| @@ -0,0 +1,316 @@ |
| 1 | +/* |
| 2 | + * copyright (c) 2021 Wu Jianhua <jianhua.wu@intel.com> |
| 3 | + * This file is part of FFmpeg. |
| 4 | + * |
| 5 | + * FFmpeg is free software; you can redistribute it and/or |
| 6 | + * modify it under the terms of the GNU Lesser General Public |
| 7 | + * License as published by the Free Software Foundation; either |
| 8 | + * version 2.1 of the License, or (at your option) any later version. |
| 9 | + * |
| 10 | + * FFmpeg is distributed in the hope that it will be useful, |
| 11 | + * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 12 | + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
| 13 | + * Lesser General Public License for more details. |
| 14 | + * |
| 15 | + * You should have received a copy of the GNU Lesser General Public |
| 16 | + * License along with FFmpeg; if not, write to the Free Software |
| 17 | + * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
| 18 | + */ |
| 19 | + |
| 20 | +#include "libavutil/random_seed.h" |
| 21 | +#include "libavutil/opt.h" |
| 22 | +#include "vulkan_filter.h" |
| 23 | +#include "internal.h" |
| 24 | + |
| 25 | +#define CGS 32 |
| 26 | + |
| 27 | +typedef struct TransposeVulkanContext { |
| 28 | + FFVulkanContext vkctx; |
| 29 | + FFVkQueueFamilyCtx qf; |
| 30 | + FFVkExecContext *exec; |
| 31 | + FFVulkanPipeline *pl; |
| 32 | + |
| 33 | + VkDescriptorImageInfo input_images[3]; |
| 34 | + VkDescriptorImageInfo output_images[3]; |
| 35 | + |
| 36 | + int initialized; |
| 37 | +} TransposeVulkanContext; |
| 38 | + |
| 39 | +static av_cold int init_filter(AVFilterContext *ctx, AVFrame *in) |
| 40 | +{ |
| 41 | + int err = 0; |
| 42 | + FFVkSPIRVShader *shd; |
| 43 | + TransposeVulkanContext *s = ctx->priv; |
| 44 | + FFVulkanContext *vkctx = &s->vkctx; |
| 45 | + const int planes = av_pix_fmt_count_planes(s->vkctx.output_format); |
| 46 | + |
| 47 | + FFVulkanDescriptorSetBinding image_descs[] = { |
| 48 | + { |
| 49 | + .name = "input_images", |
| 50 | + .type = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, |
| 51 | + .dimensions = 2, |
| 52 | + .elems = planes, |
| 53 | + .stages = VK_SHADER_STAGE_COMPUTE_BIT, |
| 54 | + .updater = s->input_images, |
| 55 | + }, |
| 56 | + { |
| 57 | + .name = "output_images", |
| 58 | + .type = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE, |
| 59 | + .mem_layout = ff_vk_shader_rep_fmt(s->vkctx.output_format), |
| 60 | + .mem_quali = "writeonly", |
| 61 | + .dimensions = 2, |
| 62 | + .elems = planes, |
| 63 | + .stages = VK_SHADER_STAGE_COMPUTE_BIT, |
| 64 | + .updater = s->output_images, |
| 65 | + }, |
| 66 | + }; |
| 67 | + |
| 68 | + image_descs[0].sampler = ff_vk_init_sampler(vkctx, 1, VK_FILTER_LINEAR); |
| 69 | + if (!image_descs[0].sampler) |
| 70 | + return AVERROR_EXTERNAL; |
| 71 | + |
| 72 | + ff_vk_qf_init(vkctx, &s->qf, VK_QUEUE_COMPUTE_BIT, 0); |
| 73 | + |
| 74 | + { |
| 75 | + s->pl = ff_vk_create_pipeline(vkctx, &s->qf); |
| 76 | + if (!s->pl) |
| 77 | + return AVERROR(ENOMEM); |
| 78 | + |
