Unified · Split
| @@ -1,5 +1,7 @@ | ||
| 1 | 1 | /* |
| 2 | 2 | * copyright (c) 2021-2022 Wu Jianhua <jianhua.wu@intel.com> |
| 3 | + * Copyright (c) Lynne | |
| 4 | + * | |
| 3 | 5 | * This file is part of FFmpeg. |
| 4 | 6 | * |
| 5 | 7 | * FFmpeg is free software; you can redistribute it and/or |
| @@ -20,6 +22,7 @@ · 14 unmodified lines | ||
| 20 | 22 | #include "libavutil/random_seed.h" |
| 21 | 23 | #include "libavutil/opt.h" |
| 22 | 24 | #include "vulkan_filter.h" |
| 25 | +#include "vulkan_spirv.h" | |
| 23 | 26 | #include "internal.h" |
| 24 | 27 | |
| 25 | 28 | #define CGS 32 |
| @@ -27,26 +30,23 @@ · 1 unmodified line | ||
| 27 | 30 | |
| 28 | 31 | typedef struct GBlurVulkanContext { |
| 29 | 32 | FFVulkanContext vkctx; |
| 30 | − FFVkQueueFamilyCtx qf; | |
| 31 | − FFVkExecContext *exec; | |
| 32 | − FFVulkanPipeline *pl_hor; | |
| 33 | − FFVulkanPipeline *pl_ver; | |
| 34 | − FFVkBuffer params_buf_hor; | |
| 35 | − FFVkBuffer params_buf_ver; | |
| 36 | − | |
| 37 | − VkDescriptorImageInfo input_images[3]; | |
| 38 | − VkDescriptorImageInfo tmp_images[3]; | |
| 39 | − VkDescriptorImageInfo output_images[3]; | |
| 40 | − VkDescriptorBufferInfo params_desc_hor; | |
| 41 | − VkDescriptorBufferInfo params_desc_ver; | |
| 42 | 33 | |
| 43 | 34 | int initialized; |
| 35 | + FFVkExecPool e; | |
| 36 | + FFVkQueueFamilyCtx qf; | |
| 37 | + VkSampler sampler; | |
| 38 | + FFVulkanPipeline pl_hor; | |
| 39 | + FFVkSPIRVShader shd_hor; | |
| 40 | + FFVkBuffer params_hor; | |
| 41 | + FFVulkanPipeline pl_ver; | |
| 42 | + FFVkSPIRVShader shd_ver; | |
| 43 | + FFVkBuffer params_ver; | |
| 44 | + | |
| 44 | 45 | int size; |
| 45 | 46 | int sizeV; |
| 46 | 47 | int planes; |
| 47 | 48 | float sigma; |
| 48 | 49 | float sigmaV; |
| 49 | − AVFrame *tmpframe; | |
| 50 | 50 | } GBlurVulkanContext; |
| 51 | 51 | |
| 52 | 52 | static const char gblur_func[] = { |
| @@ -118,16 +118,17 @@ static av_cold void init_gaussian_params(GBlurVulkanContext *s) · 65 unmodified lines | ||
| 118 | 118 | s->sizeV = s->size; |
| 119 | 119 | else |
| 120 | 120 | init_kernel_size(s, &s->sizeV); |
| 121 | − | |
| 122 | − s->tmpframe = NULL; | |
| 123 | 121 | } |
| 124 | 122 | |
| 125 | −static int init_gblur_pipeline(GBlurVulkanContext *s, FFVulkanPipeline *pl, FFVkSPIRVShader *shd, | |
| 126 | − FFVkBuffer *params_buf, VkDescriptorBufferInfo *params_desc, | |
| 127 | − int ksize, float sigma) | |
| 123 | +static int init_gblur_pipeline(GBlurVulkanContext *s, FFVulkanPipeline *pl, | |
| 124 | + FFVkSPIRVShader *shd, FFVkBuffer *params_buf, | |
| 125 | + int ksize, float sigma, FFVkSPIRVCompiler *spv) | |
| 128 | 126 | { |
| 129 | 127 | int err = 0; |
| 130 | 128 | uint8_t *kernel_mapped; |
| 129 | + uint8_t *spv_data; | |
| 130 | + size_t spv_len; | |
| 131 | + void *spv_opaque = NULL; | |
| 131 | 132 | |
| 132 | 133 | const int planes = av_pix_fmt_count_planes(s->vkctx.output_format); |
| 133 | 134 | |
