Unified · Split
| @@ -1,5 +1,7 @@ | |||
| 1 | /* | 1 | /* |
| 2 | * copyright (c) 2021 Wu Jianhua <jianhua.wu@intel.com> | 2 | * copyright (c) 2021 Wu Jianhua <jianhua.wu@intel.com> |
| 3 | * Copyright (c) Lynne | ||
| 4 | * | ||
| 3 | * This file is part of FFmpeg. | 5 | * This file is part of FFmpeg. |
| 4 | * | 6 | * |
| 5 | * FFmpeg is free software; you can redistribute it and/or | 7 | * FFmpeg is free software; you can redistribute it and/or |
| @@ -20,41 +22,60 @@ · 14 unmodified lines | |||
| 20 | #include "libavutil/random_seed.h" | 22 | #include "libavutil/random_seed.h" |
| 21 | #include "libavutil/opt.h" | 23 | #include "libavutil/opt.h" |
| 22 | #include "vulkan_filter.h" | 24 | #include "vulkan_filter.h" |
| 25 | #include "vulkan_spirv.h" | ||
| 23 | #include "internal.h" | 26 | #include "internal.h" |
| 24 | #include "transpose.h" | 27 | #include "transpose.h" |
| 25 | 28 | ||
| 26 | #define CGS 32 | ||
| 27 | |||
| 28 | typedef struct TransposeVulkanContext { | 29 | typedef struct TransposeVulkanContext { |
| 29 | FFVulkanContext vkctx; | 30 | FFVulkanContext vkctx; |
| 30 | FFVkQueueFamilyCtx qf; | ||
| 31 | FFVkExecContext *exec; | ||
| 32 | FFVulkanPipeline *pl; | ||
| 33 | 31 | ||
| 34 | VkDescriptorImageInfo input_images[3]; | 32 | int initialized; |
| 35 | VkDescriptorImageInfo output_images[3]; | 33 | FFVulkanPipeline pl; |
| 34 | FFVkExecPool e; | ||
| 35 | FFVkQueueFamilyCtx qf; | ||
| 36 | FFVkSPIRVShader shd; | ||
| 37 | VkSampler sampler; | ||
| 36 | 38 | ||
| 37 | int dir; | 39 | int dir; |
| 38 | int passthrough; | 40 | int passthrough; |
| 39 | int initialized; | ||
| 40 | } TransposeVulkanContext; | 41 | } TransposeVulkanContext; |
| 41 | 42 | ||
| 42 | static av_cold int init_filter(AVFilterContext *ctx, AVFrame *in) | 43 | static av_cold int init_filter(AVFilterContext *ctx, AVFrame *in) |
| 43 | { | 44 | { |
| 44 | int err = 0; | 45 | int err; |
| 45 | FFVkSPIRVShader *shd; | 46 | uint8_t *spv_data; |
| 47 | size_t spv_len; | ||
| 48 | void *spv_opaque = NULL; | ||
| 46 | TransposeVulkanContext *s = ctx->priv; | 49 | TransposeVulkanContext *s = ctx->priv; |
| 47 | FFVulkanContext *vkctx = &s->vkctx; | 50 | FFVulkanContext *vkctx = &s->vkctx; |
| 51 | |||
| 48 | const int planes = av_pix_fmt_count_planes(s->vkctx.output_format); | 52 | const int planes = av_pix_fmt_count_planes(s->vkctx.output_format); |
| 53 | FFVkSPIRVShader *shd = &s->shd; | ||
| 54 | FFVkSPIRVCompiler *spv; | ||
| 55 | FFVulkanDescriptorSetBinding *desc; | ||
| 56 | |||
| 57 | spv = ff_vk_spirv_init(); | ||
| 58 | if (!spv) { | ||
| 59 | av_log(ctx, AV_LOG_ERROR, "Unable to initialize SPIR-V compiler!\n"); | ||
| 60 | return AVERROR_EXTERNAL; | ||
| 61 | } | ||
