/*
* NVIDIA NVENC Support
* Copyright (C) 2015 Luca Barbato
+ * Copyright (C) 2015 Philip Langdale <philipl@overt.org>
+ * Copyright (C) 2014 Timo Rothenpieler <timo@rothenpieler.org>
*
* This file is part of Libav.
*
#include "config.h"
-#include <cuda.h>
#include <nvEncodeAPI.h>
#include <string.h>
#endif
#include "libavutil/common.h"
+#include "libavutil/hwcontext.h"
#include "libavutil/imgutils.h"
#include "libavutil/mem.h"
#include "avcodec.h"
#include "internal.h"
#include "nvenc.h"
+#if CONFIG_CUDA
+#include "libavutil/hwcontext_cuda.h"
+#endif
+
#define NVENC_CAP 0x30
#define BITSTREAM_BUFFER_SIZE 1024 * 1024
+#define IS_CBR(rc) (rc == NV_ENC_PARAMS_RC_CBR || \
+ rc == NV_ENC_PARAMS_RC_2_PASS_QUALITY || \
+ rc == NV_ENC_PARAMS_RC_2_PASS_FRAMESIZE_CAP)
#define LOAD_LIBRARY(l, path) \
do { \
} \
} while (0)
+const enum AVPixelFormat ff_nvenc_pix_fmts[] = {
+ AV_PIX_FMT_NV12,
+ AV_PIX_FMT_YUV420P,
+ AV_PIX_FMT_YUV444P,
+#if NVENCAPI_MAJOR_VERSION >= 7
+ AV_PIX_FMT_P010,
+ AV_PIX_FMT_YUV444P16,
+#endif
+#if CONFIG_CUDA
+ AV_PIX_FMT_CUDA,
+#endif
+ AV_PIX_FMT_NONE
+};
+
+#define IS_10BIT(pix_fmt) (pix_fmt == AV_PIX_FMT_P010 || \
+ pix_fmt == AV_PIX_FMT_YUV444P16)
+
+#define IS_YUV444(pix_fmt) (pix_fmt == AV_PIX_FMT_YUV444P || \
+ pix_fmt == AV_PIX_FMT_YUV444P16)
+
static const struct {
NVENCSTATUS nverr;
int averr;
PNVENCODEAPICREATEINSTANCE nvenc_create_instance;
NVENCSTATUS err;
+#if CONFIG_CUDA
+ nvel->cu_init = cuInit;
+ nvel->cu_device_get_count = cuDeviceGetCount;
+ nvel->cu_device_get = cuDeviceGet;
+ nvel->cu_device_get_name = cuDeviceGetName;
+ nvel->cu_device_compute_capability = cuDeviceComputeCapability;
+ nvel->cu_ctx_create = cuCtxCreate_v2;
+ nvel->cu_ctx_pop_current = cuCtxPopCurrent_v2;
+ nvel->cu_ctx_destroy = cuCtxDestroy_v2;
+#else
LOAD_LIBRARY(nvel->cuda, CUDA_LIBNAME);
LOAD_SYMBOL(nvel->cu_init, nvel->cuda, "cuInit");
LOAD_SYMBOL(nvel->cu_ctx_create, nvel->cuda, "cuCtxCreate_v2");
LOAD_SYMBOL(nvel->cu_ctx_pop_current, nvel->cuda, "cuCtxPopCurrent_v2");
LOAD_SYMBOL(nvel->cu_ctx_destroy, nvel->cuda, "cuCtxDestroy_v2");
+#endif
LOAD_LIBRARY(nvel->nvenc, NVENC_LIBNAME);
static int nvenc_check_capabilities(AVCodecContext *avctx)
{
+ NVENCContext *ctx = avctx->priv_data;
int ret;
ret = nvenc_check_codec_support(avctx);
}
ret = nvenc_check_cap(avctx, NV_ENC_CAPS_SUPPORT_YUV444_ENCODE);
- if (avctx->pix_fmt == AV_PIX_FMT_YUV444P && ret <= 0) {
+ if (ctx->data_pix_fmt == AV_PIX_FMT_YUV444P && ret <= 0) {
av_log(avctx, AV_LOG_VERBOSE, "YUV444P not supported\n");
return AVERROR(ENOSYS);
}
ret = nvenc_check_cap(avctx, NV_ENC_CAPS_NUM_MAX_BFRAMES);
if (ret < avctx->max_b_frames) {
- av_log(avctx, AV_LOG_VERBOSE, "Max b-frames %d exceed %d\n",
+ av_log(avctx, AV_LOG_VERBOSE, "Max B-frames %d exceed %d\n",
avctx->max_b_frames, ret);
return AVERROR(ENOSYS);
if (((major << 4) | minor) < NVENC_CAP)
goto fail;
- ret = nvel->cu_ctx_create(&ctx->cu_context, 0, cu_device);
+ ret = nvel->cu_ctx_create(&ctx->cu_context_internal, 0, cu_device);
if (ret != CUDA_SUCCESS)
goto fail;
+ ctx->cu_context = ctx->cu_context_internal;
+
ret = nvel->cu_ctx_pop_current(&dummy);
if (ret != CUDA_SUCCESS)
goto fail2;
ctx->nvenc_ctx = NULL;
fail2:
- nvel->cu_ctx_destroy(ctx->cu_context);
- ctx->cu_context = NULL;
+ nvel->cu_ctx_destroy(ctx->cu_context_internal);
+ ctx->cu_context_internal = NULL;
fail:
if (ret != 0)
{
NVENCContext *ctx = avctx->priv_data;
NVENCLibraryContext *nvel = &ctx->nvel;
- int i, nb_devices = 0;
-
- if ((nvel->cu_init(0)) != CUDA_SUCCESS) {
- av_log(avctx, AV_LOG_ERROR,
- "Cannot init CUDA\n");
- return AVERROR_UNKNOWN;
- }
-
- if ((nvel->cu_device_get_count(&nb_devices)) != CUDA_SUCCESS) {
- av_log(avctx, AV_LOG_ERROR,
- "Cannot enumerate the CUDA devices\n");
- return AVERROR_UNKNOWN;
- }
switch (avctx->codec->id) {
case AV_CODEC_ID_H264:
return AVERROR_BUG;
}
- for (i = 0; i < nb_devices; ++i) {
- if ((nvenc_check_device(avctx, i)) >= 0 && ctx->device != LIST_DEVICES)
