pCopyDstLayouts = NULL((void*)0); | |||
| 203 | s->host_image_props.copySrcLayoutCount = | |||
| 204 | s->host_image_props.copyDstLayoutCount = 0; | |||
| 205 | vk->GetPhysicalDeviceProperties2(s->hwctx->phys_dev, &s->props); | |||
| 206 | ||||
| 207 | new_size = s->host_image_props.copySrcLayoutCount + | |||
| 208 | s->host_image_props.copyDstLayoutCount; | |||
| 209 | new_size *= sizeof(*s->host_image_copy_layouts); | |||
| 210 | new_array = av_realloc(s->host_image_copy_layouts, new_size); | |||
| 211 | if (!new_array) | |||
| 212 | return AVERROR(ENOMEM)(-(12)); | |||
| 213 | ||||
| 214 | s->host_image_copy_layouts = new_array; | |||
| 215 | s->host_image_props.pCopySrcLayouts = new_array; | |||
| 216 | s->host_image_props.pCopyDstLayouts = new_array + s->host_image_props.copySrcLayoutCount; | |||
| 217 | vk->GetPhysicalDeviceProperties2(s->hwctx->phys_dev, &s->props); | |||
| 218 | } | |||
| 219 | ||||
| 220 | vk->GetPhysicalDeviceMemoryProperties(s->hwctx->phys_dev, &s->mprops); | |||
| 221 | vk->GetPhysicalDeviceFeatures2(s->hwctx->phys_dev, &s->feats); | |||
| 222 | ||||
| 223 | for (int i = 0; i < s->mprops.memoryTypeCount; i++) | |||
| 224 | s->host_cached_flag |= s->mprops.memoryTypes[i].propertyFlags & | |||
| 225 | VK_MEMORY_PROPERTY_HOST_CACHED_BIT; | |||
| 226 | ||||
| 227 | load_enabled_qfs(s); | |||
| 228 | ||||
| 229 | if (s->qf_props) | |||
| 230 | return 0; | |||
| 231 | ||||
| 232 | vk->GetPhysicalDeviceQueueFamilyProperties2(s->hwctx->phys_dev, &s->tot_nb_qfs, NULL((void*)0)); | |||
| 233 | ||||
| 234 | s->qf_props = av_calloc(s->tot_nb_qfs, sizeof(*s->qf_props)); | |||
| 235 | if (!s->qf_props) | |||
| 236 | return AVERROR(ENOMEM)(-(12)); | |||
| 237 | ||||
| 238 | s->query_props = av_calloc(s->tot_nb_qfs, sizeof(*s->query_props)); | |||
| 239 | if (!s->qf_props) { | |||
| 240 | av_freep(&s->qf_props); | |||
| 241 | return AVERROR(ENOMEM)(-(12)); | |||
| 242 | } | |||
| 243 | ||||
| 244 | s->video_props = av_calloc(s->tot_nb_qfs, sizeof(*s->video_props)); | |||
| 245 | if (!s->video_props) { | |||
| 246 | av_freep(&s->qf_props); | |||
| 247 | av_freep(&s->query_props); | |||
| 248 | return AVERROR(ENOMEM)(-(12)); | |||
| 249 | } | |||
| 250 | ||||
| 251 | for (uint32_t i = 0; i < s->tot_nb_qfs; i++) { | |||
| 252 | s->qf_props[i] = (VkQueueFamilyProperties2) { | |||
| 253 | .sType = VK_STRUCTURE_TYPE_QUEUE_FAMILY_PROPERTIES_2, | |||
| 254 | }; | |||
| 255 | ||||
| 256 | FF_VK_STRUCT_EXT(s, &s->qf_props[i], &s->query_props[i], FF_VK_EXT_VIDEO_QUEUE,do { if (((1ULL << 36) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 36))) { (&s->query_props [i])->sType = VK_STRUCTURE_TYPE_QUEUE_FAMILY_QUERY_RESULT_STATUS_PROPERTIES_KHR ; ff_vk_link_struct(&s->qf_props[i], &s->query_props [i]); } } while (0) | |||
| 257 | VK_STRUCTURE_TYPE_QUEUE_FAMILY_QUERY_RESULT_STATUS_PROPERTIES_KHR)do { if (((1ULL << 36) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 36))) { (&s->query_props [i])->sType = VK_STRUCTURE_TYPE_QUEUE_FAMILY_QUERY_RESULT_STATUS_PROPERTIES_KHR ; ff_vk_link_struct(&s->qf_props[i], &s->query_props [i]); } } while (0); | |||
| 258 | FF_VK_STRUCT_EXT(s, &s->qf_props[i], &s->video_props[i], FF_VK_EXT_VIDEO_QUEUE,do { if (((1ULL << 36) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 36))) { (&s->video_props [i])->sType = VK_STRUCTURE_TYPE_QUEUE_FAMILY_VIDEO_PROPERTIES_KHR ; ff_vk_link_struct(&s->qf_props[i], &s->video_props [i]); } } while (0) | |||
| 259 | VK_STRUCTURE_TYPE_QUEUE_FAMILY_VIDEO_PROPERTIES_KHR)do { if (((1ULL << 36) == (1ULL << 63)) || ((s)-> extensions & (1ULL << 36))) { (&s->video_props [i])->sType = VK_STRUCTURE_TYPE_QUEUE_FAMILY_VIDEO_PROPERTIES_KHR ; ff_vk_link_struct(&s->qf_props[i], &s->video_props [i]); } } while (0); | |||
| 260 | } | |||
| 261 | ||||
| 262 | vk->GetPhysicalDeviceQueueFamilyProperties2(s->hwctx->phys_dev, &s->tot_nb_qfs, s->qf_props); | |||
| 263 | ||||
| 264 | if (s->extensions & FF_VK_EXT_COOP_MATRIX(1ULL << 11)) { | |||
| 265 | vk->GetPhysicalDeviceCooperativeMatrixPropertiesKHR(s->hwctx->phys_dev, | |||
| 266 | &s->coop_mat_props_nb, NULL((void*)0)); | |||
| 267 | ||||
| 268 | if (s->coop_mat_props_nb) { | |||
| 269 | s->coop_mat_props = av_malloc_array(s->coop_mat_props_nb, | |||
| 270 | sizeof(VkCooperativeMatrixPropertiesKHR)); | |||
| 271 | for (int i = 0; i < s->coop_mat_props_nb; i++) { | |||
| 272 | s->coop_mat_props[i] = (VkCooperativeMatrixPropertiesKHR) { | |||
| 273 | .sType = VK_STRUCTURE_TYPE_COOPERATIVE_MATRIX_PROPERTIES_KHR, | |||
| 274 | }; | |||
| 275 | } | |||
| 276 | ||||
| 277 | vk->GetPhysicalDeviceCooperativeMatrixPropertiesKHR(s->hwctx->phys_dev, | |||
| 278 | &s->coop_mat_props_nb, | |||
| 279 | s->coop_mat_props); | |||
| 280 | } | |||
| 281 | } | |||
| 282 | ||||
| 283 | return 0; | |||
| 284 | } | |||
| 285 | ||||
| 286 | AVVulkanDeviceQueueFamily *ff_vk_qf_find(FFVulkanContext *s, | |||
| 287 | VkQueueFlagBits dev_family, | |||
| 288 | VkVideoCodecOperationFlagBitsKHR vid_ops) | |||
| 289 | { | |||
| 290 | for (int i = 0; i < s->hwctx->nb_qf; i++) { | |||
| 291 | if ((s->hwctx->qf[i].flags & dev_family) && | |||
| 292 | (s->hwctx->qf[i].video_caps & vid_ops) == vid_ops) { | |||
| 293 | return &s->hwctx->qf[i]; | |||
| 294 | } | |||
| 295 | } | |||
| 296 | return NULL((void*)0); | |||
| 297 | } | |||
| 298 | ||||
| 299 | void ff_vk_exec_pool_free(FFVulkanContext *s, FFVkExecPool *pool) | |||
| 300 | { | |||
| 301 | FFVulkanFunctions *vk = &s->vkfn; | |||
| 302 | ||||
| 303 | for (int i = 0; i < pool->pool_size; i++) { | |||
| 304 | FFVkExecContext *e = &pool->contexts[i]; | |||
| 305 | ||||
| 306 | if (e->fence) { | |||
| ||||
| 307 | if (e->had_submission) | |||
| 308 | vk->WaitForFences(s->hwctx->act_dev, 1, &e->fence, VK_TRUE1U, UINT64_MAX(18446744073709551615UL)); | |||
| 309 | vk->DestroyFence(s->hwctx->act_dev, e->fence, s->hwctx->alloc); | |||
| 310 | } | |||
| 311 | ||||
| 312 | ff_vk_exec_discard_deps(s, e); | |||
| 313 | ||||
| 314 | av_free(e->frame_deps); | |||
| 315 | av_free(e->sw_frame_deps); | |||
| 316 | av_free(e->buf_deps); | |||
| 317 | av_free(e->queue_family_dst); | |||
| 318 | av_free(e->layout_dst); | |||
| 319 | av_free(e->access_dst); | |||
| 320 | av_free(e->frame_update); | |||
| 321 | av_free(e->frame_locked); | |||
| 322 | av_free(e->sem_sig); | |||
| 323 | av_free(e->sem_sig_val_dst); | |||
| 324 | av_free(e->sem_wait); | |||
| 325 | } | |||
| 326 | ||||
| 327 | /* Free shader-specific data */ | |||
| 328 | for (int i = 0; i < pool->nb_reg_shd; i++) { | |||
| 329 | FFVulkanShaderData *sd = &pool->reg_shd[i]; | |||
| 330 | ||||
| 331 | if (sd->desc_pool) | |||
| 332 | vk->DestroyDescriptorPool(s->hwctx->act_dev, sd->desc_pool, | |||
| 333 | s->hwctx->alloc); | |||
| 334 | ||||
| 335 | av_freep(&sd->desc_sets); | |||
| 336 | } | |||
| 337 | pool->nb_reg_shd = 0; | |||
| 338 | ||||
| 339 | for (int i = 0; i < pool->pool_size; i++) { | |||
| 340 | if (pool->cmd_buf_pools[i]) | |||
| 341 | vk->FreeCommandBuffers(s->hwctx->act_dev, pool->cmd_buf_pools[i], | |||
| 342 | 1, &pool->cmd_bufs[i]); | |||
| 343 | ||||
| 344 | if (pool->cmd_buf_pools[i]) | |||
| 345 | vk->DestroyCommandPool(s->hwctx->act_dev, pool->cmd_buf_pools[i], s->hwctx->alloc); | |||
| 346 | } | |||
| 347 | if (pool->query_pool) | |||
| 348 | vk->DestroyQueryPool(s->hwctx->act_dev, pool->query_pool, s->hwctx->alloc); | |||
| 349 | ||||
| 350 | av_free(pool->query_data); | |||
| 351 | av_free(pool->cmd_buf_pools); | |||
| 352 | av_free(pool->cmd_bufs); | |||
| 353 | av_free(pool->contexts); | |||
| 354 | pool->pool_size = 0; | |||
| 355 | } | |||
| 356 | ||||
| 357 | int ff_vk_exec_pool_init(FFVulkanContext *s, AVVulkanDeviceQueueFamily *qf, | |||
| 358 | FFVkExecPool *pool, int nb_contexts, | |||
| 359 | int nb_queries, VkQueryType query_type, int query_64bit, | |||
| 360 | const void *query_create_pnext) | |||
| 361 | { | |||
| 362 | int err; | |||
| 363 | VkResult ret; | |||
| 364 | FFVulkanFunctions *vk = &s->vkfn; | |||
| 365 | ||||
| 366 | VkCommandPoolCreateInfo cqueue_create; | |||
| 367 | VkCommandBufferAllocateInfo cbuf_create; | |||
| 368 | ||||
| 369 | const VkQueryPoolVideoEncodeFeedbackCreateInfoKHR *ef = NULL((void*)0); | |||
| 370 | ||||
| 371 | atomic_init__c11_atomic_init(&pool->idx, 0); | |||
| 372 | ||||
| 373 | if (query_type == VK_QUERY_TYPE_VIDEO_ENCODE_FEEDBACK_KHR) { | |||
| ||||
| 374 | ef = ff_vk_find_struct(query_create_pnext, | |||
| 375 | VK_STRUCTURE_TYPE_QUERY_POOL_VIDEO_ENCODE_FEEDBACK_CREATE_INFO_KHR); | |||
| 376 | if (!ef) | |||
| 377 | return AVERROR(EINVAL)(-(22)); | |||
| 378 | } | |||
| 379 | ||||
| 380 | /* Allocate space for command buffer pools */ | |||
| 381 | pool->cmd_buf_pools = av_malloc(nb_contexts*sizeof(*pool->cmd_buf_pools)); | |||
| 382 | if (!pool->cmd_buf_pools) { | |||
| 383 | err = AVERROR(ENOMEM)(-(12)); | |||
| 384 | goto fail; | |||
| 385 | } | |||
| 386 | ||||
| 387 | /* Allocate space for command buffers */ | |||
| 388 | pool->cmd_bufs = av_malloc(nb_contexts*sizeof(*pool->cmd_bufs)); | |||
| 389 | if (!pool->cmd_bufs) { | |||
| 390 | err = AVERROR(ENOMEM)(-(12)); | |||
| 391 | goto fail; | |||
| 392 | } | |||
| 393 | ||||
| 394 | for (int i = 0; i < nb_contexts; i++) { | |||
| 395 | /* Create command pool */ | |||
| 396 | cqueue_create = (VkCommandPoolCreateInfo) { | |||
| 397 | .sType = VK_STRUCTURE_TYPE_COMMAND_POOL_CREATE_INFO, | |||
| 398 | .flags = VK_COMMAND_POOL_CREATE_TRANSIENT_BIT | | |||
| 399 | VK_COMMAND_POOL_CREATE_RESET_COMMAND_BUFFER_BIT, | |||
| 400 | .queueFamilyIndex = qf->idx, | |||
| 401 | }; | |||
| 402 | ||||
| 403 | ret = vk->CreateCommandPool(s->hwctx->act_dev, &cqueue_create, | |||
| 404 | s->hwctx->alloc, &pool->cmd_buf_pools[i]); | |||
| 405 | if (ret != VK_SUCCESS) { | |||
| 406 | av_log(s, AV_LOG_ERROR16, "Command pool creation failure: %s\n", | |||
| 407 | ff_vk_ret2str(ret)); | |||
| 408 | err = AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); | |||
| 409 | goto fail; | |||
| 410 | } | |||
| 411 | ||||
| 412 | /* Allocate command buffer */ | |||
| 413 | cbuf_create = (VkCommandBufferAllocateInfo) { | |||
| 414 | .sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO, | |||
| 415 | .level = VK_COMMAND_BUFFER_LEVEL_PRIMARY, | |||
| 416 | .commandPool = pool->cmd_buf_pools[i], | |||
| 417 | .commandBufferCount = 1, | |||
| 418 | }; | |||
| 419 | ret = vk->AllocateCommandBuffers(s->hwctx->act_dev, &cbuf_create, | |||
| 420 | &pool->cmd_bufs[i]); | |||
| 421 | if (ret != VK_SUCCESS) { | |||
| 422 | av_log(s, AV_LOG_ERROR16, "Command buffer alloc failure: %s\n", | |||
| 423 | ff_vk_ret2str(ret)); | |||
| 424 | err = AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); | |||
| 425 | goto fail; | |||
| 426 | } | |||
| 427 | } | |||
| 428 | ||||
| 429 | /* Query pool */ | |||
| 430 | if (nb_queries) { | |||
| 431 | VkQueryPoolCreateInfo query_pool_info = { | |||
| 432 | .sType = VK_STRUCTURE_TYPE_QUERY_POOL_CREATE_INFO, | |||
| 433 | .pNext = query_create_pnext, | |||
| 434 | .queryType = query_type, | |||
| 435 | .queryCount = nb_queries*nb_contexts, | |||
| 436 | }; | |||
| 437 | ret = vk->CreateQueryPool(s->hwctx->act_dev, &query_pool_info, | |||
| 438 | s->hwctx->alloc, &pool->query_pool); | |||
| 439 | if (ret != VK_SUCCESS) { | |||
| 440 | av_log(s, AV_LOG_ERROR16, "Query pool alloc failure: %s\n", | |||
| 441 | ff_vk_ret2str(ret)); | |||
| 442 | err = AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); | |||
| 443 | goto fail; | |||
| 444 | } | |||
| 445 | ||||
| 446 | pool->nb_queries = nb_queries; | |||
| 447 | pool->query_status_stride = 1 + 1; /* One result, one status by default */ | |||
| 448 | pool->query_results = nb_queries; | |||
| 449 | pool->query_statuses = nb_queries; | |||
| 450 | ||||
| 451 | /* Video encode queries produce two results per query */ | |||
| 452 | if (query_type == VK_QUERY_TYPE_VIDEO_ENCODE_FEEDBACK_KHR) { | |||
| 453 | int nb_results = av_popcountav_popcount_c(ef->encodeFeedbackFlags); | |||
| 454 | pool->query_status_stride = nb_results + 1; | |||
| 455 | pool->query_results *= nb_results; | |||
| 456 | } else if (query_type == VK_QUERY_TYPE_RESULT_STATUS_ONLY_KHR) { | |||
| 457 | pool->query_status_stride = 1; | |||
| 458 | pool->query_results = 0; | |||
| 459 | } | |||
| 460 | ||||
| 461 | pool->qd_size = (pool->query_results + pool->query_statuses)*(query_64bit ? 8 : 4); | |||
| 462 | ||||
| 463 | /* Allocate space for the query data */ | |||
| 464 | pool->query_data = av_calloc(nb_contexts, pool->qd_size); | |||
| 465 | if (!pool->query_data) { | |||
| 466 | err = AVERROR(ENOMEM)(-(12)); | |||
| 467 | goto fail; | |||
| 468 | } | |||
| 469 | } | |||
| 470 | ||||
| 471 | /* Allocate space for the contexts */ | |||
| 472 | pool->contexts = av_calloc(nb_contexts, sizeof(*pool->contexts)); | |||
| 473 | if (!pool->contexts) { | |||
| 474 | err = AVERROR(ENOMEM)(-(12)); | |||
| 475 | goto fail; | |||
| 476 | } | |||
| 477 | ||||
| 478 | pool->pool_size = nb_contexts; | |||
| 479 | ||||
| 480 | /* Init contexts */ | |||
| 481 | for (int i = 0; i < pool->pool_size; i++) { | |||
| 482 | FFVkExecContext *e = &pool->contexts[i]; | |||
| 483 | VkFenceCreateInfo fence_create = { | |||
| 484 | .sType = VK_STRUCTURE_TYPE_FENCE_CREATE_INFO, | |||
| 485 | .flags = VK_FENCE_CREATE_SIGNALED_BIT, | |||
| 486 | }; | |||
| 487 | ||||
| 488 | /* Fence */ | |||
| 489 | ret = vk->CreateFence(s->hwctx->act_dev, &fence_create, s->hwctx->alloc, | |||
| 490 | &e->fence); | |||
| 491 | if (ret != VK_SUCCESS) { | |||
| 492 | av_log(s, AV_LOG_ERROR16, "Failed to create submission fence: %s\n", | |||
| 493 | ff_vk_ret2str(ret)); | |||
| 494 | return AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); | |||
| 495 | } | |||
| 496 | ||||
| 497 | e->idx = i; | |||
| 498 | e->parent = pool; | |||
| 499 | ||||
| 500 | /* Query data */ | |||
| 501 | e->query_data = ((uint8_t *)pool->query_data) + pool->qd_size*i; | |||
| 502 | e->query_idx = nb_queries*i; | |||
| 503 | ||||
| 504 | /* Command buffer */ | |||
| 505 | e->buf = pool->cmd_bufs[i]; | |||
| 506 | ||||
| 507 | /* Queue index distribution */ | |||
| 508 | e->qi = i % qf->num; | |||
| 509 | e->qf = qf->idx; | |||
| 510 | vk->GetDeviceQueue(s->hwctx->act_dev, qf->idx, e->qi, &e->queue); | |||
| 511 | } | |||
| 512 | ||||
| 513 | return 0; | |||
| 514 | ||||
| 515 | fail: | |||
| 516 | ff_vk_exec_pool_free(s, pool); | |||
| 517 | return err; | |||
| 518 | } | |||
| 519 | ||||
| 520 | VkResult ff_vk_exec_get_query(FFVulkanContext *s, FFVkExecContext *e, | |||
| 521 | void **data, VkQueryResultFlagBits flags) | |||
| 522 | { | |||
| 523 | FFVulkanFunctions *vk = &s->vkfn; | |||
| 524 | const FFVkExecPool *pool = e->parent; | |||
| 525 | VkQueryResultFlags qf = flags & ~(VK_QUERY_RESULT_64_BIT | | |||
| 526 | VK_QUERY_RESULT_WITH_STATUS_BIT_KHR); | |||
| 527 | ||||
| 528 | if (!e->query_data) { | |||
| 529 | av_log(s, AV_LOG_ERROR16, "Requested a query with a NULL query_data pointer!\n"); | |||
| 530 | return VK_INCOMPLETE; | |||
| 531 | } | |||
| 532 | ||||
| 533 | qf |= pool->query_64bit ? | |||
| 534 | VK_QUERY_RESULT_64_BIT : 0x0; | |||
| 535 | qf |= pool->query_statuses ? | |||
| 536 | VK_QUERY_RESULT_WITH_STATUS_BIT_KHR : 0x0; | |||
| 537 | ||||
| 538 | if (data) | |||
| 539 | *data = e->query_data; | |||
| 540 | ||||
| 541 | return vk->GetQueryPoolResults(s->hwctx->act_dev, pool->query_pool, | |||
| 542 | e->query_idx, | |||
| 543 | pool->nb_queries, | |||
| 544 | pool->qd_size, e->query_data, | |||
| 545 | pool->qd_size, qf); | |||
| 546 | } | |||