| 62 | |||
| 63 | ff_vk_qf_init(vkctx, &s->qf, VK_QUEUE_COMPUTE_BIT); | ||
| 64 | RET(ff_vk_exec_pool_init(vkctx, &s->qf, &s->e, s->qf.nb_queues*4, 0, 0, 0, NULL)); | ||
| 65 | RET(ff_vk_init_sampler(vkctx, &s->sampler, 1, VK_FILTER_LINEAR)); | ||
| 66 | RET(ff_vk_shader_init(&s->pl, &s->shd, "transpose_compute", | ||
| 67 | VK_SHADER_STAGE_COMPUTE_BIT, 0)); | ||
| 49 | 68 | ||
| 50 | FFVulkanDescriptorSetBinding image_descs[] = { | 69 | ff_vk_shader_set_compute_sizes(&s->shd, 32, 1, 1); |
| 70 | |||
| 71 | desc = (FFVulkanDescriptorSetBinding []) { | ||
| 51 | { | 72 | { |
| 52 | .name = "input_images", | 73 | .name = "input_images", |
| 53 | .type = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, | 74 | .type = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, |
| 54 | .dimensions = 2, | 75 | .dimensions = 2, |
| 55 | .elems = planes, | 76 | .elems = planes, |
| 56 | .stages = VK_SHADER_STAGE_COMPUTE_BIT, | 77 | .stages = VK_SHADER_STAGE_COMPUTE_BIT, |
| 57 | .updater = s->input_images, | 78 | .samplers = DUP_SAMPLER(s->sampler), |
| 58 | }, | 79 | }, |
| 59 | { | 80 | { |
| 60 | .name = "output_images", | 81 | .name = "output_images", |
| @@ -64,154 +85,49 @@ static av_cold int init_filter(AVFilterContext *ctx, AVFrame *in) · 3 unmodified lines | |||
| 64 | .dimensions = 2, | 85 | .dimensions = 2, |
| 65 | .elems = planes, | 86 | .elems = planes, |
| 66 | .stages = VK_SHADER_STAGE_COMPUTE_BIT, | 87 | .stages = VK_SHADER_STAGE_COMPUTE_BIT, |
| 67 | .updater = s->output_images, | ||
| 68 | }, | 88 | }, |
| 69 | }; | 89 | }; |
| 70 | 90 | ||
| 71 | image_descs[0].sampler = ff_vk_init_sampler(vkctx, 1, VK_FILTER_LINEAR); | 91 | RET(ff_vk_pipeline_descriptor_set_add(vkctx, &s->pl, shd, desc, 2, 0, 0)); |
| 72 | if (!image_descs[0].sampler) | ||
| 73 | return AVERROR_EXTERNAL; | ||
| 74 | |||
| 75 | ff_vk_qf_init(vkctx, &s->qf, VK_QUEUE_COMPUTE_BIT, 0); | ||
| 76 | |||
| 77 | { | ||
| 78 | s->pl = ff_vk_create_pipeline(vkctx, &s->qf); | ||
| 79 | if (!s->pl) | ||
| 80 | return AVERROR(ENOMEM); | ||
| 81 | |||
| 82 | shd = ff_vk_init_shader(s->pl, "transpose_compute", image_descs[0].stages); | ||
| 83 | if (!shd) | ||
| 84 | return AVERROR(ENOMEM); | ||
| 85 | |||
| 86 | ff_vk_set_compute_shader_sizes(shd, (int [3]){ CGS, 1, 1 }); | ||
| 87 | RET(ff_vk_add_descriptor_set(vkctx, s->pl, shd, image_descs, FF_ARRAY_ELEMS(image_descs), 0)); | ||
| 88 | |||
| 89 | GLSLC(0, void main() ); | ||
| 90 | GLSLC(0, { ); | ||
| 91 | GLSLC(1, ivec2 size; ); | ||
| 92 | GLSLC(1, ivec2 pos = ivec2(gl_GlobalInvocationID.xy); ); | ||
| 93 | for (int i = 0; i < planes; i++) { | ||
| 94 | GLSLC(0, ); | ||
| 95 | GLSLF(1, size = imageSize(output_images[%i]); ,i); | ||
