avfilter/scale_cuda: add CUDA scale filter

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

5 files changed, 771 insertions, 1 deletions

diff --git a/libavfilter/Makefile b/libavfilter/Makefile
index f7dfe8ad54..f177fdb42b 100644
--- a/libavfilter/Makefile
+++ b/libavfilter/Makefile
@@ -267,6 +267,7 @@ OBJS-$(CONFIG_REVERSE_FILTER)                += f_reverse.o
 OBJS-$(CONFIG_ROTATE_FILTER)                 += vf_rotate.o
 OBJS-$(CONFIG_SAB_FILTER)                    += vf_sab.o
 OBJS-$(CONFIG_SCALE_FILTER)                  += vf_scale.o scale.o
+OBJS-$(CONFIG_SCALE_CUDA_FILTER)             += vf_scale_cuda.o vf_scale_cuda.ptx.o
 OBJS-$(CONFIG_SCALE_NPP_FILTER)              += vf_scale_npp.o scale.o
 OBJS-$(CONFIG_SCALE_QSV_FILTER)              += vf_scale_qsv.o
 OBJS-$(CONFIG_SCALE_VAAPI_FILTER)            += vf_scale_vaapi.o scale.o
diff --git a/libavfilter/allfilters.c b/libavfilter/allfilters.c
index cd35ae4c9c..a8939b9094 100644
--- a/libavfilter/allfilters.c
+++ b/libavfilter/allfilters.c
@@ -278,6 +278,7 @@ static void register_all(void)
     REGISTER_FILTER(ROTATE,         rotate,         vf);
     REGISTER_FILTER(SAB,            sab,            vf);
     REGISTER_FILTER(SCALE,          scale,          vf);
+    REGISTER_FILTER(SCALE_CUDA,     scale_cuda,     vf);
     REGISTER_FILTER(SCALE_NPP,      scale_npp,      vf);
     REGISTER_FILTER(SCALE_QSV,      scale_qsv,      vf);
     REGISTER_FILTER(SCALE_VAAPI,    scale_vaapi,    vf);
diff --git a/libavfilter/version.h b/libavfilter/version.h
index ebfa644d1c..853333d8f8 100644
--- a/libavfilter/version.h
+++ b/libavfilter/version.h
@@ -31,7 +31,7 @@
 
 #define LIBAVFILTER_VERSION_MAJOR   6
 #define LIBAVFILTER_VERSION_MINOR  89
-#define LIBAVFILTER_VERSION_MICRO 100
+#define LIBAVFILTER_VERSION_MICRO 101
 
 #define LIBAVFILTER_VERSION_INT AV_VERSION_INT(LIBAVFILTER_VERSION_MAJOR, \
                                                LIBAVFILTER_VERSION_MINOR, \
diff --git a/libavfilter/vf_scale_cuda.c b/libavfilter/vf_scale_cuda.c
new file mode 100644
index 0000000000..23ac27a7dc
--- /dev/null
+++ b/libavfilter/vf_scale_cuda.c
@@ -0,0 +1,556 @@
+/*
+* Copyright (c) 2017, NVIDIA CORPORATION. All rights reserved.
+*
+* Permission is hereby granted, free of charge, to any person obtaining a
+* copy of this software and associated documentation files (the "Software"),
+* to deal in the Software without restriction, including without limitation
+* the rights to use, copy, modify, merge, publish, distribute, sublicense,
+* and/or sell copies of the Software, and to permit persons to whom the
+* Software is furnished to do so, subject to the following conditions:
+*
+* The above copyright notice and this permission notice shall be included in
+* all copies or substantial portions of the Software.
+*
+* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+* THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+* DEALINGS IN THE SOFTWARE.
+*/
+
+#include <cuda.h>
+#include <stdio.h>
+#include <string.h>
+
+#include "libavutil/avstring.h"
+#include "libavutil/common.h"
+#include "libavutil/hwcontext.h"
+#include "libavutil/hwcontext_cuda_internal.h"
+#include "libavutil/internal.h"
+#include "libavutil/opt.h"
+#include "libavutil/pixdesc.h"
+
+#include "avfilter.h"
+#include "formats.h"
+#include "internal.h"
+#include "scale.h"
+#include "video.h"
+
+static const enum AVPixelFormat supported_formats[] = {
+    AV_PIX_FMT_YUV420P,
+    AV_PIX_FMT_NV12,
+    AV_PIX_FMT_YUV444P,
+    AV_PIX_FMT_P010,
+    AV_PIX_FMT_P016
+};
+
+#define DIV_UP(a, b) ( ((a) + (b) - 1) / (b) )
+#define ALIGN_UP(a, b) (((a) + (b) - 1) & ~((b) - 1))
+#define NUM_BUFFERS 2
+#define BLOCKX 32
+#define BLOCKY 16
+
+typedef struct CUDAScaleContext {
+    const AVClass *class;
+    enum AVPixelFormat in_fmt;
+    enum AVPixelFormat out_fmt;
+
+    struct {
+        int width;
+        int height;
+    } planes_in[3], planes_out[3];
+
+    AVBufferRef *frames_ctx;
+    AVFrame     *frame;
+
+    AVFrame *tmp_frame;
+    int passthrough;
+
+    /**
+     * Output sw format. AV_PIX_FMT_NONE for no conversion.
