h264dec: add a NVDEC hwaccel

Some parts of the code are based on a patch by
Timo Rothenpieler <timo@rothenpieler.org>

Merges Libav commit b9129ec4668c511e0a79e25c6f25d748cee172c9.

Due to the name clash with our cuvid decoder, rename it to nvdec.

This commit also changes the Libav code to dynamic loading of the
cuda/cuvid libraries.

Signed-off-by: Timo Rothenpieler <timo@rothenpieler.org>

11 files changed, 691 insertions, 3 deletions

diff --git a/Changelog b/Changelog
index 3914672831..ba2951db25 100644
--- a/Changelog
+++ b/Changelog
@@ -13,6 +13,7 @@ version <next>:
 - PCE support for extended channel layouts in the AAC encoder
 - native aptX encoder and decoder
 - Raw aptX muxer and demuxer
+- NVIDIA NVDEC-accelerated H.264 hwaccel decoding
 
 
 version 3.4:
diff --git a/configure b/configure
index f087ba61b1..1b90d8e9a1 100755
--- a/configure
+++ b/configure
@@ -313,6 +313,7 @@ External library support:
   --enable-libmfx          enable Intel MediaSDK (AKA Quick Sync Video) code via libmfx [no]
   --enable-libnpp          enable Nvidia Performance Primitives-based code [no]
   --enable-mmal            enable Broadcom Multi-Media Abstraction Layer (Raspberry Pi) via MMAL [no]
+  --disable-nvdec          disable Nvidia video decoding acceleration (via hwaccel) [autodetect]
   --disable-nvenc          disable Nvidia video encoding code [autodetect]
   --enable-omx             enable OpenMAX IL code [no]
   --enable-omx-rpi         enable OpenMAX IL code for Raspberry Pi [no]
@@ -1647,6 +1648,7 @@ HWACCEL_AUTODETECT_LIBRARY_LIST="
     cuvid
     d3d11va
     dxva2
+    nvdec
     nvenc
     vaapi
     vdpau
@@ -2672,6 +2674,8 @@ h264_dxva2_hwaccel_deps="dxva2"
 h264_dxva2_hwaccel_select="h264_decoder"
 h264_mediacodec_hwaccel_deps="mediacodec"
 h264_mmal_hwaccel_deps="mmal"
+h264_nvdec_hwaccel_deps="cuda nvdec"
+h264_nvdec_hwaccel_select="h264_decoder"
 h264_qsv_hwaccel_deps="libmfx"
 h264_vaapi_hwaccel_deps="vaapi"
 h264_vaapi_hwaccel_select="h264_decoder"
@@ -5940,6 +5944,8 @@ done
 enabled cuda_sdk          && require cuda_sdk cuda.h cuCtxCreate -lcuda
 enabled cuvid             && { enabled cuda ||
                                die "ERROR: CUVID requires CUDA"; }
+enabled nvdec             && { enabled cuda ||
+                               die "ERROR: NVDEC hwaccel requires CUDA"; }
 enabled chromaprint       && require chromaprint chromaprint.h chromaprint_get_version -lchromaprint
 enabled decklink          && { require_header DeckLinkAPI.h &&
                                { check_cpp_condition DeckLinkAPIVersion.h "BLACKMAGIC_DECKLINK_API_VERSION >= 0x0a060100" || die "ERROR: Decklink API version must be >= 10.6.1."; } }
@@ -6295,11 +6301,11 @@ if enabled x86; then
         mingw32*|mingw64*|win32|win64|linux|cygwin*)
             ;;
         *)
-            disable cuda cuvid nvenc
+            disable cuda cuvid nvdec nvenc
             ;;
     esac
 else
-    disable cuda cuvid nvenc
+    disable cuda cuvid nvdec nvenc
 fi
 
 enabled nvenc &&
diff --git a/fftools/ffmpeg.h b/fftools/ffmpeg.h
index 50fc8d5767..e0977e1bf1 100644
--- a/fftools/ffmpeg.h
+++ b/fftools/ffmpeg.h
@@ -68,6 +68,7 @@ enum HWAccelID {
     HWACCEL_VAAPI,
     HWACCEL_CUVID,
     HWACCEL_D3D11VA,
+    HWACCEL_NVDEC,
 };
 
 typedef struct HWAccel {
diff --git a/fftools/ffmpeg_opt.c b/fftools/ffmpeg_opt.c
index ca6f10d5ca..e50895e64c 100644