- return 0;
- }
+ if (avctx->pix_fmt == AV_PIX_FMT_CUDA) {
+#if CONFIG_CUDA
+ AVHWFramesContext *frames_ctx;
+ AVCUDADeviceContext *device_hwctx;
+ int ret;
- if (ctx->device == LIST_DEVICES)
- return AVERROR_EXIT;
+ if (!avctx->hw_frames_ctx)
+ return AVERROR(EINVAL);
- return AVERROR(ENOSYS);
+ frames_ctx = (AVHWFramesContext*)avctx->hw_frames_ctx->data;
+ device_hwctx = frames_ctx->device_ctx->hwctx;
+
+ ctx->cu_context = device_hwctx->cuda_ctx;
+
+ ret = nvenc_open_session(avctx);
+ if (ret < 0)
+ return ret;
+
+ ret = nvenc_check_capabilities(avctx);
+ if (ret < 0)
+ return ret;
+#else
+ return AVERROR_BUG;
+#endif
+ } else {
+ int i, nb_devices = 0;
+
+ if ((nvel->cu_init(0)) != CUDA_SUCCESS) {
+ av_log(avctx, AV_LOG_ERROR,
+ "Cannot init CUDA\n");
+ return AVERROR_UNKNOWN;
+ }
+
+ if ((nvel->cu_device_get_count(&nb_devices)) != CUDA_SUCCESS) {
+ av_log(avctx, AV_LOG_ERROR,
+ "Cannot enumerate the CUDA devices\n");
+ return AVERROR_UNKNOWN;
+ }
+
+
+ for (i = 0; i < nb_devices; ++i) {
+ if ((nvenc_check_device(avctx, i)) >= 0 && ctx->device != LIST_DEVICES)
+ return 0;
+ }
+
+ if (ctx->device == LIST_DEVICES)
+ return AVERROR_EXIT;
+
+ return AVERROR(ENOSYS);
+ }
+
+ return 0;
}
typedef struct GUIDTuple {
int flags;
} GUIDTuple;
+#define PRESET_ALIAS(alias, name, ...) \
+ [PRESET_ ## alias] = { NV_ENC_PRESET_ ## name ## _GUID, __VA_ARGS__ }
+
+#define PRESET(name, ...) PRESET_ALIAS(name, name, __VA_ARGS__)
+
static int nvec_map_preset(NVENCContext *ctx)
{
GUIDTuple presets[] = {
- { NV_ENC_PRESET_DEFAULT_GUID },
- { NV_ENC_PRESET_HP_GUID },
- { NV_ENC_PRESET_HQ_GUID },
- { NV_ENC_PRESET_BD_GUID },
- { NV_ENC_PRESET_LOW_LATENCY_DEFAULT_GUID, NVENC_LOWLATENCY },
- { NV_ENC_PRESET_LOW_LATENCY_HP_GUID, NVENC_LOWLATENCY },
- { NV_ENC_PRESET_LOW_LATENCY_HQ_GUID, NVENC_LOWLATENCY },
- { NV_ENC_PRESET_LOSSLESS_DEFAULT_GUID, NVENC_LOSSLESS },
- { NV_ENC_PRESET_LOSSLESS_HP_GUID, NVENC_LOSSLESS },
+ PRESET(DEFAULT),
+ PRESET(HP),
+ PRESET(HQ),
+ PRESET(BD),
+ PRESET(LOW_LATENCY_DEFAULT, NVENC_LOWLATENCY),
+ PRESET(LOW_LATENCY_HP, NVENC_LOWLATENCY),
+ PRESET(LOW_LATENCY_HQ, NVENC_LOWLATENCY),
+ PRESET(LOSSLESS_DEFAULT, NVENC_LOSSLESS),
+ PRESET(LOSSLESS_HP, NVENC_LOSSLESS),
+ PRESET_ALIAS(SLOW, HQ, NVENC_TWO_PASSES),
+ PRESET_ALIAS(MEDIUM, HQ, NVENC_ONE_PASS),
+ PRESET_ALIAS(FAST, HP, NVENC_ONE_PASS),
{ { 0 } }
};
return AVERROR(EINVAL);
}
+#undef PRESET
+#undef PRESET_ALIAS
+
static void set_constqp(AVCodecContext *avctx, NV_ENC_RC_PARAMS *rc)
{
rc->rateControlMode = NV_ENC_PARAMS_RC_CONSTQP;
}
}
+static void set_lossless(AVCodecContext *avctx, NV_ENC_RC_PARAMS *rc)
+{
+ rc->rateControlMode = NV_ENC_PARAMS_RC_CONSTQP;
+ rc->constQP.qpInterB = 0;
+ rc->constQP.qpInterP = 0;
+ rc->constQP.qpIntra = 0;
+}
+
static void nvenc_override_rate_control(AVCodecContext *avctx,
NV_ENC_RC_PARAMS *rc)
{
if (ctx->rc > 0) {
nvenc_override_rate_control(avctx, rc);
+ } else if (ctx->flags & NVENC_LOSSLESS) {
+ set_lossless(avctx, rc);
} else if (avctx->global_quality > 0) {
set_constqp(avctx, rc);
} else if (avctx->qmin >= 0 && avctx->qmax >= 0) {
if (rc->averageBitRate > 0)
avctx->bit_rate = rc->averageBitRate;
+
+#if NVENCAPI_MAJOR_VERSION >= 7
+ if (ctx->aq) {
+ ctx->config.rcParams.enableAQ = 1;
+ ctx->config.rcParams.aqStrength = ctx->aq_strength;
+ av_log(avctx, AV_LOG_VERBOSE, "AQ enabled.\n");
+ }
+
+ if (ctx->temporal_aq) {
+ ctx->config.rcParams.enableTemporalAQ = 1;
+ av_log(avctx, AV_LOG_VERBOSE, "Temporal AQ enabled.\n");
+ }
+
+ if (ctx->rc_lookahead) {
+ int lkd_bound = FFMIN(ctx->nb_surfaces, ctx->async_depth) -
+ ctx->config.frameIntervalP - 4;
+
+ if (lkd_bound < 0) {
+ av_log(avctx, AV_LOG_WARNING,
+ "Lookahead not enabled. Increase buffer delay (-delay).\n");
+ } else {
+ ctx->config.rcParams.enableLookahead = 1;
+ ctx->config.rcParams.lookaheadDepth = av_clip(ctx->rc_lookahead, 0, lkd_bound);
+ ctx->config.rcParams.disableIadapt = ctx->no_scenecut;