| 547 | ||||
| 548 | FFVkExecContext *ff_vk_exec_get(FFVulkanContext *s, FFVkExecPool *pool) | |||
| 549 | { | |||
| 550 | return &pool->contexts[atomic_fetch_add(&pool->idx, 1)__c11_atomic_fetch_add(&pool->idx, 1, 5) % pool->pool_size]; | |||
| 551 | } | |||
| 552 | ||||
| 553 | void ff_vk_exec_wait(FFVulkanContext *s, FFVkExecContext *e) | |||
| 554 | { | |||
| 555 | FFVulkanFunctions *vk = &s->vkfn; | |||
| 556 | vk->WaitForFences(s->hwctx->act_dev, 1, &e->fence, VK_TRUE1U, UINT64_MAX(18446744073709551615UL)); | |||
| 557 | ff_vk_exec_discard_deps(s, e); | |||
| 558 | } | |||
| 559 | ||||
| 560 | int ff_vk_exec_start(FFVulkanContext *s, FFVkExecContext *e) | |||
| 561 | { | |||
| 562 | VkResult ret; | |||
| 563 | FFVulkanFunctions *vk = &s->vkfn; | |||
| 564 | const FFVkExecPool *pool = e->parent; | |||
| 565 | ||||
| 566 | VkCommandBufferBeginInfo cmd_start = { | |||
| 567 | .sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO, | |||
| 568 | .flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT, | |||
| 569 | }; | |||
| 570 | ||||
| 571 | /* Wait for the fence to be signalled */ | |||
| 572 | vk->WaitForFences(s->hwctx->act_dev, 1, &e->fence, VK_TRUE1U, UINT64_MAX(18446744073709551615UL)); | |||
| 573 | vk->ResetFences(s->hwctx->act_dev, 1, &e->fence); | |||
| 574 | ||||
| 575 | /* Discard queue dependencies */ | |||
| 576 | ff_vk_exec_discard_deps(s, e); | |||
| 577 | ||||
| 578 | ret = vk->BeginCommandBuffer(e->buf, &cmd_start); | |||
| 579 | if (ret != VK_SUCCESS) { | |||
| 580 | av_log(s, AV_LOG_ERROR16, "Failed to start command recoding: %s\n", | |||
| 581 | ff_vk_ret2str(ret)); | |||
| 582 | return AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); | |||
| 583 | } | |||
| 584 | ||||
| 585 | if (pool->nb_queries) | |||
| 586 | vk->CmdResetQueryPool(e->buf, pool->query_pool, | |||
| 587 | e->query_idx, pool->nb_queries); | |||
| 588 | ||||
| 589 | return 0; | |||
| 590 | } | |||
| 591 | ||||
| 592 | void ff_vk_exec_discard_deps(FFVulkanContext *s, FFVkExecContext *e) | |||
| 593 | { | |||
| 594 | for (int j = 0; j < e->nb_buf_deps; j++) | |||
| 595 | av_buffer_unref(&e->buf_deps[j]); | |||
| 596 | e->nb_buf_deps = 0; | |||
| 597 | ||||
| 598 | for (int j = 0; j < e->nb_sw_frame_deps; j++) | |||
| 599 | av_frame_free(&e->sw_frame_deps[j]); | |||
| 600 | e->nb_sw_frame_deps = 0; | |||
| 601 | ||||
| 602 | for (int j = 0; j < e->nb_frame_deps; j++) { | |||
| 603 | AVFrame *f = e->frame_deps[j]; | |||
| 604 | if (e->frame_locked[j]) { | |||
| 605 | AVHWFramesContext *hwfc = (AVHWFramesContext *)f->hw_frames_ctx->data; | |||
| 606 | AVVulkanFramesContext *vkfc = hwfc->hwctx; | |||
| 607 | AVVkFrame *vkf = (AVVkFrame *)f->data[0]; | |||
| 608 | vkfc->unlock_frame(hwfc, vkf); | |||
| 609 | e->frame_locked[j] = 0; | |||
| 610 | } | |||
| 611 | e->frame_update[j] = 0; | |||
| 612 | } | |||
| 613 | e->nb_frame_deps = 0; | |||
| 614 | ||||
| 615 | e->sem_wait_cnt = 0; | |||
| 616 | e->sem_sig_cnt = 0; | |||
| 617 | e->sem_sig_val_dst_cnt = 0; | |||
| 618 | } | |||
| 619 | ||||
| 620 | int ff_vk_exec_add_dep_buf(FFVulkanContext *s, FFVkExecContext *e, | |||
| 621 | AVBufferRef **deps, int nb_deps, int ref) | |||
| 622 | { | |||
| 623 | AVBufferRef **dst = av_fast_realloc(e->buf_deps, &e->buf_deps_alloc_size, | |||
| 624 | (e->nb_buf_deps + nb_deps) * sizeof(*dst)); | |||
| 625 | if (!dst) { | |||
| 626 | ff_vk_exec_discard_deps(s, e); | |||
| 627 | return AVERROR(ENOMEM)(-(12)); | |||
| 628 | } | |||
| 629 | ||||
| 630 | e->buf_deps = dst; | |||
| 631 | ||||
| 632 | for (int i = 0; i < nb_deps; i++) { | |||
| 633 | if (!deps[i]) | |||
| 634 | continue; | |||
| 635 | ||||
| 636 | e->buf_deps[e->nb_buf_deps] = ref ? av_buffer_ref(deps[i]) : deps[i]; | |||
| 637 | if (!e->buf_deps[e->nb_buf_deps]) { | |||
| 638 | ff_vk_exec_discard_deps(s, e); | |||
| 639 | return AVERROR(ENOMEM)(-(12)); | |||
| 640 | } | |||
| 641 | e->nb_buf_deps++; | |||
| 642 | } | |||
| 643 | ||||
| 644 | return 0; | |||
| 645 | } | |||
| 646 | ||||
| 647 | int ff_vk_exec_add_dep_sw_frame(FFVulkanContext *s, FFVkExecContext *e, | |||
| 648 | AVFrame *f) | |||
| 649 | { | |||
| 650 | AVFrame **dst = av_fast_realloc(e->sw_frame_deps, &e->sw_frame_deps_alloc_size, | |||
| 651 | (e->nb_sw_frame_deps + 1) * sizeof(*dst)); | |||
| 652 | if (!dst) { | |||
| 653 | ff_vk_exec_discard_deps(s, e); | |||
| 654 | return AVERROR(ENOMEM)(-(12)); | |||
| 655 | } | |||
| 656 | ||||
| 657 | e->sw_frame_deps = dst; | |||
| 658 | ||||
| 659 | e->sw_frame_deps[e->nb_sw_frame_deps] = av_frame_clone(f); | |||
| 660 | if (!e->sw_frame_deps[e->nb_sw_frame_deps]) { | |||
| 661 | ff_vk_exec_discard_deps(s, e); | |||
| 662 | return AVERROR(ENOMEM)(-(12)); | |||
| 663 | } | |||
| 664 | ||||
| 665 | e->nb_sw_frame_deps++; | |||
| 666 | ||||
| 667 | return 0; | |||
| 668 | } | |||
| 669 | ||||
| 670 | #define ARR_REALLOC(str, arr, alloc_s, cnt)do { arr = av_fast_realloc(str->arr, alloc_s, (cnt + 1)*sizeof (*arr)); if (!arr) { ff_vk_exec_discard_deps(s, e); return (- (12)); } str->arr = arr; } while (0) \ | |||
| 671 | do { \ | |||
| 672 | arr = av_fast_realloc(str->arr, alloc_s, (cnt + 1)*sizeof(*arr)); \ | |||
| 673 | if (!arr) { \ | |||
| 674 | ff_vk_exec_discard_deps(s, e); \ | |||
| 675 | return AVERROR(ENOMEM)(-(12)); \ | |||
| 676 | } \ | |||
| 677 | str->arr = arr; \ | |||
| 678 | } while (0) | |||
| 679 | ||||
| 680 | typedef struct TempSyncCtx { | |||
| 681 | int nb_sem; | |||
| 682 | VkSemaphore sem[]; | |||
| 683 | } TempSyncCtx; | |||
| 684 | ||||
| 685 | static void destroy_tmp_semaphores(void *opaque, uint8_t *data) | |||
| 686 | { | |||
| 687 | FFVulkanContext *s = opaque; | |||
| 688 | FFVulkanFunctions *vk = &s->vkfn; | |||
| 689 | TempSyncCtx *ts = (TempSyncCtx *)data; | |||
| 690 | ||||
| 691 | for (int i = 0; i < ts->nb_sem; i++) | |||
| 692 | vk->DestroySemaphore(s->hwctx->act_dev, ts->sem[i], s->hwctx->alloc); | |||
| 693 | ||||
| 694 | av_free(ts); | |||
| 695 | } | |||
| 696 | ||||
| 697 | int ff_vk_exec_add_dep_wait_sem(FFVulkanContext *s, FFVkExecContext *e, | |||
| 698 | VkSemaphore sem, uint64_t val, | |||
| 699 | VkPipelineStageFlagBits2 stage) | |||
| 700 | { | |||
| 701 | VkSemaphoreSubmitInfo *sem_wait; | |||
| 702 | ARR_REALLOC(e, sem_wait, &e->sem_wait_alloc, e->sem_wait_cnt)do { sem_wait = av_fast_realloc(e->sem_wait, &e->sem_wait_alloc , (e->sem_wait_cnt + 1)*sizeof(*sem_wait)); if (!sem_wait) { ff_vk_exec_discard_deps(s, e); return (-(12)); } e->sem_wait = sem_wait; } while (0); | |||
| 703 | ||||
| 704 | e->sem_wait[e->sem_wait_cnt++] = (VkSemaphoreSubmitInfo) { | |||
| 705 | .sType = VK_STRUCTURE_TYPE_SEMAPHORE_SUBMIT_INFO, | |||
| 706 | .semaphore = sem, | |||
| 707 | .value = val, | |||
| 708 | .stageMask = stage, | |||
| 709 | }; | |||
| 710 | ||||
| 711 | return 0; | |||
| 712 | } | |||
| 713 | ||||
| 714 | int ff_vk_exec_add_dep_bool_sem(FFVulkanContext *s, FFVkExecContext *e, | |||
| 715 | VkSemaphore *sem, int nb, | |||
| 716 | VkPipelineStageFlagBits2 stage, | |||
| 717 | int wait) | |||
| 718 | { | |||
| 719 | int err; | |||
| 720 | size_t buf_size; | |||
| 721 | AVBufferRef *buf; | |||
| 722 | TempSyncCtx *ts; | |||
| 723 | FFVulkanFunctions *vk = &s->vkfn; | |||
| 724 | ||||
| 725 | /* Do not transfer ownership if we're signalling a binary semaphore, | |||
| 726 | * since we're probably exporting it. */ | |||
| 727 | if (!wait) { | |||
| 728 | for (int i = 0; i < nb; i++) { | |||
| 729 | VkSemaphoreSubmitInfo *sem_sig; | |||
| 730 | ARR_REALLOC(e, sem_sig, &e->sem_sig_alloc, e->sem_sig_cnt)do { sem_sig = av_fast_realloc(e->sem_sig, &e->sem_sig_alloc , (e->sem_sig_cnt + 1)*sizeof(*sem_sig)); if (!sem_sig) { ff_vk_exec_discard_deps (s, e); return (-(12)); } e->sem_sig = sem_sig; } while (0 ); | |||
| 731 | ||||
| 732 | e->sem_sig[e->sem_sig_cnt++] = (VkSemaphoreSubmitInfo) { | |||
| 733 | .sType = VK_STRUCTURE_TYPE_SEMAPHORE_SUBMIT_INFO, | |||
| 734 | .semaphore = sem[i], | |||
| 735 | .stageMask = stage, | |||
| 736 | }; | |||
| 737 | } | |||
| 738 | ||||
| 739 | return 0; | |||
| 740 | } | |||
| 741 | ||||
| 742 | buf_size = sizeof(*ts) + sizeof(VkSemaphore)*nb; | |||
| 743 | ts = av_mallocz(buf_size); | |||
| 744 | if (!ts) { | |||
| 745 | err = AVERROR(ENOMEM)(-(12)); | |||
| 746 | goto fail; | |||
| 747 | } | |||
| 748 | ||||
| 749 | memcpy(ts->sem, sem, nb*sizeof(*sem)); | |||
| 750 | ts->nb_sem = nb; | |||
| 751 | ||||
| 752 | buf = av_buffer_create((uint8_t *)ts, buf_size, destroy_tmp_semaphores, s, 0); | |||
| 753 | if (!buf) { | |||
| 754 | av_free(ts); | |||
| 755 | err = AVERROR(ENOMEM)(-(12)); | |||
| 756 | goto fail; | |||
| 757 | } | |||
| 758 | ||||
| 759 | err = ff_vk_exec_add_dep_buf(s, e, &buf, 1, 0); | |||
| 760 | if (err < 0) { | |||
| 761 | av_buffer_unref(&buf); | |||
| 762 | return err; | |||
| 763 | } | |||
| 764 | ||||
| 765 | for (int i = 0; i < nb; i++) { | |||
| 766 | err = ff_vk_exec_add_dep_wait_sem(s, e, sem[i], 0, stage); | |||
| 767 | if (err < 0) | |||
| 768 | return err; | |||
| 769 | } | |||
| 770 | ||||
| 771 | return 0; | |||
| 772 | ||||
| 773 | fail: | |||
| 774 | for (int i = 0; i < nb; i++) | |||
| 775 | vk->DestroySemaphore(s->hwctx->act_dev, sem[i], s->hwctx->alloc); | |||
| 776 | ||||
| 777 | return err; | |||
| 778 | } | |||
| 779 | ||||
| 780 | int ff_vk_exec_add_dep_frame(FFVulkanContext *s, FFVkExecContext *e, AVFrame *f, | |||
| 781 | VkPipelineStageFlagBits2 wait_stage, | |||
| 782 | VkPipelineStageFlagBits2 signal_stage) | |||
| 783 | { | |||
| 784 | uint8_t *frame_locked; | |||
| 785 | uint8_t *frame_update; | |||
| 786 | AVFrame **frame_deps; | |||
| 787 | AVBufferRef **buf_deps; | |||
| 788 | VkImageLayout *layout_dst; | |||
| 789 | uint32_t *queue_family_dst; | |||
| 790 | VkAccessFlagBits *access_dst; | |||
| 791 | ||||
| 792 | AVHWFramesContext *hwfc = (AVHWFramesContext *)f->hw_frames_ctx->data; | |||
| 793 | AVVulkanFramesContext *vkfc = hwfc->hwctx; | |||
| 794 | AVVkFrame *vkf = (AVVkFrame *)f->data[0]; | |||
| 795 | int nb_images = ff_vk_count_images(vkf); | |||
| 796 | ||||
| 797 | /* Don't add duplicates */ | |||
| 798 | for (int i = 0; i < e->nb_frame_deps; i++) | |||
| 799 | if (e->frame_deps[i]->data[0] == f->data[0]) | |||
| 800 | return 1; | |||
| 801 | ||||
| 802 | ARR_REALLOC(e, layout_dst, &e->layout_dst_alloc, e->nb_frame_deps)do { layout_dst = av_fast_realloc(e->layout_dst, &e-> layout_dst_alloc, (e->nb_frame_deps + 1)*sizeof(*layout_dst )); if (!layout_dst) { ff_vk_exec_discard_deps(s, e); return ( -(12)); } e->layout_dst = layout_dst; } while (0); | |||
| 803 | ARR_REALLOC(e, queue_family_dst, &e->queue_family_dst_alloc, e->nb_frame_deps)do { queue_family_dst = av_fast_realloc(e->queue_family_dst , &e->queue_family_dst_alloc, (e->nb_frame_deps + 1 )*sizeof(*queue_family_dst)); if (!queue_family_dst) { ff_vk_exec_discard_deps (s, e); return (-(12)); } e->queue_family_dst = queue_family_dst ; } while (0); | |||
| 804 | ARR_REALLOC(e, access_dst, &e->access_dst_alloc, e->nb_frame_deps)do { access_dst = av_fast_realloc(e->access_dst, &e-> access_dst_alloc, (e->nb_frame_deps + 1)*sizeof(*access_dst )); if (!access_dst) { ff_vk_exec_discard_deps(s, e); return ( -(12)); } e->access_dst = access_dst; } while (0); | |||
| 805 | ||||
| 806 | ARR_REALLOC(e, frame_locked, &e->frame_locked_alloc_size, e->nb_frame_deps)do { frame_locked = av_fast_realloc(e->frame_locked, & e->frame_locked_alloc_size, (e->nb_frame_deps + 1)*sizeof (*frame_locked)); if (!frame_locked) { ff_vk_exec_discard_deps (s, e); return (-(12)); } e->frame_locked = frame_locked; } while (0); | |||
| 807 | ARR_REALLOC(e, frame_update, &e->frame_update_alloc_size, e->nb_frame_deps)do { frame_update = av_fast_realloc(e->frame_update, & e->frame_update_alloc_size, (e->nb_frame_deps + 1)*sizeof (*frame_update)); if (!frame_update) { ff_vk_exec_discard_deps (s, e); return (-(12)); } e->frame_update = frame_update; } while (0); | |||
| 808 | ARR_REALLOC(e, frame_deps, &e->frame_deps_alloc_size, e->nb_frame_deps)do { frame_deps = av_fast_realloc(e->frame_deps, &e-> frame_deps_alloc_size, (e->nb_frame_deps + 1)*sizeof(*frame_deps )); if (!frame_deps) { ff_vk_exec_discard_deps(s, e); return ( -(12)); } e->frame_deps = frame_deps; } while (0); | |||
| 809 | ||||
| 810 | /* prepare_frame in hwcontext_vulkan.c uses the regular frame management | |||
| 811 | * code but has no frame yet, and it doesn't need to actually store a ref | |||
| 812 | * to the frame. */ | |||
| 813 | if (f->buf[0]) { | |||
| 814 | ARR_REALLOC(e, buf_deps, &e->buf_deps_alloc_size, e->nb_buf_deps)do { buf_deps = av_fast_realloc(e->buf_deps, &e->buf_deps_alloc_size , (e->nb_buf_deps + 1)*sizeof(*buf_deps)); if (!buf_deps) { ff_vk_exec_discard_deps(s, e); return (-(12)); } e->buf_deps = buf_deps; } while (0); | |||
| 815 | e->buf_deps[e->nb_buf_deps] = av_buffer_ref(f->buf[0]); | |||
| 816 | if (!e->buf_deps[e->nb_buf_deps]) { | |||
| 817 | ff_vk_exec_discard_deps(s, e); | |||
| 818 | return AVERROR(ENOMEM)(-(12)); | |||
| 819 | } | |||
| 820 | e->nb_buf_deps++; | |||
| 821 | } | |||
| 822 | ||||
| 823 | e->frame_deps[e->nb_frame_deps] = f; | |||
| 824 | ||||
| 825 | vkfc->lock_frame(hwfc, vkf); | |||
| 826 | e->frame_locked[e->nb_frame_deps] = 1; | |||
| 827 | e->frame_update[e->nb_frame_deps] = 0; | |||
| 828 | e->nb_frame_deps++; | |||
| 829 | ||||
| 830 | for (int i = 0; i < nb_images; i++) { | |||
| 831 | VkSemaphoreSubmitInfo *sem_wait; | |||
| 832 | VkSemaphoreSubmitInfo *sem_sig; | |||
| 833 | uint64_t **sem_sig_val_dst; | |||
| 834 | ||||
| 835 | ARR_REALLOC(e, sem_wait, &e->sem_wait_alloc, e->sem_wait_cnt)do { sem_wait = av_fast_realloc(e->sem_wait, &e->sem_wait_alloc , (e->sem_wait_cnt + 1)*sizeof(*sem_wait)); if (!sem_wait) { ff_vk_exec_discard_deps(s, e); return (-(12)); } e->sem_wait = sem_wait; } while (0); | |||
| 836 | ARR_REALLOC(e, sem_sig, &e->sem_sig_alloc, e->sem_sig_cnt)do { sem_sig = av_fast_realloc(e->sem_sig, &e->sem_sig_alloc , (e->sem_sig_cnt + 1)*sizeof(*sem_sig)); if (!sem_sig) { ff_vk_exec_discard_deps (s, e); return (-(12)); } e->sem_sig = sem_sig; } while (0 ); | |||
| 837 | ARR_REALLOC(e, sem_sig_val_dst, &e->sem_sig_val_dst_alloc, e->sem_sig_val_dst_cnt)do { sem_sig_val_dst = av_fast_realloc(e->sem_sig_val_dst, &e->sem_sig_val_dst_alloc, (e->sem_sig_val_dst_cnt + 1)*sizeof(*sem_sig_val_dst)); if (!sem_sig_val_dst) { ff_vk_exec_discard_deps (s, e); return (-(12)); } e->sem_sig_val_dst = sem_sig_val_dst ; } while (0); | |||
| 838 | ||||
| 839 | e->sem_wait[e->sem_wait_cnt++] = (VkSemaphoreSubmitInfo) { | |||
| 840 | .sType = VK_STRUCTURE_TYPE_SEMAPHORE_SUBMIT_INFO, | |||
| 841 | .semaphore = vkf->sem[i], | |||
| 842 | .value = vkf->sem_value[i], | |||
| 843 | .stageMask = wait_stage, | |||
| 844 | }; | |||
| 845 | ||||
| 846 | e->sem_sig[e->sem_sig_cnt++] = (VkSemaphoreSubmitInfo) { | |||
| 847 | .sType = VK_STRUCTURE_TYPE_SEMAPHORE_SUBMIT_INFO, | |||
| 848 | .semaphore = vkf->sem[i], | |||
| 849 | .value = vkf->sem_value[i] + 1, | |||
| 850 | .stageMask = signal_stage, | |||
| 851 | }; | |||
| 852 | ||||
| 853 | e->sem_sig_val_dst[e->sem_sig_val_dst_cnt] = &vkf->sem_value[i]; | |||
| 854 | e->sem_sig_val_dst_cnt++; | |||
| 855 | } | |||
| 856 | ||||
| 857 | return 0; | |||
| 858 | } | |||
| 859 | ||||
| 860 | void ff_vk_exec_update_frame(FFVulkanContext *s, FFVkExecContext *e, AVFrame *f, | |||
| 861 | VkImageMemoryBarrier2 *bar, uint32_t *nb_img_bar) | |||
| 862 | { | |||
| 863 | int i; | |||
| 864 | for (i = 0; i < e->nb_frame_deps; i++) | |||
| 865 | if (e->frame_deps[i]->data[0] == f->data[0]) | |||