| @@ -137,7 +138,6 @@ static int init_gblur_pipeline(GBlurVulkanContext *s, FFVulkanPipeline *pl, FFVk · 3 unmodified lines | ||
| 137 | 138 | .mem_quali = "readonly", |
| 138 | 139 | .mem_layout = "std430", |
| 139 | 140 | .stages = VK_SHADER_STAGE_COMPUTE_BIT, |
| 140 | − .updater = NULL, | |
| 141 | 141 | .buf_content = NULL, |
| 142 | 142 | }; |
| 143 | 143 | |
| @@ -145,10 +145,9 @@ static int init_gblur_pipeline(GBlurVulkanContext *s, FFVulkanPipeline *pl, FFVk · 1 unmodified line | ||
| 145 | 145 | if (!kernel_def) |
| 146 | 146 | return AVERROR(ENOMEM); |
| 147 | 147 | |
| 148 | − buf_desc.updater = params_desc; | |
| 149 | 148 | buf_desc.buf_content = kernel_def; |
| 150 | 149 | |
| 151 | − RET(ff_vk_add_descriptor_set(&s->vkctx, pl, shd, &buf_desc, 1, 0)); | |
| 150 | + RET(ff_vk_pipeline_descriptor_set_add(&s->vkctx, pl, shd, &buf_desc, 1, 1, 0)); | |
| 152 | 151 | |
| 153 | 152 | GLSLD( gblur_func ); |
| 154 | 153 | GLSLC(0, void main() ); |
| @@ -157,38 +156,43 @@ static int init_gblur_pipeline(GBlurVulkanContext *s, FFVulkanPipeline *pl, FFVk · 2 unmodified lines | ||
| 157 | 156 | GLSLC(1, const ivec2 pos = ivec2(gl_GlobalInvocationID.xy); ); |
| 158 | 157 | for (int i = 0; i < planes; i++) { |
| 159 | 158 | GLSLC(0, ); |
| 160 | − GLSLF(1, size = imageSize(output_images[%i]); ,i); | |
| 161 | − GLSLC(1, if (IS_WITHIN(pos, size)) { ); | |
| 159 | + GLSLF(1, size = imageSize(output_images[%i]); ,i); | |
| 160 | + GLSLC(1, if (!IS_WITHIN(pos, size)) ); | |
| 161 | + GLSLC(2, return; ); | |
| 162 | 162 | if (s->planes & (1 << i)) { |
| 163 | − GLSLF(2, gblur(pos, %i); ,i); | |
| 163 | + GLSLF(1, gblur(pos, %i); ,i); | |
| 164 | 164 | } else { |
| 165 | − GLSLF(2, vec4 res = texture(input_images[%i], pos); ,i); | |
| 166 | − GLSLF(2, imageStore(output_images[%i], pos, res); ,i); | |
| 165 | + GLSLF(1, vec4 res = texture(input_images[%i], pos); ,i); | |
| 166 | + GLSLF(1, imageStore(output_images[%i], pos, res); ,i); | |
| 167 | 167 | } |
| 168 | − GLSLC(1, } ); | |
| 169 | 168 | } |
| 170 | 169 | GLSLC(0, } ); |
| 171 | 170 | |
| 172 | − RET(ff_vk_compile_shader(&s->vkctx, shd, "main")); | |
| 171 | + RET(spv->compile_shader(spv, s, shd, &spv_data, &spv_len, "main", | |
| 172 | + &spv_opaque)); | |
| 173 | + RET(ff_vk_shader_create(&s->vkctx, shd, spv_data, spv_len, "main")); | |
| 173 | 174 | |
| 174 | − RET(ff_vk_init_pipeline_layout(&s->vkctx, pl)); | |
| 175 | − RET(ff_vk_init_compute_pipeline(&s->vkctx, pl)); | |
| 175 | + RET(ff_vk_init_compute_pipeline(&s->vkctx, pl, shd)); | |
| 176 | + RET(ff_vk_exec_pipeline_register(&s->vkctx, &s->e, pl)); | |
| 176 | 177 | |
| 177 | 178 | RET(ff_vk_create_buf(&s->vkctx, params_buf, sizeof(float) * ksize, NULL, NULL, |
| 178 | − VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT)); | |
| 179 | − RET(ff_vk_map_buffers(&s->vkctx, params_buf, &kernel_mapped, 1, 0)); | |
| 179 | + VK_BUFFER_USAGE_SHADER_DEVICE_ADDRESS_BIT | | |
| 180 | + VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, | |
| 181 | + VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT)); | |