+ ctx->config.rcParams.disableBadapt = !ctx->b_adapt;
+ av_log(avctx, AV_LOG_VERBOSE,
+ "Lookahead enabled: depth %d, scenecut %s, B-adapt %s.\n",
+ ctx->config.rcParams.lookaheadDepth,
+ ctx->config.rcParams.disableIadapt ? "disabled" : "enabled",
+ ctx->config.rcParams.disableBadapt ? "disabled" : "enabled");
+ }
+ }
+
+ if (ctx->strict_gop) {
+ ctx->config.rcParams.strictGOPTarget = 1;
+ av_log(avctx, AV_LOG_VERBOSE, "Strict GOP target enabled.\n");
+ }
+
+ if (ctx->nonref_p)
+ ctx->config.rcParams.enableNonRefP = 1;
+
+ if (ctx->zerolatency)
+ ctx->config.rcParams.zeroReorderDelay = 1;
+
+ if (ctx->quality)
+ ctx->config.rcParams.targetQuality = ctx->quality;
+#endif /* NVENCAPI_MAJOR_VERSION >= 7 */
}
static int nvenc_setup_h264_config(AVCodecContext *avctx)
NV_ENC_CONFIG_H264 *h264 = &cc->encodeCodecConfig.h264Config;
NV_ENC_CONFIG_H264_VUI_PARAMETERS *vui = &h264->h264VUIParameters;
- vui->colourDescriptionPresentFlag = 1;
- vui->videoSignalTypePresentFlag = 1;
+ vui->colourDescriptionPresentFlag = avctx->colorspace != AVCOL_SPC_UNSPECIFIED ||
+ avctx->color_primaries != AVCOL_PRI_UNSPECIFIED ||
+ avctx->color_trc != AVCOL_TRC_UNSPECIFIED;
vui->colourMatrix = avctx->colorspace;
vui->colourPrimaries = avctx->color_primaries;
vui->videoFullRangeFlag = avctx->color_range == AVCOL_RANGE_JPEG;
+ vui->videoSignalTypePresentFlag = vui->colourDescriptionPresentFlag ||
+ vui->videoFullRangeFlag;
+
h264->disableSPSPPS = (avctx->flags & AV_CODEC_FLAG_GLOBAL_HEADER) ? 1 : 0;
h264->repeatSPSPPS = (avctx->flags & AV_CODEC_FLAG_GLOBAL_HEADER) ? 0 : 1;
+ h264->outputAUD = 1;
h264->maxNumRefFrames = avctx->refs;
h264->idrPeriod = cc->gopLength;
+ h264->sliceMode = 3;
+ h264->sliceModeData = FFMAX(avctx->slices, 1);
+
+ if (ctx->flags & NVENC_LOSSLESS)
+ h264->qpPrimeYZeroTransformBypassFlag = 1;
+
+ if (IS_CBR(cc->rcParams.rateControlMode)) {
+ h264->outputBufferingPeriodSEI = 1;
+ h264->outputPictureTimingSEI = 1;
+ }
+
if (ctx->profile)
avctx->profile = ctx->profile;
- if (avctx->pix_fmt == AV_PIX_FMT_YUV444P)
+ if (ctx->data_pix_fmt == AV_PIX_FMT_YUV444P)
h264->chromaFormatIDC = 3;
else
h264->chromaFormatIDC = 1;
break;
}
+ if (ctx->data_pix_fmt == AV_PIX_FMT_YUV444P) {
+ cc->profileGUID = NV_ENC_H264_PROFILE_HIGH_444_GUID;
+ avctx->profile = FF_PROFILE_H264_HIGH_444_PREDICTIVE;
+ }
+
h264->level = ctx->level;
return 0;
NVENCContext *ctx = avctx->priv_data;
NV_ENC_CONFIG *cc = &ctx->config;
NV_ENC_CONFIG_HEVC *hevc = &cc->encodeCodecConfig.hevcConfig;
+ NV_ENC_CONFIG_HEVC_VUI_PARAMETERS *vui = &hevc->hevcVUIParameters;
+
+ vui->colourDescriptionPresentFlag = avctx->colorspace != AVCOL_SPC_UNSPECIFIED ||
+ avctx->color_primaries != AVCOL_PRI_UNSPECIFIED ||
+ avctx->color_trc != AVCOL_TRC_UNSPECIFIED;
+
+ vui->colourMatrix = avctx->colorspace;
+ vui->colourPrimaries = avctx->color_primaries;
+ vui->transferCharacteristics = avctx->color_trc;
+
+ vui->videoFullRangeFlag = avctx->color_range == AVCOL_RANGE_JPEG;
+
+ vui->videoSignalTypePresentFlag = vui->colourDescriptionPresentFlag ||
+ vui->videoFullRangeFlag;
hevc->disableSPSPPS = (avctx->flags & AV_CODEC_FLAG_GLOBAL_HEADER) ? 1 : 0;
hevc->repeatSPSPPS = (avctx->flags & AV_CODEC_FLAG_GLOBAL_HEADER) ? 0 : 1;
+ hevc->outputAUD = 1;
hevc->maxNumRefFramesInDPB = avctx->refs;
hevc->idrPeriod = cc->gopLength;
- /* No other profile is supported in the current SDK version 5 */
- cc->profileGUID = NV_ENC_HEVC_PROFILE_MAIN_GUID;
- avctx->profile = FF_PROFILE_HEVC_MAIN;
+ if (IS_CBR(cc->rcParams.rateControlMode)) {
+ hevc->outputBufferingPeriodSEI = 1;
+ hevc->outputPictureTimingSEI = 1;
+ }
+
+ switch (ctx->profile) {
+ case NV_ENC_HEVC_PROFILE_MAIN:
+ cc->profileGUID = NV_ENC_HEVC_PROFILE_MAIN_GUID;
+ avctx->profile = FF_PROFILE_HEVC_MAIN;
+ break;
+#if NVENCAPI_MAJOR_VERSION >= 7
+ case NV_ENC_HEVC_PROFILE_MAIN_10:
+ cc->profileGUID = NV_ENC_HEVC_PROFILE_MAIN10_GUID;
+ avctx->profile = FF_PROFILE_HEVC_MAIN_10;
+ break;
+ case NV_ENC_HEVC_PROFILE_REXT:
+ cc->profileGUID = NV_ENC_HEVC_PROFILE_FREXT_GUID;
+ avctx->profile = FF_PROFILE_HEVC_REXT;
+ break;
+#endif /* NVENCAPI_MAJOR_VERSION >= 7 */
+ }
+
+ // force setting profile for various input formats
+ switch (ctx->data_pix_fmt) {
+ case AV_PIX_FMT_YUV420P:
+ case AV_PIX_FMT_NV12:
+ cc->profileGUID = NV_ENC_HEVC_PROFILE_MAIN_GUID;
+ avctx->profile = FF_PROFILE_HEVC_MAIN;
+ break;
+#if NVENCAPI_MAJOR_VERSION >= 7
+ case AV_PIX_FMT_P010:
+ cc->profileGUID = NV_ENC_HEVC_PROFILE_MAIN10_GUID;
+ avctx->profile = FF_PROFILE_HEVC_MAIN_10;
+ break;
+ case AV_PIX_FMT_YUV444P:
+ case AV_PIX_FMT_YUV444P16:
+ cc->profileGUID = NV_ENC_HEVC_PROFILE_FREXT_GUID;
+ avctx->profile = FF_PROFILE_HEVC_REXT;
+ break;
+#endif /* NVENCAPI_MAJOR_VERSION >= 7 */
+ }
+
+#if NVENCAPI_MAJOR_VERSION >= 7
+ hevc->chromaFormatIDC = IS_YUV444(ctx->data_pix_fmt) ? 3 : 1;
+ hevc->pixelBitDepthMinus8 = IS_10BIT(ctx->data_pix_fmt) ? 2 : 0;
+#endif /* NVENCAPI_MAJOR_VERSION >= 7 */
+
+ hevc->sliceMode = 3;
+ hevc->sliceModeData = FFMAX(avctx->slices, 1);
if (ctx->level) {
hevc->level = ctx->level;
ctx->params.darWidth = avctx->width;
}
+ // De-compensate for hardware, dubiously, trying to compensate for
+ // playback at 704 pixel width.
+ if (avctx->width == 720 && (avctx->height == 480 || avctx->height == 576)) {
+ av_reduce(&ctx->params.darWidth, &ctx->params.darHeight,
+ ctx->params.darWidth * 44,
+ ctx->params.darHeight * 45,
+ 1024 * 1024);
+ }
+
ctx->params.frameRateNum = avctx->time_base.den;
ctx->params.frameRateDen = avctx->time_base.num * avctx->ticks_per_frame;
if (avctx->gop_size > 0) {
if (avctx->max_b_frames > 0) {
- ctx->last_dts = -2;
/* 0 is intra-only,
* 1 is I/P only,
- * 2 is one B Frame,
- * 3 two B frames, and so on. */
+ * 2 is one B-Frame,
+ * 3 two B-frames, and so on. */
ctx->config.frameIntervalP = avctx->max_b_frames + 1;
} else if (avctx->max_b_frames == 0) {
ctx->config.frameIntervalP = 1;
if (ctx->config.frameIntervalP > 1)
avctx->max_b_frames = ctx->config.frameIntervalP - 1;
+ ctx->initial_pts[0] = AV_NOPTS_VALUE;
+ ctx->initial_pts[1] = AV_NOPTS_VALUE;
+
nvenc_setup_rate_control(avctx);
if (avctx->flags & AV_CODEC_FLAG_INTERLACED_DCT) {
ret = nv->nvEncInitializeEncoder(ctx->nvenc_ctx, &ctx->params);
if (ret != NV_ENC_SUCCESS)
- return nvenc_print_error(avctx, ret, "Cannot initialize the decoder");
+ return nvenc_print_error(avctx, ret, "InitializeEncoder failed");
cpb_props = ff_add_cpb_side_data(avctx);
if (!cpb_props)
NVENCContext *ctx = avctx->priv_data;
NV_ENCODE_API_FUNCTION_LIST *nv = &ctx->nvel.nvenc_funcs;
int ret;
- NV_ENC_CREATE_INPUT_BUFFER in_buffer = { 0 };
NV_ENC_CREATE_BITSTREAM_BUFFER out_buffer = { 0 };
- in_buffer.version = NV_ENC_CREATE_INPUT_BUFFER_VER;
- out_buffer.version = NV_ENC_CREATE_BITSTREAM_BUFFER_VER;
-
- in_buffer.width = avctx->width;
- in_buffer.height = avctx->height;
-
- in_buffer.memoryHeap = NV_ENC_MEMORY_HEAP_SYSMEM_UNCACHED;
-
- switch (avctx->pix_fmt) {
+ switch (ctx->data_pix_fmt) {
case AV_PIX_FMT_YUV420P:
- in_buffer.bufferFmt = NV_ENC_BUFFER_FORMAT_YV12_PL;
+ ctx->frames[idx].format = NV_ENC_BUFFER_FORMAT_YV12_PL;
break;
case AV_PIX_FMT_NV12:
- in_buffer.bufferFmt = NV_ENC_BUFFER_FORMAT_NV12_PL;
+ ctx->frames[idx].format = NV_ENC_BUFFER_FORMAT_NV12_PL;
break;
case AV_PIX_FMT_YUV444P:
- in_buffer.bufferFmt = NV_ENC_BUFFER_FORMAT_YUV444_PL;
+ ctx->frames[idx].format = NV_ENC_BUFFER_FORMAT_YUV444_PL;
break;
+#if NVENCAPI_MAJOR_VERSION >= 7
+ case AV_PIX_FMT_P010:
+ ctx->frames[idx].format = NV_ENC_BUFFER_FORMAT_YUV420_10BIT;
+ break;
+ case AV_PIX_FMT_YUV444P16:
+ ctx->frames[idx].format = NV_ENC_BUFFER_FORMAT_YUV444_10BIT;
+ break;
+#endif /* NVENCAPI_MAJOR_VERSION >= 7 */
default:
return AVERROR_BUG;
}
- ret = nv->nvEncCreateInputBuffer(ctx->nvenc_ctx, &in_buffer);
- if (ret != NV_ENC_SUCCESS)
- return nvenc_print_error(avctx, ret, "CreateInputBuffer failed");
+ if (avctx->pix_fmt == AV_PIX_FMT_CUDA) {