| 866 | break; | |||
| 867 | av_assert0(i < e->nb_frame_deps)do { if (!(i < e->nb_frame_deps)) { av_log(((void*)0), 0 , "Assertion %s failed at %s:%d\n", "i < e->nb_frame_deps" , "/root/firefox-clang/media/ffvpx/libavutil/vulkan.c", 867); abort(); } } while (0); | |||
| 868 | ||||
| 869 | /* Don't update duplicates */ | |||
| 870 | if (nb_img_bar && !e->frame_update[i]) | |||
| 871 | (*nb_img_bar)++; | |||
| 872 | ||||
| 873 | e->queue_family_dst[i] = bar->dstQueueFamilyIndex; | |||
| 874 | e->access_dst[i] = bar->dstAccessMask; | |||
| 875 | e->layout_dst[i] = bar->newLayout; | |||
| 876 | e->frame_update[i] = 1; | |||
| 877 | } | |||
| 878 | ||||
| 879 | int ff_vk_exec_mirror_sem_value(FFVulkanContext *s, FFVkExecContext *e, | |||
| 880 | VkSemaphore *dst, uint64_t *dst_val, | |||
| 881 | AVFrame *f) | |||
| 882 | { | |||
| 883 | uint64_t **sem_sig_val_dst; | |||
| 884 | AVVkFrame *vkf = (AVVkFrame *)f->data[0]; | |||
| 885 | ||||
| 886 | /* Reject unknown frames */ | |||
| 887 | int i; | |||
| 888 | for (i = 0; i < e->nb_frame_deps; i++) | |||
| 889 | if (e->frame_deps[i]->data[0] == f->data[0]) | |||
| 890 | break; | |||
| 891 | if (i == e->nb_frame_deps) | |||
| 892 | return AVERROR(EINVAL)(-(22)); | |||
| 893 | ||||
| 894 | ARR_REALLOC(e, sem_sig_val_dst, &e->sem_sig_val_dst_alloc, e->sem_sig_val_dst_cnt)do { sem_sig_val_dst = av_fast_realloc(e->sem_sig_val_dst, &e->sem_sig_val_dst_alloc, (e->sem_sig_val_dst_cnt + 1)*sizeof(*sem_sig_val_dst)); if (!sem_sig_val_dst) { ff_vk_exec_discard_deps (s, e); return (-(12)); } e->sem_sig_val_dst = sem_sig_val_dst ; } while (0); | |||
| 895 | ||||
| 896 | *dst = vkf->sem[0]; | |||
| 897 | *dst_val = vkf->sem_value[0]; | |||
| 898 | ||||
| 899 | e->sem_sig_val_dst[e->sem_sig_val_dst_cnt] = dst_val; | |||
| 900 | e->sem_sig_val_dst_cnt++; | |||
| 901 | ||||
| 902 | return 0; | |||
| 903 | } | |||
| 904 | ||||
| 905 | int ff_vk_exec_submit(FFVulkanContext *s, FFVkExecContext *e) | |||
| 906 | { | |||
| 907 | VkResult ret; | |||
| 908 | FFVulkanFunctions *vk = &s->vkfn; | |||
| 909 | VkCommandBufferSubmitInfo cmd_buf_info = (VkCommandBufferSubmitInfo) { | |||
| 910 | .sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_SUBMIT_INFO, | |||
| 911 | .commandBuffer = e->buf, | |||
| 912 | }; | |||
| 913 | VkSubmitInfo2 submit_info = (VkSubmitInfo2) { | |||
| 914 | .sType = VK_STRUCTURE_TYPE_SUBMIT_INFO_2, | |||
| 915 | .pCommandBufferInfos = &cmd_buf_info, | |||
| 916 | .commandBufferInfoCount = 1, | |||
| 917 | .pWaitSemaphoreInfos = e->sem_wait, | |||
| 918 | .waitSemaphoreInfoCount = e->sem_wait_cnt, | |||
| 919 | .pSignalSemaphoreInfos = e->sem_sig, | |||
| 920 | .signalSemaphoreInfoCount = e->sem_sig_cnt, | |||
| 921 | }; | |||
| 922 | ||||
| 923 | ret = vk->EndCommandBuffer(e->buf); | |||
| 924 | if (ret != VK_SUCCESS) { | |||
| 925 | av_log(s, AV_LOG_ERROR16, "Unable to finish command buffer: %s\n", | |||
| 926 | ff_vk_ret2str(ret)); | |||
| 927 | ff_vk_exec_discard_deps(s, e); | |||
| 928 | return AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); | |||
| 929 | } | |||
| 930 | ||||
| 931 | #if FF_API_VULKAN_SYNC_QUEUES(60 < 62) | |||
| 932 | FF_DISABLE_DEPRECATION_WARNINGSGCC diagnostic push
GCC diagnostic ignored "-Wdeprecated-declarations" | |||
| 933 | s->hwctx->lock_queue(s->device, e->qf, e->qi); | |||
| 934 | FF_ENABLE_DEPRECATION_WARNINGSGCC diagnostic pop | |||
| 935 | #endif | |||
| 936 | ret = vk->QueueSubmit2(e->queue, 1, &submit_info, e->fence); | |||
| 937 | #if FF_API_VULKAN_SYNC_QUEUES(60 < 62) | |||
| 938 | FF_DISABLE_DEPRECATION_WARNINGSGCC diagnostic push
GCC diagnostic ignored "-Wdeprecated-declarations" | |||
| 939 | s->hwctx->unlock_queue(s->device, e->qf, e->qi); | |||
| 940 | FF_ENABLE_DEPRECATION_WARNINGSGCC diagnostic pop | |||
| 941 | #endif | |||
| 942 | ||||
| 943 | if (ret != VK_SUCCESS) { | |||
| 944 | av_log(s, AV_LOG_ERROR16, "Unable to submit command buffer: %s\n", | |||
| 945 | ff_vk_ret2str(ret)); | |||
| 946 | ff_vk_exec_discard_deps(s, e); | |||
| 947 | return AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); | |||
| 948 | } | |||
| 949 | ||||
| 950 | for (int i = 0; i < e->sem_sig_val_dst_cnt; i++) | |||
| 951 | *e->sem_sig_val_dst[i] += 1; | |||
| 952 | ||||
| 953 | /* Unlock all frames */ | |||
| 954 | for (int j = 0; j < e->nb_frame_deps; j++) { | |||
| 955 | if (e->frame_locked[j]) { | |||
| 956 | AVFrame *f = e->frame_deps[j]; | |||
| 957 | AVHWFramesContext *hwfc = (AVHWFramesContext *)f->hw_frames_ctx->data; | |||
| 958 | AVVulkanFramesContext *vkfc = hwfc->hwctx; | |||
| 959 | AVVkFrame *vkf = (AVVkFrame *)f->data[0]; | |||
| 960 | ||||
| 961 | if (e->frame_update[j]) { | |||
| 962 | int nb_images = ff_vk_count_images(vkf); | |||
| 963 | for (int i = 0; i < nb_images; i++) { | |||
| 964 | vkf->layout[i] = e->layout_dst[j]; | |||
| 965 | vkf->access[i] = e->access_dst[j]; | |||
| 966 | vkf->queue_family[i] = e->queue_family_dst[j]; | |||
| 967 | } | |||
| 968 | } | |||
| 969 | vkfc->unlock_frame(hwfc, vkf); | |||
| 970 | e->frame_locked[j] = 0; | |||
| 971 | } | |||
| 972 | } | |||
| 973 | ||||
| 974 | e->had_submission = 1; | |||
| 975 | ||||
| 976 | return 0; | |||
| 977 | } | |||
| 978 | ||||
| 979 | int ff_vk_alloc_mem(FFVulkanContext *s, VkMemoryRequirements *req, | |||
| 980 | VkMemoryPropertyFlagBits req_flags, void *alloc_extension, | |||
| 981 | VkMemoryPropertyFlagBits *mem_flags, VkDeviceMemory *mem) | |||
| 982 | { | |||
| 983 | VkResult ret; | |||
| 984 | int index = -1; | |||
| 985 | FFVulkanFunctions *vk = &s->vkfn; | |||
| 986 | ||||
| 987 | VkMemoryAllocateInfo alloc_info = { | |||
| 988 | .sType = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO, | |||
| 989 | .pNext = alloc_extension, | |||
| 990 | }; | |||
| 991 | ||||
| 992 | alloc_info.allocationSize = req->size; | |||
| 993 | ||||
| 994 | /* The vulkan spec requires memory types to be sorted in the "optimal" | |||
| 995 | * order, so the first matching type we find will be the best/fastest one */ | |||
| 996 | for (int i = 0; i < s->mprops.memoryTypeCount; i++) { | |||
| 997 | /* The memory type must be supported by the requirements (bitfield) */ | |||
| 998 | if (!(req->memoryTypeBits & (1 << i))) | |||
| 999 | continue; | |||
| 1000 | ||||
| 1001 | /* The memory type flags must include our properties */ | |||
| 1002 | if ((req_flags != UINT32_MAX(4294967295U)) && | |||
| 1003 | ((s->mprops.memoryTypes[i].propertyFlags & req_flags) != req_flags)) | |||
| 1004 | continue; | |||
| 1005 | ||||
| 1006 | /* Found a suitable memory type */ | |||
| 1007 | index = i; | |||
| 1008 | break; | |||
| 1009 | } | |||
| 1010 | ||||
| 1011 | if (index < 0) { | |||
| 1012 | av_log(s, AV_LOG_ERROR16, "No memory type found for flags 0x%x\n", | |||
| 1013 | req_flags); | |||
| 1014 | return AVERROR(EINVAL)(-(22)); | |||
| 1015 | } | |||
| 1016 | ||||
| 1017 | alloc_info.memoryTypeIndex = index; | |||
| 1018 | ||||
| 1019 | ret = vk->AllocateMemory(s->hwctx->act_dev, &alloc_info, | |||
| 1020 | s->hwctx->alloc, mem); | |||
| 1021 | if (ret != VK_SUCCESS) | |||
| 1022 | return AVERROR(ENOMEM)(-(12)); | |||
| 1023 | ||||
| 1024 | if (mem_flags) | |||
| 1025 | *mem_flags |= s->mprops.memoryTypes[index].propertyFlags; | |||
| 1026 | ||||
| 1027 | return 0; | |||
| 1028 | } | |||
| 1029 | ||||
| 1030 | int ff_vk_create_buf(FFVulkanContext *s, FFVkBuffer *buf, size_t size, | |||
| 1031 | void *pNext, void *alloc_pNext, | |||
| 1032 | VkBufferUsageFlags usage, VkMemoryPropertyFlagBits flags) | |||
| 1033 | { | |||
| 1034 | int err; | |||
| 1035 | VkResult ret; | |||
| 1036 | int use_ded_mem; | |||
| 1037 | FFVulkanFunctions *vk = &s->vkfn; | |||
| 1038 | ||||
| 1039 | VkBufferCreateInfo buf_spawn = { | |||
| 1040 | .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO, | |||
| 1041 | .pNext = pNext, | |||
| 1042 | .usage = usage, | |||
| 1043 | .sharingMode = VK_SHARING_MODE_EXCLUSIVE, | |||
| 1044 | .size = flags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT ? | |||
| 1045 | FFALIGN(size, s->props.properties.limits.minMemoryMapAlignment)(((size)+(s->props.properties.limits.minMemoryMapAlignment )-1)&~((s->props.properties.limits.minMemoryMapAlignment )-1)) : | |||
| 1046 | size, | |||
| 1047 | }; | |||
| 1048 | ||||
| 1049 | VkMemoryAllocateFlagsInfo alloc_flags = { | |||
| 1050 | .sType = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_FLAGS_INFO, | |||
| 1051 | .flags = VK_MEMORY_ALLOCATE_DEVICE_ADDRESS_BIT, | |||
| 1052 | }; | |||
| 1053 | VkBufferMemoryRequirementsInfo2 req_desc = { | |||
| 1054 | .sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_REQUIREMENTS_INFO_2, | |||
| 1055 | }; | |||
| 1056 | VkMemoryDedicatedAllocateInfo ded_alloc = { | |||
| 1057 | .sType = VK_STRUCTURE_TYPE_MEMORY_DEDICATED_ALLOCATE_INFO, | |||
| 1058 | .pNext = alloc_pNext, | |||
| 1059 | }; | |||
| 1060 | VkMemoryDedicatedRequirements ded_req = { | |||
| 1061 | .sType = VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS, | |||
| 1062 | }; | |||
| 1063 | VkMemoryRequirements2 req = { | |||
| 1064 | .sType = VK_STRUCTURE_TYPE_MEMORY_REQUIREMENTS_2, | |||
| 1065 | .pNext = &ded_req, | |||
| 1066 | }; | |||
| 1067 | ||||
| 1068 | av_log(s, AV_LOG_DEBUG48, "Creating a buffer of %zu bytes, " | |||
| 1069 | "usage: 0x%x, flags: 0x%x\n", | |||
| 1070 | size, usage, flags); | |||
| 1071 | ||||
| 1072 | ret = vk->CreateBuffer(s->hwctx->act_dev, &buf_spawn, s->hwctx->alloc, &buf->buf); | |||
| 1073 | if (ret != VK_SUCCESS) { | |||
| 1074 | av_log(s, AV_LOG_ERROR16, "Failed to create buffer: %s\n", | |||
| 1075 | ff_vk_ret2str(ret)); | |||
| 1076 | return AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); | |||
| 1077 | } | |||
| 1078 | ||||
| 1079 | req_desc.buffer = buf->buf; | |||
| 1080 | ||||
| 1081 | vk->GetBufferMemoryRequirements2(s->hwctx->act_dev, &req_desc, &req); | |||
| 1082 | ||||
| 1083 | /* In case the implementation prefers/requires dedicated allocation */ | |||
| 1084 | use_ded_mem = ded_req.prefersDedicatedAllocation | | |||
| 1085 | ded_req.requiresDedicatedAllocation; | |||
| 1086 | if (use_ded_mem) { | |||
| 1087 | ded_alloc.buffer = buf->buf; | |||
| 1088 | ded_alloc.pNext = alloc_pNext; | |||
| 1089 | alloc_pNext = &ded_alloc; | |||
| 1090 | } | |||
| 1091 | ||||
| 1092 | if (usage & VK_BUFFER_USAGE_SHADER_DEVICE_ADDRESS_BIT) { | |||
| 1093 | alloc_flags.pNext = alloc_pNext; | |||
| 1094 | alloc_pNext = &alloc_flags; | |||
| 1095 | } | |||
| 1096 | ||||
| 1097 | err = ff_vk_alloc_mem(s, &req.memoryRequirements, flags, alloc_pNext, | |||
| 1098 | &buf->flags, &buf->mem); | |||
| 1099 | if (err) | |||
| 1100 | return err; | |||
| 1101 | ||||
| 1102 | ret = vk->BindBufferMemory(s->hwctx->act_dev, buf->buf, buf->mem, 0); | |||
| 1103 | if (ret != VK_SUCCESS) { | |||
| 1104 | av_log(s, AV_LOG_ERROR16, "Failed to bind memory to buffer: %s\n", | |||
| 1105 | ff_vk_ret2str(ret)); | |||
| 1106 | return AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); | |||
| 1107 | } | |||
| 1108 | ||||
| 1109 | if (usage & VK_BUFFER_USAGE_SHADER_DEVICE_ADDRESS_BIT) { | |||
| 1110 | VkBufferDeviceAddressInfo address_info = { | |||
| 1111 | .sType = VK_STRUCTURE_TYPE_BUFFER_DEVICE_ADDRESS_INFO, | |||
| 1112 | .buffer = buf->buf, | |||
| 1113 | }; | |||
| 1114 | buf->address = vk->GetBufferDeviceAddress(s->hwctx->act_dev, &address_info); | |||
| 1115 | } | |||
| 1116 | ||||
| 1117 | buf->size = size; | |||
| 1118 | ||||
| 1119 | return 0; | |||
| 1120 | } | |||
| 1121 | ||||
| 1122 | int ff_vk_map_buffers(FFVulkanContext *s, FFVkBuffer **buf, uint8_t *mem[], | |||
| 1123 | int nb_buffers, int invalidate) | |||
| 1124 | { | |||
| 1125 | VkResult ret; | |||
| 1126 | FFVulkanFunctions *vk = &s->vkfn; | |||
| 1127 | VkMappedMemoryRange inval_list[64]; | |||
| 1128 | int inval_count = 0; | |||
| 1129 | ||||
| 1130 | for (int i = 0; i < nb_buffers; i++) { | |||
| 1131 | void *dst; | |||
| 1132 | ret = vk->MapMemory(s->hwctx->act_dev, buf[i]->mem, 0, | |||
| 1133 | VK_WHOLE_SIZE(~0ULL), 0, &dst); | |||
| 1134 | if (ret != VK_SUCCESS) { | |||
| 1135 | av_log(s, AV_LOG_ERROR16, "Failed to map buffer memory: %s\n", | |||
| 1136 | ff_vk_ret2str(ret)); | |||
| 1137 | return AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); | |||
| 1138 | } | |||
| 1139 | buf[i]->mapped_mem = dst; | |||
| 1140 | if (mem) | |||
| 1141 | mem[i] = dst; | |||
| 1142 | } | |||
| 1143 | ||||
| 1144 | if (!invalidate) | |||
| 1145 | return 0; | |||
| 1146 | ||||
| 1147 | for (int i = 0; i < nb_buffers; i++) { | |||
| 1148 | const VkMappedMemoryRange ival_buf = { | |||
| 1149 | .sType = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE, | |||
| 1150 | .memory = buf[i]->mem, | |||
| 1151 | .size = VK_WHOLE_SIZE(~0ULL), | |||
| 1152 | }; | |||
| 1153 | if (buf[i]->flags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT) | |||
| 1154 | continue; | |||
| 1155 | inval_list[inval_count++] = ival_buf; | |||
| 1156 | } | |||
| 1157 | ||||
| 1158 | if (inval_count) { | |||
| 1159 | ret = vk->InvalidateMappedMemoryRanges(s->hwctx->act_dev, inval_count, | |||
| 1160 | inval_list); | |||
| 1161 | if (ret != VK_SUCCESS) { | |||
| 1162 | av_log(s, AV_LOG_ERROR16, "Failed to invalidate memory: %s\n", | |||
| 1163 | ff_vk_ret2str(ret)); | |||
| 1164 | return AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); | |||
| 1165 | } | |||
| 1166 | } | |||
| 1167 | ||||
| 1168 | return 0; | |||
| 1169 | } | |||
| 1170 | ||||
| 1171 | int ff_vk_flush_buffer(FFVulkanContext *s, FFVkBuffer *buf, | |||
| 1172 | VkDeviceSize offset, VkDeviceSize mem_size, | |||
| 1173 | int flush) | |||
| 1174 | { | |||
| 1175 | VkResult ret; | |||
| 1176 | FFVulkanFunctions *vk = &s->vkfn; | |||
| 1177 | ||||
| 1178 | if (buf->host_ref || buf->flags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT) | |||
| 1179 | return 0; | |||
| 1180 | ||||
| 1181 | const VkMappedMemoryRange flush_data = { | |||
| 1182 | .sType = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE, | |||
| 1183 | .memory = buf->mem, | |||
| 1184 | .offset = offset, | |||
| 1185 | .size = mem_size, | |||
| 1186 | }; | |||
| 1187 | ||||
| 1188 | if (flush) | |||
| 1189 | ret = vk->FlushMappedMemoryRanges(s->hwctx->act_dev, 1, &flush_data); | |||
| 1190 | else | |||
| 1191 | ret = vk->InvalidateMappedMemoryRanges(s->hwctx->act_dev, 1, &flush_data); | |||
| 1192 | ||||
| 1193 | if (ret != VK_SUCCESS) { | |||
| 1194 | av_log(s, AV_LOG_ERROR16, "Failed to flush memory: %s\n", | |||
| 1195 | ff_vk_ret2str(ret)); | |||
| 1196 | return AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); | |||
| 1197 | } | |||
| 1198 | ||||
| 1199 | return 0; | |||
| 1200 | } | |||
| 1201 | ||||
| 1202 | int ff_vk_unmap_buffers(FFVulkanContext *s, FFVkBuffer **buf, int nb_buffers, | |||
| 1203 | int flush) | |||
| 1204 | { | |||
| 1205 | int err = 0; | |||
| 1206 | VkResult ret; | |||
| 1207 | FFVulkanFunctions *vk = &s->vkfn; | |||
| 1208 | VkMappedMemoryRange flush_list[64]; | |||
| 1209 | int flush_count = 0; | |||
| 1210 | ||||
| 1211 | if (flush) { | |||
| 1212 | for (int i = 0; i < nb_buffers; i++) { | |||
| 1213 | const VkMappedMemoryRange flush_buf = { | |||
| 1214 | .sType = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE, | |||
| 1215 | .memory = buf[i]->mem, | |||
| 1216 | .size = VK_WHOLE_SIZE(~0ULL), | |||
| 1217 | }; | |||
| 1218 | ||||
| 1219 | av_assert0(!buf[i]->host_ref)do { if (!(!buf[i]->host_ref)) { av_log(((void*)0), 0, "Assertion %s failed at %s:%d\n" , "!buf[i]->host_ref", "/root/firefox-clang/media/ffvpx/libavutil/vulkan.c" , 1219); abort(); } } while (0); | |||