| 182 | + RET(ff_vk_map_buffer(&s->vkctx, params_buf, &kernel_mapped, 0)); | |
| 180 | 183 | |
| 181 | 184 | init_gaussian_kernel((float *)kernel_mapped, sigma, ksize); |
| 182 | 185 | |
| 183 | − RET(ff_vk_unmap_buffers(&s->vkctx, params_buf, 1, 1)); | |
| 184 | − | |
| 185 | − params_desc->buffer = params_buf->buf; | |
| 186 | − params_desc->range = VK_WHOLE_SIZE; | |
| 186 | + RET(ff_vk_unmap_buffer(&s->vkctx, params_buf, 1)); | |
| 187 | 187 | |
| 188 | − ff_vk_update_descriptor_set(&s->vkctx, pl, 1); | |
| 188 | + RET(ff_vk_set_descriptor_buffer(&s->vkctx, pl, NULL, 1, 0, 0, | |
| 189 | + params_buf->address, params_buf->size, | |
| 190 | + VK_FORMAT_UNDEFINED)); | |
| 189 | 191 | |
| 190 | 192 | fail: |
| 191 | 193 | av_free(kernel_def); |
| 194 | + if (spv_opaque) | |
| 195 | + spv->free_shader(spv, &spv_opaque); | |
| 192 | 196 | return err; |
| 193 | 197 | } |
| 194 | 198 | |
| @@ -196,16 +200,35 @@ static av_cold int init_filter(AVFilterContext *ctx, AVFrame *in) · 1 unmodified line | ||
| 196 | 200 | { |
| 197 | 201 | int err = 0; |
| 198 | 202 | GBlurVulkanContext *s = ctx->priv; |
| 199 | − FFVkSPIRVShader *shd; | |
| 203 | + FFVulkanContext *vkctx = &s->vkctx; | |
| 200 | 204 | const int planes = av_pix_fmt_count_planes(s->vkctx.output_format); |
| 201 | 205 | |
| 202 | − FFVulkanDescriptorSetBinding image_descs[] = { | |
| 206 | + FFVkSPIRVShader *shd; | |
| 207 | + FFVkSPIRVCompiler *spv; | |
| 208 | + FFVulkanDescriptorSetBinding *desc; | |
| 209 | + | |
| 210 | + spv = ff_vk_spirv_init(); | |
| 211 | + if (!spv) { | |
| 212 | + av_log(ctx, AV_LOG_ERROR, "Unable to initialize SPIR-V compiler!\n"); | |
| 213 | + return AVERROR_EXTERNAL; | |
| 214 | + } | |
| 215 | + | |
| 216 | + ff_vk_qf_init(vkctx, &s->qf, VK_QUEUE_COMPUTE_BIT); | |
| 217 | + RET(ff_vk_exec_pool_init(vkctx, &s->qf, &s->e, s->qf.nb_queues*4, 0, 0, 0, NULL)); | |
| 218 | + RET(ff_vk_init_sampler(vkctx, &s->sampler, 1, VK_FILTER_LINEAR)); | |
| 219 | + RET(ff_vk_shader_init(&s->pl_hor, &s->shd_hor, "gblur_hor_compute", | |
| 220 | + VK_SHADER_STAGE_COMPUTE_BIT, 0)); | |
| 221 | + RET(ff_vk_shader_init(&s->pl_ver, &s->shd_ver, "gblur_ver_compute", | |
| 222 | + VK_SHADER_STAGE_COMPUTE_BIT, 0)); | |
| 223 | + | |
| 224 | + desc = (FFVulkanDescriptorSetBinding []) { | |
| 203 | 225 | { |
| 204 | 226 | .name = "input_images", |
| 205 | 227 | .type = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, |
| 206 | 228 | .dimensions = 2, |
| 207 | 229 | .elems = planes, |
| 208 | 230 | .stages = VK_SHADER_STAGE_COMPUTE_BIT, |
| 231 | + .samplers = DUP_SAMPLER(s->sampler), | |
| 209 | 232 | }, |
| 210 | 233 | { |
| 211 | 234 | .name = "output_images", |
| @@ -218,215 +241,64 @@ static av_cold int init_filter(AVFilterContext *ctx, AVFrame *in) · 6 unmodified lines | ||
| 218 | 241 | }, |
| 219 | 242 | }; |
| 220 | 243 | |
| 221 | − image_descs[0].sampler = ff_vk_init_sampler(&s->vkctx, 1, VK_FILTER_LINEAR); | |
| 222 | − if (!image_descs[0].sampler) | |
