Unified · Split
| @@ -1,4 +1,6 @@ | ||
| 1 | 1 | /* |
| 2 | + * Copyright (c) Lynne | |
| 3 | + * | |
| 2 | 4 | * This file is part of FFmpeg. |
| 3 | 5 | * |
| 4 | 6 | * FFmpeg is free software; you can redistribute it and/or |
| @@ -19,12 +21,11 @@ · 14 unmodified lines | ||
| 19 | 21 | #include "libavutil/random_seed.h" |
| 20 | 22 | #include "libavutil/opt.h" |
| 21 | 23 | #include "vulkan_filter.h" |
| 24 | +#include "vulkan_spirv.h" | |
| 22 | 25 | #include "scale_eval.h" |
| 23 | 26 | #include "internal.h" |
| 24 | 27 | #include "colorspace.h" |
| 25 | 28 | |
| 26 | −#define CGROUPS (int [3]){ 32, 32, 1 } | |
| 27 | − | |
| 28 | 29 | enum ScalerFunc { |
| 29 | 30 | F_BILINEAR = 0, |
| 30 | 31 | F_NEAREST, |
| @@ -35,15 +36,17 @@ enum ScalerFunc { · 4 unmodified lines | ||
| 35 | 36 | typedef struct ScaleVulkanContext { |
| 36 | 37 | FFVulkanContext vkctx; |
| 37 | 38 | |
| 39 | + int initialized; | |
| 40 | + FFVulkanPipeline pl; | |
| 41 | + FFVkExecPool e; | |
| 38 | 42 | FFVkQueueFamilyCtx qf; |
| 39 | − FFVkExecContext *exec; | |
| 40 | − FFVulkanPipeline *pl; | |
| 41 | − FFVkBuffer params_buf; | |
| 43 | + FFVkSPIRVShader shd; | |
| 44 | + VkSampler sampler; | |
| 42 | 45 | |
| 43 | − /* Shader updators, must be in the main filter struct */ | |
| 44 | − VkDescriptorImageInfo input_images[3]; | |
| 45 | − VkDescriptorImageInfo output_images[3]; | |
| 46 | − VkDescriptorBufferInfo params_desc; | |
| 46 | + /* Push constants / options */ | |
| 47 | + struct { | |
| 48 | + float yuv_matrix[4][4]; | |
| 49 | + } opts; | |
| 47 | 50 | |
| 48 | 51 | char *out_format_string; |
| 49 | 52 | char *w_expr; |
| @@ -51,8 +54,6 @@ typedef struct ScaleVulkanContext { · 1 unmodified line | ||
| 51 | 54 | |
| 52 | 55 | enum ScalerFunc scaler; |
| 53 | 56 | enum AVColorRange out_range; |
| 54 | − | |
| 55 | − int initialized; | |
| 56 | 57 | } ScaleVulkanContext; |
| 57 | 58 | |
| 58 | 59 | static const char scale_bilinear[] = { |
| @@ -110,10 +111,15 @@ static const char write_444[] = { · 51 unmodified lines | ||
| 110 | 111 | static av_cold int init_filter(AVFilterContext *ctx, AVFrame *in) |
| 111 | 112 | { |
| 112 | 113 | int err; |
| 113 | − FFVkSampler *sampler; | |
| 114 | + uint8_t *spv_data; | |
| 115 | + size_t spv_len; | |
| 116 | + void *spv_opaque = NULL; | |
| 114 | 117 | VkFilter sampler_mode; |
| 115 | 118 | ScaleVulkanContext *s = ctx->priv; |
| 116 | 119 | FFVulkanContext *vkctx = &s->vkctx; |
| 120 | + FFVkSPIRVShader *shd = &s->shd; | |
| 121 | + FFVkSPIRVCompiler *spv; | |
| 122 | + FFVulkanDescriptorSetBinding *desc; | |
| 117 | 123 | |
| 118 | 124 | int crop_x = in->crop_left; |
| 119 | 125 | int crop_y = in->crop_top; |
| @@ -121,8 +127,6 @@ static av_cold int init_filter(AVFilterContext *ctx, AVFrame *in) · 1 unmodified line | ||
| 121 | 127 | int crop_h = in->height - (in->crop_top + in->crop_bottom); |