| 1220 | if (buf[i]->flags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT) | |||
| 1221 | continue; | |||
| 1222 | flush_list[flush_count++] = flush_buf; | |||
| 1223 | } | |||
| 1224 | } | |||
| 1225 | ||||
| 1226 | if (flush_count) { | |||
| 1227 | ret = vk->FlushMappedMemoryRanges(s->hwctx->act_dev, flush_count, | |||
| 1228 | flush_list); | |||
| 1229 | if (ret != VK_SUCCESS) { | |||
| 1230 | av_log(s, AV_LOG_ERROR16, "Failed to flush memory: %s\n", | |||
| 1231 | ff_vk_ret2str(ret)); | |||
| 1232 | err = AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); /* We still want to try to unmap them */ | |||
| 1233 | } | |||
| 1234 | } | |||
| 1235 | ||||
| 1236 | for (int i = 0; i < nb_buffers; i++) { | |||
| 1237 | vk->UnmapMemory(s->hwctx->act_dev, buf[i]->mem); | |||
| 1238 | buf[i]->mapped_mem = NULL((void*)0); | |||
| 1239 | } | |||
| 1240 | ||||
| 1241 | return err; | |||
| 1242 | } | |||
| 1243 | ||||
| 1244 | void ff_vk_free_buf(FFVulkanContext *s, FFVkBuffer *buf) | |||
| 1245 | { | |||
| 1246 | FFVulkanFunctions *vk = &s->vkfn; | |||
| 1247 | ||||
| 1248 | if (!buf || !s->hwctx) | |||
| 1249 | return; | |||
| 1250 | ||||
| 1251 | if (buf->mapped_mem && !buf->host_ref) | |||
| 1252 | ff_vk_unmap_buffer(s, buf, 0); | |||
| 1253 | if (buf->buf != VK_NULL_HANDLE((void*)0)) | |||
| 1254 | vk->DestroyBuffer(s->hwctx->act_dev, buf->buf, s->hwctx->alloc); | |||
| 1255 | if (buf->mem != VK_NULL_HANDLE((void*)0)) | |||
| 1256 | vk->FreeMemory(s->hwctx->act_dev, buf->mem, s->hwctx->alloc); | |||
| 1257 | if (buf->host_ref) | |||
| 1258 | av_buffer_unref(&buf->host_ref); | |||
| 1259 | ||||
| 1260 | buf->buf = VK_NULL_HANDLE((void*)0); | |||
| 1261 | buf->mem = VK_NULL_HANDLE((void*)0); | |||
| 1262 | buf->mapped_mem = NULL((void*)0); | |||
| 1263 | } | |||
| 1264 | ||||
| 1265 | static void free_data_buf(void *opaque, uint8_t *data) | |||
| 1266 | { | |||
| 1267 | FFVulkanContext *ctx = opaque; | |||
| 1268 | FFVkBuffer *buf = (FFVkBuffer *)data; | |||
| 1269 | ff_vk_free_buf(ctx, buf); | |||
| 1270 | av_free(data); | |||
| 1271 | } | |||
| 1272 | ||||
| 1273 | static AVBufferRef *alloc_data_buf(void *opaque, size_t size) | |||
| 1274 | { | |||
| 1275 | AVBufferRef *ref; | |||
| 1276 | uint8_t *buf = av_mallocz(size); | |||
| 1277 | if (!buf) | |||
| 1278 | return NULL((void*)0); | |||
| 1279 | ||||
| 1280 | ref = av_buffer_create(buf, size, free_data_buf, opaque, 0); | |||
| 1281 | if (!ref) | |||
| 1282 | av_free(buf); | |||
| 1283 | return ref; | |||
| 1284 | } | |||
| 1285 | ||||
| 1286 | int ff_vk_get_pooled_buffer(FFVulkanContext *ctx, AVBufferPool **buf_pool, | |||
| 1287 | AVBufferRef **buf, VkBufferUsageFlags usage, | |||
| 1288 | void *create_pNext, size_t size, | |||
| 1289 | VkMemoryPropertyFlagBits mem_props) | |||
| 1290 | { | |||
| 1291 | int err; | |||
| 1292 | AVBufferRef *ref; | |||
| 1293 | FFVkBuffer *data; | |||
| 1294 | ||||
| 1295 | *buf = NULL((void*)0); | |||
| 1296 | ||||
| 1297 | if (!(*buf_pool)) { | |||
| 1298 | *buf_pool = av_buffer_pool_init2(sizeof(FFVkBuffer), ctx, | |||
| 1299 | alloc_data_buf, NULL((void*)0)); | |||
| 1300 | if (!(*buf_pool)) | |||
| 1301 | return AVERROR(ENOMEM)(-(12)); | |||
| 1302 | } | |||
| 1303 | ||||
| 1304 | *buf = ref = av_buffer_pool_get(*buf_pool); | |||
| 1305 | if (!ref) | |||
| 1306 | return AVERROR(ENOMEM)(-(12)); | |||
| 1307 | ||||
| 1308 | data = (FFVkBuffer *)ref->data; | |||
| 1309 | ||||
| 1310 | if (data->size >= size) | |||
| 1311 | return 0; | |||
| 1312 | ||||
| 1313 | ff_vk_free_buf(ctx, data); | |||
| 1314 | memset(data, 0, sizeof(*data)); | |||
| 1315 | ||||
| 1316 | err = ff_vk_create_buf(ctx, data, size, | |||
| 1317 | create_pNext, NULL((void*)0), usage, | |||
| 1318 | mem_props); | |||
| 1319 | if (err < 0) { | |||
| 1320 | av_buffer_unref(&ref); | |||
| 1321 | *buf = NULL((void*)0); | |||
| 1322 | return err; | |||
| 1323 | } | |||
| 1324 | ||||
| 1325 | if (mem_props & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT) { | |||
| 1326 | err = ff_vk_map_buffer(ctx, data, &data->mapped_mem, 0); | |||
| 1327 | if (err < 0) { | |||
| 1328 | av_buffer_unref(&ref); | |||
| 1329 | *buf = NULL((void*)0); | |||
| 1330 | return err; | |||
| 1331 | } | |||
| 1332 | } | |||
| 1333 | ||||
| 1334 | return 0; | |||
| 1335 | } | |||
| 1336 | ||||
| 1337 | static int create_mapped_buffer(FFVulkanContext *s, | |||
| 1338 | FFVkBuffer *vkb, VkBufferUsageFlags usage, | |||
| 1339 | size_t size, | |||
| 1340 | VkExternalMemoryBufferCreateInfo *create_desc, | |||
| 1341 | VkImportMemoryHostPointerInfoEXT *import_desc, | |||
| 1342 | VkMemoryHostPointerPropertiesEXT props) | |||
| 1343 | { | |||
| 1344 | int err; | |||
| 1345 | VkResult ret; | |||
| 1346 | FFVulkanFunctions *vk = &s->vkfn; | |||
| 1347 | ||||
| 1348 | VkBufferCreateInfo buf_spawn = { | |||
| 1349 | .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO, | |||
| 1350 | .pNext = create_desc, | |||
| 1351 | .usage = usage, | |||
| 1352 | .sharingMode = VK_SHARING_MODE_EXCLUSIVE, | |||
| 1353 | .size = size, | |||
| 1354 | }; | |||
| 1355 | VkMemoryRequirements req = { | |||
| 1356 | .size = size, | |||
| 1357 | .alignment = s->hprops.minImportedHostPointerAlignment, | |||
| 1358 | .memoryTypeBits = props.memoryTypeBits, | |||
| 1359 | }; | |||
| 1360 | ||||
| 1361 | err = ff_vk_alloc_mem(s, &req, | |||
| 1362 | VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT, | |||
| 1363 | import_desc, &vkb->flags, &vkb->mem); | |||
| 1364 | if (err < 0) | |||
| 1365 | return err; | |||
| 1366 | ||||
| 1367 | ret = vk->CreateBuffer(s->hwctx->act_dev, &buf_spawn, s->hwctx->alloc, &vkb->buf); | |||
| 1368 | if (ret != VK_SUCCESS) { | |||
| 1369 | vk->FreeMemory(s->hwctx->act_dev, vkb->mem, s->hwctx->alloc); | |||
| 1370 | return AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); | |||
| 1371 | } | |||
| 1372 | ||||
| 1373 | ret = vk->BindBufferMemory(s->hwctx->act_dev, vkb->buf, vkb->mem, 0); | |||
| 1374 | if (ret != VK_SUCCESS) { | |||
| 1375 | vk->FreeMemory(s->hwctx->act_dev, vkb->mem, s->hwctx->alloc); | |||
| 1376 | vk->DestroyBuffer(s->hwctx->act_dev, vkb->buf, s->hwctx->alloc); | |||
| 1377 | return AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); | |||
| 1378 | } | |||
| 1379 | ||||
| 1380 | return 0; | |||
| 1381 | } | |||
| 1382 | ||||
| 1383 | static void destroy_avvkbuf(void *opaque, uint8_t *data) | |||
| 1384 | { | |||
| 1385 | FFVulkanContext *s = opaque; | |||
| 1386 | FFVkBuffer *buf = (FFVkBuffer *)data; | |||
| 1387 | ff_vk_free_buf(s, buf); | |||
| 1388 | av_free(buf); | |||
| 1389 | } | |||
| 1390 | ||||
| 1391 | int ff_vk_host_map_buffer(FFVulkanContext *s, AVBufferRef **dst, | |||
| 1392 | uint8_t *src_data, const AVBufferRef *src_buf, | |||
| 1393 | VkBufferUsageFlags usage) | |||
| 1394 | { | |||
| 1395 | int err; | |||
| 1396 | VkResult ret; | |||
| 1397 | FFVulkanFunctions *vk = &s->vkfn; | |||
| 1398 | ||||
| 1399 | VkExternalMemoryBufferCreateInfo create_desc = { | |||
| 1400 | .sType = VK_STRUCTURE_TYPE_EXTERNAL_MEMORY_BUFFER_CREATE_INFO, | |||
| 1401 | .handleTypes = VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT, | |||
| 1402 | }; | |||
| 1403 | VkMemoryAllocateFlagsInfo alloc_flags = { | |||
| 1404 | .sType = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_FLAGS_INFO, | |||
| 1405 | .flags = VK_MEMORY_ALLOCATE_DEVICE_ADDRESS_BIT, | |||
| 1406 | }; | |||
| 1407 | VkImportMemoryHostPointerInfoEXT import_desc = { | |||
| 1408 | .sType = VK_STRUCTURE_TYPE_IMPORT_MEMORY_HOST_POINTER_INFO_EXT, | |||
| 1409 | .handleType = VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT, | |||
| 1410 | .pNext = usage & VK_BUFFER_USAGE_SHADER_DEVICE_ADDRESS_BIT ? &alloc_flags : NULL((void*)0), | |||
| 1411 | }; | |||
| 1412 | VkMemoryHostPointerPropertiesEXT props; | |||
| 1413 | ||||
| 1414 | AVBufferRef *ref; | |||
| 1415 | FFVkBuffer *vkb; | |||
| 1416 | size_t offs; | |||
| 1417 | size_t buffer_size; | |||
| 1418 | ||||
| 1419 | *dst = NULL((void*)0); | |||
| 1420 | ||||
| 1421 | /* Get the previous point at which mapping was possible and use it */ | |||
| 1422 | offs = (uintptr_t)src_data % s->hprops.minImportedHostPointerAlignment; | |||
| 1423 | import_desc.pHostPointer = src_data - offs; | |||
| 1424 | ||||
| 1425 | props = (VkMemoryHostPointerPropertiesEXT) { | |||
| 1426 | VK_STRUCTURE_TYPE_MEMORY_HOST_POINTER_PROPERTIES_EXT, | |||
| 1427 | }; | |||
| 1428 | ret = vk->GetMemoryHostPointerPropertiesEXT(s->hwctx->act_dev, | |||
| 1429 | import_desc.handleType, | |||
| 1430 | import_desc.pHostPointer, | |||
| 1431 | &props); | |||
| 1432 | if (!(ret == VK_SUCCESS && props.memoryTypeBits)) | |||
| 1433 | return AVERROR(EINVAL)(-(22)); | |||
| 1434 | ||||
| 1435 | /* Ref the source buffer */ | |||
| 1436 | ref = av_buffer_ref(src_buf); | |||
| 1437 | if (!ref) | |||
| 1438 | return AVERROR(ENOMEM)(-(12)); | |||
| 1439 | ||||
| 1440 | /* Add the offset at the start, which gets ignored */ | |||
| 1441 | const ptrdiff_t src_offset = src_data - src_buf->data; | |||
| 1442 | buffer_size = offs + (src_buf->size - src_offset); | |||
| 1443 | buffer_size = FFALIGN(buffer_size, s->props.properties.limits.minMemoryMapAlignment)(((buffer_size)+(s->props.properties.limits.minMemoryMapAlignment )-1)&~((s->props.properties.limits.minMemoryMapAlignment )-1)); | |||
| 1444 | buffer_size = FFALIGN(buffer_size, s->hprops.minImportedHostPointerAlignment)(((buffer_size)+(s->hprops.minImportedHostPointerAlignment )-1)&~((s->hprops.minImportedHostPointerAlignment)-1)); | |||
| 1445 | ||||
| 1446 | /* Create a buffer struct */ | |||
| 1447 | vkb = av_mallocz(sizeof(*vkb)); | |||
| 1448 | if (!vkb) { | |||
| 1449 | av_buffer_unref(&ref); | |||
| 1450 | return AVERROR(ENOMEM)(-(12)); | |||
| 1451 | } | |||
| 1452 | ||||
| 1453 | err = create_mapped_buffer(s, vkb, usage, | |||
| 1454 | buffer_size, &create_desc, &import_desc, | |||
| 1455 | props); | |||
| 1456 | if (err < 0) { | |||
| 1457 | av_buffer_unref(&ref); | |||
| 1458 | av_free(vkb); | |||
| 1459 | return err; | |||
| 1460 | } | |||
| 1461 | ||||
| 1462 | if (usage & VK_BUFFER_USAGE_SHADER_DEVICE_ADDRESS_BIT) { | |||
| 1463 | VkBufferDeviceAddressInfo address_info = { | |||
| 1464 | .sType = VK_STRUCTURE_TYPE_BUFFER_DEVICE_ADDRESS_INFO, | |||
| 1465 | .buffer = vkb->buf, | |||
| 1466 | }; | |||
| 1467 | vkb->address = vk->GetBufferDeviceAddress(s->hwctx->act_dev, &address_info); | |||
| 1468 | } | |||
| 1469 | ||||
| 1470 | vkb->host_ref = ref; | |||
| 1471 | vkb->virtual_offset = offs; | |||
| 1472 | vkb->address += offs; | |||
| 1473 | vkb->mapped_mem = src_data; | |||
| 1474 | vkb->size = buffer_size - offs; | |||
| 1475 | vkb->flags |= VK_MEMORY_PROPERTY_HOST_COHERENT_BIT; | |||
| 1476 | ||||
| 1477 | /* Create a ref */ | |||
| 1478 | *dst = av_buffer_create((uint8_t *)vkb, sizeof(*vkb), | |||
| 1479 | destroy_avvkbuf, s, 0); | |||
| 1480 | if (!(*dst)) { | |||
| 1481 | destroy_avvkbuf(s, (uint8_t *)vkb); | |||
| 1482 | *dst = NULL((void*)0); | |||
| 1483 | return AVERROR(ENOMEM)(-(12)); | |||
| 1484 | } | |||
| 1485 | ||||
| 1486 | return 0; | |||
| 1487 | } | |||
| 1488 | ||||
| 1489 | int ff_vk_shader_add_push_const(FFVulkanShader *shd, int offset, int size, | |||
| 1490 | VkShaderStageFlagBits stage) | |||
| 1491 | { | |||
| 1492 | VkPushConstantRange *pc = &shd->push_consts[shd->push_consts_num++]; | |||
| 1493 | av_assert1(shd->push_consts_num < FF_VK_MAX_PUSH_CONSTS)do { if (!(shd->push_consts_num < 4)) { av_log(((void*) 0), 0, "Assertion %s failed at %s:%d\n", "shd->push_consts_num < 4" , "/root/firefox-clang/media/ffvpx/libavutil/vulkan.c", 1493) ; abort(); } } while (0); | |||
| 1494 | pc->stageFlags = stage; | |||
| 1495 | pc->offset = offset; | |||
| 1496 | pc->size = size; | |||
| 1497 | return 0; | |||
| 1498 | } | |||
| 1499 | ||||
| 1500 | int ff_vk_init_sampler(FFVulkanContext *s, VkSampler *sampler, | |||
| 1501 | int unnorm_coords, VkFilter filt) | |||
| 1502 | { | |||
| 1503 | VkResult ret; | |||
| 1504 | FFVulkanFunctions *vk = &s->vkfn; | |||
| 1505 | ||||
| 1506 | VkSamplerCreateInfo sampler_info = { | |||
| 1507 | .sType = VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO, | |||
| 1508 | .magFilter = filt, | |||
| 1509 | .minFilter = sampler_info.magFilter, | |||
| 1510 | .mipmapMode = unnorm_coords ? VK_SAMPLER_MIPMAP_MODE_NEAREST : | |||
| 1511 | VK_SAMPLER_MIPMAP_MODE_LINEAR, | |||
| 1512 | .addressModeU = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE, | |||
| 1513 | .addressModeV = sampler_info.addressModeU, | |||
| 1514 | .addressModeW = sampler_info.addressModeU, | |||
| 1515 | .anisotropyEnable = VK_FALSE0U, | |||
| 1516 | .compareOp = VK_COMPARE_OP_NEVER, | |||
| 1517 | .borderColor = VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK, | |||
| 1518 | .unnormalizedCoordinates = unnorm_coords, | |||
| 1519 | }; | |||
| 1520 | ||||
| 1521 | ret = vk->CreateSampler(s->hwctx->act_dev, &sampler_info, | |||
| 1522 | s->hwctx->alloc, sampler); | |||
| 1523 | if (ret != VK_SUCCESS) { | |||
| 1524 | av_log(s, AV_LOG_ERROR16, "Unable to init sampler: %s\n", | |||
| 1525 | ff_vk_ret2str(ret)); | |||
| 1526 | return AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); | |||
| 1527 | } | |||
| 1528 | ||||
| 1529 | return 0; | |||
| 1530 | } | |||
| 1531 | ||||
| 1532 | VkImageAspectFlags ff_vk_aspect_flag(AVFrame *f, int p) | |||
| 1533 | { | |||
| 1534 | AVVkFrame *vkf = (AVVkFrame *)f->data[0]; | |||
| 1535 | AVHWFramesContext *hwfc = (AVHWFramesContext *)f->hw_frames_ctx->data; | |||
| 1536 | int nb_images = ff_vk_count_images(vkf); | |||
| 1537 | int nb_planes = av_pix_fmt_count_planes(hwfc->sw_format); | |||
| 1538 | ||||
| 1539 | static const VkImageAspectFlags plane_aspect[] = { VK_IMAGE_ASPECT_PLANE_0_BIT, | |||
| 1540 | VK_IMAGE_ASPECT_PLANE_1_BIT, | |||
| 1541 | VK_IMAGE_ASPECT_PLANE_2_BIT, }; | |||
| 1542 | ||||
| 1543 | if (ff_vk_mt_is_np_rgb(hwfc->sw_format) || (nb_planes == nb_images)) | |||
| 1544 | return VK_IMAGE_ASPECT_COLOR_BIT; | |||
| 1545 | ||||
| 1546 | return plane_aspect[p]; | |||
| 1547 | } | |||
| 1548 | ||||
| 1549 | int ff_vk_mt_is_np_rgb(enum AVPixelFormat pix_fmt) | |||
| 1550 | { | |||
| 1551 | if (pix_fmt == AV_PIX_FMT_ABGR || pix_fmt == AV_PIX_FMT_BGRA || | |||
| 1552 | pix_fmt == AV_PIX_FMT_RGBA || pix_fmt == AV_PIX_FMT_RGB24 || | |||
| 1553 | pix_fmt == AV_PIX_FMT_BGR24 || pix_fmt == AV_PIX_FMT_RGB48AV_PIX_FMT_RGB48LE || | |||
| 1554 | pix_fmt == AV_PIX_FMT_RGBA64AV_PIX_FMT_RGBA64LE || pix_fmt == AV_PIX_FMT_RGB565AV_PIX_FMT_RGB565LE || | |||
| 1555 | pix_fmt == AV_PIX_FMT_BGR565AV_PIX_FMT_BGR565LE || pix_fmt == AV_PIX_FMT_BGR0 || | |||
| 1556 | pix_fmt == AV_PIX_FMT_0BGR || pix_fmt == AV_PIX_FMT_RGB0 || | |||
| 1557 | pix_fmt == AV_PIX_FMT_GBRP10AV_PIX_FMT_GBRP10LE || pix_fmt == AV_PIX_FMT_GBRP12AV_PIX_FMT_GBRP12LE || | |||
| 1558 | pix_fmt == AV_PIX_FMT_GBRP14AV_PIX_FMT_GBRP14LE || pix_fmt == AV_PIX_FMT_GBRP16AV_PIX_FMT_GBRP16LE || | |||
| 1559 | pix_fmt == AV_PIX_FMT_GBRAP || pix_fmt == AV_PIX_FMT_GBRAP10AV_PIX_FMT_GBRAP10LE || | |||
| 1560 | pix_fmt == AV_PIX_FMT_GBRAP12AV_PIX_FMT_GBRAP12LE || pix_fmt == AV_PIX_FMT_GBRAP14AV_PIX_FMT_GBRAP14LE || | |||
| 1561 | pix_fmt == AV_PIX_FMT_GBRAP16AV_PIX_FMT_GBRAP16LE || pix_fmt == AV_PIX_FMT_GBRAP32AV_PIX_FMT_GBRAP32LE || | |||
| 1562 | pix_fmt == AV_PIX_FMT_GBRPF16AV_PIX_FMT_GBRPF16LE || pix_fmt == AV_PIX_FMT_GBRAPF16AV_PIX_FMT_GBRAPF16LE || | |||
| 1563 | pix_fmt == AV_PIX_FMT_GBRPF32AV_PIX_FMT_GBRPF32LE || pix_fmt == AV_PIX_FMT_GBRAPF32AV_PIX_FMT_GBRAPF32LE || | |||