| 96 | GLSLC(1, if (IS_WITHIN(pos, size)) { ); | ||
| 97 | if (s->dir == TRANSPOSE_CCLOCK) | ||
| 98 | GLSLF(2, vec4 res = texture(input_images[%i], ivec2(size.y - pos.y, pos.x)); ,i); | ||
| 99 | else if (s->dir == TRANSPOSE_CLOCK_FLIP || s->dir == TRANSPOSE_CLOCK) { | ||
| 100 | GLSLF(2, vec4 res = texture(input_images[%i], ivec2(size.yx - pos.yx)); ,i); | ||
| 101 | if (s->dir == TRANSPOSE_CLOCK) | ||
| 102 | GLSLC(2, pos = ivec2(pos.x, size.y - pos.y); ); | ||
| 103 | } else | ||
| 104 | GLSLF(2, vec4 res = texture(input_images[%i], pos.yx); ,i); | ||
| 105 | GLSLF(2, imageStore(output_images[%i], pos, res); ,i); | ||
| 106 | GLSLC(1, } ); | ||
| 107 | } | ||
| 108 | GLSLC(0, } ); | ||
| 109 | |||
| 110 | RET(ff_vk_compile_shader(vkctx, shd, "main")); | ||
| 111 | RET(ff_vk_init_pipeline_layout(vkctx, s->pl)); | ||
| 112 | RET(ff_vk_init_compute_pipeline(vkctx, s->pl)); | ||
| 113 | } | ||
| 114 | |||
| 115 | RET(ff_vk_create_exec_ctx(vkctx, &s->exec, &s->qf)); | ||
| 116 | s->initialized = 1; | ||
| 117 | |||
| 118 | fail: | ||
| 119 | return err; | ||
| 120 | } | ||
| 121 | |||
| 122 | static int process_frames(AVFilterContext *avctx, AVFrame *outframe, AVFrame *inframe) | ||
| 123 | { | ||
| 124 | int err = 0; | ||
| 125 | VkCommandBuffer cmd_buf; | ||
| 126 | TransposeVulkanContext *s = avctx->priv; | ||
| 127 | FFVulkanContext *vkctx = &s->vkctx; | ||
| 128 | FFVulkanFunctions *vk = &s->vkctx.vkfn; | ||
| 129 | const int planes = av_pix_fmt_count_planes(s->vkctx.output_format); | ||
| 130 | |||
| 131 | AVVkFrame *in = (AVVkFrame *)inframe->data[0]; | ||
| 132 | AVVkFrame *out = (AVVkFrame *)outframe->data[0]; | ||
| 133 | |||
| 134 | const VkFormat *input_formats = av_vkfmt_from_pixfmt(s->vkctx.input_format); | ||
| 135 | const VkFormat *output_formats = av_vkfmt_from_pixfmt(s->vkctx.output_format); | ||
| 136 | |||
| 137 | ff_vk_start_exec_recording(vkctx, s->exec); | ||
| 138 | cmd_buf = ff_vk_get_exec_buf(s->exec); | ||
| 139 | |||
| 140 | for (int i = 0; i < planes; i++) { | ||
| 141 | RET(ff_vk_create_imageview(vkctx, s->exec, | ||
| 142 | &s->input_images[i].imageView, in->img[i], | ||
| 143 | input_formats[i], | ||
| 144 | ff_comp_identity_map)); | ||
| 145 | |||
| 146 | RET(ff_vk_create_imageview(vkctx, s->exec, | ||
| 147 | &s->output_images[i].imageView, out->img[i], | ||
| 148 | output_formats[i], | ||
| 149 | ff_comp_identity_map)); | ||
| 150 | |||
| 151 | s->input_images[i].imageLayout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL; | ||
| 152 | s->output_images[i].imageLayout = VK_IMAGE_LAYOUT_GENERAL; | ||
| 153 | } | ||
| 154 | |||
| 155 | ff_vk_update_descriptor_set(vkctx, s->pl, 0); | ||
| 156 | 92 | ||
| 93 | GLSLC(0, void main() ); | ||
| 94 | GLSLC(0, { ); | ||