+     */
+    enum AVPixelFormat format;
+
+    char *w_expr;               ///< width  expression string
+    char *h_expr;               ///< height expression string
+
+    CUcontext   cu_ctx;
+    CUevent     cu_event;
+    CUmodule    cu_module;
+    CUfunction  cu_func_uchar;
+    CUfunction  cu_func_uchar2;
+    CUfunction  cu_func_uchar4;
+    CUfunction  cu_func_ushort;
+    CUfunction  cu_func_ushort2;
+    CUfunction  cu_func_ushort4;
+    CUtexref    cu_tex_uchar;
+    CUtexref    cu_tex_uchar2;
+    CUtexref    cu_tex_uchar4;
+    CUtexref    cu_tex_ushort;
+    CUtexref    cu_tex_ushort2;
+    CUtexref    cu_tex_ushort4;
+
+    CUdeviceptr srcBuffer;
+    CUdeviceptr dstBuffer;
+    int         tex_alignment;
+} CUDAScaleContext;
+
+static av_cold int cudascale_init(AVFilterContext *ctx)
+{
+    CUDAScaleContext *s = ctx->priv;
+
+    s->format = AV_PIX_FMT_NONE;
+    s->frame = av_frame_alloc();
+    if (!s->frame)
+        return AVERROR(ENOMEM);
+
+    s->tmp_frame = av_frame_alloc();
+    if (!s->tmp_frame)
+        return AVERROR(ENOMEM);
+
+    return 0;
+}
+
+static av_cold void cudascale_uninit(AVFilterContext *ctx)
+{
+    CUDAScaleContext *s = ctx->priv;
+
+    av_frame_free(&s->frame);
+    av_buffer_unref(&s->frames_ctx);
+    av_frame_free(&s->tmp_frame);
+}
+
+static int cudascale_query_formats(AVFilterContext *ctx)
+{
+    static const enum AVPixelFormat pixel_formats[] = {
+        AV_PIX_FMT_CUDA, AV_PIX_FMT_NONE,
+    };
+    AVFilterFormats *pix_fmts = ff_make_format_list(pixel_formats);
+
+    return ff_set_common_formats(ctx, pix_fmts);
+}
+
+static av_cold int init_stage(CUDAScaleContext *s, AVBufferRef *device_ctx)
+{
+    AVBufferRef *out_ref = NULL;
+    AVHWFramesContext *out_ctx;
+    int in_sw, in_sh, out_sw, out_sh;
+    int ret, i;
+
+    av_pix_fmt_get_chroma_sub_sample(s->in_fmt,  &in_sw,  &in_sh);
+    av_pix_fmt_get_chroma_sub_sample(s->out_fmt, &out_sw, &out_sh);
+    if (!s->planes_out[0].width) {
+        s->planes_out[0].width  = s->planes_in[0].width;
+        s->planes_out[0].height = s->planes_in[0].height;
+    }
+
+    for (i = 1; i < FF_ARRAY_ELEMS(s->planes_in); i++) {
+        s->planes_in[i].width   = s->planes_in[0].width   >> in_sw;
+        s->planes_in[i].height  = s->planes_in[0].height  >> in_sh;
+        s->planes_out[i].width  = s->planes_out[0].width  >> out_sw;
+        s->planes_out[i].height = s->planes_out[0].height >> out_sh;
+    }
+
+    out_ref = av_hwframe_ctx_alloc(device_ctx);
+    if (!out_ref)
+        return AVERROR(ENOMEM);
+    out_ctx = (AVHWFramesContext*)out_ref->data;
+
+    out_ctx->format    = AV_PIX_FMT_CUDA;
+    out_ctx->sw_format = s->out_fmt;
+    out_ctx->width     = FFALIGN(s->planes_out[0].width,  32);
+    out_ctx->height    = FFALIGN(s->planes_out[0].height, 32);
+
+    ret = av_hwframe_ctx_init(out_ref);
+    if (ret < 0)
+        goto fail;
+
+    av_frame_unref(s->frame);
+    ret = av_hwframe_get_buffer(out_ref, s->frame, 0);
+    if (ret < 0)
+        goto fail;
+
+    s->frame->width  = s->planes_out[0].width;
+    s->frame->height = s->planes_out[0].height;
+
+    av_buffer_unref(&s->frames_ctx);
+    s->frames_ctx = out_ref;
+
+    return 0;
+fail:
+    av_buffer_unref(&out_ref);
+    return ret;
+}
+
+static int format_is_supported(enum AVPixelFormat fmt)
+{
+    int i;
+
+    for (i = 0; i < FF_ARRAY_ELEMS(supported_formats); i++)
+        if (supported_formats[i] == fmt)
+            return 1;
+    return 0;
+}
+
+static av_cold int init_processing_chain(AVFilterContext *ctx, int in_width, int in_height,
+                                         int out_width, int out_height)
+{
+    CUDAScaleContext *s = ctx->priv;
+
+    AVHWFramesContext *in_frames_ctx;
+
+    enum AVPixelFormat in_format;
+    enum AVPixelFormat out_format;
+    int ret;
+
+    /* check that we have a hw context */