--- a/fftools/ffmpeg_opt.c
+++ b/fftools/ffmpeg_opt.c
@@ -90,6 +90,10 @@ const HWAccel hwaccels[] = {
     { "vaapi", hwaccel_decode_init, HWACCEL_VAAPI, AV_PIX_FMT_VAAPI,
       AV_HWDEVICE_TYPE_VAAPI },
 #endif
+#if CONFIG_NVDEC
+    { "nvdec", hwaccel_decode_init, HWACCEL_NVDEC, AV_PIX_FMT_CUDA,
+       AV_HWDEVICE_TYPE_CUDA },
+#endif
 #if CONFIG_CUVID
     { "cuvid", cuvid_init, HWACCEL_CUVID, AV_PIX_FMT_CUDA,
       AV_HWDEVICE_TYPE_NONE },
diff --git a/libavcodec/Makefile b/libavcodec/Makefile
index 3704316d8c..db1f70784a 100644
--- a/libavcodec/Makefile
+++ b/libavcodec/Makefile
@@ -830,6 +830,7 @@ OBJS-$(CONFIG_ADPCM_YAMAHA_ENCODER)       += adpcmenc.o adpcm_data.o
 # hardware accelerators
 OBJS-$(CONFIG_D3D11VA)                    += dxva2.o
 OBJS-$(CONFIG_DXVA2)                      += dxva2.o
+OBJS-$(CONFIG_NVDEC)                      += nvdec.o
 OBJS-$(CONFIG_VAAPI)                      += vaapi_decode.o
 OBJS-$(CONFIG_VIDEOTOOLBOX)               += videotoolbox.o
 OBJS-$(CONFIG_VDPAU)                      += vdpau.o
@@ -838,6 +839,7 @@ OBJS-$(CONFIG_H263_VAAPI_HWACCEL)         += vaapi_mpeg4.o
 OBJS-$(CONFIG_H263_VIDEOTOOLBOX_HWACCEL)  += videotoolbox.o
 OBJS-$(CONFIG_H264_D3D11VA_HWACCEL)       += dxva2_h264.o
 OBJS-$(CONFIG_H264_DXVA2_HWACCEL)         += dxva2_h264.o
+OBJS-$(CONFIG_H264_NVDEC_HWACCEL)         += nvdec_h264.o
 OBJS-$(CONFIG_H264_QSV_HWACCEL)           += qsvdec_h2645.o
 OBJS-$(CONFIG_H264_VAAPI_HWACCEL)         += vaapi_h264.o
 OBJS-$(CONFIG_H264_VDPAU_HWACCEL)         += vdpau_h264.o
diff --git a/libavcodec/allcodecs.c b/libavcodec/allcodecs.c
index 463f7ed64e..c58f99c176 100644
--- a/libavcodec/allcodecs.c
+++ b/libavcodec/allcodecs.c
@@ -70,6 +70,7 @@ static void register_all(void)
     REGISTER_HWACCEL(H264_DXVA2,        h264_dxva2);
     REGISTER_HWACCEL(H264_MEDIACODEC,   h264_mediacodec);
     REGISTER_HWACCEL(H264_MMAL,         h264_mmal);
+    REGISTER_HWACCEL(H264_NVDEC,        h264_nvdec);
     REGISTER_HWACCEL(H264_QSV,          h264_qsv);
     REGISTER_HWACCEL(H264_VAAPI,        h264_vaapi);
     REGISTER_HWACCEL(H264_VDPAU,        h264_vdpau);
diff --git a/libavcodec/h264_slice.c b/libavcodec/h264_slice.c
index 94a1547ad6..da76b9293f 100644
--- a/libavcodec/h264_slice.c
+++ b/libavcodec/h264_slice.c
@@ -757,6 +757,7 @@ static enum AVPixelFormat get_pixel_format(H264Context *h, int force_callback)
 {
 #define HWACCEL_MAX (CONFIG_H264_DXVA2_HWACCEL + \
                      (CONFIG_H264_D3D11VA_HWACCEL * 2) + \
+                     CONFIG_H264_NVDEC_HWACCEL + \
                      CONFIG_H264_VAAPI_HWACCEL + \
                      CONFIG_H264_VIDEOTOOLBOX_HWACCEL + \
                      CONFIG_H264_VDPAU_HWACCEL)
@@ -813,6 +814,9 @@ static enum AVPixelFormat get_pixel_format(H264Context *h, int force_callback)
 #if CONFIG_H264_VDPAU_HWACCEL
         *fmt++ = AV_PIX_FMT_VDPAU;
 #endif
+#if CONFIG_H264_NVDEC_HWACCEL
+        *fmt++ = AV_PIX_FMT_CUDA;
+#endif
         if (CHROMA444(h)) {
             if (h->avctx->colorspace == AVCOL_SPC_RGB)
                 *fmt++ = AV_PIX_FMT_GBRP;
diff --git a/libavcodec/nvdec.c b/libavcodec/nvdec.c
new file mode 100644
index 0000000000..9ca9faa378
--- /dev/null
+++ b/libavcodec/nvdec.c
@@ -0,0 +1,431 @@
+/*
+ * HW decode acceleration through NVDEC
+ *
+ * Copyright (c) 2016 Anton Khirnov
+ *
+ * This file is part of Libav.