| 223 | − return AVERROR_EXTERNAL; | |
| 224 | − | |
| 225 | 244 | init_gaussian_params(s); |
| 226 | 245 | |
| 227 | − ff_vk_qf_init(&s->vkctx, &s->qf, VK_QUEUE_COMPUTE_BIT, 0); | |
| 228 | − | |
| 229 | 246 | { |
| 230 | − /* Create shader for the horizontal pass */ | |
| 231 | − image_descs[0].updater = s->input_images; | |
| 232 | − image_descs[1].updater = s->tmp_images; | |
| 233 | − | |
| 234 | − s->pl_hor = ff_vk_create_pipeline(&s->vkctx, &s->qf); | |
| 235 | − if (!s->pl_hor) { | |
| 236 | − err = AVERROR(ENOMEM); | |
| 237 | − goto fail; | |
| 238 | − } | |
| 239 | − | |
| 240 | − shd = ff_vk_init_shader(s->pl_hor, "gblur_compute_hor", image_descs[0].stages); | |
| 241 | − if (!shd) { | |
| 242 | − err = AVERROR(ENOMEM); | |
| 243 | − goto fail; | |
| 244 | − } | |
| 247 | + shd = &s->shd_hor; | |
| 248 | + ff_vk_shader_set_compute_sizes(shd, 32, 1, 1); | |
| 245 | 249 | |
| 246 | − ff_vk_set_compute_shader_sizes(shd, (int [3]){ CGS, 1, 1 }); | |
| 247 | − RET(ff_vk_add_descriptor_set(&s->vkctx, s->pl_hor, shd, image_descs, FF_ARRAY_ELEMS(image_descs), 0)); | |
| 250 | + RET(ff_vk_pipeline_descriptor_set_add(vkctx, &s->pl_hor, shd, desc, 2, 0, 0)); | |
| 248 | 251 | |
| 249 | 252 | GLSLC(0, #define OFFSET (vec2(i, 0.0))); |
| 250 | − RET(init_gblur_pipeline(s, s->pl_hor, shd, &s->params_buf_hor, &s->params_desc_hor, | |
| 251 | − s->size, s->sigma)); | |
| 253 | + RET(init_gblur_pipeline(s, &s->pl_hor, shd, &s->params_hor, s->size, s->sigma, spv)); | |
| 252 | 254 | } |
| 253 | 255 | |
| 254 | 256 | { |
| 255 | − /* Create shader for the vertical pass */ | |
| 256 | − image_descs[0].updater = s->tmp_images; | |
| 257 | − image_descs[1].updater = s->output_images; | |
| 258 | − | |
| 259 | − s->pl_ver = ff_vk_create_pipeline(&s->vkctx, &s->qf); | |
| 260 | − if (!s->pl_ver) { | |
| 261 | − err = AVERROR(ENOMEM); | |
| 262 | − goto fail; | |
| 263 | − } | |
| 257 | + shd = &s->shd_ver; | |
| 258 | + ff_vk_shader_set_compute_sizes(shd, 1, 32, 1); | |
| 264 | 259 | |
| 265 | − shd = ff_vk_init_shader(s->pl_ver, "gblur_compute_ver", image_descs[0].stages); | |
| 266 | − if (!shd) { | |
| 267 | − err = AVERROR(ENOMEM); | |
| 268 | − goto fail; | |
| 269 | − } | |
| 270 | − | |
| 271 | − ff_vk_set_compute_shader_sizes(shd, (int [3]){ 1, CGS, 1 }); | |
| 272 | − RET(ff_vk_add_descriptor_set(&s->vkctx, s->pl_ver, shd, image_descs, FF_ARRAY_ELEMS(image_descs), 0)); | |
| 260 | + RET(ff_vk_pipeline_descriptor_set_add(vkctx, &s->pl_ver, shd, desc, 2, 0, 0)); | |
| 273 | 261 | |
| 274 | 262 | GLSLC(0, #define OFFSET (vec2(0.0, i))); |
| 275 | − RET(init_gblur_pipeline(s, s->pl_ver, shd, &s->params_buf_ver, &s->params_desc_ver, | |
| 276 | − s->sizeV, s->sigmaV)); | |
| 263 | + RET(init_gblur_pipeline(s, &s->pl_ver, shd, &s->params_ver, s->sizeV, s->sigmaV, spv)); | |
| 277 | 264 | } |
| 278 | 265 | |
| 279 | − RET(ff_vk_create_exec_ctx(&s->vkctx, &s->exec, &s->qf)); | |
| 280 | − | |
| 281 | 266 | s->initialized = 1; |
