#include <cstdint>
#include <utility>
#include <vector>
+#include <unordered_set>
#include <Eigen/Core>
#include <Eigen/LU>
#include <srt/srt.h>
#endif
-#define FRAME_SIZE (8 << 20) // 8 MB.
-
using namespace std;
using namespace std::chrono;
using namespace bmusb;
if (desc->comp[0].depth != 8) continue;
// Same or better chroma resolution only.
- int chroma_w_diff = desc->log2_chroma_w - src_desc->log2_chroma_w;
- int chroma_h_diff = desc->log2_chroma_h - src_desc->log2_chroma_h;
+ int chroma_w_diff = src_desc->log2_chroma_w - desc->log2_chroma_w;
+ int chroma_h_diff = src_desc->log2_chroma_h - desc->log2_chroma_h;
if (chroma_w_diff < 0 || chroma_h_diff < 0)
continue;
}
}
-AVPixelFormat get_vaapi_hw_format(AVCodecContext *ctx, const AVPixelFormat *fmt)
+template<AVHWDeviceType type>
+AVPixelFormat get_hw_format(AVCodecContext *ctx, const AVPixelFormat *fmt)
{
- for (const AVPixelFormat *fmt_ptr = fmt; *fmt_ptr != -1; ++fmt_ptr) {
- for (int i = 0;; ++i) { // Termination condition inside loop.
- const AVCodecHWConfig *config = avcodec_get_hw_config(ctx->codec, i);
- if (config == nullptr) { // End of list.
- fprintf(stderr, "Decoder %s does not support device.\n", ctx->codec->name);
- break;
- }
- if (config->methods & AV_CODEC_HW_CONFIG_METHOD_HW_DEVICE_CTX &&
- config->device_type == AV_HWDEVICE_TYPE_VAAPI &&
- config->pix_fmt == *fmt_ptr) {
+ bool found_config_of_right_type = false;
+ for (int i = 0;; ++i) { // Termination condition inside loop.
+ const AVCodecHWConfig *config = avcodec_get_hw_config(ctx->codec, i);
+ if (config == nullptr) { // End of list.
+ break;
+ }
+ if (!(config->methods & AV_CODEC_HW_CONFIG_METHOD_HW_DEVICE_CTX) ||
+ config->device_type != type) {
+ // Not interesting for us.
+ continue;
+ }
+
+ // We have a config of the right type, but does it actually support
+ // the pixel format we want? (Seemingly, FFmpeg's way of signaling errors
+ // is to just replace the pixel format with a software-decoded one,
+ // such as yuv420p.)
+ found_config_of_right_type = true;
+ for (const AVPixelFormat *fmt_ptr = fmt; *fmt_ptr != -1; ++fmt_ptr) {
+ if (config->pix_fmt == *fmt_ptr) {
+ fprintf(stderr, "Initialized '%s' hardware decoding for codec '%s'.\n",
+ av_hwdevice_get_type_name(type), ctx->codec->name);
+ if (ctx->profile == FF_PROFILE_H264_BASELINE) {
+ fprintf(stderr, "WARNING: Stream claims to be H.264 Baseline, which is generally poorly supported in hardware decoders.\n");
+ fprintf(stderr, " Consider encoding it as Constrained Baseline, Main or High instead.\n");
+ fprintf(stderr, " Decoding might fail and fall back to software.\n");
+ }
return config->pix_fmt;
}
}
+ fprintf(stderr, "Decoder '%s' supports only these pixel formats:", ctx->codec->name);
+ unordered_set<AVPixelFormat> seen;
+ for (const AVPixelFormat *fmt_ptr = fmt; *fmt_ptr != -1; ++fmt_ptr) {
+ if (!seen.count(*fmt_ptr)) {
+ fprintf(stderr, " %s", av_get_pix_fmt_name(*fmt_ptr));
+ seen.insert(*fmt_ptr);
+ }
+ }
+ fprintf(stderr, " (wanted %s for hardware acceleration)\n", av_get_pix_fmt_name(config->pix_fmt));
+
}
- // We found no VA-API formats, so take the best software format.
+ if (!found_config_of_right_type) {
+ fprintf(stderr, "Decoder '%s' does not support device type '%s'.\n", ctx->codec->name, av_hwdevice_get_type_name(type));
+ }
+
+ // We found no VA-API formats, so take the first software format.