| 95 | GLSLC(1, ivec2 size; ); | ||
| 96 | GLSLC(1, ivec2 pos = ivec2(gl_GlobalInvocationID.xy); ); | ||
| 157 | for (int i = 0; i < planes; i++) { | 97 | for (int i = 0; i < planes; i++) { |
| 158 | VkImageMemoryBarrier barriers[] = { | 98 | GLSLC(0, ); |
| 159 | { | 99 | GLSLF(1, size = imageSize(output_images[%i]); ,i); |
| 160 | .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER, | 100 | GLSLC(1, if (IS_WITHIN(pos, size)) { ); |
| 161 | .srcAccessMask = 0, | 101 | if (s->dir == TRANSPOSE_CCLOCK) |
| 162 | .dstAccessMask = VK_ACCESS_SHADER_READ_BIT, | 102 | GLSLF(2, vec4 res = texture(input_images[%i], ivec2(size.y - pos.y, pos.x)); ,i); |
| 163 | .oldLayout = in->layout[i], | 103 | else if (s->dir == TRANSPOSE_CLOCK_FLIP || s->dir == TRANSPOSE_CLOCK) { |
| 164 | .newLayout = s->input_images[i].imageLayout, | 104 | GLSLF(2, vec4 res = texture(input_images[%i], ivec2(size.yx - pos.yx)); ,i); |
| 165 | .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED, | 105 | if (s->dir == TRANSPOSE_CLOCK) |
| 166 | .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED, | 106 | GLSLC(2, pos = ivec2(pos.x, size.y - pos.y); ); |
| 167 | .image = in->img[i], | 107 | } else |
| 168 | .subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT, | 108 | GLSLF(2, vec4 res = texture(input_images[%i], pos.yx); ,i); |
| 169 | .subresourceRange.levelCount = 1, | 109 | GLSLF(2, imageStore(output_images[%i], pos, res); ,i); |
| 170 | .subresourceRange.layerCount = 1, | 110 | GLSLC(1, } ); |
| 171 | }, | ||
| 172 | { | ||
| 173 | .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER, | ||
| 174 | .srcAccessMask = 0, | ||
| 175 | .dstAccessMask = VK_ACCESS_SHADER_WRITE_BIT, | ||
| 176 | .oldLayout = out->layout[i], | ||
| 177 | .newLayout = s->output_images[i].imageLayout, | ||
| 178 | .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED, | ||
| 179 | .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED, | ||
| 180 | .image = out->img[i], | ||
| 181 | .subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT, | ||
| 182 | .subresourceRange.levelCount = 1, | ||
| 183 | .subresourceRange.layerCount = 1, | ||
| 184 | }, | ||
| 185 | }; | ||
| 186 | |||
| 187 | vk->CmdPipelineBarrier(cmd_buf, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, | ||
| 188 | VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, 0, | ||
| 189 | 0, NULL, 0, NULL, FF_ARRAY_ELEMS(barriers), barriers); | ||
| 190 | |||
| 191 | in->layout[i] = barriers[0].newLayout; | ||
| 192 | in->access[i] = barriers[0].dstAccessMask; | ||
| 193 | |||
| 194 | out->layout[i] = barriers[1].newLayout; | ||
| 195 | out->access[i] = barriers[1].dstAccessMask; | ||
| 196 | } | 111 | } |
| 112 | GLSLC(0, } ); | ||