| 282 | 267 | |
| 283 | 268 | fail: |
| 269 | + if (spv) | |
| 270 | + spv->uninit(&spv); | |
| 271 | + | |
| 284 | 272 | return err; |
| 285 | 273 | } |
| 286 | 274 | |
| 287 | 275 | static av_cold void gblur_vulkan_uninit(AVFilterContext *avctx) |
| 288 | 276 | { |
| 289 | 277 | GBlurVulkanContext *s = avctx->priv; |
| 278 | + FFVulkanContext *vkctx = &s->vkctx; | |
| 279 | + FFVulkanFunctions *vk = &vkctx->vkfn; | |
| 290 | 280 | |
| 291 | − av_frame_free(&s->tmpframe); | |
| 281 | + ff_vk_exec_pool_free(vkctx, &s->e); | |
| 282 | + ff_vk_pipeline_free(vkctx, &s->pl_hor); | |
| 283 | + ff_vk_pipeline_free(vkctx, &s->pl_ver); | |
| 284 | + ff_vk_shader_free(vkctx, &s->shd_hor); | |
| 285 | + ff_vk_shader_free(vkctx, &s->shd_ver); | |
| 286 | + ff_vk_free_buf(vkctx, &s->params_hor); | |
| 287 | + ff_vk_free_buf(vkctx, &s->params_ver); | |
| 288 | + | |
| 289 | + if (s->sampler) | |
| 290 | + vk->DestroySampler(vkctx->hwctx->act_dev, s->sampler, | |
| 291 | + vkctx->hwctx->alloc); | |
| 292 | 292 | |
| 293 | − ff_vk_free_buf(&s->vkctx, &s->params_buf_hor); | |
| 294 | − ff_vk_free_buf(&s->vkctx, &s->params_buf_ver); | |
| 295 | 293 | ff_vk_uninit(&s->vkctx); |
| 296 | 294 | |
| 297 | 295 | s->initialized = 0; |
| 298 | 296 | } |
| 299 | 297 | |
| 300 | −static int process_frames(AVFilterContext *avctx, AVFrame *outframe, AVFrame *inframe) | |
| 301 | −{ | |
| 302 | − int err; | |
| 303 | − VkCommandBuffer cmd_buf; | |
| 304 | − GBlurVulkanContext *s = avctx->priv; | |
| 305 | − FFVulkanFunctions *vk = &s->vkctx.vkfn; | |
| 306 | − | |
| 307 | − const int planes = av_pix_fmt_count_planes(s->vkctx.output_format); | |
| 308 | − | |
| 309 | − AVVkFrame *in = (AVVkFrame *)inframe->data[0]; | |
| 310 | − AVVkFrame *out = (AVVkFrame *)outframe->data[0]; | |
| 311 | − AVVkFrame *tmp = (AVVkFrame *)s->tmpframe->data[0]; | |
| 312 | − | |
| 313 | − const VkFormat *input_formats = av_vkfmt_from_pixfmt(s->vkctx.input_format); | |
| 314 | − const VkFormat *output_formats = av_vkfmt_from_pixfmt(s->vkctx.output_format); | |
| 315 | − | |
| 316 | − ff_vk_start_exec_recording(&s->vkctx, s->exec); | |
| 317 | − cmd_buf = ff_vk_get_exec_buf(s->exec); | |
| 318 | − | |
| 319 | − for (int i = 0; i < planes; i++) { | |
| 320 | − RET(ff_vk_create_imageview(&s->vkctx, s->exec, &s->input_images[i].imageView, | |
| 321 | − in->img[i], | |
| 322 | − input_formats[i], | |
| 323 | − ff_comp_identity_map)); | |
| 324 | − | |
| 325 | − RET(ff_vk_create_imageview(&s->vkctx, s->exec, &s->tmp_images[i].imageView, | |
| 326 | − tmp->img[i], | |
| 327 | − output_formats[i], | |
| 328 | − ff_comp_identity_map)); | |
| 329 | − | |
| 330 | − RET(ff_vk_create_imageview(&s->vkctx, s->exec, &s->output_images[i].imageView, | |
| 331 | − out->img[i], | |
| 332 | − output_formats[i], | |
| 333 | − ff_comp_identity_map)); | |
| 334 | − | |
| 335 | − s->input_images[i].imageLayout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL; | |
| 336 | − s->tmp_images[i].imageLayout = VK_IMAGE_LAYOUT_GENERAL; | |