+    if (!ctx->inputs[0]->hw_frames_ctx) {
+        av_log(ctx, AV_LOG_ERROR, "No hw context provided on input\n");
+        return AVERROR(EINVAL);
+    }
+    in_frames_ctx = (AVHWFramesContext*)ctx->inputs[0]->hw_frames_ctx->data;
+    in_format     = in_frames_ctx->sw_format;
+    out_format    = (s->format == AV_PIX_FMT_NONE) ? in_format : s->format;
+
+    if (!format_is_supported(in_format)) {
+        av_log(ctx, AV_LOG_ERROR, "Unsupported input format: %s\n",
+               av_get_pix_fmt_name(in_format));
+        return AVERROR(ENOSYS);
+    }
+    if (!format_is_supported(out_format)) {
+        av_log(ctx, AV_LOG_ERROR, "Unsupported output format: %s\n",
+               av_get_pix_fmt_name(out_format));
+        return AVERROR(ENOSYS);
+    }
+
+    if (in_width == out_width && in_height == out_height)
+        s->passthrough = 1;
+
+    s->in_fmt = in_format;
+    s->out_fmt = out_format;
+
+    s->planes_in[0].width   = in_width;
+    s->planes_in[0].height  = in_height;
+    s->planes_out[0].width  = out_width;
+    s->planes_out[0].height = out_height;
+
+    ret = init_stage(s, in_frames_ctx->device_ref);
+    if (ret < 0)
+        return ret;
+
+    ctx->outputs[0]->hw_frames_ctx = av_buffer_ref(s->frames_ctx);
+    if (!ctx->outputs[0]->hw_frames_ctx)
+        return AVERROR(ENOMEM);
+
+    return 0;
+}
+
+static av_cold int cudascale_config_props(AVFilterLink *outlink)
+{
+    AVFilterContext *ctx = outlink->src;
+    AVFilterLink *inlink = outlink->src->inputs[0];
+    CUDAScaleContext *s  = ctx->priv;
+    AVHWFramesContext     *frames_ctx = (AVHWFramesContext*)inlink->hw_frames_ctx->data;
+    AVCUDADeviceContext *device_hwctx = frames_ctx->device_ctx->hwctx;
+    CUcontext dummy, cuda_ctx = device_hwctx->cuda_ctx;
+    CUresult err;
+    int w, h;
+    int ret;
+
+    extern char vf_scale_cuda_ptx[];
+
+    err = cuCtxPushCurrent(cuda_ctx);
+    if (err != CUDA_SUCCESS) {
+        av_log(ctx, AV_LOG_ERROR, "Error pushing cuda context\n");
+        ret = AVERROR_UNKNOWN;
+        goto fail;
+    }
+
+    err = cuModuleLoadData(&s->cu_module, vf_scale_cuda_ptx);
+    if (err != CUDA_SUCCESS) {
+        av_log(ctx, AV_LOG_ERROR, "Error loading module data\n");
+        ret = AVERROR_UNKNOWN;
+        goto fail;
+    }
+
+    cuModuleGetFunction(&s->cu_func_uchar, s->cu_module, "Subsample_Bilinear_uchar");
+    cuModuleGetFunction(&s->cu_func_uchar2, s->cu_module, "Subsample_Bilinear_uchar2");
+    cuModuleGetFunction(&s->cu_func_uchar4, s->cu_module, "Subsample_Bilinear_uchar4");
+    cuModuleGetFunction(&s->cu_func_ushort, s->cu_module, "Subsample_Bilinear_ushort");
+    cuModuleGetFunction(&s->cu_func_ushort2, s->cu_module, "Subsample_Bilinear_ushort2");
+    cuModuleGetFunction(&s->cu_func_ushort4, s->cu_module, "Subsample_Bilinear_ushort4");
+
+    cuModuleGetTexRef(&s->cu_tex_uchar, s->cu_module, "uchar_tex");
+    cuModuleGetTexRef(&s->cu_tex_uchar2, s->cu_module, "uchar2_tex");
+    cuModuleGetTexRef(&s->cu_tex_uchar4, s->cu_module, "uchar4_tex");
+    cuModuleGetTexRef(&s->cu_tex_ushort, s->cu_module, "ushort_tex");
+    cuModuleGetTexRef(&s->cu_tex_ushort2, s->cu_module, "ushort2_tex");
+    cuModuleGetTexRef(&s->cu_tex_ushort4, s->cu_module, "ushort4_tex");
+
+    cuTexRefSetFlags(s->cu_tex_uchar, CU_TRSF_READ_AS_INTEGER);
+    cuTexRefSetFlags(s->cu_tex_uchar2, CU_TRSF_READ_AS_INTEGER);
+    cuTexRefSetFlags(s->cu_tex_uchar4, CU_TRSF_READ_AS_INTEGER);
+    cuTexRefSetFlags(s->cu_tex_ushort, CU_TRSF_READ_AS_INTEGER);
+    cuTexRefSetFlags(s->cu_tex_ushort2, CU_TRSF_READ_AS_INTEGER);
+    cuTexRefSetFlags(s->cu_tex_ushort4, CU_TRSF_READ_AS_INTEGER);
+
+    cuTexRefSetFilterMode(s->cu_tex_uchar, CU_TR_FILTER_MODE_LINEAR);
+    cuTexRefSetFilterMode(s->cu_tex_uchar2, CU_TR_FILTER_MODE_LINEAR);
+    cuTexRefSetFilterMode(s->cu_tex_uchar4, CU_TR_FILTER_MODE_LINEAR);