+ *
+ * Libav is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * Libav is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with Libav; if not, write to the Free Software Foundation,
+ * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#include "config.h"
+
+#include "libavutil/common.h"
+#include "libavutil/error.h"
+#include "libavutil/hwcontext.h"
+#include "libavutil/hwcontext_cuda_internal.h"
+#include "libavutil/pixdesc.h"
+#include "libavutil/pixfmt.h"
+
+#include "avcodec.h"
+#include "decode.h"
+#include "nvdec.h"
+#include "internal.h"
+
+typedef struct NVDECDecoder {
+    CUvideodecoder decoder;
+
+    AVBufferRef *hw_device_ref;
+    CUcontext    cuda_ctx;
+
+    CudaFunctions *cudl;
+    CuvidFunctions *cvdl;
+} NVDECDecoder;
+
+typedef struct NVDECFramePool {
+    unsigned int dpb_size;
+    unsigned int nb_allocated;
+} NVDECFramePool;
+
+static int map_avcodec_id(enum AVCodecID id)
+{
+    switch (id) {
+    case AV_CODEC_ID_H264: return cudaVideoCodec_H264;
+    }
+    return -1;
+}
+
+static int map_chroma_format(enum AVPixelFormat pix_fmt)
+{
+    int shift_h = 0, shift_v = 0;
+
+    av_pix_fmt_get_chroma_sub_sample(pix_fmt, &shift_h, &shift_v);
+
+    if (shift_h == 1 && shift_v == 1)
+        return cudaVideoChromaFormat_420;
+    else if (shift_h == 1 && shift_v == 0)
+        return cudaVideoChromaFormat_422;
+    else if (shift_h == 0 && shift_v == 0)
+        return cudaVideoChromaFormat_444;
+
+    return -1;
+}
+
+static void nvdec_decoder_free(void *opaque, uint8_t *data)
+{
+    NVDECDecoder *decoder = (NVDECDecoder*)data;
+
+    if (decoder->decoder)
+        decoder->cvdl->cuvidDestroyDecoder(decoder->decoder);
+
+    av_buffer_unref(&decoder->hw_device_ref);
+
+    cuvid_free_functions(&decoder->cvdl);
+
+    av_freep(&decoder);
+}
+
+static int nvdec_decoder_create(AVBufferRef **out, AVBufferRef *hw_device_ref,
+                                CUVIDDECODECREATEINFO *params, void *logctx)
+{
+    AVHWDeviceContext  *hw_device_ctx = (AVHWDeviceContext*)hw_device_ref->data;
+    AVCUDADeviceContext *device_hwctx = hw_device_ctx->hwctx;
+
+    AVBufferRef *decoder_ref;
+    NVDECDecoder *decoder;
+
+    CUcontext dummy;
+    CUresult err;
+    int ret;
+
+    decoder = av_mallocz(sizeof(*decoder));
+    if (!decoder)
+        return AVERROR(ENOMEM);
+
+    decoder_ref = av_buffer_create((uint8_t*)decoder, sizeof(*decoder),
+                                   nvdec_decoder_free, NULL, AV_BUFFER_FLAG_READONLY);
+    if (!decoder_ref) {
+        av_freep(&decoder);
+        return AVERROR(ENOMEM);
+    }
+
+    decoder->hw_device_ref = av_buffer_ref(hw_device_ref);
+    if (!decoder->hw_device_ref) {
+        ret = AVERROR(ENOMEM);
+        goto fail;
+    }
+    decoder->cuda_ctx = device_hwctx->cuda_ctx;
+    decoder->cudl = device_hwctx->internal->cuda_dl;
+
+    ret = cuvid_load_functions(&decoder->cvdl);
+    if (ret < 0) {
+        av_log(logctx, AV_LOG_ERROR, "Failed loading nvcuvid.\n");
+        goto fail;
+    }
+
+    err = decoder->cudl->cuCtxPushCurrent(decoder->cuda_ctx);
+    if (err != CUDA_SUCCESS) {
+        ret = AVERROR_UNKNOWN;
+        goto fail;
+    }
+
+    err = decoder->cvdl->cuvidCreateDecoder(&decoder->decoder, params);
+
+    decoder->cudl->cuCtxPopCurrent(&dummy);
+
+    if (err != CUDA_SUCCESS) {
+        av_log(logctx, AV_LOG_ERROR, "Error creating a NVDEC decoder: %d\n", err);
+        ret = AVERROR_UNKNOWN;
+        goto fail;
+    }
+