| 122 | 128 | int in_planes = av_pix_fmt_count_planes(s->vkctx.input_format); |
| 123 | 129 | |
| 124 | − ff_vk_qf_init(vkctx, &s->qf, VK_QUEUE_COMPUTE_BIT, 0); | |
| 125 | − | |
| 126 | 130 | switch (s->scaler) { |
| 127 | 131 | case F_NEAREST: |
| 128 | 132 | sampler_mode = VK_FILTER_NEAREST; |
| @@ -132,264 +136,134 @@ static av_cold int init_filter(AVFilterContext *ctx, AVFrame *in) · 3 unmodified lines | ||
| 132 | 136 | break; |
| 133 | 137 | }; |
| 134 | 138 | |
| 135 | − /* Create a sampler */ | |
| 136 | − sampler = ff_vk_init_sampler(vkctx, 0, sampler_mode); | |
| 137 | − if (!sampler) | |
| 139 | + spv = ff_vk_spirv_init(); | |
| 140 | + if (!spv) { | |
| 141 | + av_log(ctx, AV_LOG_ERROR, "Unable to initialize SPIR-V compiler!\n"); | |
| 138 | 142 | return AVERROR_EXTERNAL; |
| 143 | + } | |
| 139 | 144 | |
| 140 | − s->pl = ff_vk_create_pipeline(vkctx, &s->qf); | |
| 141 | − if (!s->pl) | |
| 142 | − return AVERROR(ENOMEM); | |
| 143 | − | |
| 144 | − { /* Create the shader */ | |
| 145 | − FFVulkanDescriptorSetBinding desc_i[2] = { | |
| 146 | − { | |
| 147 | − .name = "input_img", | |
| 148 | − .type = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, | |
| 149 | − .dimensions = 2, | |
| 150 | − .elems = in_planes, | |
| 151 | − .stages = VK_SHADER_STAGE_COMPUTE_BIT, | |
| 152 | − .updater = s->input_images, | |
| 153 | − .sampler = sampler, | |
| 154 | − }, | |
| 155 | − { | |
| 156 | − .name = "output_img", | |
| 157 | − .type = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE, | |
| 158 | − .mem_layout = ff_vk_shader_rep_fmt(s->vkctx.output_format), | |
| 159 | − .mem_quali = "writeonly", | |
| 160 | − .dimensions = 2, | |
| 161 | − .elems = av_pix_fmt_count_planes(s->vkctx.output_format), | |
| 162 | − .stages = VK_SHADER_STAGE_COMPUTE_BIT, | |
| 163 | − .updater = s->output_images, | |
| 164 | − }, | |
| 165 | − }; | |
| 166 | − | |
| 167 | − FFVulkanDescriptorSetBinding desc_b = { | |
| 168 | − .name = "params", | |
| 169 | − .type = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER, | |
| 170 | − .mem_quali = "readonly", | |
| 171 | − .mem_layout = "std430", | |
| 172 | − .stages = VK_SHADER_STAGE_COMPUTE_BIT, | |
| 173 | − .updater = &s->params_desc, | |
| 174 | − .buf_content = "mat4 yuv_matrix;", | |
| 175 | − }; | |
| 176 | − | |
| 177 | − FFVkSPIRVShader *shd = ff_vk_init_shader(s->pl, "scale_compute", | |
| 178 | − VK_SHADER_STAGE_COMPUTE_BIT); | |
| 179 | − if (!shd) | |
| 180 | − return AVERROR(ENOMEM); | |
| 181 | − | |
| 182 | − ff_vk_set_compute_shader_sizes(shd, CGROUPS); | |
| 183 | − | |
| 184 | − RET(ff_vk_add_descriptor_set(vkctx, s->pl, shd, desc_i, FF_ARRAY_ELEMS(desc_i), 0)); /* set 0 */ | |
| 185 | − RET(ff_vk_add_descriptor_set(vkctx, s->pl, shd, &desc_b, 1, 0)); /* set 1 */ | |
| 186 | − | |
| 187 | − GLSLD( scale_bilinear ); | |
| 188 | − | |
| 189 | − if (s->vkctx.output_format != s->vkctx.input_format) { | |