| 79 | + shd = ff_vk_init_shader(s->pl, "transpose_compute", image_descs[0].stages); |
| 80 | + if (!shd) |
| 81 | + return AVERROR(ENOMEM); |
| 82 | + |
| 83 | + ff_vk_set_compute_shader_sizes(shd, (int [3]){ CGS, 1, 1 }); |
| 84 | + RET(ff_vk_add_descriptor_set(vkctx, s->pl, shd, image_descs, FF_ARRAY_ELEMS(image_descs), 0)); |
| 85 | + |
| 86 | + GLSLC(0, void main() ); |
| 87 | + GLSLC(0, { ); |
| 88 | + GLSLC(1, ivec2 size; ); |
| 89 | + GLSLC(1, const ivec2 pos = ivec2(gl_GlobalInvocationID.xy); ); |
| 90 | + for (int i = 0; i < planes; i++) { |
| 91 | + GLSLC(0, ); |
| 92 | + GLSLF(1, size = imageSize(output_images[%i]); ,i); |
| 93 | + GLSLC(1, if (IS_WITHIN(pos, size)) { ); |
| 94 | + GLSLF(2, vec4 res = texture(input_images[%i], pos.yx); ,i); |
| 95 | + GLSLF(2, imageStore(output_images[%i], pos, res); ,i); |
| 96 | + GLSLC(1, } ); |
| 97 | + } |
| 98 | + GLSLC(0, } ); |
| 99 | + |
| 100 | + RET(ff_vk_compile_shader(vkctx, shd, "main")); |
| 101 | + RET(ff_vk_init_pipeline_layout(vkctx, s->pl)); |
| 102 | + RET(ff_vk_init_compute_pipeline(vkctx, s->pl)); |
| 103 | + } |
| 104 | + |
| 105 | + RET(ff_vk_create_exec_ctx(vkctx, &s->exec, &s->qf)); |
| 106 | + s->initialized = 1; |
| 107 | + |
| 108 | +fail: |
| 109 | + return err; |
| 110 | +} |
| 111 | + |
| 112 | +static int process_frames(AVFilterContext *avctx, AVFrame *outframe, AVFrame *inframe) |
| 113 | +{ |
| 114 | + int err = 0; |
| 115 | + VkCommandBuffer cmd_buf; |
| 116 | + TransposeVulkanContext *s = avctx->priv; |
| 117 | + FFVulkanContext *vkctx = &s->vkctx; |
| 118 | + FFVulkanFunctions *vk = &s->vkctx.vkfn; |
| 119 | + const int planes = av_pix_fmt_count_planes(s->vkctx.output_format); |
| 120 | + |
| 121 | + AVVkFrame *in = (AVVkFrame *)inframe->data[0]; |
| 122 | + AVVkFrame *out = (AVVkFrame *)outframe->data[0]; |
| 123 | + |
| 124 | + const VkFormat *input_formats = av_vkfmt_from_pixfmt(s->vkctx.input_format); |
| 125 | + const VkFormat *output_formats = av_vkfmt_from_pixfmt(s->vkctx.output_format); |
| 126 | + |
| 127 | + ff_vk_start_exec_recording(vkctx, s->exec); |
| 128 | + cmd_buf = ff_vk_get_exec_buf(s->exec); |
| 129 | + |
| 130 | + for (int i = 0; i < planes; i++) { |
| 131 | + RET(ff_vk_create_imageview(vkctx, s->exec, |
| 132 | + &s->input_images[i].imageView, in->img[i], |
| 133 | + input_formats[i], |
| 134 | + ff_comp_identity_map)); |
| 135 | + |
| 136 | + RET(ff_vk_create_imageview(vkctx, s->exec, |
| 137 | + &s->output_images[i].imageView, out->img[i], |
| 138 | + output_formats[i], |
| 139 | + ff_comp_identity_map)); |
| 140 | + |
| 141 | + s->input_images[i].imageLayout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL; |
| 142 | + s->output_images[i].imageLayout = VK_IMAGE_LAYOUT_GENERAL; |
| 143 | + } |
| 144 | + |
| 145 | + ff_vk_update_descriptor_set(vkctx, s->pl, 0); |
| 146 | + |
| 147 | + for (int i = 0; i < planes; i++) { |
| 148 | + VkImageMemoryBarrier barriers[] = { |
| 149 | + { |
| 150 | + .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER, |
| 151 | + .srcAccessMask = 0, |