+    *out = decoder_ref;
+
+    return 0;
+fail:
+    av_buffer_unref(&decoder_ref);
+    return ret;
+}
+
+static AVBufferRef *nvdec_decoder_frame_alloc(void *opaque, int size)
+{
+    NVDECFramePool *pool = opaque;
+    AVBufferRef *ret;
+
+    if (pool->nb_allocated >= pool->dpb_size)
+        return NULL;
+
+    ret = av_buffer_alloc(sizeof(unsigned int));
+    if (!ret)
+        return NULL;
+
+    *(unsigned int*)ret->data = pool->nb_allocated++;
+
+    return ret;
+}
+
+int ff_nvdec_decode_uninit(AVCodecContext *avctx)
+{
+    NVDECContext *ctx = avctx->internal->hwaccel_priv_data;
+
+    av_freep(&ctx->bitstream);
+    ctx->bitstream_len       = 0;
+    ctx->bitstream_allocated = 0;
+
+    av_freep(&ctx->slice_offsets);
+    ctx->nb_slices               = 0;
+    ctx->slice_offsets_allocated = 0;
+
+    av_buffer_unref(&ctx->decoder_ref);
+    av_buffer_pool_uninit(&ctx->decoder_pool);
+
+    return 0;
+}
+
+int ff_nvdec_decode_init(AVCodecContext *avctx, unsigned int dpb_size)
+{
+    NVDECContext *ctx = avctx->internal->hwaccel_priv_data;
+
+    NVDECFramePool      *pool;
+    AVHWFramesContext   *frames_ctx;
+    const AVPixFmtDescriptor *sw_desc;
+
+    CUVIDDECODECREATEINFO params = { 0 };
+
+    int cuvid_codec_type, cuvid_chroma_format;
+    int ret = 0;
+
+    sw_desc = av_pix_fmt_desc_get(avctx->sw_pix_fmt);
+    if (!sw_desc)
+        return AVERROR_BUG;
+
+    cuvid_codec_type = map_avcodec_id(avctx->codec_id);
+    if (cuvid_codec_type < 0) {
+        av_log(avctx, AV_LOG_ERROR, "Unsupported codec ID\n");
+        return AVERROR_BUG;
+    }
+
+    cuvid_chroma_format = map_chroma_format(avctx->sw_pix_fmt);
+    if (cuvid_chroma_format < 0) {
+        av_log(avctx, AV_LOG_ERROR, "Unsupported chroma format\n");
+        return AVERROR(ENOSYS);
+    }
+
+    if (avctx->thread_type & FF_THREAD_FRAME)
+        dpb_size += avctx->thread_count;
+
+    if (!avctx->hw_frames_ctx) {
+        AVHWFramesContext *frames_ctx;
+
+        if (!avctx->hw_device_ctx) {
+            av_log(avctx, AV_LOG_ERROR, "A hardware device or frames context "
+                   "is required for CUVID decoding.\n");
+            return AVERROR(EINVAL);
+        }
+
+        avctx->hw_frames_ctx = av_hwframe_ctx_alloc(avctx->hw_device_ctx);
+        if (!avctx->hw_frames_ctx)
+            return AVERROR(ENOMEM);
+        frames_ctx = (AVHWFramesContext*)avctx->hw_frames_ctx->data;
+
+        frames_ctx->format            = AV_PIX_FMT_CUDA;
+        frames_ctx->width             = avctx->coded_width;
+        frames_ctx->height            = avctx->coded_height;
+        frames_ctx->sw_format         = AV_PIX_FMT_NV12;
+        frames_ctx->sw_format         = sw_desc->comp[0].depth > 8 ?
+                                        AV_PIX_FMT_P010 : AV_PIX_FMT_NV12;
+        frames_ctx->initial_pool_size = dpb_size;
+
+        ret = av_hwframe_ctx_init(avctx->hw_frames_ctx);
+        if (ret < 0) {
+            av_log(avctx, AV_LOG_ERROR, "Error initializing internal frames context\n");
+            return ret;
+        }
+    }
+    frames_ctx = (AVHWFramesContext*)avctx->hw_frames_ctx->data;
+
+    params.ulWidth             = avctx->coded_width;
+    params.ulHeight            = avctx->coded_height;
+    params.ulTargetWidth       = avctx->coded_width;
+    params.ulTargetHeight      = avctx->coded_height;
+    params.bitDepthMinus8      = sw_desc->comp[0].depth - 8;
+    params.OutputFormat        = params.bitDepthMinus8 ?