| 190 | − GLSLD( rgb2yuv ); | |
| 191 | − } | |
| 145 | + ff_vk_qf_init(vkctx, &s->qf, VK_QUEUE_COMPUTE_BIT); | |
| 146 | + RET(ff_vk_exec_pool_init(vkctx, &s->qf, &s->e, s->qf.nb_queues*4, 0, 0, 0, NULL)); | |
| 147 | + RET(ff_vk_init_sampler(vkctx, &s->sampler, 0, sampler_mode)); | |
| 148 | + RET(ff_vk_shader_init(&s->pl, &s->shd, "scale_compute", | |
| 149 | + VK_SHADER_STAGE_COMPUTE_BIT, 0)); | |
| 150 | + | |
| 151 | + ff_vk_shader_set_compute_sizes(&s->shd, 32, 32, 1); | |
| 152 | + | |
| 153 | + GLSLC(0, layout(push_constant, std430) uniform pushConstants { ); | |
| 154 | + GLSLC(1, mat4 yuv_matrix; ); | |
| 155 | + GLSLC(0, }; ); | |
| 156 | + GLSLC(0, ); | |
| 157 | + | |
| 158 | + ff_vk_add_push_constant(&s->pl, 0, sizeof(s->opts), | |
| 159 | + VK_SHADER_STAGE_COMPUTE_BIT); | |
| 160 | + | |
| 161 | + desc = (FFVulkanDescriptorSetBinding []) { | |
| 162 | + { | |
| 163 | + .name = "input_img", | |
| 164 | + .type = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, | |
| 165 | + .dimensions = 2, | |
| 166 | + .elems = in_planes, | |
| 167 | + .stages = VK_SHADER_STAGE_COMPUTE_BIT, | |
| 168 | + .samplers = DUP_SAMPLER(s->sampler), | |
| 169 | + }, | |
| 170 | + { | |
| 171 | + .name = "output_img", | |
| 172 | + .type = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE, | |
| 173 | + .mem_layout = ff_vk_shader_rep_fmt(s->vkctx.output_format), | |
| 174 | + .mem_quali = "writeonly", | |
| 175 | + .dimensions = 2, | |
| 176 | + .elems = av_pix_fmt_count_planes(s->vkctx.output_format), | |
| 177 | + .stages = VK_SHADER_STAGE_COMPUTE_BIT, | |
| 178 | + }, | |
| 179 | + }; | |
| 192 | 180 | |
| 193 | − switch (s->vkctx.output_format) { | |
| 194 | − case AV_PIX_FMT_NV12: GLSLD(write_nv12); break; | |
| 195 | − case AV_PIX_FMT_YUV420P: GLSLD( write_420); break; | |
| 196 | − case AV_PIX_FMT_YUV444P: GLSLD( write_444); break; | |
| 197 | − default: break; | |
| 198 | − } | |
| 181 | + RET(ff_vk_pipeline_descriptor_set_add(vkctx, &s->pl, shd, desc, 2, 0, 0)); | |
| 199 | 182 | |
| 200 | − GLSLC(0, void main() ); | |
| 201 | − GLSLC(0, { ); | |
| 202 | − GLSLC(1, ivec2 size; ); | |
| 203 | − GLSLC(1, ivec2 pos = ivec2(gl_GlobalInvocationID.xy); ); | |
| 204 | − GLSLF(1, vec2 in_d = vec2(%i, %i); ,in->width, in->height); | |
| 205 | − GLSLF(1, vec2 c_r = vec2(%i, %i) / in_d; ,crop_w, crop_h); | |
| 206 | − GLSLF(1, vec2 c_o = vec2(%i, %i) / in_d; ,crop_x,crop_y); | |
| 207 | − GLSLC(0, ); | |
| 208 | − | |
| 209 | − if (s->vkctx.output_format == s->vkctx.input_format) { | |
| 210 | − for (int i = 0; i < desc_i[1].elems; i++) { | |
| 211 | − GLSLF(1, size = imageSize(output_img[%i]); ,i); | |
| 212 | − GLSLC(1, if (IS_WITHIN(pos, size)) { ); | |
| 213 | − switch (s->scaler) { | |
| 214 | − case F_NEAREST: | |
| 215 | − case F_BILINEAR: | |
| 216 | − GLSLF(2, vec4 res = scale_bilinear(%i, pos, c_r, c_o); ,i); | |