+    cuTexRefSetFilterMode(s->cu_tex_ushort, CU_TR_FILTER_MODE_LINEAR);
+    cuTexRefSetFilterMode(s->cu_tex_ushort2, CU_TR_FILTER_MODE_LINEAR);
+    cuTexRefSetFilterMode(s->cu_tex_ushort4, CU_TR_FILTER_MODE_LINEAR);
+
+    cuCtxPopCurrent(&dummy);
+
+    if ((ret = ff_scale_eval_dimensions(s,
+                                        s->w_expr, s->h_expr,
+                                        inlink, outlink,
+                                        &w, &h)) < 0)
+        goto fail;
+
+    if (((int64_t)h * inlink->w) > INT_MAX  ||
+        ((int64_t)w * inlink->h) > INT_MAX)
+        av_log(ctx, AV_LOG_ERROR, "Rescaled value for width or height is too big.\n");
+
+    outlink->w = w;
+    outlink->h = h;
+
+    ret = init_processing_chain(ctx, inlink->w, inlink->h, w, h);
+    if (ret < 0)
+        return ret;
+
+    av_log(ctx, AV_LOG_VERBOSE, "w:%d h:%d -> w:%d h:%d\n",
+           inlink->w, inlink->h, outlink->w, outlink->h);
+
+    if (inlink->sample_aspect_ratio.num) {
+        outlink->sample_aspect_ratio = av_mul_q((AVRational){outlink->h*inlink->w,
+                                                             outlink->w*inlink->h},
+                                                inlink->sample_aspect_ratio);
+    } else {
+        outlink->sample_aspect_ratio = inlink->sample_aspect_ratio;
+    }
+
+    return 0;
+
+fail:
+    return ret;
+}
+
+static int call_resize_kernel(CUDAScaleContext *s, CUfunction func, CUtexref tex, int channels,
+                              uint8_t *src_dptr, int src_width, int src_height, int src_pitch,
+                              uint8_t *dst_dptr, int dst_width, int dst_height, int dst_pitch,
+                              int pixel_size)
+{
+    CUdeviceptr src_devptr = (CUdeviceptr)src_dptr;
+    CUdeviceptr dst_devptr = (CUdeviceptr)dst_dptr;
+    void *args_uchar[] = { &dst_devptr, &dst_width, &dst_height, &dst_pitch, &src_width, &src_height };
+    CUDA_ARRAY_DESCRIPTOR desc;
+
+    desc.Width  = src_width;
+    desc.Height = src_height;
+    desc.NumChannels = channels;
+    if (pixel_size == 1) {
+        desc.Format = CU_AD_FORMAT_UNSIGNED_INT8;
+    } else {
+        desc.Format = CU_AD_FORMAT_UNSIGNED_INT16;
+    }
+
+    cuTexRefSetAddress2D_v3(tex, &desc, src_devptr, src_pitch * pixel_size);
+    cuLaunchKernel(func, DIV_UP(dst_width, BLOCKX), DIV_UP(dst_height, BLOCKY), 1, BLOCKX, BLOCKY, 1, 0, 0, args_uchar, NULL);
+
+    return 0;
+}
+
+static int scalecuda_resize(AVFilterContext *ctx,
+                            AVFrame *out, AVFrame *in)
+{
+    AVHWFramesContext *in_frames_ctx = (AVHWFramesContext*)in->hw_frames_ctx->data;
+    CUDAScaleContext *s = ctx->priv;
+
+    switch (in_frames_ctx->sw_format) {
+    case AV_PIX_FMT_YUV420P:
+        call_resize_kernel(s, s->cu_func_uchar, s->cu_tex_uchar, 1,
+                           in->data[0], in->width, in->height, in->linesize[0],
+                           out->data[0], out->width, out->height, out->linesize[0],
+                           1);
+        call_resize_kernel(s, s->cu_func_uchar, s->cu_tex_uchar, 1,
+                           in->data[0]+in->linesize[0]*in->height, in->width/2, in->height/2, in->linesize[0]/2,
+                           out->data[0]+out->linesize[0]*out->height, out->width/2, out->height/2, out->linesize[0]/2,
+                           1);
+        call_resize_kernel(s, s->cu_func_uchar, s->cu_tex_uchar, 1,
+                           in->data[0]+ ALIGN_UP((in->linesize[0]*in->height*5)/4, s->tex_alignment), in->width/2, in->height/2, in->linesize[0]/2,
+                           out->data[0]+(out->linesize[0]*out->height*5)/4, out->width/2, out->height/2, out->linesize[0]/2,
+                           1);
+        break;
+    case AV_PIX_FMT_YUV444P:
+        call_resize_kernel(s, s->cu_func_uchar, s->cu_tex_uchar, 1,