+                                 cudaVideoSurfaceFormat_P016 : cudaVideoSurfaceFormat_NV12;
+    params.CodecType           = cuvid_codec_type;
+    params.ChromaFormat        = cuvid_chroma_format;
+    params.ulNumDecodeSurfaces = dpb_size;
+    params.ulNumOutputSurfaces = 1;
+
+    ret = nvdec_decoder_create(&ctx->decoder_ref, frames_ctx->device_ref, &params, avctx);
+    if (ret < 0)
+        return ret;
+
+    pool = av_mallocz(sizeof(*pool));
+    if (!pool) {
+        ret = AVERROR(ENOMEM);
+        goto fail;
+    }
+    pool->dpb_size = dpb_size;
+
+    ctx->decoder_pool = av_buffer_pool_init2(sizeof(int), pool,
+                                             nvdec_decoder_frame_alloc, av_free);
+    if (!ctx->decoder_pool) {
+        ret = AVERROR(ENOMEM);
+        goto fail;
+    }
+
+    return 0;
+fail:
+    ff_nvdec_decode_uninit(avctx);
+    return ret;
+}
+
+static void nvdec_fdd_priv_free(void *priv)
+{
+    NVDECFrame *cf = priv;
+
+    if (!cf)
+        return;
+
+    av_buffer_unref(&cf->idx_ref);
+    av_buffer_unref(&cf->decoder_ref);
+
+    av_freep(&priv);
+}
+
+static int nvdec_retrieve_data(void *logctx, AVFrame *frame)
+{
+    FrameDecodeData  *fdd = (FrameDecodeData*)frame->private_ref->data;
+    NVDECFrame        *cf = (NVDECFrame*)fdd->hwaccel_priv;
+    NVDECDecoder *decoder = (NVDECDecoder*)cf->decoder_ref->data;
+
+    CUVIDPROCPARAMS vpp = { .progressive_frame = 1 };
+
+    CUresult err;
+    CUcontext dummy;
+    CUdeviceptr devptr;
+
+    unsigned int pitch, i;
+    unsigned int offset = 0;
+    int ret = 0;
+
+    err = decoder->cudl->cuCtxPushCurrent(decoder->cuda_ctx);
+    if (err != CUDA_SUCCESS)
+        return AVERROR_UNKNOWN;
+
+    err = decoder->cvdl->cuvidMapVideoFrame(decoder->decoder, cf->idx, &devptr,
+                                            &pitch, &vpp);
+    if (err != CUDA_SUCCESS) {
+        av_log(logctx, AV_LOG_ERROR, "Error mapping a picture with CUVID: %d\n",
+               err);
+        ret = AVERROR_UNKNOWN;
+        goto finish;
+    }
+
+    for (i = 0; frame->data[i]; i++) {
+        CUDA_MEMCPY2D cpy = {
+            .srcMemoryType = CU_MEMORYTYPE_DEVICE,
+            .dstMemoryType = CU_MEMORYTYPE_DEVICE,
+            .srcDevice     = devptr,
+            .dstDevice     = (CUdeviceptr)frame->data[i],
+            .srcPitch      = pitch,
+            .dstPitch      = frame->linesize[i],
+            .srcY          = offset,
+            .WidthInBytes  = FFMIN(pitch, frame->linesize[i]),
+            .Height        = frame->height >> (i ? 1 : 0),
+        };
+
+        err = decoder->cudl->cuMemcpy2D(&cpy);
+        if (err != CUDA_SUCCESS) {
+            av_log(logctx, AV_LOG_ERROR, "Error copying decoded frame: %d\n",
+                   err);
+            ret = AVERROR_UNKNOWN;
+            goto copy_fail;
+        }
+
+        offset += cpy.Height;
+    }
+
+copy_fail:
+    decoder->cvdl->cuvidUnmapVideoFrame(decoder->decoder, devptr);
+
+finish:
+    decoder->cudl->cuCtxPopCurrent(&dummy);
+    return ret;
+}
+
+int ff_nvdec_start_frame(AVCodecContext *avctx, AVFrame *frame)
+{
+    NVDECContext *ctx = avctx->internal->hwaccel_priv_data;
+    FrameDecodeData *fdd = (FrameDecodeData*)frame->private_ref->data;
+    NVDECFrame *cf = NULL;
+    int ret;
+
+    ctx->bitstream_len = 0;
+    ctx->nb_slices     = 0;
+
+    if (fdd->hwaccel_priv)
+        return 0;
+
+    cf = av_mallocz(sizeof(*cf));