| 217 | − GLSLF(2, imageStore(output_img[%i], pos, res); ,i); | |
| 218 | − break; | |
| 219 | − }; | |
| 220 | − GLSLC(1, } ); | |
| 221 | − } | |
| 222 | − } else { | |
| 223 | − GLSLC(1, vec4 res = scale_bilinear(0, pos, c_r, c_o); ); | |
| 224 | − GLSLF(1, res = rgb2yuv(res, %i); ,s->out_range == AVCOL_RANGE_JPEG); | |
| 225 | − switch (s->vkctx.output_format) { | |
| 226 | − case AV_PIX_FMT_NV12: GLSLC(1, write_nv12(res, pos); ); break; | |
| 227 | − case AV_PIX_FMT_YUV420P: GLSLC(1, write_420(res, pos); ); break; | |
| 228 | − case AV_PIX_FMT_YUV444P: GLSLC(1, write_444(res, pos); ); break; | |
| 229 | − default: return AVERROR(EINVAL); | |
| 230 | − } | |
| 231 | − } | |
| 183 | + GLSLD( scale_bilinear ); | |
| 184 | + | |
| 185 | + if (s->vkctx.output_format != s->vkctx.input_format) { | |
| 186 | + GLSLD( rgb2yuv ); | |
| 187 | + } | |
| 232 | 188 | |
| 233 | − GLSLC(0, } ); | |
| 189 | + switch (s->vkctx.output_format) { | |
| 190 | + case AV_PIX_FMT_NV12: GLSLD(write_nv12); break; | |
| 191 | + case AV_PIX_FMT_YUV420P: GLSLD( write_420); break; | |
| 192 | + case AV_PIX_FMT_YUV444P: GLSLD( write_444); break; | |
| 193 | + default: break; | |
| 194 | + } | |
| 234 | 195 | |
| 235 | − RET(ff_vk_compile_shader(vkctx, shd, "main")); | |
| 196 | + GLSLC(0, void main() ); | |
| 197 | + GLSLC(0, { ); | |
| 198 | + GLSLC(1, ivec2 size; ); | |
| 199 | + GLSLC(1, ivec2 pos = ivec2(gl_GlobalInvocationID.xy); ); | |
| 200 | + GLSLF(1, vec2 in_d = vec2(%i, %i); ,in->width, in->height); | |
| 201 | + GLSLF(1, vec2 c_r = vec2(%i, %i) / in_d; ,crop_w, crop_h); | |
| 202 | + GLSLF(1, vec2 c_o = vec2(%i, %i) / in_d; ,crop_x,crop_y); | |
| 203 | + GLSLC(0, ); | |
| 204 | + | |
| 205 | + if (s->vkctx.output_format == s->vkctx.input_format) { | |
| 206 | + for (int i = 0; i < desc[i].elems; i++) { | |
| 207 | + GLSLF(1, size = imageSize(output_img[%i]); ,i); | |
| 208 | + GLSLC(1, if (IS_WITHIN(pos, size)) { ); | |
| 209 | + switch (s->scaler) { | |
| 210 | + case F_NEAREST: | |
| 211 | + case F_BILINEAR: | |
| 212 | + GLSLF(2, vec4 res = scale_bilinear(%i, pos, c_r, c_o); ,i); | |
| 213 | + GLSLF(2, imageStore(output_img[%i], pos, res); ,i); | |
| 214 | + break; | |
| 215 | + }; | |
| 216 | + GLSLC(1, } ); | |
| 217 | + } | |
| 218 | + } else { | |
| 219 | + GLSLC(1, vec4 res = scale_bilinear(0, pos, c_r, c_o); ); | |
| 220 | + GLSLF(1, res = rgb2yuv(res, %i); ,s->out_range == AVCOL_RANGE_JPEG); | |
| 221 | + switch (s->vkctx.output_format) { | |
| 222 | + case AV_PIX_FMT_NV12: GLSLC(1, write_nv12(res, pos); ); break; | |
| 223 | + case AV_PIX_FMT_YUV420P: GLSLC(1, write_420(res, pos); ); break; | |
| 224 | + case AV_PIX_FMT_YUV444P: GLSLC(1, write_444(res, pos); ); break; | |
| 225 | + default: return AVERROR(EINVAL); | |
| 226 | + } | |
| 236 | 227 | } |
| 237 | 228 | |
| 238 | − RET(ff_vk_init_pipeline_layout(vkctx, s->pl)); | |