| 1564 | pix_fmt == AV_PIX_FMT_X2RGB10AV_PIX_FMT_X2RGB10LE || pix_fmt == AV_PIX_FMT_X2BGR10AV_PIX_FMT_X2BGR10LE || | |||
| 1565 | pix_fmt == AV_PIX_FMT_RGBAF32AV_PIX_FMT_RGBAF32LE || pix_fmt == AV_PIX_FMT_RGBF32AV_PIX_FMT_RGBF32LE || | |||
| 1566 | pix_fmt == AV_PIX_FMT_RGBA128AV_PIX_FMT_RGBA128LE || pix_fmt == AV_PIX_FMT_RGB96AV_PIX_FMT_RGB96LE || | |||
| 1567 | pix_fmt == AV_PIX_FMT_GBRP || pix_fmt == AV_PIX_FMT_BAYER_RGGB16AV_PIX_FMT_BAYER_RGGB16LE) | |||
| 1568 | return 1; | |||
| 1569 | return 0; | |||
| 1570 | } | |||
| 1571 | ||||
| 1572 | void ff_vk_set_perm(enum AVPixelFormat pix_fmt, int lut[4], int inv) | |||
| 1573 | { | |||
| 1574 | switch (pix_fmt) { | |||
| 1575 | case AV_PIX_FMT_GBRP: | |||
| 1576 | case AV_PIX_FMT_GBRAP: | |||
| 1577 | case AV_PIX_FMT_GBRAP10AV_PIX_FMT_GBRAP10LE: | |||
| 1578 | case AV_PIX_FMT_GBRAP12AV_PIX_FMT_GBRAP12LE: | |||
| 1579 | case AV_PIX_FMT_GBRAP14AV_PIX_FMT_GBRAP14LE: | |||
| 1580 | case AV_PIX_FMT_GBRAP16AV_PIX_FMT_GBRAP16LE: | |||
| 1581 | case AV_PIX_FMT_GBRAPF16AV_PIX_FMT_GBRAPF16LE: | |||
| 1582 | case AV_PIX_FMT_GBRP10AV_PIX_FMT_GBRP10LE: | |||
| 1583 | case AV_PIX_FMT_GBRP12AV_PIX_FMT_GBRP12LE: | |||
| 1584 | case AV_PIX_FMT_GBRP14AV_PIX_FMT_GBRP14LE: | |||
| 1585 | case AV_PIX_FMT_GBRP16AV_PIX_FMT_GBRP16LE: | |||
| 1586 | case AV_PIX_FMT_GBRPF16AV_PIX_FMT_GBRPF16LE: | |||
| 1587 | case AV_PIX_FMT_GBRPF32AV_PIX_FMT_GBRPF32LE: | |||
| 1588 | case AV_PIX_FMT_GBRAP32AV_PIX_FMT_GBRAP32LE: | |||
| 1589 | case AV_PIX_FMT_GBRAPF32AV_PIX_FMT_GBRAPF32LE: | |||
| 1590 | lut[0] = 1; | |||
| 1591 | lut[1] = 2; | |||
| 1592 | lut[2] = 0; | |||
| 1593 | lut[3] = 3; | |||
| 1594 | break; | |||
| 1595 | case AV_PIX_FMT_X2BGR10AV_PIX_FMT_X2BGR10LE: | |||
| 1596 | lut[0] = 0; | |||
| 1597 | lut[1] = 2; | |||
| 1598 | lut[2] = 1; | |||
| 1599 | lut[3] = 3; | |||
| 1600 | break; | |||
| 1601 | default: | |||
| 1602 | lut[0] = 0; | |||
| 1603 | lut[1] = 1; | |||
| 1604 | lut[2] = 2; | |||
| 1605 | lut[3] = 3; | |||
| 1606 | break; | |||
| 1607 | } | |||
| 1608 | ||||
| 1609 | if (inv) { | |||
| 1610 | int lut_tmp[4] = { lut[0], lut[1], lut[2], lut[3] }; | |||
| 1611 | for (int i = 0; i < 4; i++) | |||
| 1612 | lut[lut_tmp[i]] = i; | |||
| 1613 | } | |||
| 1614 | ||||
| 1615 | return; | |||
| 1616 | } | |||
| 1617 | ||||
| 1618 | const char *ff_vk_shader_rep_fmt(enum AVPixelFormat pix_fmt, | |||
| 1619 | enum FFVkShaderRepFormat rep_fmt) | |||
| 1620 | { | |||
| 1621 | switch (pix_fmt) { | |||
| 1622 | case AV_PIX_FMT_RGBA: | |||
| 1623 | case AV_PIX_FMT_BGRA: | |||
| 1624 | case AV_PIX_FMT_RGB24: | |||
| 1625 | case AV_PIX_FMT_BGR24: | |||
| 1626 | case AV_PIX_FMT_BGR0: | |||
| 1627 | case AV_PIX_FMT_RGB0: | |||
| 1628 | case AV_PIX_FMT_RGB565AV_PIX_FMT_RGB565LE: | |||
| 1629 | case AV_PIX_FMT_BGR565AV_PIX_FMT_BGR565LE: | |||
| 1630 | case AV_PIX_FMT_UYVA: | |||
| 1631 | case AV_PIX_FMT_YUYV422: | |||
| 1632 | case AV_PIX_FMT_UYVY422: { | |||
| 1633 | const char *rep_tab[] = { | |||
| 1634 | [FF_VK_REP_NATIVE] = "rgba8ui", | |||
| 1635 | [FF_VK_REP_FLOAT] = "rgba8", | |||
| 1636 | [FF_VK_REP_INT] = "rgba8i", | |||
| 1637 | [FF_VK_REP_UINT] = "rgba8ui", | |||
| 1638 | }; | |||
| 1639 | return rep_tab[rep_fmt]; | |||
| 1640 | } | |||
| 1641 | case AV_PIX_FMT_X2RGB10AV_PIX_FMT_X2RGB10LE: | |||
| 1642 | case AV_PIX_FMT_X2BGR10AV_PIX_FMT_X2BGR10LE: | |||
| 1643 | case AV_PIX_FMT_Y210AV_PIX_FMT_Y210LE: | |||
| 1644 | case AV_PIX_FMT_XV30AV_PIX_FMT_XV30LE: { | |||
| 1645 | const char *rep_tab[] = { | |||
| 1646 | [FF_VK_REP_NATIVE] = "rgb10_a2ui", | |||
| 1647 | [FF_VK_REP_FLOAT] = "rgb10_a2", | |||
| 1648 | [FF_VK_REP_INT] = NULL((void*)0), | |||
| 1649 | [FF_VK_REP_UINT] = "rgb10_a2ui", | |||
| 1650 | }; | |||
| 1651 | return rep_tab[rep_fmt]; | |||
| 1652 | } | |||
| 1653 | case AV_PIX_FMT_RGB48AV_PIX_FMT_RGB48LE: | |||
| 1654 | case AV_PIX_FMT_RGBA64AV_PIX_FMT_RGBA64LE: | |||
| 1655 | case AV_PIX_FMT_Y212AV_PIX_FMT_Y212LE: | |||
| 1656 | case AV_PIX_FMT_Y216AV_PIX_FMT_Y216LE: | |||
| 1657 | case AV_PIX_FMT_XV36AV_PIX_FMT_XV36LE: | |||
| 1658 | case AV_PIX_FMT_XV48AV_PIX_FMT_XV48LE: { | |||
| 1659 | const char *rep_tab[] = { | |||
| 1660 | [FF_VK_REP_NATIVE] = "rgba16ui", | |||
| 1661 | [FF_VK_REP_FLOAT] = "rgba16", | |||
| 1662 | [FF_VK_REP_INT] = "rgba16i", | |||
| 1663 | [FF_VK_REP_UINT] = "rgba16ui", | |||
| 1664 | }; | |||
| 1665 | return rep_tab[rep_fmt]; | |||
| 1666 | } | |||
| 1667 | case AV_PIX_FMT_RGBF32AV_PIX_FMT_RGBF32LE: | |||
| 1668 | case AV_PIX_FMT_RGBAF32AV_PIX_FMT_RGBAF32LE: { | |||
| 1669 | const char *rep_tab[] = { | |||
| 1670 | [FF_VK_REP_NATIVE] = "rgba32f", | |||
| 1671 | [FF_VK_REP_FLOAT] = "rgba32f", | |||
| 1672 | [FF_VK_REP_INT] = "rgba32i", | |||
| 1673 | [FF_VK_REP_UINT] = "rgba32ui", | |||
| 1674 | }; | |||
| 1675 | return rep_tab[rep_fmt]; | |||
| 1676 | } | |||
| 1677 | case AV_PIX_FMT_RGB96AV_PIX_FMT_RGB96LE: | |||
| 1678 | case AV_PIX_FMT_RGBA128AV_PIX_FMT_RGBA128LE: { | |||
| 1679 | const char *rep_tab[] = { | |||
| 1680 | [FF_VK_REP_NATIVE] = "rgba32ui", | |||
| 1681 | [FF_VK_REP_FLOAT] = NULL((void*)0), | |||
| 1682 | [FF_VK_REP_INT] = "rgba32i", | |||
| 1683 | [FF_VK_REP_UINT] = "rgba32ui", | |||
| 1684 | }; | |||
| 1685 | return rep_tab[rep_fmt]; | |||
| 1686 | } | |||
| 1687 | case AV_PIX_FMT_GBRP: | |||
| 1688 | case AV_PIX_FMT_GRAY8: | |||
| 1689 | case AV_PIX_FMT_GBRAP: | |||
| 1690 | case AV_PIX_FMT_YUV420P: | |||
| 1691 | case AV_PIX_FMT_YUV422P: | |||
| 1692 | case AV_PIX_FMT_YUV444P: | |||
| 1693 | case AV_PIX_FMT_YUVA420P: | |||
| 1694 | case AV_PIX_FMT_YUVA422P: | |||
| 1695 | case AV_PIX_FMT_YUVA444P: { | |||
| 1696 | const char *rep_tab[] = { | |||
| 1697 | [FF_VK_REP_NATIVE] = "r8ui", | |||
| 1698 | [FF_VK_REP_FLOAT] = "r8", | |||
| 1699 | [FF_VK_REP_INT] = "r8i", | |||
| 1700 | [FF_VK_REP_UINT] = "r8ui", | |||
| 1701 | }; | |||
| 1702 | return rep_tab[rep_fmt]; | |||
| 1703 | }; | |||
| 1704 | case AV_PIX_FMT_GRAY10AV_PIX_FMT_GRAY10LE: | |||
| 1705 | case AV_PIX_FMT_GRAY12AV_PIX_FMT_GRAY12LE: | |||
| 1706 | case AV_PIX_FMT_GRAY14AV_PIX_FMT_GRAY14LE: | |||
| 1707 | case AV_PIX_FMT_GRAY16AV_PIX_FMT_GRAY16LE: | |||
| 1708 | case AV_PIX_FMT_GBRAP10AV_PIX_FMT_GBRAP10LE: | |||
| 1709 | case AV_PIX_FMT_GBRAP12AV_PIX_FMT_GBRAP12LE: | |||
| 1710 | case AV_PIX_FMT_GBRAP14AV_PIX_FMT_GBRAP14LE: | |||
| 1711 | case AV_PIX_FMT_GBRAP16AV_PIX_FMT_GBRAP16LE: | |||
| 1712 | case AV_PIX_FMT_GBRAPF16AV_PIX_FMT_GBRAPF16LE: | |||
| 1713 | case AV_PIX_FMT_GBRP10AV_PIX_FMT_GBRP10LE: | |||
| 1714 | case AV_PIX_FMT_GBRP12AV_PIX_FMT_GBRP12LE: | |||
| 1715 | case AV_PIX_FMT_GBRP14AV_PIX_FMT_GBRP14LE: | |||
| 1716 | case AV_PIX_FMT_GBRP16AV_PIX_FMT_GBRP16LE: | |||
| 1717 | case AV_PIX_FMT_GBRPF16AV_PIX_FMT_GBRPF16LE: | |||
| 1718 | case AV_PIX_FMT_YUV420P10AV_PIX_FMT_YUV420P10LE: | |||
| 1719 | case AV_PIX_FMT_YUV420P12AV_PIX_FMT_YUV420P12LE: | |||
| 1720 | case AV_PIX_FMT_YUV420P16AV_PIX_FMT_YUV420P16LE: | |||
| 1721 | case AV_PIX_FMT_YUV422P10AV_PIX_FMT_YUV422P10LE: | |||
| 1722 | case AV_PIX_FMT_YUV422P12AV_PIX_FMT_YUV422P12LE: | |||
| 1723 | case AV_PIX_FMT_YUV422P16AV_PIX_FMT_YUV422P16LE: | |||
| 1724 | case AV_PIX_FMT_YUV444P10AV_PIX_FMT_YUV444P10LE: | |||
| 1725 | case AV_PIX_FMT_YUV444P12AV_PIX_FMT_YUV444P12LE: | |||
| 1726 | case AV_PIX_FMT_YUV444P16AV_PIX_FMT_YUV444P16LE: | |||
| 1727 | case AV_PIX_FMT_YUVA420P10AV_PIX_FMT_YUVA420P10LE: | |||
| 1728 | case AV_PIX_FMT_YUVA420P16AV_PIX_FMT_YUVA420P16LE: | |||
| 1729 | case AV_PIX_FMT_YUVA422P10AV_PIX_FMT_YUVA422P10LE: | |||
| 1730 | case AV_PIX_FMT_YUVA422P12AV_PIX_FMT_YUVA422P12LE: | |||
| 1731 | case AV_PIX_FMT_YUVA422P16AV_PIX_FMT_YUVA422P16LE: | |||
| 1732 | case AV_PIX_FMT_YUVA444P10AV_PIX_FMT_YUVA444P10LE: | |||
| 1733 | case AV_PIX_FMT_YUVA444P12AV_PIX_FMT_YUVA444P12LE: | |||
| 1734 | case AV_PIX_FMT_YUVA444P16AV_PIX_FMT_YUVA444P16LE: | |||
| 1735 | case AV_PIX_FMT_BAYER_RGGB16AV_PIX_FMT_BAYER_RGGB16LE: { | |||
| 1736 | const char *rep_tab[] = { | |||
| 1737 | [FF_VK_REP_NATIVE] = "r16ui", | |||
| 1738 | [FF_VK_REP_FLOAT] = "r16f", | |||
| 1739 | [FF_VK_REP_INT] = "r16i", | |||
| 1740 | [FF_VK_REP_UINT] = "r16ui", | |||
| 1741 | }; | |||
| 1742 | return rep_tab[rep_fmt]; | |||
| 1743 | }; | |||
| 1744 | case AV_PIX_FMT_GRAY32AV_PIX_FMT_GRAY32LE: | |||
| 1745 | case AV_PIX_FMT_GRAYF32AV_PIX_FMT_GRAYF32LE: | |||
| 1746 | case AV_PIX_FMT_GBRPF32AV_PIX_FMT_GBRPF32LE: | |||
| 1747 | case AV_PIX_FMT_GBRAPF32AV_PIX_FMT_GBRAPF32LE: { | |||
| 1748 | const char *rep_tab[] = { | |||
| 1749 | [FF_VK_REP_NATIVE] = "r32f", | |||
| 1750 | [FF_VK_REP_FLOAT] = "r32f", | |||
| 1751 | [FF_VK_REP_INT] = "r32i", | |||
| 1752 | [FF_VK_REP_UINT] = "r32ui", | |||
| 1753 | }; | |||
| 1754 | return rep_tab[rep_fmt]; | |||
| 1755 | }; | |||
| 1756 | case AV_PIX_FMT_GBRAP32AV_PIX_FMT_GBRAP32LE: { | |||
| 1757 | const char *rep_tab[] = { | |||
| 1758 | [FF_VK_REP_NATIVE] = "r32ui", | |||
| 1759 | [FF_VK_REP_FLOAT] = NULL((void*)0), | |||
| 1760 | [FF_VK_REP_INT] = "r32i", | |||
| 1761 | [FF_VK_REP_UINT] = "r32ui", | |||
| 1762 | }; | |||
| 1763 | return rep_tab[rep_fmt]; | |||
| 1764 | }; | |||
| 1765 | case AV_PIX_FMT_NV12: | |||
| 1766 | case AV_PIX_FMT_NV16: | |||
| 1767 | case AV_PIX_FMT_NV24: { | |||
| 1768 | const char *rep_tab[] = { | |||
| 1769 | [FF_VK_REP_NATIVE] = "rg8ui", | |||
| 1770 | [FF_VK_REP_FLOAT] = "rg8", | |||
| 1771 | [FF_VK_REP_INT] = "rg8i", | |||
| 1772 | [FF_VK_REP_UINT] = "rg8ui", | |||
| 1773 | }; | |||
| 1774 | return rep_tab[rep_fmt]; | |||
| 1775 | }; | |||
| 1776 | case AV_PIX_FMT_P010AV_PIX_FMT_P010LE: | |||
| 1777 | case AV_PIX_FMT_P210AV_PIX_FMT_P210LE: | |||
| 1778 | case AV_PIX_FMT_P410AV_PIX_FMT_P410LE: { | |||
| 1779 | const char *rep_tab[] = { | |||
| 1780 | [FF_VK_REP_NATIVE] = "rgb10_a2ui", | |||
| 1781 | [FF_VK_REP_FLOAT] = "rgb10_a2", | |||
| 1782 | [FF_VK_REP_INT] = NULL((void*)0), | |||
| 1783 | [FF_VK_REP_UINT] = "rgb10_a2ui", | |||
| 1784 | }; | |||
| 1785 | return rep_tab[rep_fmt]; | |||
| 1786 | }; | |||
| 1787 | case AV_PIX_FMT_P012AV_PIX_FMT_P012LE: | |||
| 1788 | case AV_PIX_FMT_P016AV_PIX_FMT_P016LE: | |||
| 1789 | case AV_PIX_FMT_P212AV_PIX_FMT_P212LE: | |||
| 1790 | case AV_PIX_FMT_P216AV_PIX_FMT_P216LE: | |||
| 1791 | case AV_PIX_FMT_P412AV_PIX_FMT_P412LE: | |||
| 1792 | case AV_PIX_FMT_P416AV_PIX_FMT_P416LE: { | |||
| 1793 | const char *rep_tab[] = { | |||
| 1794 | [FF_VK_REP_NATIVE] = "rg16ui", | |||
| 1795 | [FF_VK_REP_FLOAT] = "rg16", | |||
| 1796 | [FF_VK_REP_INT] = "rg16i", | |||
| 1797 | [FF_VK_REP_UINT] = "rg16ui", | |||
| 1798 | }; | |||
| 1799 | return rep_tab[rep_fmt]; | |||
| 1800 | }; | |||
| 1801 | default: | |||
| 1802 | return "rgba32f"; | |||
| 1803 | } | |||
| 1804 | } | |||
| 1805 | ||||
| 1806 | typedef struct ImageViewCtx { | |||
| 1807 | int nb_views; | |||
| 1808 | VkImageView views[]; | |||
| 1809 | } ImageViewCtx; | |||
| 1810 | ||||
| 1811 | static void destroy_imageviews(void *opaque, uint8_t *data) | |||
| 1812 | { | |||
| 1813 | FFVulkanContext *s = opaque; | |||
| 1814 | FFVulkanFunctions *vk = &s->vkfn; | |||
| 1815 | ImageViewCtx *iv = (ImageViewCtx *)data; | |||
| 1816 | ||||
| 1817 | for (int i = 0; i < iv->nb_views; i++) | |||
| 1818 | vk->DestroyImageView(s->hwctx->act_dev, iv->views[i], s->hwctx->alloc); | |||
| 1819 | ||||
| 1820 | av_free(iv); | |||
| 1821 | } | |||
| 1822 | ||||
| 1823 | static VkFormat map_fmt_to_rep(VkFormat fmt, enum FFVkShaderRepFormat rep_fmt) | |||
| 1824 | { | |||
| 1825 | #define REPS_FMT(fmt) \ | |||
| 1826 | [FF_VK_REP_NATIVE] = fmt ## _UINT, \ | |||
| 1827 | [FF_VK_REP_FLOAT] = fmt ## _UNORM, \ | |||
| 1828 | [FF_VK_REP_INT] = fmt ## _SINT, \ | |||
| 1829 | [FF_VK_REP_UINT] = fmt ## _UINT, | |||
| 1830 | ||||
| 1831 | #define REPS_FMT_PACK(fmt, num) \ | |||
| 1832 | [FF_VK_REP_NATIVE] = fmt ## _UINT_PACK ## num, \ | |||
| 1833 | [FF_VK_REP_FLOAT] = fmt ## _UNORM_PACK ## num, \ | |||
| 1834 | [FF_VK_REP_INT] = fmt ## _SINT_PACK ## num, \ | |||
| 1835 | [FF_VK_REP_UINT] = fmt ## _UINT_PACK ## num, | |||
| 1836 | ||||
| 1837 | const VkFormat fmts_map[][4] = { | |||
| 1838 | { REPS_FMT_PACK(VK_FORMAT_A2B10G10R10, 32) }, | |||
| 1839 | { REPS_FMT_PACK(VK_FORMAT_A2R10G10B10, 32) }, | |||
| 1840 | { | |||
| 1841 | VK_FORMAT_B5G6R5_UNORM_PACK16, | |||
| 1842 | VK_FORMAT_B5G6R5_UNORM_PACK16, | |||
| 1843 | VK_FORMAT_UNDEFINED, | |||
| 1844 | VK_FORMAT_UNDEFINED, | |||
| 1845 | }, | |||
| 1846 | { | |||
| 1847 | VK_FORMAT_R5G6B5_UNORM_PACK16, | |||
| 1848 | VK_FORMAT_R5G6B5_UNORM_PACK16, | |||
| 1849 | VK_FORMAT_UNDEFINED, | |||
| 1850 | VK_FORMAT_UNDEFINED, | |||
| 1851 | }, | |||
| 1852 | { REPS_FMT(VK_FORMAT_B8G8R8) }, | |||
| 1853 | { REPS_FMT(VK_FORMAT_B8G8R8A8) }, | |||
| 1854 | { REPS_FMT(VK_FORMAT_R8) }, | |||
| 1855 | { REPS_FMT(VK_FORMAT_R8G8) }, | |||
| 1856 | { REPS_FMT(VK_FORMAT_R8G8B8) }, | |||
| 1857 | { REPS_FMT(VK_FORMAT_R8G8B8A8) }, | |||
| 1858 | { REPS_FMT(VK_FORMAT_R16) }, | |||
| 1859 | { REPS_FMT(VK_FORMAT_R16G16) }, | |||
| 1860 | { REPS_FMT(VK_FORMAT_R16G16B16) }, | |||
| 1861 | { REPS_FMT(VK_FORMAT_R16G16B16A16) }, | |||
| 1862 | { | |||
| 1863 | VK_FORMAT_R32_UINT, | |||
| 1864 | VK_FORMAT_R32_SFLOAT, | |||
| 1865 | VK_FORMAT_R32_SINT, | |||
| 1866 | VK_FORMAT_R32_UINT, | |||
| 1867 | }, | |||
| 1868 | { | |||
| 1869 | VK_FORMAT_R32G32B32_SFLOAT, | |||
| 1870 | VK_FORMAT_R32G32B32_SFLOAT, | |||
| 1871 | VK_FORMAT_UNDEFINED, | |||
| 1872 | VK_FORMAT_UNDEFINED, | |||
| 1873 | }, | |||
| 1874 | { | |||
| 1875 | VK_FORMAT_R32G32B32A32_SFLOAT, | |||
| 1876 | VK_FORMAT_R32G32B32A32_SFLOAT, | |||
| 1877 | VK_FORMAT_UNDEFINED, | |||
| 1878 | VK_FORMAT_UNDEFINED, | |||
| 1879 | }, | |||
| 1880 | { | |||
| 1881 | VK_FORMAT_R32G32B32_UINT, | |||
| 1882 | VK_FORMAT_UNDEFINED, | |||
| 1883 | VK_FORMAT_R32G32B32_SINT, | |||
| 1884 | VK_FORMAT_R32G32B32_UINT, | |||
| 1885 | }, | |||
| 1886 | { | |||
| 1887 | VK_FORMAT_R32G32B32A32_UINT, | |||
| 1888 | VK_FORMAT_UNDEFINED, | |||
| 1889 | VK_FORMAT_R32G32B32A32_SINT, | |||
| 1890 | VK_FORMAT_R32G32B32A32_UINT, | |||
| 1891 | }, | |||
| 1892 | { | |||
| 1893 | VK_FORMAT_R16_SFLOAT, | |||
| 1894 | VK_FORMAT_R16_SFLOAT, | |||
| 1895 | VK_FORMAT_UNDEFINED, | |||
| 1896 | VK_FORMAT_UNDEFINED, | |||
| 1897 | }, | |||
| 1898 | }; | |||
| 1899 | #undef REPS_FMT_PACK | |||
| 1900 | #undef REPS_FMT | |||
| 1901 | ||||
| 1902 | if (fmt == VK_FORMAT_UNDEFINED) | |||
| 1903 | return VK_FORMAT_UNDEFINED; | |||
| 1904 | ||||
| 1905 | for (int i = 0; i < FF_ARRAY_ELEMS(fmts_map)(sizeof(fmts_map) / sizeof((fmts_map)[0])); i++) { | |||
| 1906 | if (fmts_map[i][FF_VK_REP_NATIVE] == fmt || | |||
| 1907 | fmts_map[i][FF_VK_REP_FLOAT] == fmt || | |||
| 1908 | fmts_map[i][FF_VK_REP_INT] == fmt || | |||
| 1909 | fmts_map[i][FF_VK_REP_UINT] == fmt) | |||
| 1910 | return fmts_map[i][rep_fmt]; | |||
| 1911 | } | |||
| 1912 | ||||
| 1913 | return VK_FORMAT_UNDEFINED; | |||
| 1914 | } | |||
| 1915 | ||||
| 1916 | int ff_vk_create_imageview(FFVulkanContext *s, | |||
| 1917 | VkImageView *img_view, VkImageAspectFlags *aspect, | |||
| 1918 | AVFrame *f, int plane, enum FFVkShaderRepFormat rep_fmt) | |||
| 1919 | { | |||
| 1920 | VkResult ret; | |||
| 1921 | FFVulkanFunctions *vk = &s->vkfn; | |||
| 1922 | AVHWFramesContext *hwfc = (AVHWFramesContext *)f->hw_frames_ctx->data; | |||
| 1923 | AVVulkanFramesContext *vkfc = hwfc->hwctx; | |||
| 1924 | const VkFormat *rep_fmts = av_vkfmt_from_pixfmt(hwfc->sw_format); | |||
| 1925 | AVVkFrame *vkf = (AVVkFrame *)f->data[0]; | |||
| 1926 | const int nb_images = ff_vk_count_images(vkf); | |||
| 1927 | ||||