| 337 | − s->output_images[i].imageLayout = VK_IMAGE_LAYOUT_GENERAL; | |
| 338 | − } | |
| 339 | − | |
| 340 | − ff_vk_update_descriptor_set(&s->vkctx, s->pl_hor, 0); | |
| 341 | − ff_vk_update_descriptor_set(&s->vkctx, s->pl_ver, 0); | |
| 342 | − | |
| 343 | − for (int i = 0; i < planes; i++) { | |
| 344 | − VkImageMemoryBarrier barriers[] = { | |
| 345 | − { | |
| 346 | − .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER, | |
| 347 | − .srcAccessMask = 0, | |
| 348 | − .dstAccessMask = VK_ACCESS_SHADER_READ_BIT, | |
| 349 | − .oldLayout = in->layout[i], | |
| 350 | − .newLayout = s->input_images[i].imageLayout, | |
| 351 | − .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED, | |
| 352 | − .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED, | |
| 353 | − .image = in->img[i], | |
| 354 | − .subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT, | |
| 355 | − .subresourceRange.levelCount = 1, | |
| 356 | − .subresourceRange.layerCount = 1, | |
| 357 | − }, | |
| 358 | − { | |
| 359 | − .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER, | |
| 360 | − .srcAccessMask = 0, | |
| 361 | − .dstAccessMask = VK_ACCESS_SHADER_WRITE_BIT | VK_ACCESS_SHADER_READ_BIT, | |
| 362 | − .oldLayout = tmp->layout[i], | |
| 363 | − .newLayout = s->tmp_images[i].imageLayout, | |
| 364 | − .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED, | |
| 365 | − .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED, | |
| 366 | − .image = tmp->img[i], | |
| 367 | − .subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT, | |
| 368 | − .subresourceRange.levelCount = 1, | |
| 369 | − .subresourceRange.layerCount = 1, | |
| 370 | − }, | |
| 371 | − { | |
| 372 | − .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER, | |
| 373 | − .srcAccessMask = 0, | |
| 374 | − .dstAccessMask = VK_ACCESS_SHADER_WRITE_BIT, | |
| 375 | − .oldLayout = out->layout[i], | |
| 376 | − .newLayout = s->output_images[i].imageLayout, | |
| 377 | − .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED, | |
| 378 | − .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED, | |
| 379 | − .image = out->img[i], | |
| 380 | − .subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT, | |
| 381 | − .subresourceRange.levelCount = 1, | |
| 382 | − .subresourceRange.layerCount = 1, | |
| 383 | − }, | |
| 384 | − }; | |
| 385 | − | |
| 386 | − vk->CmdPipelineBarrier(cmd_buf, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, | |
| 387 | − VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, 0, | |
| 388 | − 0, NULL, 0, NULL, FF_ARRAY_ELEMS(barriers), barriers); | |
| 389 | − | |
| 390 | − in->layout[i] = barriers[0].newLayout; | |
| 391 | − in->access[i] = barriers[0].dstAccessMask; | |
| 392 | − | |
| 393 | − tmp->layout[i] = barriers[1].newLayout; | |
| 394 | − tmp->access[i] = barriers[1].dstAccessMask; | |
| 395 | − | |
| 396 | − out->layout[i] = barriers[2].newLayout; | |
| 397 | − out->access[i] = barriers[2].dstAccessMask; | |