+ ctx->frames[idx].in_ref = av_frame_alloc();
+ if (!ctx->frames[idx].in_ref)
+ return AVERROR(ENOMEM);
+ } else {
+ NV_ENC_CREATE_INPUT_BUFFER in_buffer = { 0 };
+
+ in_buffer.version = NV_ENC_CREATE_INPUT_BUFFER_VER;
+
+ in_buffer.width = avctx->width;
+ in_buffer.height = avctx->height;
+
+ in_buffer.bufferFmt = ctx->frames[idx].format;
+ in_buffer.memoryHeap = NV_ENC_MEMORY_HEAP_SYSMEM_UNCACHED;
- ctx->in[idx].in = in_buffer.inputBuffer;
- ctx->in[idx].format = in_buffer.bufferFmt;
+ ret = nv->nvEncCreateInputBuffer(ctx->nvenc_ctx, &in_buffer);
+ if (ret != NV_ENC_SUCCESS)
+ return nvenc_print_error(avctx, ret, "CreateInputBuffer failed");
+ ctx->frames[idx].in = in_buffer.inputBuffer;
+ }
+
+ out_buffer.version = NV_ENC_CREATE_BITSTREAM_BUFFER_VER;
/* 1MB is large enough to hold most output frames.
- * NVENC increases this automaticaly if it's not enough. */
+ * NVENC increases this automatically if it is not enough. */
out_buffer.size = BITSTREAM_BUFFER_SIZE;
out_buffer.memoryHeap = NV_ENC_MEMORY_HEAP_SYSMEM_UNCACHED;
if (ret != NV_ENC_SUCCESS)
return nvenc_print_error(avctx, ret, "CreateBitstreamBuffer failed");
- ctx->out[idx].out = out_buffer.bitstreamBuffer;
- ctx->out[idx].busy = 0;
+ ctx->frames[idx].out = out_buffer.bitstreamBuffer;
return 0;
}
ctx->nb_surfaces = FFMAX(4 + avctx->max_b_frames,
ctx->nb_surfaces);
+ ctx->async_depth = FFMIN(ctx->async_depth, ctx->nb_surfaces - 1);
- ctx->in = av_mallocz(ctx->nb_surfaces * sizeof(*ctx->in));
- if (!ctx->in)
- return AVERROR(ENOMEM);
- ctx->out = av_mallocz(ctx->nb_surfaces * sizeof(*ctx->out));
- if (!ctx->out)
+ ctx->frames = av_mallocz_array(ctx->nb_surfaces, sizeof(*ctx->frames));
+ if (!ctx->frames)
return AVERROR(ENOMEM);
ctx->timestamps = av_fifo_alloc(ctx->nb_surfaces * sizeof(int64_t));
if (!ctx->timestamps)
return AVERROR(ENOMEM);
- ctx->pending = av_fifo_alloc(ctx->nb_surfaces * sizeof(ctx->out));
+ ctx->pending = av_fifo_alloc(ctx->nb_surfaces * sizeof(*ctx->frames));
if (!ctx->pending)
return AVERROR(ENOMEM);
- ctx->ready = av_fifo_alloc(ctx->nb_surfaces * sizeof(ctx->out));
+ ctx->ready = av_fifo_alloc(ctx->nb_surfaces * sizeof(*ctx->frames));
if (!ctx->ready)
return AVERROR(ENOMEM);
av_fifo_free(ctx->pending);
av_fifo_free(ctx->ready);
- if (ctx->in) {
+ if (ctx->frames) {
for (i = 0; i < ctx->nb_surfaces; ++i) {
- nv->nvEncDestroyInputBuffer(ctx->nvenc_ctx, ctx->in[i].in);
- nv->nvEncDestroyBitstreamBuffer(ctx->nvenc_ctx, ctx->out[i].out);
+ if (avctx->pix_fmt != AV_PIX_FMT_CUDA) {
+ nv->nvEncDestroyInputBuffer(ctx->nvenc_ctx, ctx->frames[i].in);
+ } else if (ctx->frames[i].in) {
+ nv->nvEncUnmapInputResource(ctx->nvenc_ctx, ctx->frames[i].in_map.mappedResource);
+ }
+
+ av_frame_free(&ctx->frames[i].in_ref);
+ nv->nvEncDestroyBitstreamBuffer(ctx->nvenc_ctx, ctx->frames[i].out);
}
}
+ for (i = 0; i < ctx->nb_registered_frames; i++) {
+ if (ctx->registered_frames[i].regptr)
+ nv->nvEncUnregisterResource(ctx->nvenc_ctx, ctx->registered_frames[i].regptr);
+ }
+ ctx->nb_registered_frames = 0;
- av_freep(&ctx->in);
- av_freep(&ctx->out);
+ av_freep(&ctx->frames);
if (ctx->nvenc_ctx)
nv->nvEncDestroyEncoder(ctx->nvenc_ctx);
- if (ctx->cu_context)
- ctx->nvel.cu_ctx_destroy(ctx->cu_context);
+ if (ctx->cu_context_internal)
+ ctx->nvel.cu_ctx_destroy(ctx->cu_context_internal);
if (ctx->nvel.nvenc)
dlclose(ctx->nvel.nvenc);
+#if !CONFIG_CUDA
if (ctx->nvel.cuda)
dlclose(ctx->nvel.cuda);
+#endif
return 0;
}
av_cold int ff_nvenc_encode_init(AVCodecContext *avctx)
{
+ NVENCContext *ctx = avctx->priv_data;
int ret;
+ if (avctx->pix_fmt == AV_PIX_FMT_CUDA) {
+ AVHWFramesContext *frames_ctx;
+ if (!avctx->hw_frames_ctx) {
+ av_log(avctx, AV_LOG_ERROR,
+ "hw_frames_ctx must be set when using GPU frames as input\n");
+ return AVERROR(EINVAL);
+ }
+ frames_ctx = (AVHWFramesContext*)avctx->hw_frames_ctx->data;
+ ctx->data_pix_fmt = frames_ctx->sw_format;
+ } else {
+ ctx->data_pix_fmt = avctx->pix_fmt;
+ }
+
if ((ret = nvenc_load_libraries(avctx)) < 0)