+ for (const AVPixelFormat *fmt_ptr = fmt; *fmt_ptr != -1; ++fmt_ptr) {
+ if ((av_pix_fmt_desc_get(*fmt_ptr)->flags & AV_PIX_FMT_FLAG_HWACCEL) == 0) {
+ fprintf(stderr, "Falling back to software format %s.\n", av_get_pix_fmt_name(*fmt_ptr));
+ return *fmt_ptr;
+ }
+ }
+
+ // Fallback: Just return anything. (Should never really happen.)
return fmt[0];
}
}
// Seemingly, it's not too easy to make something that just initializes
- // “whatever goes”, so we don't get VDPAU or CUDA here without enumerating
- // through several different types. VA-API will do for now.
+ // “whatever goes”, so we don't get CUDA or VULKAN or whatever here
+ // without enumerating through several different types.
+ // VA-API and VDPAU will do for now. We prioritize VDPAU for the
+ // simple reason that there's a VA-API-via-VDPAU emulation for NVidia
+ // cards that seems to work, but just hangs when trying to transfer the frame.
+ //
+ // Note that we don't actually check codec support beforehand,
+ // so if you have a low-end VDPAU device but a high-end VA-API device,
+ // you lose out on the extra codec support from the latter.
AVBufferRef *hw_device_ctx = nullptr;
- if (av_hwdevice_ctx_create(&hw_device_ctx, AV_HWDEVICE_TYPE_VAAPI, nullptr, nullptr, 0) < 0) {
- fprintf(stderr, "Failed to initialize VA-API for FFmpeg acceleration. Decoding video in software.\n");
- } else {
+ if (av_hwdevice_ctx_create(&hw_device_ctx, AV_HWDEVICE_TYPE_VDPAU, nullptr, nullptr, 0) >= 0) {
+ video_codec_ctx->hw_device_ctx = av_buffer_ref(hw_device_ctx);
+ video_codec_ctx->get_format = get_hw_format<AV_HWDEVICE_TYPE_VDPAU>;
+ } else if (av_hwdevice_ctx_create(&hw_device_ctx, AV_HWDEVICE_TYPE_VAAPI, nullptr, nullptr, 0) >= 0) {
video_codec_ctx->hw_device_ctx = av_buffer_ref(hw_device_ctx);
- video_codec_ctx->get_format = get_vaapi_hw_format;
+ video_codec_ctx->get_format = get_hw_format<AV_HWDEVICE_TYPE_VAAPI>;
+ } else {
+ fprintf(stderr, "Failed to initialize VA-API or VDPAU for FFmpeg acceleration. Decoding video in software.\n");
}
if (avcodec_open2(video_codec_ctx.get(), video_codec, nullptr) < 0) {
// Decode video, if we have a frame.
int err = avcodec_receive_frame(video_codec_ctx, video_avframe.get());
if (err == 0) {
- if (video_avframe->format == AV_PIX_FMT_VAAPI) {
+ if (video_avframe->format == AV_PIX_FMT_VAAPI ||
+ video_avframe->format == AV_PIX_FMT_VDPAU) {
// Get the frame down to the CPU. (TODO: See if we can keep it
// on the GPU all the way, since it will be going up again later.
// However, this only works if the OpenGL GPU is the same one.)
current_frame_ycbcr_format = decode_ycbcr_format(desc, frame, is_mjpeg, &last_colorspace, &last_chroma_location);
}
+
+ // FIXME: Currently, if the video is too high-res for one of the allocated
+ // frames, we simply refuse to scale it here to avoid crashes. It would be better
+ // if we could somehow signal getting larger frames, especially as 4K is a thing now.
+ if (video_frame->len > FRAME_SIZE) {
+ fprintf(stderr, "%s: Decoded frame would be larger than supported FRAME_SIZE (%zu > %u), not decoding.\n", pathname.c_str(), video_frame->len, FRAME_SIZE);
+ *error = true;
+ return video_frame;
+ }
+
sws_scale(sws_ctx.get(), frame->data, frame->linesize, 0, frame->height, pic_data, linesizes);
return video_frame;