+                           in->data[0], in->width, in->height, in->linesize[0],
+                           out->data[0], out->width, out->height, out->linesize[0],
+                           1);
+        call_resize_kernel(s, s->cu_func_uchar, s->cu_tex_uchar, 1,
+                           in->data[0]+in->linesize[0]*in->height, in->width, in->height, in->linesize[0],
+                           out->data[0]+out->linesize[0]*out->height, out->width, out->height, out->linesize[0],
+                           1);
+        call_resize_kernel(s, s->cu_func_uchar, s->cu_tex_uchar, 1,
+                           in->data[0]+in->linesize[0]*in->height*2, in->width, in->height, in->linesize[0],
+                           out->data[0]+out->linesize[0]*out->height*2, out->width, out->height, out->linesize[0],
+                           1);
+        break;
+    case AV_PIX_FMT_NV12:
+        call_resize_kernel(s, s->cu_func_uchar, s->cu_tex_uchar, 1,
+                           in->data[0], in->width, in->height, in->linesize[0],
+                           out->data[0], out->width, out->height, out->linesize[0],
+                           1);
+        call_resize_kernel(s, s->cu_func_uchar2, s->cu_tex_uchar2, 2,
+                           in->data[1], in->width/2, in->height/2, in->linesize[1],
+                           out->data[0] + out->linesize[0] * ((out->height + 31) & ~0x1f), out->width/2, out->height/2, out->linesize[1]/2,
+                           1);
+        break;
+    case AV_PIX_FMT_P010LE:
+        call_resize_kernel(s, s->cu_func_ushort, s->cu_tex_ushort, 1,
+                           in->data[0], in->width, in->height, in->linesize[0]/2,
+                           out->data[0], out->width, out->height, out->linesize[0]/2,
+                           2);
+        call_resize_kernel(s, s->cu_func_ushort2, s->cu_tex_ushort2, 2,
+                           in->data[1], in->width / 2, in->height / 2, in->linesize[1]/2,
+                           out->data[0] + out->linesize[0] * ((out->height + 31) & ~0x1f), out->width / 2, out->height / 2, out->linesize[1] / 4,
+                           2);
+        break;
+    case AV_PIX_FMT_P016LE:
+        call_resize_kernel(s, s->cu_func_ushort, s->cu_tex_ushort, 1,
+                           in->data[0], in->width, in->height, in->linesize[0] / 2,
+                           out->data[0], out->width, out->height, out->linesize[0] / 2,
+                           2);
+        call_resize_kernel(s, s->cu_func_ushort2, s->cu_tex_ushort2, 2,
+                           in->data[1], in->width / 2, in->height / 2, in->linesize[1] / 2,
+                           out->data[0] + out->linesize[0] * ((out->height + 31) & ~0x1f), out->width / 2, out->height / 2, out->linesize[1] / 4,
+                           2);
+        break;
+    default:
+        return AVERROR_BUG;
+    }
+
+    return 0;
+}
+
+static int cudascale_scale(AVFilterContext *ctx, AVFrame *out, AVFrame *in)
+{
+    CUDAScaleContext *s = ctx->priv;
+    AVFrame *src = in;
+    int ret;
+
+    ret = scalecuda_resize(ctx, s->frame, src);
+    if (ret < 0)
+        return ret;
+
+    src = s->frame;
+    ret = av_hwframe_get_buffer(src->hw_frames_ctx, s->tmp_frame, 0);
+    if (ret < 0)
+        return ret;
+
+    av_frame_move_ref(out, s->frame);
+    av_frame_move_ref(s->frame, s->tmp_frame);
+
+    ret = av_frame_copy_props(out, in);
+    if (ret < 0)
+        return ret;
+
+    return 0;
+}
+
+static int cudascale_filter_frame(AVFilterLink *link, AVFrame *in)
+{
+    AVFilterContext              *ctx = link->dst;
+    CUDAScaleContext               *s = ctx->priv;
+    AVFilterLink             *outlink = ctx->outputs[0];
+    AVHWFramesContext     *frames_ctx = (AVHWFramesContext*)s->frames_ctx->data;
