Originally, the decoder had a single execution pool, with one execution context per thread. Execution pools were always intended to be thread-safe, as long as there were enough execution contexts in the pool to satisfy all threads. Due to synchronization issues, the threading part was removed at some point, and, for decoding, each thread had its own execution pool. Having a single execution pool per context is hacky, not to mention wasteful. Most importantly, we *cannot* associate single shaders across multiple execution pools for a single application. This means that we cannot use shaders to either apply film grain, or use this framework for software-defined decoders. The recent commits added threading capabilities back to the execution pool, and the number of contexts in each pool was increased. This was done with the assumption that the execution pool was singular, which it was not. This led to increased parallelism and number of frames in flight, which is taxing on memory. This commit finally restores proper threading behaviour. The validation layer has isses that are reported and addressed in the earlier commit.
Unified · Split
| @@ -83,25 +83,6 @@ int ff_vk_update_thread_context(AVCodecContext *dst, const AVCodecContext *src) · 82 unmodified lines | ||
| 83 | 83 | FFVulkanDecodeContext *src_ctx = src->internal->hwaccel_priv_data; |
| 84 | 84 | FFVulkanDecodeContext *dst_ctx = dst->internal->hwaccel_priv_data; |
| 85 | 85 | |
| 86 | − if (!dst_ctx->exec_pool.cmd_bufs) { | |
| 87 | − FFVulkanDecodeShared *ctx = src_ctx->shared_ctx; | |
| 88 | − | |
| 89 | − const VkVideoProfileInfoKHR *profile = get_video_profile(ctx, dst->codec_id); | |
| 90 | − if (!profile) { | |
| 91 | − av_log(dst, AV_LOG_ERROR, "Video profile missing from frames context!\n"); | |
| 92 | − return AVERROR(EINVAL); | |
| 93 | − } | |
| 94 | − | |
| 95 | − err = ff_vk_exec_pool_init(&ctx->s, &ctx->qf, | |
| 96 | − &dst_ctx->exec_pool, | |
| 97 | − src_ctx->exec_pool.pool_size, | |
| 98 | − src_ctx->exec_pool.nb_queries, | |
| 99 | − VK_QUERY_TYPE_RESULT_STATUS_ONLY_KHR, 0, | |
| 100 | − profile); | |
| 101 | − if (err < 0) | |
| 102 | − return err; | |
| 103 | − } | |
| 104 | − | |
| 105 | 86 | av_refstruct_replace(&dst_ctx->shared_ctx, src_ctx->shared_ctx); |
| 106 | 87 | |
| 107 | 88 | if (src_ctx->session_params) { |
| @@ -293,8 +274,11 @@ void ff_vk_decode_flush(AVCodecContext *avctx) · 185 unmodified lines | ||
| 293 | 274 | }; |
| 294 | 275 | |
| 295 | 276 | VkCommandBuffer cmd_buf; |
| 296 | − FFVkExecContext *exec = ff_vk_exec_get(&ctx->s, &dec->exec_pool); | |
| 297 | − int had_submission = exec->had_submission; | |
| 277 | + FFVkExecContext *exec; | |
| 278 | + int had_submission; | |
| 279 | + | |
| 280 | + exec = ff_vk_exec_get(&ctx->s, &ctx->exec_pool); | |
| 281 | + had_submission = exec->had_submission; | |
| 298 | 282 | ff_vk_exec_start(&ctx->s, exec); |
| 299 | 283 | cmd_buf = exec->buf; |
| 300 | 284 | |
| @@ -345,7 +329,7 @@ int ff_vk_decode_frame(AVCodecContext *avctx, · 44 unmodified lines | ||
| 345 | 329 | size_t data_size = FFALIGN(vp->slices_size, |
| 346 | 330 | ctx->caps.minBitstreamBufferSizeAlignment); |
| 347 | 331 | |
| 348 | − FFVkExecContext *exec = ff_vk_exec_get(&ctx->s, &dec->exec_pool); | |
| 332 | + FFVkExecContext *exec = ff_vk_exec_get(&ctx->s, &ctx->exec_pool); | |
| 349 | 333 | |
| 350 | 334 | /* The current decoding reference has to be bound as an inactive reference */ |
| 351 | 335 | VkVideoReferenceSlotInfoKHR *cur_vk_ref; |
| @@ -354,7 +338,7 @@ int ff_vk_decode_frame(AVCodecContext *avctx, · 2 unmodified lines | ||
| 354 | 338 | cur_vk_ref[0].slotIndex = -1; |
| 355 | 339 | decode_start.referenceSlotCount++; |
| 356 | 340 | |