| 1928 | VkImageViewUsageCreateInfo view_usage_info = { | |||
| 1929 | .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_USAGE_CREATE_INFO, | |||
| 1930 | .usage = vkfc->usage & | |||
| 1931 | (~(VK_IMAGE_USAGE_VIDEO_ENCODE_SRC_BIT_KHR | | |||
| 1932 | VK_IMAGE_USAGE_VIDEO_DECODE_DST_BIT_KHR)), | |||
| 1933 | }; | |||
| 1934 | VkImageViewCreateInfo view_create_info = { | |||
| 1935 | .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO, | |||
| 1936 | .pNext = &view_usage_info, | |||
| 1937 | .image = vkf->img[FFMIN(plane, nb_images - 1)((plane) > (nb_images - 1) ? (nb_images - 1) : (plane))], | |||
| 1938 | .viewType = VK_IMAGE_VIEW_TYPE_2D, | |||
| 1939 | .format = map_fmt_to_rep(rep_fmts[plane], rep_fmt), | |||
| 1940 | .components = ff_comp_identity_map, | |||
| 1941 | .subresourceRange = { | |||
| 1942 | .aspectMask = ff_vk_aspect_flag(f, plane), | |||
| 1943 | .levelCount = 1, | |||
| 1944 | .layerCount = 1, | |||
| 1945 | }, | |||
| 1946 | }; | |||
| 1947 | if (view_create_info.format == VK_FORMAT_UNDEFINED) { | |||
| 1948 | av_log(s, AV_LOG_ERROR16, "Unable to find a compatible representation " | |||
| 1949 | "of format %i and mode %i\n", | |||
| 1950 | rep_fmts[plane], rep_fmt); | |||
| 1951 | return AVERROR(EINVAL)(-(22)); | |||
| 1952 | } | |||
| 1953 | ||||
| 1954 | ret = vk->CreateImageView(s->hwctx->act_dev, &view_create_info, | |||
| 1955 | s->hwctx->alloc, img_view); | |||
| 1956 | if (ret != VK_SUCCESS) { | |||
| 1957 | av_log(s, AV_LOG_ERROR16, "Failed to create imageview: %s\n", | |||
| 1958 | ff_vk_ret2str(ret)); | |||
| 1959 | return AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); | |||
| 1960 | } | |||
| 1961 | ||||
| 1962 | *aspect = view_create_info.subresourceRange.aspectMask; | |||
| 1963 | ||||
| 1964 | return 0; | |||
| 1965 | } | |||
| 1966 | ||||
| 1967 | int ff_vk_create_imageviews(FFVulkanContext *s, FFVkExecContext *e, | |||
| 1968 | VkImageView views[AV_NUM_DATA_POINTERS8], | |||
| 1969 | AVFrame *f, enum FFVkShaderRepFormat rep_fmt) | |||
| 1970 | { | |||
| 1971 | int err; | |||
| 1972 | VkResult ret; | |||
| 1973 | AVBufferRef *buf; | |||
| 1974 | FFVulkanFunctions *vk = &s->vkfn; | |||
| 1975 | AVHWFramesContext *hwfc = (AVHWFramesContext *)f->hw_frames_ctx->data; | |||
| 1976 | AVVulkanFramesContext *vkfc = hwfc->hwctx; | |||
| 1977 | const VkFormat *rep_fmts = av_vkfmt_from_pixfmt(hwfc->sw_format); | |||
| 1978 | AVVkFrame *vkf = (AVVkFrame *)f->data[0]; | |||
| 1979 | const int nb_images = ff_vk_count_images(vkf); | |||
| 1980 | const int nb_planes = av_pix_fmt_count_planes(hwfc->sw_format); | |||
| 1981 | ||||
| 1982 | ImageViewCtx *iv; | |||
| 1983 | const size_t buf_size = sizeof(*iv) + nb_planes*sizeof(VkImageView); | |||
| 1984 | iv = av_mallocz(buf_size); | |||
| 1985 | if (!iv) | |||
| 1986 | return AVERROR(ENOMEM)(-(12)); | |||
| 1987 | ||||
| 1988 | for (int i = 0; i < nb_planes; i++) { | |||
| 1989 | VkImageViewUsageCreateInfo view_usage_info = { | |||
| 1990 | .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_USAGE_CREATE_INFO, | |||
| 1991 | .usage = vkfc->usage & | |||
| 1992 | (~(VK_IMAGE_USAGE_VIDEO_ENCODE_SRC_BIT_KHR | | |||
| 1993 | VK_IMAGE_USAGE_VIDEO_DECODE_DST_BIT_KHR)), | |||
| 1994 | }; | |||
| 1995 | VkImageViewCreateInfo view_create_info = { | |||
| 1996 | .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO, | |||
| 1997 | .pNext = &view_usage_info, | |||
| 1998 | .image = vkf->img[FFMIN(i, nb_images - 1)((i) > (nb_images - 1) ? (nb_images - 1) : (i))], | |||
| 1999 | .viewType = VK_IMAGE_VIEW_TYPE_2D, | |||
| 2000 | .format = map_fmt_to_rep(rep_fmts[i], rep_fmt), | |||
| 2001 | .components = ff_comp_identity_map, | |||
| 2002 | .subresourceRange = { | |||
| 2003 | .aspectMask = ff_vk_aspect_flag(f, i), | |||
| 2004 | .levelCount = 1, | |||
| 2005 | .layerCount = 1, | |||
| 2006 | }, | |||
| 2007 | }; | |||
| 2008 | if (view_create_info.format == VK_FORMAT_UNDEFINED) { | |||
| 2009 | av_log(s, AV_LOG_ERROR16, "Unable to find a compatible representation " | |||
| 2010 | "of format %i and mode %i\n", | |||
| 2011 | rep_fmts[i], rep_fmt); | |||
| 2012 | err = AVERROR(EINVAL)(-(22)); | |||
| 2013 | goto fail; | |||
| 2014 | } | |||
| 2015 | ||||
| 2016 | ret = vk->CreateImageView(s->hwctx->act_dev, &view_create_info, | |||
| 2017 | s->hwctx->alloc, &iv->views[i]); | |||
| 2018 | if (ret != VK_SUCCESS) { | |||
| 2019 | av_log(s, AV_LOG_ERROR16, "Failed to create imageview: %s\n", | |||
| 2020 | ff_vk_ret2str(ret)); | |||
| 2021 | err = AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); | |||
| 2022 | goto fail; | |||
| 2023 | } | |||
| 2024 | ||||
| 2025 | iv->nb_views++; | |||
| 2026 | } | |||
| 2027 | ||||
| 2028 | buf = av_buffer_create((uint8_t *)iv, buf_size, destroy_imageviews, s, 0); | |||
| 2029 | if (!buf) { | |||
| 2030 | err = AVERROR(ENOMEM)(-(12)); | |||
| 2031 | goto fail; | |||
| 2032 | } | |||
| 2033 | ||||
| 2034 | /* Add to queue dependencies */ | |||
| 2035 | err = ff_vk_exec_add_dep_buf(s, e, &buf, 1, 0); | |||
| 2036 | if (err < 0) | |||
| 2037 | av_buffer_unref(&buf); | |||
| 2038 | ||||
| 2039 | memcpy(views, iv->views, nb_planes*sizeof(*views)); | |||
| 2040 | ||||
| 2041 | return err; | |||
| 2042 | ||||
| 2043 | fail: | |||
| 2044 | for (int i = 0; i < iv->nb_views; i++) | |||
| 2045 | vk->DestroyImageView(s->hwctx->act_dev, iv->views[i], s->hwctx->alloc); | |||
| 2046 | av_free(iv); | |||
| 2047 | return err; | |||
| 2048 | } | |||
| 2049 | ||||
| 2050 | void ff_vk_frame_barrier(FFVulkanContext *s, FFVkExecContext *e, | |||
| 2051 | AVFrame *pic, VkImageMemoryBarrier2 *bar, int *nb_bar, | |||
| 2052 | VkPipelineStageFlags2 src_stage, | |||
| 2053 | VkPipelineStageFlags2 dst_stage, | |||
| 2054 | VkAccessFlagBits2 new_access, | |||
| 2055 | VkImageLayout new_layout, | |||
| 2056 | uint32_t new_qf) | |||
| 2057 | { | |||
| 2058 | int found = -1; | |||
| 2059 | AVVkFrame *vkf = (AVVkFrame *)pic->data[0]; | |||
| 2060 | const int nb_images = ff_vk_count_images(vkf); | |||
| 2061 | for (int i = 0; i < e->nb_frame_deps; i++) | |||
| 2062 | if (e->frame_deps[i]->data[0] == pic->data[0]) { | |||
| 2063 | if (e->frame_update[i]) | |||
| 2064 | found = i; | |||
| 2065 | break; | |||
| 2066 | } | |||
| 2067 | ||||
| 2068 | for (int i = 0; i < nb_images; i++) { | |||
| 2069 | bar[*nb_bar] = (VkImageMemoryBarrier2) { | |||
| 2070 | .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER_2, | |||
| 2071 | .pNext = NULL((void*)0), | |||
| 2072 | .srcStageMask = src_stage, | |||
| 2073 | .dstStageMask = dst_stage, | |||
| 2074 | .srcAccessMask = found >= 0 ? e->access_dst[found] : vkf->access[i], | |||
| 2075 | .dstAccessMask = new_access, | |||
| 2076 | .oldLayout = found >= 0 ? e->layout_dst[found] : vkf->layout[0], | |||
| 2077 | .newLayout = new_layout, | |||
| 2078 | .srcQueueFamilyIndex = found >= 0 ? e->queue_family_dst[found] : vkf->queue_family[0], | |||
| 2079 | .dstQueueFamilyIndex = new_qf, | |||
| 2080 | .image = vkf->img[i], | |||
| 2081 | .subresourceRange = (VkImageSubresourceRange) { | |||
| 2082 | .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT, | |||
| 2083 | .layerCount = 1, | |||
| 2084 | .levelCount = 1, | |||
| 2085 | }, | |||
| 2086 | }; | |||
| 2087 | *nb_bar += 1; | |||
| 2088 | } | |||
| 2089 | ||||
| 2090 | ff_vk_exec_update_frame(s, e, pic, &bar[*nb_bar - nb_images], NULL((void*)0)); | |||
| 2091 | } | |||
| 2092 | ||||
| 2093 | int ff_vk_shader_load(FFVulkanShader *shd, | |||
| 2094 | VkPipelineStageFlags stage, VkSpecializationInfo *spec, | |||
| 2095 | uint32_t wg_size[3], uint32_t required_subgroup_size) | |||
| 2096 | { | |||
| 2097 | shd->stage = stage; | |||
| 2098 | shd->precompiled = 1; | |||
| 2099 | shd->specialization_info = spec; | |||
| 2100 | memcpy(shd->lg_size, wg_size, 3*sizeof(uint32_t)); | |||
| 2101 | ||||
| 2102 | switch (shd->stage) { | |||
| 2103 | case VK_SHADER_STAGE_ANY_HIT_BIT_KHR: | |||
| 2104 | case VK_SHADER_STAGE_CALLABLE_BIT_KHR: | |||
| 2105 | case VK_SHADER_STAGE_CLOSEST_HIT_BIT_KHR: | |||
| 2106 | case VK_SHADER_STAGE_INTERSECTION_BIT_KHR: | |||
| 2107 | case VK_SHADER_STAGE_MISS_BIT_KHR: | |||
| 2108 | case VK_SHADER_STAGE_RAYGEN_BIT_KHR: | |||
| 2109 | shd->bind_point = VK_PIPELINE_BIND_POINT_RAY_TRACING_KHR; | |||
| 2110 | break; | |||
| 2111 | case VK_SHADER_STAGE_COMPUTE_BIT: | |||
| 2112 | shd->bind_point = VK_PIPELINE_BIND_POINT_COMPUTE; | |||
| 2113 | break; | |||
| 2114 | default: | |||
| 2115 | shd->bind_point = VK_PIPELINE_BIND_POINT_GRAPHICS; | |||
| 2116 | break; | |||
| 2117 | }; | |||
| 2118 | ||||
| 2119 | return 0; | |||
| 2120 | } | |||
| 2121 | ||||
| 2122 | int ff_vk_shader_init(FFVulkanContext *s, FFVulkanShader *shd, const char *name, | |||
| 2123 | VkPipelineStageFlags stage, | |||
| 2124 | const char *extensions[], int nb_extensions, | |||
| 2125 | int lg_x, int lg_y, int lg_z, | |||
| 2126 | uint32_t required_subgroup_size) | |||
| 2127 | { | |||
| 2128 | ff_vk_shader_load(shd, stage, NULL((void*)0), | |||
| 2129 | (uint32_t []) { lg_x, lg_y, lg_z }, required_subgroup_size); | |||
| 2130 | ||||
| 2131 | shd->name = name; | |||
| 2132 | shd->precompiled = 0; | |||
| 2133 | av_bprint_init(&shd->src, 0, AV_BPRINT_SIZE_UNLIMITED((unsigned)-1)); | |||
| 2134 | ||||
| 2135 | if (required_subgroup_size) { | |||
| 2136 | shd->subgroup_info.sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_REQUIRED_SUBGROUP_SIZE_CREATE_INFO; | |||
| 2137 | shd->subgroup_info.requiredSubgroupSize = required_subgroup_size; | |||
| 2138 | } | |||
| 2139 | ||||
| 2140 | av_bprintf(&shd->src, "/* %s shader: %s */\n", | |||
| 2141 | (stage == VK_SHADER_STAGE_TASK_BIT_EXT || | |||
| 2142 | stage == VK_SHADER_STAGE_MESH_BIT_EXT) ? | |||
| 2143 | "Mesh" : | |||
| 2144 | (shd->bind_point == VK_PIPELINE_BIND_POINT_RAY_TRACING_KHR) ? | |||
| 2145 | "Raytrace" : | |||
| 2146 | (shd->bind_point == VK_PIPELINE_BIND_POINT_COMPUTE) ? | |||
| 2147 | "Compute" : "Graphics", | |||
| 2148 | name); | |||
| 2149 | GLSLF(0, #version %i ,460)do { av_bprintf(&shd->src, "#version %i" "\n", 460); } while (0); | |||
| 2150 | GLSLC(0, )do { av_bprintf(&shd->src, "" "\n"); } while (0); | |||
| 2151 | ||||
| 2152 | /* Common utilities */ | |||
| 2153 | GLSLC(0, #define IS_WITHIN(v1, v2) ((v1.x < v2.x) && (v1.y < v2.y)) )do { av_bprintf(&shd->src, "#define IS_WITHIN(v1, v2) ((v1.x < v2.x) && (v1.y < v2.y))" "\n"); } while (0); | |||
| 2154 | GLSLC(0, )do { av_bprintf(&shd->src, "" "\n"); } while (0); | |||
| 2155 | GLSLC(0, #extension GL_EXT_scalar_block_layout : require )do { av_bprintf(&shd->src, "#extension GL_EXT_scalar_block_layout : require" "\n"); } while (0); | |||
| 2156 | GLSLC(0, #extension GL_EXT_shader_explicit_arithmetic_types : require )do { av_bprintf(&shd->src, "#extension GL_EXT_shader_explicit_arithmetic_types : require" "\n"); } while (0); | |||
| 2157 | GLSLC(0, #extension GL_EXT_control_flow_attributes : require )do { av_bprintf(&shd->src, "#extension GL_EXT_control_flow_attributes : require" "\n"); } while (0); | |||
| 2158 | GLSLC(0, #extension GL_EXT_shader_image_load_formatted : require )do { av_bprintf(&shd->src, "#extension GL_EXT_shader_image_load_formatted : require" "\n"); } while (0); | |||
| 2159 | if (s->extensions & FF_VK_EXT_EXPECT_ASSUME(1ULL << 16)) { | |||
| 2160 | GLSLC(0, #extension GL_EXT_expect_assume : require )do { av_bprintf(&shd->src, "#extension GL_EXT_expect_assume : require" "\n"); } while (0); | |||
| 2161 | } else { | |||
| 2162 | GLSLC(0, #define assumeEXT(x) (x) )do { av_bprintf(&shd->src, "#define assumeEXT(x) (x)" "\n" ); } while (0); | |||
| 2163 | GLSLC(0, #define expectEXT(x, c) (x) )do { av_bprintf(&shd->src, "#define expectEXT(x, c) (x)" "\n"); } while (0); | |||
| 2164 | } | |||
| 2165 | if ((s->extensions & FF_VK_EXT_DEBUG_UTILS(1ULL << 5)) && | |||
| 2166 | (s->extensions & FF_VK_EXT_RELAXED_EXTENDED_INSTR(1ULL << 15))) { | |||
| 2167 | GLSLC(0, #extension GL_EXT_debug_printf : require )do { av_bprintf(&shd->src, "#extension GL_EXT_debug_printf : require" "\n"); } while (0); | |||
| 2168 | GLSLC(0, #define DEBUG )do { av_bprintf(&shd->src, "#define 1" "\n"); } while ( 0); | |||
| 2169 | } | |||
| 2170 | ||||
| 2171 | if (stage == VK_SHADER_STAGE_TASK_BIT_EXT || | |||
| 2172 | stage == VK_SHADER_STAGE_MESH_BIT_EXT) | |||
| 2173 | GLSLC(0, #extension GL_EXT_mesh_shader : require )do { av_bprintf(&shd->src, "#extension GL_EXT_mesh_shader : require" "\n"); } while (0); | |||
| 2174 | ||||
| 2175 | for (int i = 0; i < nb_extensions; i++) | |||
| 2176 | GLSLF(0, #extension %s : %s ,extensions[i], "require")do { av_bprintf(&shd->src, "#extension %s : %s" "\n", extensions [i], "require"); } while (0); | |||
| 2177 | GLSLC(0, )do { av_bprintf(&shd->src, "" "\n"); } while (0); | |||
| 2178 | ||||
| 2179 | GLSLF(0, layout (local_size_x = %i, local_size_y = %i, local_size_z = %i) in;do { av_bprintf(&shd->src, "layout (local_size_x = %i, local_size_y = %i, local_size_z = %i) in;" "\n", shd->lg_size[0], shd->lg_size[1], shd->lg_size [2]); } while (0) | |||
| 2180 | , shd->lg_size[0], shd->lg_size[1], shd->lg_size[2])do { av_bprintf(&shd->src, "layout (local_size_x = %i, local_size_y = %i, local_size_z = %i) in;" "\n", shd->lg_size[0], shd->lg_size[1], shd->lg_size [2]); } while (0); | |||
| 2181 | GLSLC(0, )do { av_bprintf(&shd->src, "" "\n"); } while (0); | |||
| 2182 | ||||
| 2183 | return 0; | |||
| 2184 | } | |||
| 2185 | ||||
| 2186 | void ff_vk_shader_print(void *ctx, FFVulkanShader *shd, int prio) | |||
| 2187 | { | |||
| 2188 | int line = 0; | |||
| 2189 | const char *p = shd->src.str; | |||
| 2190 | const char *start = p; | |||
| 2191 | const size_t len = strlen(p); | |||
| 2192 | ||||
| 2193 | AVBPrint buf; | |||
| 2194 | av_bprint_init(&buf, 0, AV_BPRINT_SIZE_UNLIMITED((unsigned)-1)); | |||
| 2195 | ||||
| 2196 | for (int i = 0; i < len; i++) { | |||
| 2197 | if (p[i] == '\n') { | |||
| 2198 | av_bprintf(&buf, "%i\t", ++line); | |||
| 2199 | av_bprint_append_data(&buf, start, &p[i] - start + 1); | |||
| 2200 | start = &p[i + 1]; | |||
| 2201 | } | |||
| 2202 | } | |||
| 2203 | ||||
| 2204 | av_log(ctx, prio, "Shader %s: \n%s", shd->name, buf.str); | |||
| 2205 | av_bprint_finalize(&buf, NULL((void*)0)); | |||
| 2206 | } | |||
| 2207 | ||||
| 2208 | static int init_pipeline_layout(FFVulkanContext *s, FFVulkanShader *shd) | |||
| 2209 | { | |||
| 2210 | VkResult ret; | |||
| 2211 | FFVulkanFunctions *vk = &s->vkfn; | |||
| 2212 | VkPipelineLayoutCreateInfo pipeline_layout_info; | |||
| 2213 | ||||
| 2214 | /* Finally create the pipeline layout */ | |||
| 2215 | pipeline_layout_info = (VkPipelineLayoutCreateInfo) { | |||
| 2216 | .sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO, | |||
| 2217 | .pSetLayouts = shd->desc_layout, | |||
| 2218 | .setLayoutCount = shd->nb_descriptor_sets, | |||
| 2219 | .pushConstantRangeCount = shd->push_consts_num, | |||
| 2220 | .pPushConstantRanges = shd->push_consts, | |||
| 2221 | }; | |||
| 2222 | ||||
| 2223 | ret = vk->CreatePipelineLayout(s->hwctx->act_dev, &pipeline_layout_info, | |||
| 2224 | s->hwctx->alloc, &shd->pipeline_layout); | |||
| 2225 | if (ret != VK_SUCCESS) { | |||
| 2226 | av_log(s, AV_LOG_ERROR16, "Unable to init pipeline layout: %s\n", | |||
| 2227 | ff_vk_ret2str(ret)); | |||
| 2228 | return AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); | |||
| 2229 | } | |||
| 2230 | ||||
| 2231 | return 0; | |||
| 2232 | } | |||
| 2233 | ||||
| 2234 | static int create_shader_module(FFVulkanContext *s, FFVulkanShader *shd, | |||
| 2235 | VkShaderModule *mod, | |||