+    AVCUDADeviceContext *device_hwctx = frames_ctx->device_ctx->hwctx;
+
+    AVFrame *out = NULL;
+    CUresult err;
+    CUcontext dummy;
+    int ret = 0;
+
+    out = av_frame_alloc();
+    if (!out) {
+        ret = AVERROR(ENOMEM);
+        goto fail;
+    }
+
+    err = cuCtxPushCurrent(device_hwctx->cuda_ctx);
+    if (err != CUDA_SUCCESS) {
+        ret = AVERROR_UNKNOWN;
+        goto fail;
+    }
+
+    ret = cudascale_scale(ctx, out, in);
+
+    cuCtxPopCurrent(&dummy);
+    if (ret < 0)
+        goto fail;
+
+    av_reduce(&out->sample_aspect_ratio.num, &out->sample_aspect_ratio.den,
+              (int64_t)in->sample_aspect_ratio.num * outlink->h * link->w,
+              (int64_t)in->sample_aspect_ratio.den * outlink->w * link->h,
+              INT_MAX);
+
+    av_frame_free(&in);
+    return ff_filter_frame(outlink, out);
+fail:
+    av_frame_free(&in);
+    av_frame_free(&out);
+    return ret;
+}
+
+#define OFFSET(x) offsetof(CUDAScaleContext, x)
+#define FLAGS (AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_VIDEO_PARAM)
+static const AVOption options[] = {
+    { "w",      "Output video width",  OFFSET(w_expr),     AV_OPT_TYPE_STRING, { .str = "iw"   }, .flags = FLAGS },
+    { "h",      "Output video height", OFFSET(h_expr),     AV_OPT_TYPE_STRING, { .str = "ih"   }, .flags = FLAGS },
+    { NULL },
+};
+
+static const AVClass cudascale_class = {
+    .class_name = "cudascale",
+    .item_name  = av_default_item_name,
+    .option     = options,
+    .version    = LIBAVUTIL_VERSION_INT,
+};
+
+static const AVFilterPad cudascale_inputs[] = {
+    {
+        .name        = "default",
+        .type        = AVMEDIA_TYPE_VIDEO,
+        .filter_frame = cudascale_filter_frame,
+    },
+    { NULL }
+};
+
+static const AVFilterPad cudascale_outputs[] = {
+    {
+        .name         = "default",
+        .type         = AVMEDIA_TYPE_VIDEO,
+        .config_props = cudascale_config_props,
+    },
+    { NULL }
+};
+
+AVFilter ff_vf_scale_cuda = {
+    .name      = "scale_cuda",
+    .description = NULL_IF_CONFIG_SMALL("GPU accelerated video resizer"),
+
+    .init          = cudascale_init,
+    .uninit        = cudascale_uninit,
+    .query_formats = cudascale_query_formats,
+
+    .priv_size = sizeof(CUDAScaleContext),
+    .priv_class = &cudascale_class,
+
+    .inputs    = cudascale_inputs,
+    .outputs   = cudascale_outputs,
+
+    .flags_internal = FF_FILTER_FLAG_HWFRAME_AWARE,
+};
diff --git a/libavfilter/vf_scale_cuda.cu b/libavfilter/vf_scale_cuda.cu
new file mode 100644
index 0000000000..5f5ec81989
--- /dev/null
+++ b/libavfilter/vf_scale_cuda.cu
@@ -0,0 +1,212 @@
+/*
+ * Copyright (c) 2017, NVIDIA CORPORATION. All rights reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+ * DEALINGS IN THE SOFTWARE.
+ */
+
+extern "C" {
+
+texture<unsigned char, 2> uchar_tex;
+texture<uchar2, 2>  uchar2_tex;
+texture<uchar4, 2>  uchar4_tex;
+texture<unsigned short, 2> ushort_tex;
+texture<ushort2, 2>  ushort2_tex;
+texture<ushort4, 2>  ushort4_tex;
+
+__global__ void Subsample_Bilinear_uchar(unsigned char *dst,
+                                    int dst_width, int dst_height, int dst_pitch,
+                                    int src_width, int src_height)
+{
+    int xo = blockIdx.x * blockDim.x + threadIdx.x;
+    int yo = blockIdx.y * blockDim.y + threadIdx.y;
+
+    if (yo < dst_height && xo < dst_width)
+    {
+        float hscale = (float)src_width / (float)dst_width;
+        float vscale = (float)src_height / (float)dst_height;
+        float xi = (xo + 0.5f) * hscale;
+        float yi = (yo + 0.5f) * vscale;
+        // 3-tap filter weights are {wh,1.0,wh} and {wv,1.0,wv}
+        float wh = min(max(0.5f * (hscale - 1.0f), 0.0f), 1.0f);
+        float wv = min(max(0.5f * (vscale - 1.0f), 0.0f), 1.0f);
+        // Convert weights to two bilinear weights -> {wh,1.0,wh} -> {wh,0.5,0} + {0,0.5,wh}