| 239 | − RET(ff_vk_init_compute_pipeline(vkctx, s->pl)); | |
| 229 | + GLSLC(0, } ); | |
| 240 | 230 | |
| 241 | 231 | if (s->vkctx.output_format != s->vkctx.input_format) { |
| 242 | 232 | const AVLumaCoefficients *lcoeffs; |
| 243 | 233 | double tmp_mat[3][3]; |
| 244 | 234 | |
| 245 | − struct { | |
| 246 | − float yuv_matrix[4][4]; | |
| 247 | − } *par; | |
| 248 | − | |
| 249 | 235 | lcoeffs = av_csp_luma_coeffs_from_avcsp(in->colorspace); |
| 250 | 236 | if (!lcoeffs) { |
| 251 | 237 | av_log(ctx, AV_LOG_ERROR, "Unsupported colorspace\n"); |
| 252 | 238 | return AVERROR(EINVAL); |
| 253 | 239 | } |
| 254 | 240 | |
| 255 | − RET(ff_vk_create_buf(vkctx, &s->params_buf, | |
| 256 | − sizeof(*par), NULL, NULL, | |
| 257 | − VK_BUFFER_USAGE_STORAGE_BUFFER_BIT, | |
| 258 | − VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT)); | |
| 259 | − | |
| 260 | − RET(ff_vk_map_buffers(vkctx, &s->params_buf, (uint8_t **)&par, 1, 0)); | |
| 261 | − | |
| 262 | 241 | ff_fill_rgb2yuv_table(lcoeffs, tmp_mat); |
| 263 | 242 | |
| 264 | − memset(par, 0, sizeof(*par)); | |
| 265 | − | |
| 266 | 243 | for (int y = 0; y < 3; y++) |
| 267 | 244 | for (int x = 0; x < 3; x++) |
| 268 | − par->yuv_matrix[x][y] = tmp_mat[x][y]; | |
| 269 | − | |
| 270 | − par->yuv_matrix[3][3] = 1.0; | |
| 271 | − | |
| 272 | − RET(ff_vk_unmap_buffers(vkctx, &s->params_buf, 1, 1)); | |
| 273 | − | |
| 274 | − s->params_desc.buffer = s->params_buf.buf; | |
| 275 | − s->params_desc.range = VK_WHOLE_SIZE; | |
| 276 | − | |
| 277 | − ff_vk_update_descriptor_set(vkctx, s->pl, 1); | |
| 245 | + s->opts.yuv_matrix[x][y] = tmp_mat[x][y]; | |
| 246 | + s->opts.yuv_matrix[3][3] = 1.0; | |
| 278 | 247 | } |
| 279 | 248 | |
| 280 | − /* Execution context */ | |
| 281 | − RET(ff_vk_create_exec_ctx(vkctx, &s->exec, &s->qf)); | |
| 249 | + RET(spv->compile_shader(spv, ctx, shd, &spv_data, &spv_len, "main", | |
| 250 | + &spv_opaque)); | |
| 251 | + RET(ff_vk_shader_create(vkctx, shd, spv_data, spv_len, "main")); | |
| 252 | + | |
| 253 | + RET(ff_vk_init_compute_pipeline(vkctx, &s->pl, shd)); | |
| 254 | + RET(ff_vk_exec_pipeline_register(vkctx, &s->e, &s->pl)); | |
| 282 | 255 | |
| 283 | 256 | s->initialized = 1; |
| 284 | 257 | |
| 285 | 258 | return 0; |
| 286 | 259 | |
| 287 | 260 | fail: |
| 288 | − return err; | |
| 289 | −} | |
| 290 | − | |
| 291 | −static int process_frames(AVFilterContext *avctx, AVFrame *out_f, AVFrame *in_f) | |
| 292 | −{ | |
| 293 | − int err = 0; | |
| 294 | − VkCommandBuffer cmd_buf; | |
| 295 | − ScaleVulkanContext *s = avctx->priv; | |
| 296 | − FFVulkanContext *vkctx = &s->vkctx; | |
| 297 | − FFVulkanFunctions *vk = &vkctx->vkfn; | |
| 298 | − AVVkFrame *in = (AVVkFrame *)in_f->data[0]; | |
| 299 | − AVVkFrame *out = (AVVkFrame *)out_f->data[0]; | |
| 300 | − VkImageMemoryBarrier barriers[AV_NUM_DATA_POINTERS*2]; | |
| 301 | − int barrier_count = 0; | |