| 152 | + .dstAccessMask = VK_ACCESS_SHADER_READ_BIT, |
| 153 | + .oldLayout = in->layout[i], |
| 154 | + .newLayout = s->input_images[i].imageLayout, |
| 155 | + .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED, |
| 156 | + .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED, |
| 157 | + .image = in->img[i], |
| 158 | + .subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT, |
| 159 | + .subresourceRange.levelCount = 1, |
| 160 | + .subresourceRange.layerCount = 1, |
| 161 | + }, |
| 162 | + { |
| 163 | + .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER, |
| 164 | + .srcAccessMask = 0, |
| 165 | + .dstAccessMask = VK_ACCESS_SHADER_WRITE_BIT, |
| 166 | + .oldLayout = out->layout[i], |
| 167 | + .newLayout = s->output_images[i].imageLayout, |
| 168 | + .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED, |
| 169 | + .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED, |
| 170 | + .image = out->img[i], |
| 171 | + .subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT, |
| 172 | + .subresourceRange.levelCount = 1, |
| 173 | + .subresourceRange.layerCount = 1, |
| 174 | + }, |
| 175 | + }; |
| 176 | + |
| 177 | + vk->CmdPipelineBarrier(cmd_buf, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, |
| 178 | + VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, 0, |
| 179 | + 0, NULL, 0, NULL, FF_ARRAY_ELEMS(barriers), barriers); |
| 180 | + |
| 181 | + in->layout[i] = barriers[0].newLayout; |
| 182 | + in->access[i] = barriers[0].dstAccessMask; |
| 183 | + |
| 184 | + out->layout[i] = barriers[1].newLayout; |
| 185 | + out->access[i] = barriers[1].dstAccessMask; |
| 186 | + } |
| 187 | + |
| 188 | + ff_vk_bind_pipeline_exec(vkctx, s->exec, s->pl); |
| 189 | + vk->CmdDispatch(cmd_buf, FFALIGN(s->vkctx.output_width, CGS)/CGS, |
| 190 | + s->vkctx.output_height, 1); |
| 191 | + |
| 192 | + ff_vk_add_exec_dep(vkctx, s->exec, inframe, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT); |
| 193 | + ff_vk_add_exec_dep(vkctx, s->exec, outframe, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT); |
| 194 | + |
| 195 | + err = ff_vk_submit_exec_queue(vkctx, s->exec); |
| 196 | + if (err) |
| 197 | + return err; |
| 198 | + |
| 199 | + ff_vk_qf_rotate(&s->qf); |
| 200 | + |
| 201 | + return 0; |
| 202 | + |
| 203 | +fail: |
| 204 | + ff_vk_discard_exec_deps(s->exec); |
| 205 | + return err; |
| 206 | +} |
| 207 | + |
| 208 | +static int filter_frame(AVFilterLink *inlink, AVFrame *in) |
| 209 | +{ |
| 210 | + int err; |
| 211 | + AVFrame *out = NULL; |
| 212 | + AVFilterContext *ctx = inlink->dst; |
| 213 | + TransposeVulkanContext *s = ctx->priv; |
| 214 | + AVFilterLink *outlink = ctx->outputs[0]; |
| 215 | + |
| 216 | + out = ff_get_video_buffer(outlink, outlink->w, outlink->h); |
| 217 | + if (!out) { |
| 218 | + err = AVERROR(ENOMEM); |
| 219 | + goto fail; |
| 220 | + } |
| 221 | + |
| 222 | + if (!s->initialized) |
| 223 | + RET(init_filter(ctx, in)); |
| 224 | + |
| 225 | + RET(process_frames(ctx, out, in)); |
| 226 | + |
| 227 | + RET(av_frame_copy_props(out, in)); |