| 398 | − } | |
| 399 | − | |
| 400 | − ff_vk_bind_pipeline_exec(&s->vkctx, s->exec, s->pl_hor); | |
| 401 | − | |
| 402 | − vk->CmdDispatch(cmd_buf, FFALIGN(s->vkctx.output_width, CGS)/CGS, | |
| 403 | − s->vkctx.output_height, 1); | |
| 404 | − | |
| 405 | − ff_vk_bind_pipeline_exec(&s->vkctx, s->exec, s->pl_ver); | |
| 406 | − | |
| 407 | − vk->CmdDispatch(cmd_buf,s->vkctx.output_width, | |
| 408 | − FFALIGN(s->vkctx.output_height, CGS)/CGS, 1); | |
| 409 | − | |
| 410 | − ff_vk_add_exec_dep(&s->vkctx, s->exec, inframe, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT); | |
| 411 | − ff_vk_add_exec_dep(&s->vkctx, s->exec, outframe, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT); | |
| 412 | − | |
| 413 | − err = ff_vk_submit_exec_queue(&s->vkctx, s->exec); | |
| 414 | − if (err) | |
| 415 | − return err; | |
| 416 | − | |
| 417 | − ff_vk_qf_rotate(&s->qf); | |
| 418 | − | |
| 419 | − return 0; | |
| 420 | − | |
| 421 | −fail: | |
| 422 | − ff_vk_discard_exec_deps(s->exec); | |
| 423 | − return err; | |
| 424 | −} | |
| 425 | − | |
| 426 | 298 | static int gblur_vulkan_filter_frame(AVFilterLink *link, AVFrame *in) |
| 427 | 299 | { |
| 428 | 300 | int err; |
| 429 | − AVFrame *out = NULL; | |
| 301 | + AVFrame *tmp = NULL, *out = NULL; | |
| 430 | 302 | AVFilterContext *ctx = link->dst; |
| 431 | 303 | GBlurVulkanContext *s = ctx->priv; |
| 432 | 304 | AVFilterLink *outlink = ctx->outputs[0]; |
| @@ -437,28 +309,32 @@ static int gblur_vulkan_filter_frame(AVFilterLink *link, AVFrame *in) · 4 unmodified lines | ||
| 437 | 309 | goto fail; |
| 438 | 310 | } |
| 439 | 311 | |
| 440 | − if (!s->initialized) { | |
| 441 | − RET(init_filter(ctx, in)); | |
| 442 | − s->tmpframe = ff_get_video_buffer(outlink, outlink->w, outlink->h); | |
| 443 | − if (!s->tmpframe) { | |
| 444 | − err = AVERROR(ENOMEM); | |
| 445 | − goto fail; | |
| 446 | − } | |
| 312 | + tmp = ff_get_video_buffer(outlink, outlink->w, outlink->h); | |
| 313 | + if (!tmp) { | |
| 314 | + err = AVERROR(ENOMEM); | |
| 315 | + goto fail; | |
| 447 | 316 | } |
| 448 | 317 | |
| 449 | − RET(process_frames(ctx, out, in)); | |
| 318 | + if (!s->initialized) | |
| 319 | + RET(init_filter(ctx, in)); | |
| 450 | 320 | |
| 451 | − RET(av_frame_copy_props(out, in)); | |
| 321 | + RET(ff_vk_filter_process_2pass(&s->vkctx, &s->e, | |
| 322 | + (FFVulkanPipeline *[2]){ &s->pl_hor, &s->pl_ver }, | |
| 323 | + out, tmp, in, s->sampler, NULL, 0)); | |
| 324 | + | |
| 325 | + err = av_frame_copy_props(out, in); | |
| 326 | + if (err < 0) | |
| 327 | + goto fail; | |
| 452 | 328 | |
| 453 | 329 | av_frame_free(&in); |
| 330 | + av_frame_free(&tmp); | |
| 454 | 331 | |
| 455 | 332 | return ff_filter_frame(outlink, out); |
| 456 | 333 | |
| 457 | 334 | fail: |
| 458 | 335 | av_frame_free(&in); |
| 336 | + av_frame_free(&tmp); | |
| 459 | 337 | av_frame_free(&out); |
| 460 | − av_frame_free(&s->tmpframe); | |
| 461 | − | |
| 462 | 338 | return err; |
| 463 | 339 | } |
| 464 | 340 | |