return ret;
return 0;
}
-static NVENCInputSurface *get_input_surface(NVENCContext *ctx)
+static NVENCFrame *get_free_frame(NVENCContext *ctx)
{
int i;
for (i = 0; i < ctx->nb_surfaces; i++) {
- if (!ctx->in[i].locked) {
- ctx->in[i].locked = 1;
- return &ctx->in[i];
- }
- }
-
- return NULL;
-}
-
-static NVENCOutputSurface *get_output_surface(NVENCContext *ctx)
-{
- int i;
-
- for (i = 0; i < ctx->nb_surfaces; i++) {
- if (!ctx->out[i].busy) {
- return &ctx->out[i];
+ if (!ctx->frames[i].locked) {
+ ctx->frames[i].locked = 1;
+ return &ctx->frames[i];
}
}
frame->data[1], frame->linesize[1],
frame->width, frame->height >> 1);
break;
+ case AV_PIX_FMT_P010:
+ av_image_copy_plane(buf, in->pitch,
+ frame->data[0], frame->linesize[0],
+ frame->width << 1, frame->height);
+ buf += off;
+
+ av_image_copy_plane(buf, in->pitch,
+ frame->data[1], frame->linesize[1],
+ frame->width << 1, frame->height >> 1);
+ break;
case AV_PIX_FMT_YUV444P:
av_image_copy_plane(buf, in->pitch,
frame->data[0], frame->linesize[0],
frame->data[2], frame->linesize[2],
frame->width, frame->height);
break;
+ case AV_PIX_FMT_YUV444P16:
+ av_image_copy_plane(buf, in->pitch,
+ frame->data[0], frame->linesize[0],
+ frame->width << 1, frame->height);
+ buf += off;
+
+ av_image_copy_plane(buf, in->pitch,
+ frame->data[1], frame->linesize[1],
+ frame->width << 1, frame->height);
+ buf += off;
+
+ av_image_copy_plane(buf, in->pitch,
+ frame->data[2], frame->linesize[2],
+ frame->width << 1, frame->height);
+ break;
default:
return AVERROR_BUG;
}
return 0;
}
-static int nvenc_enqueue_frame(AVCodecContext *avctx, const AVFrame *frame,
- NVENCInputSurface **in_surf)
+static int nvenc_find_free_reg_resource(AVCodecContext *avctx)
+{
+ NVENCContext *ctx = avctx->priv_data;
+ NV_ENCODE_API_FUNCTION_LIST *nv = &ctx->nvel.nvenc_funcs;
+ int i;
+
+ if (ctx->nb_registered_frames == FF_ARRAY_ELEMS(ctx->registered_frames)) {
+ for (i = 0; i < ctx->nb_registered_frames; i++) {
+ if (!ctx->registered_frames[i].mapped) {
+ if (ctx->registered_frames[i].regptr) {
+ nv->nvEncUnregisterResource(ctx->nvenc_ctx,
+ ctx->registered_frames[i].regptr);
+ ctx->registered_frames[i].regptr = NULL;
+ }
+ return i;
+ }
+ }
+ } else {
+ return ctx->nb_registered_frames++;
+ }
+
+ av_log(avctx, AV_LOG_ERROR, "Too many registered CUDA frames\n");
+ return AVERROR(ENOMEM);
+}
+
+static int nvenc_register_frame(AVCodecContext *avctx, const AVFrame *frame)
+{
+ NVENCContext *ctx = avctx->priv_data;
+ NV_ENCODE_API_FUNCTION_LIST *nv = &ctx->nvel.nvenc_funcs;
+ AVHWFramesContext *frames_ctx = (AVHWFramesContext*)avctx->hw_frames_ctx->data;
+ NV_ENC_REGISTER_RESOURCE reg;
+ int i, idx, ret;
+
+ for (i = 0; i < ctx->nb_registered_frames; i++) {
+ if (ctx->registered_frames[i].ptr == (CUdeviceptr)frame->data[0])
+ return i;
+ }
+
+ idx = nvenc_find_free_reg_resource(avctx);
+ if (idx < 0)
+ return idx;
+
+ reg.version = NV_ENC_REGISTER_RESOURCE_VER;
+ reg.resourceType = NV_ENC_INPUT_RESOURCE_TYPE_CUDADEVICEPTR;
+ reg.width = frames_ctx->width;
+ reg.height = frames_ctx->height;
+ reg.bufferFormat = ctx->frames[0].format;
+ reg.pitch = frame->linesize[0];
+ reg.resourceToRegister = frame->data[0];
+
+ ret = nv->nvEncRegisterResource(ctx->nvenc_ctx, ®);
+ if (ret != NV_ENC_SUCCESS) {
+ nvenc_print_error(avctx, ret, "Error registering an input resource");
+ return AVERROR_UNKNOWN;
+ }
+
+ ctx->registered_frames[idx].ptr = (CUdeviceptr)frame->data[0];
+ ctx->registered_frames[idx].regptr = reg.registeredResource;
+ return idx;
+}
+
+static int nvenc_upload_frame(AVCodecContext *avctx, const AVFrame *frame,
+ NVENCFrame *nvenc_frame)
{
NVENCContext *ctx = avctx->priv_data;
NV_ENCODE_API_FUNCTION_LIST *nv = &ctx->nvel.nvenc_funcs;
- NV_ENC_LOCK_INPUT_BUFFER params = { 0 };
- NVENCInputSurface *in = get_input_surface(ctx);
int ret;
- if (!in)
- return AVERROR_BUG;
+ if (avctx->pix_fmt == AV_PIX_FMT_CUDA) {
+ int reg_idx;