| 2236 | const uint8_t *spirv, size_t spirv_len) | |||
| 2237 | { | |||
| 2238 | VkResult ret; | |||
| 2239 | FFVulkanFunctions *vk = &s->vkfn; | |||
| 2240 | ||||
| 2241 | VkShaderModuleCreateInfo shader_module_info = { | |||
| 2242 | .sType = VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO, | |||
| 2243 | .pNext = NULL((void*)0), | |||
| 2244 | .flags = 0x0, | |||
| 2245 | .pCode = (void *)spirv, | |||
| 2246 | .codeSize = spirv_len, | |||
| 2247 | }; | |||
| 2248 | ||||
| 2249 | ret = vk->CreateShaderModule(s->hwctx->act_dev, &shader_module_info, | |||
| 2250 | s->hwctx->alloc, mod); | |||
| 2251 | if (ret != VK_SUCCESS) { | |||
| 2252 | av_log(s, AV_LOG_ERROR16, "Error creating shader module: %s\n", | |||
| 2253 | ff_vk_ret2str(ret)); | |||
| 2254 | return AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); | |||
| 2255 | } | |||
| 2256 | ||||
| 2257 | return 0; | |||
| 2258 | } | |||
| 2259 | ||||
| 2260 | static int init_compute_pipeline(FFVulkanContext *s, FFVulkanShader *shd, | |||
| 2261 | VkShaderModule mod, const char *entrypoint) | |||
| 2262 | { | |||
| 2263 | VkResult ret; | |||
| 2264 | FFVulkanFunctions *vk = &s->vkfn; | |||
| 2265 | ||||
| 2266 | VkComputePipelineCreateInfo pipeline_create_info = { | |||
| 2267 | .sType = VK_STRUCTURE_TYPE_COMPUTE_PIPELINE_CREATE_INFO, | |||
| 2268 | .flags = 0x0, | |||
| 2269 | .layout = shd->pipeline_layout, | |||
| 2270 | .stage = (VkPipelineShaderStageCreateInfo) { | |||
| 2271 | .sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO, | |||
| 2272 | .pNext = shd->subgroup_info.requiredSubgroupSize ? | |||
| 2273 | &shd->subgroup_info : NULL((void*)0), | |||
| 2274 | .pName = entrypoint, | |||
| 2275 | .flags = shd->subgroup_info.requiredSubgroupSize ? | |||
| 2276 | VK_PIPELINE_SHADER_STAGE_CREATE_REQUIRE_FULL_SUBGROUPS_BIT : 0x0, | |||
| 2277 | .stage = shd->stage, | |||
| 2278 | .module = mod, | |||
| 2279 | .pSpecializationInfo = shd->specialization_info, | |||
| 2280 | }, | |||
| 2281 | }; | |||
| 2282 | ||||
| 2283 | ret = vk->CreateComputePipelines(s->hwctx->act_dev, VK_NULL_HANDLE((void*)0), 1, | |||
| 2284 | &pipeline_create_info, | |||
| 2285 | s->hwctx->alloc, &shd->pipeline); | |||
| 2286 | if (ret != VK_SUCCESS) { | |||
| 2287 | av_log(s, AV_LOG_ERROR16, "Unable to init compute pipeline: %s\n", | |||
| 2288 | ff_vk_ret2str(ret)); | |||
| 2289 | return AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); | |||
| 2290 | } | |||
| 2291 | ||||
| 2292 | return 0; | |||
| 2293 | } | |||
| 2294 | ||||
| 2295 | static int create_shader_object(FFVulkanContext *s, FFVulkanShader *shd, | |||
| 2296 | const uint8_t *spirv, size_t spirv_len, | |||
| 2297 | size_t *binary_size, const char *entrypoint) | |||
| 2298 | { | |||
| 2299 | VkResult ret; | |||
| 2300 | FFVulkanFunctions *vk = &s->vkfn; | |||
| 2301 | ||||
| 2302 | VkShaderCreateInfoEXT shader_obj_create = { | |||
| 2303 | .sType = VK_STRUCTURE_TYPE_SHADER_CREATE_INFO_EXT, | |||
| 2304 | .flags = shd->subgroup_info.requiredSubgroupSize ? | |||
| 2305 | VK_SHADER_CREATE_REQUIRE_FULL_SUBGROUPS_BIT_EXT : 0x0, | |||
| 2306 | .stage = shd->stage, | |||
| 2307 | .nextStage = 0, | |||
| 2308 | .codeType = VK_SHADER_CODE_TYPE_SPIRV_EXT, | |||
| 2309 | .pCode = spirv, | |||
| 2310 | .codeSize = spirv_len, | |||
| 2311 | .pName = entrypoint, | |||
| 2312 | .pSetLayouts = shd->desc_layout, | |||
| 2313 | .setLayoutCount = shd->nb_descriptor_sets, | |||
| 2314 | .pushConstantRangeCount = shd->push_consts_num, | |||
| 2315 | .pPushConstantRanges = shd->push_consts, | |||
| 2316 | .pSpecializationInfo = shd->specialization_info, | |||
| 2317 | }; | |||
| 2318 | ||||
| 2319 | ret = vk->CreateShadersEXT(s->hwctx->act_dev, 1, &shader_obj_create, | |||
| 2320 | s->hwctx->alloc, &shd->object); | |||
| 2321 | if (ret != VK_SUCCESS) { | |||
| 2322 | av_log(s, AV_LOG_ERROR16, "Unable to create shader object: %s\n", | |||
| 2323 | ff_vk_ret2str(ret)); | |||
| 2324 | return AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); | |||
| 2325 | } | |||
| 2326 | ||||
| 2327 | if (vk->GetShaderBinaryDataEXT(s->hwctx->act_dev, shd->object, | |||
| 2328 | binary_size, NULL((void*)0)) != VK_SUCCESS) | |||
| 2329 | return AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); | |||
| 2330 | ||||
| 2331 | return 0; | |||
| 2332 | } | |||
| 2333 | ||||
| 2334 | static int init_descriptors(FFVulkanContext *s, FFVulkanShader *shd) | |||
| 2335 | { | |||
| 2336 | VkResult ret; | |||
| 2337 | FFVulkanFunctions *vk = &s->vkfn; | |||
| 2338 | ||||
| 2339 | int has_singular = 0; | |||
| 2340 | int max_descriptors = 0; | |||
| 2341 | for (int i = 0; i < shd->nb_descriptor_sets; i++) { | |||
| 2342 | max_descriptors = FFMAX(max_descriptors, shd->desc_set[i].nb_bindings)((max_descriptors) > (shd->desc_set[i].nb_bindings) ? ( max_descriptors) : (shd->desc_set[i].nb_bindings)); | |||
| 2343 | if (shd->desc_set[i].singular) | |||
| 2344 | has_singular = 1; | |||
| 2345 | } | |||
| 2346 | shd->use_push = (s->extensions & FF_VK_EXT_PUSH_DESCRIPTOR(1ULL << 14)) && | |||
| 2347 | (max_descriptors <= s->push_desc_props.maxPushDescriptors) && | |||
| 2348 | (shd->nb_descriptor_sets == 1) && | |||
| 2349 | (has_singular == 0); | |||
| 2350 | ||||
| 2351 | for (int i = 0; i < shd->nb_descriptor_sets; i++) { | |||
| 2352 | FFVulkanDescriptorSet *set = &shd->desc_set[i]; | |||
| 2353 | VkDescriptorSetLayoutCreateInfo desc_layout_create = { | |||
| 2354 | .sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO, | |||
| 2355 | .bindingCount = set->nb_bindings, | |||
| 2356 | .pBindings = set->binding, | |||
| 2357 | .flags = (shd->use_push) ? | |||
| 2358 | VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR : | |||
| 2359 | 0x0, | |||
| 2360 | }; | |||
| 2361 | ||||
| 2362 | ret = vk->CreateDescriptorSetLayout(s->hwctx->act_dev, | |||
| 2363 | &desc_layout_create, | |||
| 2364 | s->hwctx->alloc, | |||
| 2365 | &shd->desc_layout[i]); | |||
| 2366 | if (ret != VK_SUCCESS) { | |||
| 2367 | av_log(s, AV_LOG_ERROR16, "Unable to create descriptor set layout: %s", | |||
| 2368 | ff_vk_ret2str(ret)); | |||
| 2369 | return AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); | |||
| 2370 | } | |||
| 2371 | } | |||
| 2372 | ||||
| 2373 | return 0; | |||
| 2374 | } | |||
| 2375 | ||||
| 2376 | int ff_vk_shader_link(FFVulkanContext *s, FFVulkanShader *shd, | |||
| 2377 | const char *spirv, size_t spirv_len, | |||
| 2378 | const char *entrypoint) | |||
| 2379 | { | |||
| 2380 | int err; | |||
| 2381 | FFVulkanFunctions *vk = &s->vkfn; | |||
| 2382 | VkSpecializationMapEntry spec_entries[3]; | |||
| 2383 | VkSpecializationInfo spec_info; | |||
| 2384 | size_t input_size = spirv_len, binary_size = 0; | |||
| 2385 | ||||
| 2386 | if (shd->precompiled) { | |||
| 2387 | if (!shd->specialization_info) { | |||
| 2388 | spec_info = (VkSpecializationInfo) { | |||
| 2389 | .pMapEntries = spec_entries, | |||
| 2390 | .mapEntryCount = 0, | |||
| 2391 | .pData = shd->lg_size, | |||
| 2392 | .dataSize = 0, | |||
| 2393 | }; | |||
| 2394 | shd->specialization_info = &spec_info; | |||
| 2395 | } | |||
| 2396 | ||||
| 2397 | VkSpecializationMapEntry *spe = (void *)shd->specialization_info->pMapEntries; | |||
| 2398 | for (int i = 0; i < 3; i++) { | |||
| 2399 | spe[shd->specialization_info->mapEntryCount++] = (VkSpecializationMapEntry) { | |||
| 2400 | .constantID = 253 + i, | |||
| 2401 | .offset = shd->specialization_info->dataSize + i*sizeof(uint32_t), | |||
| 2402 | .size = sizeof(uint32_t), | |||
| 2403 | }; | |||
| 2404 | } | |||
| 2405 | ||||
| 2406 | uint8_t *spd = (uint8_t *)shd->specialization_info->pData; | |||
| 2407 | memcpy(&spd[shd->specialization_info->dataSize], | |||
| 2408 | shd->lg_size, 3*sizeof(uint32_t)); | |||
| 2409 | shd->specialization_info->dataSize += 3*sizeof(uint32_t); | |||
| 2410 | ||||
| 2411 | #if CONFIG_SHADER_COMPRESSION0 | |||
| 2412 | uint8_t *out; | |||
| 2413 | size_t out_len; | |||
| 2414 | int ret = ff_zlib_expand(s, &out, &out_len, spirv, spirv_len); | |||
| 2415 | if (ret < 0) | |||
| 2416 | return ret; | |||
| 2417 | spirv = out; | |||
| 2418 | spirv_len = out_len; | |||
| 2419 | #endif | |||
| 2420 | } | |||
| 2421 | ||||
| 2422 | err = init_descriptors(s, shd); | |||
| 2423 | if (err < 0) | |||
| 2424 | goto end; | |||
| 2425 | ||||
| 2426 | err = init_pipeline_layout(s, shd); | |||
| 2427 | if (err < 0) | |||
| 2428 | goto end; | |||
| 2429 | ||||
| 2430 | if (s->extensions & FF_VK_EXT_SHADER_OBJECT(1ULL << 13)) { | |||
| 2431 | err = create_shader_object(s, shd, spirv, spirv_len, | |||
| 2432 | &binary_size, entrypoint); | |||
| 2433 | if (err < 0) | |||
| 2434 | goto end; | |||
| 2435 | } else { | |||
| 2436 | VkShaderModule mod; | |||
| 2437 | err = create_shader_module(s, shd, &mod, spirv, spirv_len); | |||
| 2438 | if (err < 0) | |||
| 2439 | goto end; | |||
| 2440 | ||||
| 2441 | switch (shd->bind_point) { | |||
| 2442 | case VK_PIPELINE_BIND_POINT_COMPUTE: | |||
| 2443 | err = init_compute_pipeline(s, shd, mod, entrypoint); | |||
| 2444 | break; | |||
| 2445 | default: | |||
| 2446 | av_log(s, AV_LOG_ERROR16, "Unsupported shader type: %i\n", | |||
| 2447 | shd->bind_point); | |||
| 2448 | err = AVERROR(EINVAL)(-(22)); | |||
| 2449 | goto end; | |||
| 2450 | break; | |||
| 2451 | }; | |||
| 2452 | ||||
| 2453 | vk->DestroyShaderModule(s->hwctx->act_dev, mod, s->hwctx->alloc); | |||
| 2454 | if (err < 0) | |||
| 2455 | goto end; | |||
| 2456 | } | |||
| 2457 | ||||
| 2458 | if (shd->name) | |||
| 2459 | av_log(s, AV_LOG_VERBOSE40, "Shader %s linked, size:", shd->name); | |||
| 2460 | else | |||
| 2461 | av_log(s, AV_LOG_VERBOSE40, "Shader linked, size:"); | |||
| 2462 | ||||
| 2463 | if (input_size != spirv_len) | |||
| 2464 | av_log(s, AV_LOG_VERBOSE40, " %zu compressed,", input_size); | |||
| 2465 | av_log(s, AV_LOG_VERBOSE40, " %zu SPIR-V", spirv_len); | |||
| 2466 | if (binary_size != spirv_len) | |||
| 2467 | av_log(s, AV_LOG_VERBOSE40, ", %zu binary", spirv_len); | |||
| 2468 | av_log(s, AV_LOG_VERBOSE40, "\n"); | |||
| 2469 | ||||
| 2470 | end: | |||
| 2471 | if (shd->precompiled) { | |||
| 2472 | shd->specialization_info->mapEntryCount -= 3; | |||
| 2473 | shd->specialization_info->dataSize -= 3*sizeof(uint32_t); | |||
| 2474 | } | |||
| 2475 | #if CONFIG_SHADER_COMPRESSION0 | |||
| 2476 | if (shd->precompiled) | |||
| 2477 | av_free((void *)spirv); | |||
| 2478 | #endif | |||
| 2479 | return err; | |||
| 2480 | } | |||
| 2481 | ||||
| 2482 | static const struct descriptor_props { | |||
| 2483 | size_t struct_size; /* Size of the opaque which updates the descriptor */ | |||
| 2484 | const char *type; | |||
| 2485 | int is_uniform; | |||
| 2486 | int mem_quali; /* Can use a memory qualifier */ | |||
| 2487 | int dim_needed; /* Must indicate dimension */ | |||
| 2488 | int buf_content; /* Must indicate buffer contents */ | |||
| 2489 | } descriptor_props[] = { | |||
| 2490 | [VK_DESCRIPTOR_TYPE_SAMPLER] = { sizeof(VkDescriptorImageInfo), "sampler", 1, 0, 0, 0, }, | |||
| 2491 | [VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE] = { sizeof(VkDescriptorImageInfo), "texture", 1, 0, 1, 0, }, | |||
| 2492 | [VK_DESCRIPTOR_TYPE_STORAGE_IMAGE] = { sizeof(VkDescriptorImageInfo), "image", 1, 1, 1, 0, }, | |||
| 2493 | [VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT] = { sizeof(VkDescriptorImageInfo), "subpassInput", 1, 0, 0, 0, }, | |||
| 2494 | [VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER] = { sizeof(VkDescriptorImageInfo), "sampler", 1, 0, 1, 0, }, | |||
| 2495 | [VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER] = { sizeof(VkDescriptorBufferInfo), NULL((void*)0), 1, 0, 0, 1, }, | |||
| 2496 | [VK_DESCRIPTOR_TYPE_STORAGE_BUFFER] = { sizeof(VkDescriptorBufferInfo), "buffer", 0, 1, 0, 1, }, | |||
| 2497 | [VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC] = { sizeof(VkDescriptorBufferInfo), NULL((void*)0), 1, 0, 0, 1, }, | |||
| 2498 | [VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC] = { sizeof(VkDescriptorBufferInfo), "buffer", 0, 1, 0, 1, }, | |||
| 2499 | [VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER] = { sizeof(VkBufferView), "samplerBuffer", 1, 0, 0, 0, }, | |||
| 2500 | [VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER] = { sizeof(VkBufferView), "imageBuffer", 1, 0, 0, 0, }, | |||
| 2501 | }; | |||
| 2502 | ||||
| 2503 | int ff_vk_shader_add_descriptor_set(FFVulkanContext *s, FFVulkanShader *shd, | |||
| 2504 | const FFVulkanDescriptorSetBinding *desc, int nb, | |||
| 2505 | int singular, int print_to_shader_only) | |||
| 2506 | { | |||
| 2507 | if (print_to_shader_only) | |||
| 2508 | goto print; | |||
| 2509 | ||||
| 2510 | FFVulkanDescriptorSet *set = &shd->desc_set[shd->nb_descriptor_sets++]; | |||
| 2511 | av_assert1(shd->nb_descriptor_sets < FF_VK_MAX_DESCRIPTOR_SETS)do { if (!(shd->nb_descriptor_sets < 4)) { av_log(((void *)0), 0, "Assertion %s failed at %s:%d\n", "shd->nb_descriptor_sets < 4" , "/root/firefox-clang/media/ffvpx/libavutil/vulkan.c", 2511) ; abort(); } } while (0); | |||
| 2512 | av_assert1(nb < FF_VK_MAX_DESCRIPTOR_BINDINGS)do { if (!(nb < 16)) { av_log(((void*)0), 0, "Assertion %s failed at %s:%d\n" , "nb < 16", "/root/firefox-clang/media/ffvpx/libavutil/vulkan.c" , 2512); abort(); } } while (0); | |||
| 2513 | ||||
| 2514 | for (int i = 0; i < nb; i++) { | |||
| 2515 | set->binding[i].binding = i; | |||
| 2516 | set->binding[i].descriptorType = desc[i].type; | |||
| 2517 | set->binding[i].descriptorCount = FFMAX(desc[i].elems, 1)((desc[i].elems) > (1) ? (desc[i].elems) : (1)); | |||
| 2518 | set->binding[i].stageFlags = desc[i].stages; | |||
| 2519 | set->binding[i].pImmutableSamplers = desc[i].samplers; | |||
| 2520 | } | |||
| 2521 | ||||
| 2522 | for (int i = 0; i < nb; i++) { | |||
| 2523 | int j; | |||
| 2524 | for (j = 0; j < shd->nb_desc_pool_size; j++) | |||
| 2525 | if (shd->desc_pool_size[j].type == desc[i].type) | |||
| 2526 | break; | |||
| 2527 | if (j >= shd->nb_desc_pool_size) { | |||
| 2528 | shd->nb_desc_pool_size++; | |||
| 2529 | av_assert1(shd->nb_desc_pool_size < FF_VK_MAX_DESCRIPTOR_TYPES)do { if (!(shd->nb_desc_pool_size < 16)) { av_log(((void *)0), 0, "Assertion %s failed at %s:%d\n", "shd->nb_desc_pool_size < 16" , "/root/firefox-clang/media/ffvpx/libavutil/vulkan.c", 2529) ; abort(); } } while (0); | |||
| 2530 | } | |||
| 2531 | ||||
| 2532 | shd->desc_pool_size[j].type = desc[i].type; | |||
| 2533 | shd->desc_pool_size[j].descriptorCount += FFMAX(desc[i].elems, 1)((desc[i].elems) > (1) ? (desc[i].elems) : (1)); | |||
| 2534 | } | |||
| 2535 | ||||
| 2536 | set->singular = singular; | |||
| 2537 | set->nb_bindings = nb; | |||
| 2538 | ||||
| 2539 | if (shd->precompiled) | |||
| 2540 | return 0; | |||
| 2541 | ||||
| 2542 | print: | |||
| 2543 | /* Write shader info */ | |||
| 2544 | for (int i = 0; i < nb; i++) { | |||
| 2545 | const struct descriptor_props *prop = &descriptor_props[desc[i].type]; | |||
| 2546 | GLSLA("layout (set = %i, binding = %i", FFMAX(shd->nb_descriptor_sets - 1, 0), i)do { av_bprintf(&shd->src, "layout (set = %i, binding = %i" , ((shd->nb_descriptor_sets - 1) > (0) ? (shd->nb_descriptor_sets - 1) : (0)), i); } while (0); | |||
| 2547 | ||||
| 2548 | if (desc[i].mem_layout && | |||
| 2549 | (desc[i].type != VK_DESCRIPTOR_TYPE_STORAGE_IMAGE)) | |||
| 2550 | GLSLA(", %s", desc[i].mem_layout)do { av_bprintf(&shd->src, ", %s", desc[i].mem_layout) ; } while (0); | |||
| 2551 | ||||
| 2552 | GLSLA(")")do { av_bprintf(&shd->src, ")"); } while (0); | |||