| 302 | − const int planes = av_pix_fmt_count_planes(s->vkctx.input_format); | |
| 303 | − const VkFormat *input_formats = av_vkfmt_from_pixfmt(s->vkctx.input_format); | |
| 304 | − const VkFormat *output_formats = av_vkfmt_from_pixfmt(s->vkctx.output_format); | |
| 305 | − | |
| 306 | − /* Update descriptors and init the exec context */ | |
| 307 | − ff_vk_start_exec_recording(vkctx, s->exec); | |
| 308 | − cmd_buf = ff_vk_get_exec_buf(s->exec); | |
| 309 | − | |
| 310 | − for (int i = 0; i < planes; i++) { | |
| 311 | − RET(ff_vk_create_imageview(vkctx, s->exec, | |
| 312 | − &s->input_images[i].imageView, in->img[i], | |
| 313 | − input_formats[i], | |
| 314 | − ff_comp_identity_map)); | |
| 315 | − | |
| 316 | − RET(ff_vk_create_imageview(vkctx, s->exec, | |
| 317 | − &s->output_images[i].imageView, out->img[i], | |
| 318 | − output_formats[i], | |
| 319 | − ff_comp_identity_map)); | |
| 320 | − | |
| 321 | − s->input_images[i].imageLayout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL; | |
| 322 | − s->output_images[i].imageLayout = VK_IMAGE_LAYOUT_GENERAL; | |
| 323 | − } | |
| 324 | − | |
| 325 | − ff_vk_update_descriptor_set(vkctx, s->pl, 0); | |
| 326 | − | |
| 327 | − for (int i = 0; i < planes; i++) { | |
| 328 | − VkImageMemoryBarrier bar = { | |
| 329 | − .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER, | |
| 330 | − .srcAccessMask = 0, | |
| 331 | − .dstAccessMask = VK_ACCESS_SHADER_READ_BIT, | |
| 332 | − .oldLayout = in->layout[i], | |
| 333 | − .newLayout = s->input_images[i].imageLayout, | |
| 334 | − .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED, | |
| 335 | − .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED, | |
| 336 | − .image = in->img[i], | |
| 337 | − .subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT, | |
| 338 | − .subresourceRange.levelCount = 1, | |
| 339 | − .subresourceRange.layerCount = 1, | |
| 340 | − }; | |
| 341 | − | |
| 342 | − memcpy(&barriers[barrier_count++], &bar, sizeof(VkImageMemoryBarrier)); | |
| 343 | − | |
| 344 | − in->layout[i] = bar.newLayout; | |
| 345 | − in->access[i] = bar.dstAccessMask; | |
| 346 | − } | |
| 347 | − | |
| 348 | − for (int i = 0; i < av_pix_fmt_count_planes(s->vkctx.output_format); i++) { | |
| 349 | − VkImageMemoryBarrier bar = { | |
| 350 | − .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER, | |
| 351 | − .srcAccessMask = 0, | |
| 352 | − .dstAccessMask = VK_ACCESS_SHADER_WRITE_BIT, | |
| 353 | − .oldLayout = out->layout[i], | |
| 354 | − .newLayout = s->output_images[i].imageLayout, | |
| 355 | − .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED, | |
| 356 | − .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED, | |
| 357 | − .image = out->img[i], | |
| 358 | − .subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT, | |
| 359 | − .subresourceRange.levelCount = 1, | |
| 360 | − .subresourceRange.layerCount = 1, | |
| 361 | − }; | |