| 357 | − if (dec->exec_pool.nb_queries && exec->had_submission) { | |
| 341 | + if (ctx->exec_pool.nb_queries && exec->had_submission) { | |
| 358 | 342 | uint32_t *result; |
| 359 | 343 | ret = ff_vk_exec_get_query(&ctx->s, exec, (void **)&result, |
| 360 | 344 | VK_QUERY_RESULT_WAIT_BIT); |
| @@ -525,14 +509,14 @@ int ff_vk_decode_frame(AVCodecContext *avctx, · 164 unmodified lines | ||
| 525 | 509 | vk->CmdBeginVideoCodingKHR(cmd_buf, &decode_start); |
| 526 | 510 | |
| 527 | 511 | /* Start status query */ |
| 528 | − if (dec->exec_pool.nb_queries) | |
| 529 | − vk->CmdBeginQuery(cmd_buf, dec->exec_pool.query_pool, exec->query_idx + 0, 0); | |
| 512 | + if (ctx->exec_pool.nb_queries) | |
| 513 | + vk->CmdBeginQuery(cmd_buf, ctx->exec_pool.query_pool, exec->query_idx + 0, 0); | |
| 530 | 514 | |
| 531 | 515 | vk->CmdDecodeVideoKHR(cmd_buf, &vp->decode_info); |
| 532 | 516 | |
| 533 | 517 | /* End status query */ |
| 534 | − if (dec->exec_pool.nb_queries) | |
| 535 | − vk->CmdEndQuery(cmd_buf, dec->exec_pool.query_pool, exec->query_idx + 0); | |
| 518 | + if (ctx->exec_pool.nb_queries) | |
| 519 | + vk->CmdEndQuery(cmd_buf, ctx->exec_pool.query_pool, exec->query_idx + 0); | |
| 536 | 520 | |
| 537 | 521 | vk->CmdEndVideoCodingKHR(cmd_buf, &decode_end); |
| 538 | 522 | |
| @@ -577,6 +561,9 @@ static void free_common(AVRefStructOpaque unused, void *obj) · 38 unmodified lines | ||
| 577 | 561 | FFVulkanContext *s = &ctx->s; |
| 578 | 562 | FFVulkanFunctions *vk = &ctx->s.vkfn; |
| 579 | 563 | |
| 564 | + /* Wait on and free execution pool */ | |
| 565 | + ff_vk_exec_pool_free(&ctx->s, &ctx->exec_pool); | |
| 566 | + | |
| 580 | 567 | /* This also frees all references from this pool */ |
| 581 | 568 | av_frame_free(&ctx->common.layered_frame); |
| 582 | 569 | |
| @@ -1066,10 +1053,6 @@ int ff_vk_decode_create_params(AVBufferRef **par_ref, void *logctx, FFVulkanDeco · 483 unmodified lines | ||
| 1066 | 1053 | int ff_vk_decode_uninit(AVCodecContext *avctx) |
| 1067 | 1054 | { |
| 1068 | 1055 | FFVulkanDecodeContext *dec = avctx->internal->hwaccel_priv_data; |
| 1069 | − FFVulkanDecodeShared *ctx = dec->shared_ctx; | |
| 1070 | − | |
| 1071 | − /* Wait on and free execution pool */ | |
| 1072 | − ff_vk_exec_pool_free(&ctx->s, &dec->exec_pool); | |
| 1073 | 1056 | |
| 1074 | 1057 | av_freep(&dec->hevc_headers); |
| 1075 | 1058 | av_buffer_unref(&dec->session_params); |
| @@ -1159,7 +1142,7 @@ int ff_vk_decode_init(AVCodecContext *avctx) · 83 unmodified lines | ||
| 1159 | 1142 | /* Create decode exec context for this specific main thread. |
| 1160 | 1143 | * 2 async contexts per thread was experimentally determined to be optimal |
| 1161 | 1144 | * for a majority of streams. */ |
| 1162 | − err = ff_vk_exec_pool_init(s, &ctx->qf, &dec->exec_pool, | |
| 1145 | + err = ff_vk_exec_pool_init(s, &ctx->qf, &ctx->exec_pool, | |
| 1163 | 1146 | FFMAX(2*ctx->qf.nb_queues, avctx->thread_count), |
| 1164 | 1147 | nb_q, VK_QUERY_TYPE_RESULT_STATUS_ONLY_KHR, 0, |
| 1165 | 1148 | profile); |
| @@ -47,6 +47,7 @@ typedef struct FFVulkanDecodeShared { · 46 unmodified lines | ||
| 47 | 47 | FFVulkanContext s; |
| 48 | 48 | FFVkVideoCommon common; |
| 49 | 49 | FFVkQueueFamilyCtx qf; |
| 50 | + FFVkExecPool exec_pool; | |
| 50 | 51 | |
| 51 | 52 | AVBufferPool *buf_pool; |
| 52 | 53 | |
| @@ -59,7 +60,6 @@ typedef struct FFVulkanDecodeShared { · 6 unmodified lines | ||
| 59 | 60 | typedef struct FFVulkanDecodeContext { |
| 60 | 61 | FFVulkanDecodeShared *shared_ctx; |
| 61 | 62 | AVBufferRef *session_params; |
| 62 | − FFVkExecPool exec_pool; | |
| 63 | 63 | |
| 64 | 64 | int dedicated_dpb; /* Oddity #1 - separate DPB images */ |
| 65 | 65 | int external_fg; /* Oddity #2 - hardware can't apply film grain */ |