+        float dx = wh / (0.5f + wh);
+        float dy = wv / (0.5f + wv);
+        int y0 = tex2D(uchar_tex, xi-dx, yi-dy);
+        int y1 = tex2D(uchar_tex, xi+dx, yi-dy);
+        int y2 = tex2D(uchar_tex, xi-dx, yi+dy);
+        int y3 = tex2D(uchar_tex, xi+dx, yi+dy);
+        dst[yo*dst_pitch+xo] = (unsigned char)((y0+y1+y2+y3+2) >> 2);
+    }
+}
+
+__global__ void Subsample_Bilinear_uchar2(uchar2 *dst,
+                                    int dst_width, int dst_height, int dst_pitch2,
+                                    int src_width, int src_height)
+{
+    int xo = blockIdx.x * blockDim.x + threadIdx.x;
+    int yo = blockIdx.y * blockDim.y + threadIdx.y;
+
+    if (yo < dst_height && xo < dst_width)
+    {
+        float hscale = (float)src_width / (float)dst_width;
+        float vscale = (float)src_height / (float)dst_height;
+        float xi = (xo + 0.5f) * hscale;
+        float yi = (yo + 0.5f) * vscale;
+        // 3-tap filter weights are {wh,1.0,wh} and {wv,1.0,wv}
+        float wh = min(max(0.5f * (hscale - 1.0f), 0.0f), 1.0f);
+        float wv = min(max(0.5f * (vscale - 1.0f), 0.0f), 1.0f);
+        // Convert weights to two bilinear weights -> {wh,1.0,wh} -> {wh,0.5,0} + {0,0.5,wh}
+        float dx = wh / (0.5f + wh);
+        float dy = wv / (0.5f + wv);
+        uchar2 c0 = tex2D(uchar2_tex, xi-dx, yi-dy);
+        uchar2 c1 = tex2D(uchar2_tex, xi+dx, yi-dy);
+        uchar2 c2 = tex2D(uchar2_tex, xi-dx, yi+dy);
+        uchar2 c3 = tex2D(uchar2_tex, xi+dx, yi+dy);
+        int2 uv;
+        uv.x = ((int)c0.x+(int)c1.x+(int)c2.x+(int)c3.x+2) >> 2;
+        uv.y = ((int)c0.y+(int)c1.y+(int)c2.y+(int)c3.y+2) >> 2;
+        dst[yo*dst_pitch2+xo] = make_uchar2((unsigned char)uv.x, (unsigned char)uv.y);
+    }
+}
+
+__global__ void Subsample_Bilinear_uchar4(uchar4 *dst,
+                                    int dst_width, int dst_height, int dst_pitch,
+                                    int src_width, int src_height)
+{
+    int xo = blockIdx.x * blockDim.x + threadIdx.x;
+    int yo = blockIdx.y * blockDim.y + threadIdx.y;
+
+    if (yo < dst_height && xo < dst_width)
+    {
+        float hscale = (float)src_width / (float)dst_width;
+        float vscale = (float)src_height / (float)dst_height;
+        float xi = (xo + 0.5f) * hscale;
+        float yi = (yo + 0.5f) * vscale;
+        // 3-tap filter weights are {wh,1.0,wh} and {wv,1.0,wv}
+        float wh = min(max(0.5f * (hscale - 1.0f), 0.0f), 1.0f);
+        float wv = min(max(0.5f * (vscale - 1.0f), 0.0f), 1.0f);
+        // Convert weights to two bilinear weights -> {wh,1.0,wh} -> {wh,0.5,0} + {0,0.5,wh}
+        float dx = wh / (0.5f + wh);
+        float dy = wv / (0.5f + wv);
+        uchar4 c0 = tex2D(uchar4_tex, xi-dx, yi-dy);
+        uchar4 c1 = tex2D(uchar4_tex, xi+dx, yi-dy);
+        uchar4 c2 = tex2D(uchar4_tex, xi-dx, yi+dy);
+        uchar4 c3 = tex2D(uchar4_tex, xi+dx, yi+dy);
+        int4 res;
+        res.x =  ((int)c0.x+(int)c1.x+(int)c2.x+(int)c3.x+2) >> 2;
+        res.y =  ((int)c0.y+(int)c1.y+(int)c2.y+(int)c3.y+2) >> 2;
+        res.z =  ((int)c0.z+(int)c1.z+(int)c2.z+(int)c3.z+2) >> 2;
+        res.w =  ((int)c0.w+(int)c1.w+(int)c2.w+(int)c3.w+2) >> 2;
+        dst[yo*dst_pitch+xo] = make_uchar4(
+            (unsigned char)res.x, (unsigned char)res.y, (unsigned char)res.z, (unsigned char)res.w);
+    }
+}
+
+__global__ void Subsample_Bilinear_ushort(unsigned short *dst,
+                                    int dst_width, int dst_height, int dst_pitch,
+                                    int src_width, int src_height)
+{
+    int xo = blockIdx.x * blockDim.x + threadIdx.x;
+    int yo = blockIdx.y * blockDim.y + threadIdx.y;
+
+    if (yo < dst_height && xo < dst_width)
+    {
+        float hscale = (float)src_width / (float)dst_width;