| 2553 | ||||
| 2554 | if (prop->is_uniform) | |||
| 2555 | GLSLA(" uniform")do { av_bprintf(&shd->src, " uniform"); } while (0); | |||
| 2556 | ||||
| 2557 | if (prop->mem_quali && desc[i].mem_quali) | |||
| 2558 | GLSLA(" %s", desc[i].mem_quali)do { av_bprintf(&shd->src, " %s", desc[i].mem_quali); } while (0); | |||
| 2559 | ||||
| 2560 | if (prop->type) { | |||
| 2561 | GLSLA(" ")do { av_bprintf(&shd->src, " "); } while (0); | |||
| 2562 | if (desc[i].type == VK_DESCRIPTOR_TYPE_STORAGE_IMAGE) { | |||
| 2563 | if (desc[i].mem_layout) { | |||
| 2564 | int len = strlen(desc[i].mem_layout); | |||
| 2565 | if (desc[i].mem_layout[len - 1] == 'i' && | |||
| 2566 | desc[i].mem_layout[len - 2] == 'u') { | |||
| 2567 | GLSLA("u")do { av_bprintf(&shd->src, "u"); } while (0); | |||
| 2568 | } else if (desc[i].mem_layout[len - 1] == 'i') { | |||
| 2569 | GLSLA("i")do { av_bprintf(&shd->src, "i"); } while (0); | |||
| 2570 | } | |||
| 2571 | } | |||
| 2572 | } | |||
| 2573 | GLSLA("%s", prop->type)do { av_bprintf(&shd->src, "%s", prop->type); } while (0); | |||
| 2574 | } | |||
| 2575 | ||||
| 2576 | if (prop->dim_needed) | |||
| 2577 | GLSLA("%iD", desc[i].dimensions)do { av_bprintf(&shd->src, "%iD", desc[i].dimensions); } while (0); | |||
| 2578 | ||||
| 2579 | GLSLA(" %s", desc[i].name)do { av_bprintf(&shd->src, " %s", desc[i].name); } while (0); | |||
| 2580 | ||||
| 2581 | if (prop->buf_content) { | |||
| 2582 | GLSLA(" {\n ")do { av_bprintf(&shd->src, " {\n "); } while (0); | |||
| 2583 | if (desc[i].buf_elems) { | |||
| 2584 | GLSLA("%s", desc[i].buf_content)do { av_bprintf(&shd->src, "%s", desc[i].buf_content); } while (0); | |||
| 2585 | GLSLA("[%i];", desc[i].buf_elems)do { av_bprintf(&shd->src, "[%i];", desc[i].buf_elems) ; } while (0); | |||
| 2586 | } else { | |||
| 2587 | GLSLA("%s", desc[i].buf_content)do { av_bprintf(&shd->src, "%s", desc[i].buf_content); } while (0); | |||
| 2588 | } | |||
| 2589 | GLSLA("\n}")do { av_bprintf(&shd->src, "\n}"); } while (0); | |||
| 2590 | } | |||
| 2591 | ||||
| 2592 | if (desc[i].elems > 0) | |||
| 2593 | GLSLA("[%i]", desc[i].elems)do { av_bprintf(&shd->src, "[%i]", desc[i].elems); } while (0); | |||
| 2594 | ||||
| 2595 | GLSLA(";")do { av_bprintf(&shd->src, ";"); } while (0); | |||
| 2596 | GLSLA("\n")do { av_bprintf(&shd->src, "\n"); } while (0); | |||
| 2597 | } | |||
| 2598 | GLSLA("\n")do { av_bprintf(&shd->src, "\n"); } while (0); | |||
| 2599 | ||||
| 2600 | return 0; | |||
| 2601 | } | |||
| 2602 | ||||
| 2603 | int ff_vk_shader_register_exec(FFVulkanContext *s, FFVkExecPool *pool, | |||
| 2604 | FFVulkanShader *shd) | |||
| 2605 | { | |||
| 2606 | if (!shd->nb_descriptor_sets) | |||
| 2607 | return 0; | |||
| 2608 | ||||
| 2609 | FFVulkanShaderData *sd = &pool->reg_shd[pool->nb_reg_shd++]; | |||
| 2610 | av_assert1(pool->nb_reg_shd < FF_VK_MAX_SHADERS)do { if (!(pool->nb_reg_shd < 16)) { av_log(((void*)0), 0, "Assertion %s failed at %s:%d\n", "pool->nb_reg_shd < 16" , "/root/firefox-clang/media/ffvpx/libavutil/vulkan.c", 2610) ; abort(); } } while (0); | |||
| 2611 | ||||
| 2612 | sd->shd = shd; | |||
| 2613 | sd->nb_descriptor_sets = shd->nb_descriptor_sets; | |||
| 2614 | ||||
| 2615 | if (!shd->use_push) { | |||
| 2616 | VkResult ret; | |||
| 2617 | FFVulkanFunctions *vk = &s->vkfn; | |||
| 2618 | VkDescriptorSetLayout *tmp_layouts; | |||
| 2619 | VkDescriptorSetAllocateInfo set_alloc_info; | |||
| 2620 | VkDescriptorPoolCreateInfo pool_create_info; | |||
| 2621 | ||||
| 2622 | for (int i = 0; i < shd->nb_desc_pool_size; i++) | |||
| 2623 | shd->desc_pool_size[i].descriptorCount *= pool->pool_size; | |||
| 2624 | ||||
| 2625 | pool_create_info = (VkDescriptorPoolCreateInfo) { | |||
| 2626 | .sType = VK_STRUCTURE_TYPE_DESCRIPTOR_POOL_CREATE_INFO, | |||
| 2627 | .flags = 0, | |||
| 2628 | .pPoolSizes = shd->desc_pool_size, | |||
| 2629 | .poolSizeCount = shd->nb_desc_pool_size, | |||
| 2630 | .maxSets = sd->nb_descriptor_sets*pool->pool_size, | |||
| 2631 | }; | |||
| 2632 | ||||
| 2633 | ret = vk->CreateDescriptorPool(s->hwctx->act_dev, &pool_create_info, | |||
| 2634 | s->hwctx->alloc, &sd->desc_pool); | |||
| 2635 | if (ret != VK_SUCCESS) { | |||
| 2636 | av_log(s, AV_LOG_ERROR16, "Unable to create descriptor pool: %s\n", | |||
| 2637 | ff_vk_ret2str(ret)); | |||
| 2638 | return AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); | |||
| 2639 | } | |||
| 2640 | ||||
| 2641 | tmp_layouts = av_malloc_array(pool_create_info.maxSets, sizeof(*tmp_layouts)); | |||
| 2642 | if (!tmp_layouts) | |||
| 2643 | return AVERROR(ENOMEM)(-(12)); | |||
| 2644 | ||||
| 2645 | /* Colate each execution context's descriptor set layouts */ | |||
| 2646 | for (int i = 0; i < pool->pool_size; i++) | |||
| 2647 | for (int j = 0; j < sd->nb_descriptor_sets; j++) | |||
| 2648 | tmp_layouts[i*sd->nb_descriptor_sets + j] = shd->desc_layout[j]; | |||
| 2649 | ||||
| 2650 | set_alloc_info = (VkDescriptorSetAllocateInfo) { | |||
| 2651 | .sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_ALLOCATE_INFO, | |||
| 2652 | .descriptorPool = sd->desc_pool, | |||
| 2653 | .pSetLayouts = tmp_layouts, | |||
| 2654 | .descriptorSetCount = pool_create_info.maxSets, | |||
| 2655 | }; | |||
| 2656 | ||||
| 2657 | sd->desc_sets = av_malloc_array(pool_create_info.maxSets, | |||
| 2658 | sizeof(*tmp_layouts)); | |||
| 2659 | if (!sd->desc_sets) { | |||
| 2660 | av_free(tmp_layouts); | |||
| 2661 | return AVERROR(ENOMEM)(-(12)); | |||
| 2662 | } | |||
| 2663 | ret = vk->AllocateDescriptorSets(s->hwctx->act_dev, &set_alloc_info, | |||
| 2664 | sd->desc_sets); | |||
| 2665 | av_free(tmp_layouts); | |||
| 2666 | if (ret != VK_SUCCESS) { | |||
| 2667 | av_log(s, AV_LOG_ERROR16, "Unable to allocate descriptor set: %s\n", | |||
| 2668 | ff_vk_ret2str(ret)); | |||
| 2669 | av_freep(&sd->desc_sets); | |||
| 2670 | return AVERROR_EXTERNAL(-(int)(('E') | (('X') << 8) | (('T') << 16) | (( unsigned)(' ') << 24))); | |||
| 2671 | } | |||
| 2672 | } | |||
| 2673 | ||||
| 2674 | return 0; | |||
| 2675 | } | |||
| 2676 | ||||
| 2677 | static inline const FFVulkanShaderData *get_shd_data(FFVkExecContext *e, | |||
| 2678 | const FFVulkanShader *shd) | |||
| 2679 | { | |||
| 2680 | for (int i = 0; i < e->parent->nb_reg_shd; i++) | |||
| 2681 | if (e->parent->reg_shd[i].shd == shd) | |||
| 2682 | return &e->parent->reg_shd[i]; | |||
| 2683 | return NULL((void*)0); | |||
| 2684 | } | |||
| 2685 | ||||
| 2686 | static inline void update_set_pool_write(FFVulkanContext *s, FFVkExecContext *e, | |||
| 2687 | FFVulkanShader *shd, int set, | |||
| 2688 | VkWriteDescriptorSet *write_info) | |||
| 2689 | { | |||
| 2690 | FFVulkanFunctions *vk = &s->vkfn; | |||
| 2691 | FFVulkanDescriptorSet *desc_set = &shd->desc_set[set]; | |||
| 2692 | const FFVulkanShaderData *sd = get_shd_data(e, shd); | |||
| 2693 | ||||
| 2694 | if (desc_set->singular) { | |||
| 2695 | for (int i = 0; i < e->parent->pool_size; i++) { | |||
| 2696 | write_info->dstSet = sd->desc_sets[i*sd->nb_descriptor_sets + set]; | |||
| 2697 | vk->UpdateDescriptorSets(s->hwctx->act_dev, 1, write_info, 0, NULL((void*)0)); | |||
| 2698 | } | |||
| 2699 | } else { | |||
| 2700 | if (shd->use_push) { | |||
| 2701 | vk->CmdPushDescriptorSetKHR(e->buf, | |||
| 2702 | shd->bind_point, | |||
| 2703 | shd->pipeline_layout, | |||
| 2704 | set, 1, | |||
| 2705 | write_info); | |||
| 2706 | } else { | |||
| 2707 | write_info->dstSet = sd->desc_sets[e->idx*sd->nb_descriptor_sets + set]; | |||
| 2708 | vk->UpdateDescriptorSets(s->hwctx->act_dev, 1, write_info, 0, NULL((void*)0)); | |||
| 2709 | } | |||
| 2710 | } | |||
| 2711 | } | |||
| 2712 | ||||
| 2713 | int ff_vk_shader_update_img(FFVulkanContext *s, FFVkExecContext *e, | |||
| 2714 | FFVulkanShader *shd, int set, int bind, int offs, | |||
| 2715 | VkImageView view, VkImageLayout layout, | |||
| 2716 | VkSampler sampler) | |||
| 2717 | { | |||
| 2718 | FFVulkanDescriptorSet *desc_set = &shd->desc_set[set]; | |||
| 2719 | ||||
| 2720 | VkDescriptorImageInfo desc_pool_write_info_img = { | |||
| 2721 | .sampler = sampler, | |||
| 2722 | .imageView = view, | |||
| 2723 | .imageLayout = layout, | |||
| 2724 | }; | |||
| 2725 | VkWriteDescriptorSet desc_pool_write_info = { | |||
| 2726 | .sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET, | |||
| 2727 | .dstBinding = bind, | |||
| 2728 | .descriptorCount = 1, | |||
| 2729 | .dstArrayElement = offs, | |||
| 2730 | .descriptorType = desc_set->binding[bind].descriptorType, | |||
| 2731 | .pImageInfo = &desc_pool_write_info_img, | |||
| 2732 | }; | |||
| 2733 | update_set_pool_write(s, e, shd, set, &desc_pool_write_info); | |||
| 2734 | ||||
| 2735 | return 0; | |||
| 2736 | } | |||
| 2737 | ||||
| 2738 | void ff_vk_shader_update_img_array(FFVulkanContext *s, FFVkExecContext *e, | |||
| 2739 | FFVulkanShader *shd, AVFrame *f, | |||
| 2740 | VkImageView *views, int set, int binding, | |||
| 2741 | VkImageLayout layout, VkSampler sampler) | |||
| 2742 | { | |||
| 2743 | AVHWFramesContext *hwfc = (AVHWFramesContext *)f->hw_frames_ctx->data; | |||
| 2744 | const int nb_planes = av_pix_fmt_count_planes(hwfc->sw_format); | |||
| 2745 | ||||
| 2746 | for (int i = 0; i < nb_planes; i++) | |||
| 2747 | ff_vk_shader_update_img(s, e, shd, set, binding, i, | |||
| 2748 | views[i], layout, sampler); | |||
| 2749 | } | |||
| 2750 | ||||
| 2751 | int ff_vk_shader_update_desc_buffer(FFVulkanContext *s, FFVkExecContext *e, | |||
| 2752 | FFVulkanShader *shd, | |||
| 2753 | int set, int bind, int elem, | |||
| 2754 | FFVkBuffer *buf, VkDeviceSize offset, VkDeviceSize len, | |||
| 2755 | VkFormat fmt) | |||
| 2756 | { | |||
| 2757 | FFVulkanDescriptorSet *desc_set = &shd->desc_set[set]; | |||
| 2758 | ||||
| 2759 | VkDescriptorBufferInfo desc_pool_write_info_buf = { | |||
| 2760 | .buffer = buf->buf, | |||
| 2761 | .offset = buf->virtual_offset + offset, | |||
| 2762 | .range = len, | |||
| 2763 | }; | |||
| 2764 | VkWriteDescriptorSet desc_pool_write_info = { | |||
| 2765 | .sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET, | |||
| 2766 | .dstBinding = bind, | |||
| 2767 | .descriptorCount = 1, | |||
| 2768 | .dstArrayElement = elem, | |||
| 2769 | .descriptorType = desc_set->binding[bind].descriptorType, | |||
| 2770 | .pBufferInfo = &desc_pool_write_info_buf, | |||
| 2771 | }; | |||
| 2772 | update_set_pool_write(s, e, shd, set, &desc_pool_write_info); | |||
| 2773 | ||||
| 2774 | return 0; | |||
| 2775 | } | |||
| 2776 | ||||
| 2777 | void ff_vk_shader_update_push_const(FFVulkanContext *s, FFVkExecContext *e, | |||
| 2778 | FFVulkanShader *shd, | |||
| 2779 | VkShaderStageFlagBits stage, | |||
| 2780 | int offset, size_t size, void *src) | |||
| 2781 | { | |||
| 2782 | FFVulkanFunctions *vk = &s->vkfn; | |||
| 2783 | vk->CmdPushConstants(e->buf, shd->pipeline_layout, | |||
| 2784 | stage, offset, size, src); | |||
| 2785 | } | |||
| 2786 | ||||
| 2787 | void ff_vk_exec_bind_shader(FFVulkanContext *s, FFVkExecContext *e, | |||
| 2788 | const FFVulkanShader *shd) | |||
| 2789 | { | |||
| 2790 | FFVulkanFunctions *vk = &s->vkfn; | |||
| 2791 | const FFVulkanShaderData *sd = get_shd_data(e, shd); | |||
| 2792 | ||||
| 2793 | if (s->extensions & FF_VK_EXT_SHADER_OBJECT(1ULL << 13)) { | |||
| 2794 | VkShaderStageFlagBits stages = shd->stage; | |||
| 2795 | vk->CmdBindShadersEXT(e->buf, 1, &stages, &shd->object); | |||
| 2796 | } else { | |||
| 2797 | vk->CmdBindPipeline(e->buf, shd->bind_point, shd->pipeline); | |||
| 2798 | } | |||
| 2799 | ||||
| 2800 | if (sd && sd->nb_descriptor_sets) { | |||
| 2801 | if (!shd->use_push) { | |||
| 2802 | vk->CmdBindDescriptorSets(e->buf, shd->bind_point, shd->pipeline_layout, | |||
| 2803 | 0, sd->nb_descriptor_sets, | |||
| 2804 | &sd->desc_sets[e->idx*sd->nb_descriptor_sets], | |||
| 2805 | 0, NULL((void*)0)); | |||
| 2806 | } | |||
| 2807 | } | |||
| 2808 | } | |||
| 2809 | ||||
| 2810 | void ff_vk_shader_free(FFVulkanContext *s, FFVulkanShader *shd) | |||
| 2811 | { | |||
| 2812 | FFVulkanFunctions *vk = &s->vkfn; | |||
| 2813 | ||||
| 2814 | av_bprint_finalize(&shd->src, NULL((void*)0)); | |||
| 2815 | ||||
| 2816 | #if 0 | |||
| 2817 | if (shd->shader.module) | |||
| 2818 | vk->DestroyShaderModule(s->hwctx->act_dev, shd->shader.module, | |||
| 2819 | s->hwctx->alloc); | |||
| 2820 | #endif | |||
| 2821 | ||||
| 2822 | if (shd->object) | |||
| 2823 | vk->DestroyShaderEXT(s->hwctx->act_dev, shd->object, s->hwctx->alloc); | |||
| 2824 | if (shd->pipeline) | |||
| 2825 | vk->DestroyPipeline(s->hwctx->act_dev, shd->pipeline, s->hwctx->alloc); | |||
| 2826 | if (shd->pipeline_layout) | |||
| 2827 | vk->DestroyPipelineLayout(s->hwctx->act_dev, shd->pipeline_layout, | |||
| 2828 | s->hwctx->alloc); | |||
| 2829 | ||||
| 2830 | for (int i = 0; i < shd->nb_descriptor_sets; i++) | |||
| 2831 | if (shd->desc_layout[i]) | |||
| 2832 | vk->DestroyDescriptorSetLayout(s->hwctx->act_dev, shd->desc_layout[i], | |||
| 2833 | s->hwctx->alloc); | |||
| 2834 | } | |||
| 2835 | ||||
| 2836 | void ff_vk_uninit(FFVulkanContext *s) | |||
| 2837 | { | |||
| 2838 | av_freep(&s->query_props); | |||
| 2839 | av_freep(&s->qf_props); | |||
| 2840 | av_freep(&s->video_props); | |||
| 2841 | av_freep(&s->coop_mat_props); | |||
| 2842 | av_freep(&s->host_image_copy_layouts); | |||
| 2843 | ||||
| 2844 | av_buffer_unref(&s->device_ref); | |||
| 2845 | av_buffer_unref(&s->frames_ref); | |||
| 2846 | } | |||
| 2847 | ||||
| 2848 | int ff_vk_init(FFVulkanContext *s, void *log_parent, | |||
| 2849 | AVBufferRef *device_ref, AVBufferRef *frames_ref) | |||
| 2850 | { | |||
| 2851 | int err; | |||
| 2852 | ||||
| 2853 | static const AVClass vulkan_context_class = { | |||
| 2854 | .class_name = "vk", | |||
| 2855 | .version = LIBAVUTIL_VERSION_INT((60)<<16 | (29)<<8 | (100)), | |||
| 2856 | .parent_log_context_offset = offsetof(FFVulkanContext, log_parent)__builtin_offsetof(FFVulkanContext, log_parent), | |||
| 2857 | }; | |||
| 2858 | ||||
| 2859 | memset(s, 0, sizeof(*s)); | |||
| 2860 | s->log_parent = log_parent; | |||
| 2861 | s->class = &vulkan_context_class; | |||
| 2862 | ||||
| 2863 | if (frames_ref) { | |||
| 2864 | s->frames_ref = av_buffer_ref(frames_ref); | |||
| 2865 | if (!s->frames_ref) | |||
| 2866 | return AVERROR(ENOMEM)(-(12)); | |||
| 2867 | ||||
| 2868 | s->frames = (AVHWFramesContext *)s->frames_ref->data; | |||
| 2869 | s->hwfc = s->frames->hwctx; | |||
| 2870 | ||||
| 2871 | device_ref = s->frames->device_ref; | |||
| 2872 | } | |||
| 2873 | ||||
| 2874 | s->device_ref = av_buffer_ref(device_ref); | |||
| 2875 | if (!s->device_ref) { | |||
| 2876 | ff_vk_uninit(s); | |||
| 2877 | return AVERROR(ENOMEM)(-(12)); | |||
| 2878 | } | |||
| 2879 | ||||
| 2880 | s->device = (AVHWDeviceContext *)s->device_ref->data; | |||
| 2881 | s->hwctx = s->device->hwctx; | |||
| 2882 | ||||
| 2883 | s->extensions = ff_vk_extensions_to_mask(s->hwctx->enabled_dev_extensions, | |||
| 2884 | s->hwctx->nb_enabled_dev_extensions); | |||
| 2885 | s->extensions |= ff_vk_extensions_to_mask(s->hwctx->enabled_inst_extensions, | |||
| 2886 | s->hwctx->nb_enabled_inst_extensions); | |||
| 2887 | ||||
| 2888 | err = ff_vk_load_functions(s->device, &s->vkfn, s->extensions, 1, 1); | |||
| 2889 | if (err < 0) { | |||
| 2890 | ff_vk_uninit(s); | |||
| 2891 | return err; | |||
| 2892 | } | |||
| 2893 | ||||
| 2894 | err = ff_vk_load_props(s); | |||
| 2895 | if (err < 0) { | |||
| 2896 | ff_vk_uninit(s); | |||
| 2897 | return err; | |||
| 2898 | } | |||
| 2899 | ||||
| 2900 | return 0; | |||
| 2901 | } |