+    if (!cf)
+        return AVERROR(ENOMEM);
+
+    cf->decoder_ref = av_buffer_ref(ctx->decoder_ref);
+    if (!cf->decoder_ref)
+        goto fail;
+
+    cf->idx_ref = av_buffer_pool_get(ctx->decoder_pool);
+    if (!cf->idx_ref) {
+        av_log(avctx, AV_LOG_ERROR, "No decoder surfaces left\n");
+        ret = AVERROR(ENOMEM);
+        goto fail;
+    }
+    cf->idx = *(unsigned int*)cf->idx_ref->data;
+
+    fdd->hwaccel_priv      = cf;
+    fdd->hwaccel_priv_free = nvdec_fdd_priv_free;
+    fdd->post_process      = nvdec_retrieve_data;
+
+    return 0;
+fail:
+    nvdec_fdd_priv_free(cf);
+    return ret;
+
+}
+
+int ff_nvdec_end_frame(AVCodecContext *avctx)
+{
+    NVDECContext     *ctx = avctx->internal->hwaccel_priv_data;
+    NVDECDecoder *decoder = (NVDECDecoder*)ctx->decoder_ref->data;
+    CUVIDPICPARAMS    *pp = &ctx->pic_params;
+
+    CUresult err;
+    CUcontext dummy;
+
+    int ret = 0;
+
+    pp->nBitstreamDataLen = ctx->bitstream_len;
+    pp->pBitstreamData    = ctx->bitstream;
+    pp->nNumSlices        = ctx->nb_slices;
+    pp->pSliceDataOffsets = ctx->slice_offsets;
+
+    err = decoder->cudl->cuCtxPushCurrent(decoder->cuda_ctx);
+    if (err != CUDA_SUCCESS)
+        return AVERROR_UNKNOWN;
+
+    err = decoder->cvdl->cuvidDecodePicture(decoder->decoder, &ctx->pic_params);
+    if (err != CUDA_SUCCESS) {
+        av_log(avctx, AV_LOG_ERROR, "Error decoding a picture with NVDEC: %d\n",
+               err);
+        ret = AVERROR_UNKNOWN;
+        goto finish;
+    }
+
+finish:
+    decoder->cudl->cuCtxPopCurrent(&dummy);
+
+    return ret;
+}
diff --git a/libavcodec/nvdec.h b/libavcodec/nvdec.h
new file mode 100644
index 0000000000..75ed5d492d
--- /dev/null
+++ b/libavcodec/nvdec.h
@@ -0,0 +1,62 @@
+/*
+ * HW decode acceleration through NVDEC
+ *
+ * Copyright (c) 2016 Anton Khirnov
+ *
+ * This file is part of Libav.
+ *
+ * Libav is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * Libav is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with Libav; if not, write to the Free Software Foundation,
+ * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#ifndef AVCODEC_NVDEC_H
+#define AVCODEC_NVDEC_H
+
+#include "compat/cuda/dynlink_loader.h"
+
+#include <stdint.h>
+
+#include "libavutil/buffer.h"
+#include "libavutil/frame.h"
+
+#include "avcodec.h"
+
+typedef struct NVDECFrame {
+    unsigned int idx;
+    AVBufferRef *idx_ref;
+    AVBufferRef *decoder_ref;
+} NVDECFrame;
+
+typedef struct NVDECContext {
+    CUVIDPICPARAMS pic_params;
+
+    AVBufferPool *decoder_pool;
+
+    AVBufferRef  *decoder_ref;
+
+    uint8_t      *bitstream;
+    int           bitstream_len;
+    unsigned int  bitstream_allocated;
+
+    unsigned     *slice_offsets;
+    int           nb_slices;
+    unsigned int  slice_offsets_allocated;
+} NVDECContext;
+
+int ff_nvdec_decode_init(AVCodecContext *avctx, unsigned int dpb_size);
+int ff_nvdec_decode_uninit(AVCodecContext *avctx);
+int ff_nvdec_start_frame(AVCodecContext *avctx, AVFrame *frame);
+int ff_nvdec_end_frame(AVCodecContext *avctx);
+
+#endif /* AVCODEC_NVDEC_H */
diff --git a/libavcodec/nvdec_h264.c b/libavcodec/nvdec_h264.c
new file mode 100644
index 0000000000..4a6054089b
--- /dev/null
+++ b/libavcodec/nvdec_h264.c
@@ -0,0 +1,176 @@
+/*
+ * MPEG-4 Part 10 / AVC / H.264 HW decode acceleration through NVDEC
+ *
+ * Copyright (c) 2016 Anton Khirnov
+ *
+ * This file is part of Libav.