+        float vscale = (float)src_height / (float)dst_height;
+        float xi = (xo + 0.5f) * hscale;
+        float yi = (yo + 0.5f) * vscale;
+        // 3-tap filter weights are {wh,1.0,wh} and {wv,1.0,wv}
+        float wh = min(max(0.5f * (hscale - 1.0f), 0.0f), 1.0f);
+        float wv = min(max(0.5f * (vscale - 1.0f), 0.0f), 1.0f);
+        // Convert weights to two bilinear weights -> {wh,1.0,wh} -> {wh,0.5,0} + {0,0.5,wh}
+        float dx = wh / (0.5f + wh);
+        float dy = wv / (0.5f + wv);
+        int y0 = tex2D(ushort_tex, xi-dx, yi-dy);
+        int y1 = tex2D(ushort_tex, xi+dx, yi-dy);
+        int y2 = tex2D(ushort_tex, xi-dx, yi+dy);
+        int y3 = tex2D(ushort_tex, xi+dx, yi+dy);
+        dst[yo*dst_pitch+xo] = (unsigned short)((y0+y1+y2+y3+2) >> 2);
+    }
+}
+
+__global__ void Subsample_Bilinear_ushort2(ushort2 *dst,
+                                    int dst_width, int dst_height, int dst_pitch2,
+                                    int src_width, int src_height)
+{
+    int xo = blockIdx.x * blockDim.x + threadIdx.x;
+    int yo = blockIdx.y * blockDim.y + threadIdx.y;
+
+    if (yo < dst_height && xo < dst_width)
+    {
+        float hscale = (float)src_width / (float)dst_width;
+        float vscale = (float)src_height / (float)dst_height;
+        float xi = (xo + 0.5f) * hscale;
+        float yi = (yo + 0.5f) * vscale;
+        // 3-tap filter weights are {wh,1.0,wh} and {wv,1.0,wv}
+        float wh = min(max(0.5f * (hscale - 1.0f), 0.0f), 1.0f);
+        float wv = min(max(0.5f * (vscale - 1.0f), 0.0f), 1.0f);
+        // Convert weights to two bilinear weights -> {wh,1.0,wh} -> {wh,0.5,0} + {0,0.5,wh}
+        float dx = wh / (0.5f + wh);
+        float dy = wv / (0.5f + wv);
+        ushort2 c0 = tex2D(ushort2_tex, xi-dx, yi-dy);
+        ushort2 c1 = tex2D(ushort2_tex, xi+dx, yi-dy);
+        ushort2 c2 = tex2D(ushort2_tex, xi-dx, yi+dy);
+        ushort2 c3 = tex2D(ushort2_tex, xi+dx, yi+dy);
+        int2 uv;
+        uv.x = ((int)c0.x+(int)c1.x+(int)c2.x+(int)c3.x+2) >> 2;
+        uv.y = ((int)c0.y+(int)c1.y+(int)c2.y+(int)c3.y+2) >> 2;
+        dst[yo*dst_pitch2+xo] = make_ushort2((unsigned short)uv.x, (unsigned short)uv.y);
+    }
+}
+
+__global__ void Subsample_Bilinear_ushort4(ushort4 *dst,
+                                    int dst_width, int dst_height, int dst_pitch,
+                                    int src_width, int src_height)
+{
+    int xo = blockIdx.x * blockDim.x + threadIdx.x;
+    int yo = blockIdx.y * blockDim.y + threadIdx.y;
+
+    if (yo < dst_height && xo < dst_width)
+    {
+        float hscale = (float)src_width / (float)dst_width;
+        float vscale = (float)src_height / (float)dst_height;
+        float xi = (xo + 0.5f) * hscale;
+        float yi = (yo + 0.5f) * vscale;
+        // 3-tap filter weights are {wh,1.0,wh} and {wv,1.0,wv}
+        float wh = min(max(0.5f * (hscale - 1.0f), 0.0f), 1.0f);
+        float wv = min(max(0.5f * (vscale - 1.0f), 0.0f), 1.0f);
+        // Convert weights to two bilinear weights -> {wh,1.0,wh} -> {wh,0.5,0} + {0,0.5,wh}
+        float dx = wh / (0.5f + wh);
+        float dy = wv / (0.5f + wv);
+        ushort4 c0 = tex2D(ushort4_tex, xi-dx, yi-dy);
+        ushort4 c1 = tex2D(ushort4_tex, xi+dx, yi-dy);
+        ushort4 c2 = tex2D(ushort4_tex, xi-dx, yi+dy);
+        ushort4 c3 = tex2D(ushort4_tex, xi+dx, yi+dy);
+        int4 res;
+        res.x =  ((int)c0.x+(int)c1.x+(int)c2.x+(int)c3.x+2) >> 2;
+        res.y =  ((int)c0.y+(int)c1.y+(int)c2.y+(int)c3.y+2) >> 2;
+        res.z =  ((int)c0.z+(int)c1.z+(int)c2.z+(int)c3.z+2) >> 2;
+        res.w =  ((int)c0.w+(int)c1.w+(int)c2.w+(int)c3.w+2) >> 2;
+        dst[yo*dst_pitch+xo] = make_ushort4(
+            (unsigned short)res.x, (unsigned short)res.y, (unsigned short)res.z, (unsigned short)res.w);
+    }
+}
+
+}