| 197 | 113 | ||
| 198 | ff_vk_bind_pipeline_exec(vkctx, s->exec, s->pl); | 114 | RET(spv->compile_shader(spv, ctx, shd, &spv_data, &spv_len, "main", |
| 199 | vk->CmdDispatch(cmd_buf, FFALIGN(s->vkctx.output_width, CGS)/CGS, | 115 | &spv_opaque)); |
| 200 | s->vkctx.output_height, 1); | 116 | RET(ff_vk_shader_create(vkctx, shd, spv_data, spv_len, "main")); |
| 201 | 117 | ||
| 202 | ff_vk_add_exec_dep(vkctx, s->exec, inframe, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT); | 118 | RET(ff_vk_init_compute_pipeline(vkctx, &s->pl, shd)); |
| 203 | ff_vk_add_exec_dep(vkctx, s->exec, outframe, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT); | 119 | RET(ff_vk_exec_pipeline_register(vkctx, &s->e, &s->pl)); |
| 204 | 120 | ||
| 205 | err = ff_vk_submit_exec_queue(vkctx, s->exec); | 121 | s->initialized = 1; |
| 206 | if (err) | ||
| 207 | return err; | ||
| 208 | |||
| 209 | ff_vk_qf_rotate(&s->qf); | ||
| 210 | 122 | ||
| 211 | return 0; | 123 | return 0; |
| 212 | 124 | ||
| 213 | fail: | 125 | fail: |
| 214 | ff_vk_discard_exec_deps(s->exec); | 126 | if (spv_opaque) |
| 127 | spv->free_shader(spv, &spv_opaque); | ||
| 128 | if (spv) | ||
| 129 | spv->uninit(&spv); | ||
| 130 | |||
| 215 | return err; | 131 | return err; |
| 216 | } | 132 | } |
| 217 | 133 | ||
| @@ -235,7 +151,8 @@ static int filter_frame(AVFilterLink *inlink, AVFrame *in) · 17 unmodified lines | |||
| 235 | if (!s->initialized) | 151 | if (!s->initialized) |
| 236 | RET(init_filter(ctx, in)); | 152 | RET(init_filter(ctx, in)); |
| 237 | 153 | ||
| 238 | RET(process_frames(ctx, out, in)); | 154 | RET(ff_vk_filter_process_simple(&s->vkctx, &s->e, &s->pl, out, in, |
| 155 | s->sampler, NULL, 0)); | ||
| 239 | 156 | ||
| 240 | RET(av_frame_copy_props(out, in)); | 157 | RET(av_frame_copy_props(out, in)); |
| 241 | 158 | ||
| @@ -259,6 +176,17 @@ fail: · 17 unmodified lines | |||
| 259 | static av_cold void transpose_vulkan_uninit(AVFilterContext *avctx) | 176 | static av_cold void transpose_vulkan_uninit(AVFilterContext *avctx) |
| 260 | { | 177 | { |
| 261 | TransposeVulkanContext *s = avctx->priv; | 178 | TransposeVulkanContext *s = avctx->priv; |
| 179 | FFVulkanContext *vkctx = &s->vkctx; | ||
| 180 | FFVulkanFunctions *vk = &vkctx->vkfn; | ||
| 181 | |||
| 182 | ff_vk_exec_pool_free(vkctx, &s->e); | ||
| 183 | ff_vk_pipeline_free(vkctx, &s->pl); | ||
| 184 | ff_vk_shader_free(vkctx, &s->shd); | ||
| 185 | |||
| 186 | if (s->sampler) | ||
| 187 | vk->DestroySampler(vkctx->hwctx->act_dev, s->sampler, | ||
| 188 | vkctx->hwctx->alloc); | ||
| 189 | |||
| 262 | ff_vk_uninit(&s->vkctx); | 190 | ff_vk_uninit(&s->vkctx); |
| 263 | 191 | ||
| 264 | s->initialized = 0; | 192 | s->initialized = 0; |