+ *
+ * Libav is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * Libav is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with Libav; if not, write to the Free Software Foundation,
+ * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#include <stdint.h>
+#include <string.h>
+
+#include "avcodec.h"
+#include "nvdec.h"
+#include "decode.h"
+#include "internal.h"
+#include "h264dec.h"
+
+static void dpb_add(const H264Context *h, CUVIDH264DPBENTRY *dst, const H264Picture *src,
+                    int frame_idx)
+{
+    FrameDecodeData *fdd = (FrameDecodeData*)src->f->private_ref->data;
+    const NVDECFrame *cf = fdd->hwaccel_priv;
+
+    dst->PicIdx             = cf ? cf->idx : -1;
+    dst->FrameIdx           = frame_idx;
+    dst->is_long_term       = src->long_ref;
+    dst->not_existing       = 0;
+    dst->used_for_reference = src->reference & 3;
+    dst->FieldOrderCnt[0]   = src->field_poc[0];
+    dst->FieldOrderCnt[1]   = src->field_poc[1];
+}
+
+static int nvdec_h264_start_frame(AVCodecContext *avctx,
+                                  const uint8_t *buffer, uint32_t size)
+{
+    const H264Context *h = avctx->priv_data;
+    const PPS *pps = h->ps.pps;
+    const SPS *sps = h->ps.sps;
+
+    NVDECContext       *ctx = avctx->internal->hwaccel_priv_data;
+    CUVIDPICPARAMS      *pp = &ctx->pic_params;
+    CUVIDH264PICPARAMS *ppc = &pp->CodecSpecific.h264;
+    FrameDecodeData *fdd;
+    NVDECFrame *cf;
+
+    int i, dpb_size, ret;
+
+    ret = ff_nvdec_start_frame(avctx, h->cur_pic_ptr->f);
+    if (ret < 0)
+        return ret;
+
+    fdd = (FrameDecodeData*)h->cur_pic_ptr->f->private_ref->data;
+    cf  = (NVDECFrame*)fdd->hwaccel_priv;
+
+    *pp = (CUVIDPICPARAMS) {
+        .PicWidthInMbs     = h->mb_width,
+        .FrameHeightInMbs  = h->mb_height,
+        .CurrPicIdx        = cf->idx,
+        .field_pic_flag    = FIELD_PICTURE(h),
+        .bottom_field_flag = h->picture_structure == PICT_BOTTOM_FIELD,
+        .second_field      = FIELD_PICTURE(h) && !h->first_field,
+        .ref_pic_flag      = h->nal_ref_idc != 0,
+        .intra_pic_flag    = 0,
+
+        .CodecSpecific.h264 = {
+            .log2_max_frame_num_minus4            = sps->log2_max_frame_num - 4,
+            .pic_order_cnt_type                   = sps->poc_type,
+            .log2_max_pic_order_cnt_lsb_minus4    = FFMAX(sps->log2_max_poc_lsb - 4, 0),
+            .delta_pic_order_always_zero_flag     = sps->delta_pic_order_always_zero_flag,
+            .frame_mbs_only_flag                  = sps->frame_mbs_only_flag,
+            .direct_8x8_inference_flag            = sps->direct_8x8_inference_flag,
+            .num_ref_frames                       = sps->ref_frame_count,
+            .residual_colour_transform_flag       = sps->residual_color_transform_flag,
+            .bit_depth_luma_minus8                = sps->bit_depth_luma - 8,
+            .bit_depth_chroma_minus8              = sps->bit_depth_chroma - 8,
+            .qpprime_y_zero_transform_bypass_flag = sps->transform_bypass,
+
+            .entropy_coding_mode_flag               = pps->cabac,
+            .pic_order_present_flag                 = pps->pic_order_present,
+            .num_ref_idx_l0_active_minus1           = pps->ref_count[0] - 1,
+            .num_ref_idx_l1_active_minus1           = pps->ref_count[1] - 1,
+            .weighted_pred_flag                     = pps->weighted_pred,
+            .weighted_bipred_idc                    = pps->weighted_bipred_idc,
+            .pic_init_qp_minus26                    = pps->init_qp - 26,
+            .deblocking_filter_control_present_flag = pps->deblocking_filter_parameters_present,