| 362 | − | |
| 363 | − memcpy(&barriers[barrier_count++], &bar, sizeof(VkImageMemoryBarrier)); | |
| 364 | − | |
| 365 | − out->layout[i] = bar.newLayout; | |
| 366 | − out->access[i] = bar.dstAccessMask; | |
| 367 | − } | |
| 368 | − | |
| 369 | − vk->CmdPipelineBarrier(cmd_buf, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, | |
| 370 | − VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, 0, | |
| 371 | − 0, NULL, 0, NULL, barrier_count, barriers); | |
| 372 | − | |
| 373 | − ff_vk_bind_pipeline_exec(vkctx, s->exec, s->pl); | |
| 374 | − | |
| 375 | − vk->CmdDispatch(cmd_buf, | |
| 376 | − FFALIGN(vkctx->output_width, CGROUPS[0])/CGROUPS[0], | |
| 377 | − FFALIGN(vkctx->output_height, CGROUPS[1])/CGROUPS[1], 1); | |
| 378 | − | |
| 379 | − ff_vk_add_exec_dep(vkctx, s->exec, in_f, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT); | |
| 380 | − ff_vk_add_exec_dep(vkctx, s->exec, out_f, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT); | |
| 381 | − | |
| 382 | − err = ff_vk_submit_exec_queue(vkctx, s->exec); | |
| 383 | − if (err) | |
| 384 | − return err; | |
| 385 | − | |
| 386 | − ff_vk_qf_rotate(&s->qf); | |
| 261 | + if (spv_opaque) | |
| 262 | + spv->free_shader(spv, &spv_opaque); | |
| 263 | + if (spv) | |
| 264 | + spv->uninit(&spv); | |
| 387 | 265 | |
| 388 | 266 | return err; |
| 389 | − | |
| 390 | −fail: | |
| 391 | − ff_vk_discard_exec_deps(s->exec); | |
| 392 | − return err; | |
| 393 | 267 | } |
| 394 | 268 | |
| 395 | 269 | static int scale_vulkan_filter_frame(AVFilterLink *link, AVFrame *in) |
| @@ -408,7 +282,8 @@ static int scale_vulkan_filter_frame(AVFilterLink *link, AVFrame *in) · 12 unmodified lines | ||
| 408 | 282 | if (!s->initialized) |
| 409 | 283 | RET(init_filter(ctx, in)); |
| 410 | 284 | |
| 411 | − RET(process_frames(ctx, out, in)); | |
| 285 | + RET(ff_vk_filter_process_simple(&s->vkctx, &s->e, &s->pl, out, in, | |
| 286 | + s->sampler, &s->opts, sizeof(s->opts))); | |
| 412 | 287 | |
| 413 | 288 | err = av_frame_copy_props(out, in); |
| 414 | 289 | if (err < 0) |
| @@ -475,8 +350,17 @@ static int scale_vulkan_config_output(AVFilterLink *outlink) · 60 unmodified lines | ||
| 475 | 350 | static void scale_vulkan_uninit(AVFilterContext *avctx) |
| 476 | 351 | { |
| 477 | 352 | ScaleVulkanContext *s = avctx->priv; |
| 353 | + FFVulkanContext *vkctx = &s->vkctx; | |
| 354 | + FFVulkanFunctions *vk = &vkctx->vkfn; | |
| 355 | + | |
| 356 | + ff_vk_exec_pool_free(vkctx, &s->e); | |
| 357 | + ff_vk_pipeline_free(vkctx, &s->pl); | |
| 358 | + ff_vk_shader_free(vkctx, &s->shd); | |
| 359 | + | |
| 360 | + if (s->sampler) | |
| 361 | + vk->DestroySampler(vkctx->hwctx->act_dev, s->sampler, | |
| 362 | + vkctx->hwctx->alloc); | |
| 478 | 363 | |
| 479 | − ff_vk_free_buf(&s->vkctx, &s->params_buf); | |
| 480 | 364 | ff_vk_uninit(&s->vkctx); |
| 481 | 365 | |
| 482 | 366 | s->initialized = 0; |