| 228 | + |
| 229 | + if (in->sample_aspect_ratio.num) |
| 230 | + out->sample_aspect_ratio = in->sample_aspect_ratio; |
| 231 | + else { |
| 232 | + out->sample_aspect_ratio.num = in->sample_aspect_ratio.den; |
| 233 | + out->sample_aspect_ratio.den = in->sample_aspect_ratio.num; |
| 234 | + } |
| 235 | + |
| 236 | + av_frame_free(&in); |
| 237 | + |
| 238 | + return ff_filter_frame(outlink, out); |
| 239 | + |
| 240 | +fail: |
| 241 | + av_frame_free(&in); |
| 242 | + av_frame_free(&out); |
| 243 | + return err; |
| 244 | +} |
| 245 | + |
| 246 | +static av_cold void transpose_vulkan_uninit(AVFilterContext *avctx) |
| 247 | +{ |
| 248 | + TransposeVulkanContext *s = avctx->priv; |
| 249 | + ff_vk_uninit(&s->vkctx); |
| 250 | + |
| 251 | + s->initialized = 0; |
| 252 | +} |
| 253 | + |
| 254 | +static int config_props_output(AVFilterLink *outlink) |
| 255 | +{ |
| 256 | + int err = 0; |
| 257 | + AVFilterContext *avctx = outlink->src; |
| 258 | + TransposeVulkanContext *s = avctx->priv; |
| 259 | + FFVulkanContext *vkctx = &s->vkctx; |
| 260 | + AVFilterLink *inlink = avctx->inputs[0]; |
| 261 | + |
| 262 | + vkctx->output_width = inlink->h; |
| 263 | + vkctx->output_height = inlink->w; |
| 264 | + |
| 265 | + RET(ff_vk_filter_config_output(outlink)); |
| 266 | + |
| 267 | + outlink->w = inlink->h; |
| 268 | + outlink->h = inlink->w; |
| 269 | + |
| 270 | + if (inlink->sample_aspect_ratio.num) |
| 271 | + outlink->sample_aspect_ratio = av_div_q((AVRational) { 1, 1 }, |
| 272 | + inlink->sample_aspect_ratio); |
| 273 | + else |
| 274 | + outlink->sample_aspect_ratio = inlink->sample_aspect_ratio; |
| 275 | + |
| 276 | + err = 0; |
| 277 | + |
| 278 | +fail: |
| 279 | + return err; |
| 280 | +} |
| 281 | + |
| 282 | +static const AVOption transpose_vulkan_options[] = { |
| 283 | + { NULL } |
| 284 | +}; |
| 285 | + |
| 286 | +AVFILTER_DEFINE_CLASS(transpose_vulkan); |
| 287 | + |
| 288 | +static const AVFilterPad transpose_vulkan_inputs[] = { |
| 289 | + { |
| 290 | + .name = "default", |
| 291 | + .type = AVMEDIA_TYPE_VIDEO, |
| 292 | + .filter_frame = &filter_frame, |
| 293 | + .config_props = &ff_vk_filter_config_input, |
| 294 | + } |
| 295 | +}; |
| 296 | + |
| 297 | +static const AVFilterPad transpose_vulkan_outputs[] = { |
| 298 | + { |
| 299 | + .name = "default", |
| 300 | + .type = AVMEDIA_TYPE_VIDEO, |
| 301 | + .config_props = &config_props_output, |
| 302 | + } |
| 303 | +}; |
| 304 | + |
| 305 | +const AVFilter ff_vf_transpose_vulkan = { |
| 306 | + .name = "transpose_vulkan", |
| 307 | + .description = NULL_IF_CONFIG_SMALL("Transpose Vulkan Filter"), |
| 308 | + .priv_size = sizeof(TransposeVulkanContext), |
| 309 | + .init = &ff_vk_filter_init, |
| 310 | + .uninit = &transpose_vulkan_uninit, |
| 311 | + FILTER_INPUTS(transpose_vulkan_inputs), |
| 312 | + FILTER_OUTPUTS(transpose_vulkan_outputs), |
| 313 | + FILTER_SINGLE_PIXFMT(AV_PIX_FMT_VULKAN), |
| 314 | + .priv_class = &transpose_vulkan_class, |
| 315 | + .flags_internal = FF_FILTER_FLAG_HWFRAME_AWARE, |
| 316 | +}; |