- params.version = NV_ENC_LOCK_INPUT_BUFFER_VER;
- params.inputBuffer = in->in;
+ ret = nvenc_register_frame(avctx, frame);
+ if (ret < 0) {
+ av_log(avctx, AV_LOG_ERROR, "Could not register an input CUDA frame\n");
+ return ret;
+ }
+ reg_idx = ret;
+ ret = av_frame_ref(nvenc_frame->in_ref, frame);
+ if (ret < 0)
+ return ret;
- ret = nv->nvEncLockInputBuffer(ctx->nvenc_ctx, ¶ms);
- if (ret != NV_ENC_SUCCESS)
- return nvenc_print_error(avctx, ret, "Cannot lock the buffer");
+ nvenc_frame->in_map.version = NV_ENC_MAP_INPUT_RESOURCE_VER;
+ nvenc_frame->in_map.registeredResource = ctx->registered_frames[reg_idx].regptr;
- ret = nvenc_copy_frame(¶ms, frame);
- if (ret < 0)
- goto fail;
+ ret = nv->nvEncMapInputResource(ctx->nvenc_ctx, &nvenc_frame->in_map);
+ if (ret != NV_ENC_SUCCESS) {
+ av_frame_unref(nvenc_frame->in_ref);
+ return nvenc_print_error(avctx, ret, "Error mapping an input resource");
+ }
- ret = nv->nvEncUnlockInputBuffer(ctx->nvenc_ctx, in->in);
- if (ret != NV_ENC_SUCCESS)
- return nvenc_print_error(avctx, ret, "Cannot unlock the buffer");
+ ctx->registered_frames[reg_idx].mapped = 1;
+ nvenc_frame->reg_idx = reg_idx;
+ nvenc_frame->in = nvenc_frame->in_map.mappedResource;
+ } else {
+ NV_ENC_LOCK_INPUT_BUFFER params = { 0 };
- *in_surf = in;
+ params.version = NV_ENC_LOCK_INPUT_BUFFER_VER;
+ params.inputBuffer = nvenc_frame->in;
- return 0;
+ ret = nv->nvEncLockInputBuffer(ctx->nvenc_ctx, ¶ms);
+ if (ret != NV_ENC_SUCCESS)
+ return nvenc_print_error(avctx, ret, "Cannot lock the buffer");
-fail:
- nv->nvEncUnlockInputBuffer(ctx->nvenc_ctx, in->in);
+ ret = nvenc_copy_frame(¶ms, frame);
+ if (ret < 0) {
+ nv->nvEncUnlockInputBuffer(ctx->nvenc_ctx, nvenc_frame->in);
+ return ret;
+ }
- return ret;
+ ret = nv->nvEncUnlockInputBuffer(ctx->nvenc_ctx, nvenc_frame->in);
+ if (ret != NV_ENC_SUCCESS)
+ return nvenc_print_error(avctx, ret, "Cannot unlock the buffer");
+ }
+
+ return 0;
}
static void nvenc_codec_specific_pic_params(AVCodecContext *avctx,
return av_fifo_generic_read(f, pts, sizeof(*pts), NULL);
}
-static inline int nvenc_enqueue_surface(AVFifoBuffer *f,
- NVENCOutputSurface *surf)
-{
- surf->busy = 1;
- return av_fifo_generic_write(f, &surf, sizeof(surf), NULL);
-}
-
-static inline int nvenc_dequeue_surface(AVFifoBuffer *f,
- NVENCOutputSurface **surf)
-{
- return av_fifo_generic_read(f, surf, sizeof(*surf), NULL);
-}
-
-static int nvenc_set_timestamp(NVENCContext *ctx,
+static int nvenc_set_timestamp(AVCodecContext *avctx,
NV_ENC_LOCK_BITSTREAM *params,
AVPacket *pkt)
{
+ NVENCContext *ctx = avctx->priv_data;
+
pkt->pts = params->outputTimeStamp;
pkt->duration = params->outputDuration;
+ /* generate the first dts by linearly extrapolating the
+ * first two pts values to the past */
+ if (avctx->max_b_frames > 0 && !ctx->first_packet_output &&
+ ctx->initial_pts[1] != AV_NOPTS_VALUE) {
+ int64_t ts0 = ctx->initial_pts[0], ts1 = ctx->initial_pts[1];
+ int64_t delta;
+
+ if ((ts0 < 0 && ts1 > INT64_MAX + ts0) ||
+ (ts0 > 0 && ts1 < INT64_MIN + ts0))
+ return AVERROR(ERANGE);
+ delta = ts1 - ts0;
+
+ if ((delta < 0 && ts0 > INT64_MAX + delta) ||
+ (delta > 0 && ts0 < INT64_MIN + delta))
+ return AVERROR(ERANGE);
+ pkt->dts = ts0 - delta;
+
+ ctx->first_packet_output = 1;
+ return 0;
+ }
return nvenc_dequeue_timestamp(ctx->timestamps, &pkt->dts);
}
-static int nvenc_get_frame(AVCodecContext *avctx, AVPacket *pkt)
+static int nvenc_get_output(AVCodecContext *avctx, AVPacket *pkt)
{
NVENCContext *ctx = avctx->priv_data;
NV_ENCODE_API_FUNCTION_LIST *nv = &ctx->nvel.nvenc_funcs;
NV_ENC_LOCK_BITSTREAM params = { 0 };
- NVENCOutputSurface *out = NULL;
+ NVENCFrame *frame;
int ret;
- ret = nvenc_dequeue_surface(ctx->pending, &out);
+ ret = av_fifo_generic_read(ctx->ready, &frame, sizeof(frame), NULL);
if (ret)
return ret;
params.version = NV_ENC_LOCK_BITSTREAM_VER;
- params.outputBitstream = out->out;
+ params.outputBitstream = frame->out;
ret = nv->nvEncLockBitstream(ctx->nvenc_ctx, ¶ms);