+            .redundant_pic_cnt_present_flag         = pps->redundant_pic_cnt_present,
+            .transform_8x8_mode_flag                = pps->transform_8x8_mode,
+            .MbaffFrameFlag                         = sps->mb_aff && !FIELD_PICTURE(h),
+            .constrained_intra_pred_flag            = pps->constrained_intra_pred,
+            .chroma_qp_index_offset                 = pps->chroma_qp_index_offset[0],
+            .second_chroma_qp_index_offset          = pps->chroma_qp_index_offset[1],
+            .ref_pic_flag                           = h->nal_ref_idc != 0,
+            .frame_num                              = h->poc.frame_num,
+            .CurrFieldOrderCnt[0]                   = h->cur_pic_ptr->field_poc[0],
+            .CurrFieldOrderCnt[1]                   = h->cur_pic_ptr->field_poc[1],
+        },
+    };
+
+    memcpy(ppc->WeightScale4x4,    pps->scaling_matrix4,    sizeof(ppc->WeightScale4x4));
+    memcpy(ppc->WeightScale8x8[0], pps->scaling_matrix8[0], sizeof(ppc->WeightScale8x8[0]));
+    memcpy(ppc->WeightScale8x8[1], pps->scaling_matrix8[3], sizeof(ppc->WeightScale8x8[0]));
+
+    dpb_size = 0;
+    for (i = 0; i < h->short_ref_count; i++)
+        dpb_add(h, &ppc->dpb[dpb_size++], h->short_ref[i], h->short_ref[i]->frame_num);
+    for (i = 0; i < 16; i++) {
+        if (h->long_ref[i])
+            dpb_add(h, &ppc->dpb[dpb_size++], h->long_ref[i], i);
+    }
+
+    for (i = dpb_size; i < FF_ARRAY_ELEMS(ppc->dpb); i++)
+        ppc->dpb[i].PicIdx = -1;
+
+    return 0;
+}
+
+static int nvdec_h264_decode_slice(AVCodecContext *avctx, const uint8_t *buffer,
+                                   uint32_t size)
+{
+    NVDECContext *ctx = avctx->internal->hwaccel_priv_data;
+    void *tmp;
+
+    tmp = av_fast_realloc(ctx->bitstream, &ctx->bitstream_allocated,
+                          ctx->bitstream_len + size + 3);
+    if (!tmp)
+        return AVERROR(ENOMEM);
+    ctx->bitstream = tmp;
+
+    tmp = av_fast_realloc(ctx->slice_offsets, &ctx->slice_offsets_allocated,
+                          (ctx->nb_slices + 1) * sizeof(*ctx->slice_offsets));
+    if (!tmp)
+        return AVERROR(ENOMEM);
+    ctx->slice_offsets = tmp;
+
+    AV_WB24(ctx->bitstream + ctx->bitstream_len, 1);
+    memcpy(ctx->bitstream + ctx->bitstream_len + 3, buffer, size);
+    ctx->slice_offsets[ctx->nb_slices] = ctx->bitstream_len ;
+    ctx->bitstream_len += size + 3;
+    ctx->nb_slices++;
+
+    return 0;
+}
+
+static int nvdec_h264_decode_init(AVCodecContext *avctx)
+{
+    const H264Context *h = avctx->priv_data;
+    const SPS       *sps = h->ps.sps;
+    return ff_nvdec_decode_init(avctx, sps->ref_frame_count + sps->num_reorder_frames);
+}
+
+AVHWAccel ff_h264_nvdec_hwaccel = {
+    .name                 = "h264_nvdec",
+    .type                 = AVMEDIA_TYPE_VIDEO,
+    .id                   = AV_CODEC_ID_H264,
+    .pix_fmt              = AV_PIX_FMT_CUDA,
+    .start_frame          = nvdec_h264_start_frame,
+    .end_frame            = ff_nvdec_end_frame,
+    .decode_slice         = nvdec_h264_decode_slice,
+    .init                 = nvdec_h264_decode_init,
+    .uninit               = ff_nvdec_decode_uninit,
+    .priv_data_size       = sizeof(NVDECContext),
+};
diff --git a/libavcodec/version.h b/libavcodec/version.h
index 6c0d7a8328..e7323764ec 100644
--- a/libavcodec/version.h
+++ b/libavcodec/version.h
@@ -28,7 +28,7 @@
 #include "libavutil/version.h"
 
 #define LIBAVCODEC_VERSION_MAJOR  58
-#define LIBAVCODEC_VERSION_MINOR   1
+#define LIBAVCODEC_VERSION_MINOR   2
 #define LIBAVCODEC_VERSION_MICRO 100
 
 #define LIBAVCODEC_VERSION_INT  AV_VERSION_INT(LIBAVCODEC_VERSION_MAJOR, \