if (ret < 0)
memcpy(pkt->data, params.bitstreamBufferPtr, pkt->size);
- ret = nv->nvEncUnlockBitstream(ctx->nvenc_ctx, out->out);
+ ret = nv->nvEncUnlockBitstream(ctx->nvenc_ctx, frame->out);
if (ret < 0)
return nvenc_print_error(avctx, ret, "Cannot unlock the bitstream");
- out->busy = out->in->locked = 0;
+ if (avctx->pix_fmt == AV_PIX_FMT_CUDA) {
+ nv->nvEncUnmapInputResource(ctx->nvenc_ctx, frame->in_map.mappedResource);
+ av_frame_unref(frame->in_ref);
+ ctx->registered_frames[frame->reg_idx].mapped = 0;
+
+ frame->in = NULL;
+ }
+
+ frame->locked = 0;
- ret = nvenc_set_timestamp(ctx, ¶ms, pkt);
+ ret = nvenc_set_timestamp(avctx, ¶ms, pkt);
if (ret < 0)
return ret;
return 0;
}
+static int output_ready(AVCodecContext *avctx, int flush)
+{
+ NVENCContext *ctx = avctx->priv_data;
+ int nb_ready, nb_pending;
+
+ /* when B-frames are enabled, we wait for two initial timestamps to
+ * calculate the first dts */
+ if (!flush && avctx->max_b_frames > 0 &&
+ (ctx->initial_pts[0] == AV_NOPTS_VALUE || ctx->initial_pts[1] == AV_NOPTS_VALUE))
+ return 0;
+
+ nb_ready = av_fifo_size(ctx->ready) / sizeof(NVENCFrame*);
+ nb_pending = av_fifo_size(ctx->pending) / sizeof(NVENCFrame*);
+ if (flush)
+ return nb_ready > 0;
+ return (nb_ready > 0) && (nb_ready + nb_pending >= ctx->async_depth);
+}
+
int ff_nvenc_encode_frame(AVCodecContext *avctx, AVPacket *pkt,
const AVFrame *frame, int *got_packet)
{
NVENCContext *ctx = avctx->priv_data;
NV_ENCODE_API_FUNCTION_LIST *nv = &ctx->nvel.nvenc_funcs;
NV_ENC_PIC_PARAMS params = { 0 };
- NVENCInputSurface *in = NULL;
- NVENCOutputSurface *out = NULL;
- int ret;
+ NVENCFrame *nvenc_frame = NULL;
+ int enc_ret, ret;
params.version = NV_ENC_PIC_PARAMS_VER;
if (frame) {
- ret = nvenc_enqueue_frame(avctx, frame, &in);
- if (ret < 0)
- return ret;
- out = get_output_surface(ctx);
- if (!out)
+ nvenc_frame = get_free_frame(ctx);
+ if (!nvenc_frame) {
+ av_log(avctx, AV_LOG_ERROR, "No free surfaces\n");
return AVERROR_BUG;
+ }
- out->in = in;
+ ret = nvenc_upload_frame(avctx, frame, nvenc_frame);
+ if (ret < 0)
+ return ret;
- params.inputBuffer = in->in;
- params.bufferFmt = in->format;
+ params.inputBuffer = nvenc_frame->in;
+ params.bufferFmt = nvenc_frame->format;
params.inputWidth = frame->width;
params.inputHeight = frame->height;
- params.outputBitstream = out->out;
+ params.outputBitstream = nvenc_frame->out;
params.inputTimeStamp = frame->pts;
if (avctx->flags & AV_CODEC_FLAG_INTERLACED_DCT) {
ret = nvenc_enqueue_timestamp(ctx->timestamps, frame->pts);
if (ret < 0)
return ret;
+
+ if (ctx->initial_pts[0] == AV_NOPTS_VALUE)
+ ctx->initial_pts[0] = frame->pts;
+ else if (ctx->initial_pts[1] == AV_NOPTS_VALUE)
+ ctx->initial_pts[1] = frame->pts;
} else {
params.encodePicFlags = NV_ENC_PIC_FLAG_EOS;
}
- ret = nv->nvEncEncodePicture(ctx->nvenc_ctx, ¶ms);
- if (ret != NV_ENC_SUCCESS &&
- ret != NV_ENC_ERR_NEED_MORE_INPUT)
- return nvenc_print_error(avctx, ret, "Error encoding the frame");
+ enc_ret = nv->nvEncEncodePicture(ctx->nvenc_ctx, ¶ms);
+ if (enc_ret != NV_ENC_SUCCESS &&
+ enc_ret != NV_ENC_ERR_NEED_MORE_INPUT)
+ return nvenc_print_error(avctx, enc_ret, "Error encoding the frame");
- if (out) {
- ret = nvenc_enqueue_surface(ctx->pending, out);
+ if (nvenc_frame) {
+ ret = av_fifo_generic_write(ctx->pending, &nvenc_frame, sizeof(nvenc_frame), NULL);
if (ret < 0)
return ret;
}
- if (ret != NV_ENC_ERR_NEED_MORE_INPUT &&
- av_fifo_size(ctx->pending)) {
- ret = nvenc_get_frame(avctx, pkt);
+ /* all the pending buffers are now ready for output */
+ if (enc_ret == NV_ENC_SUCCESS) {
+ while (av_fifo_size(ctx->pending) > 0) {
+ av_fifo_generic_read(ctx->pending, &nvenc_frame, sizeof(nvenc_frame), NULL);
+ av_fifo_generic_write(ctx->ready, &nvenc_frame, sizeof(nvenc_frame), NULL);
+ }
+ }
+
+ if (output_ready(avctx, !frame)) {
+ ret = nvenc_get_output(avctx, pkt);
if (ret < 0)
return ret;
*got_packet = 1;
projects / ffmpeg / blobdiff