lavfi: add an NVIDIA NPP-based scaling filter

7 files changed, 707 insertions, 1 deletions

diff --git a/Changelog b/Changelog
index 4a7d2a117b..692c865f52 100644
--- a/Changelog
+++ b/Changelog
@@ -52,6 +52,7 @@ version <next>:
 - G.723.1 muxer and encoder
 - compressed SWF
 - VAAPI-accelerated format conversion and scaling
+- libnpp/CUDA-accelerated format conversion and scaling
 
 
 version 11:
diff --git a/configure b/configure
index f016ebd170..a68a5d7354 100755
--- a/configure
+++ b/configure
@@ -195,6 +195,7 @@ External library support:
   --enable-libkvazaar      enable HEVC encoding via libkvazaar [no]
   --enable-libmfx          enable HW acceleration through libmfx
   --enable-libmp3lame      enable MP3 encoding via libmp3lame [no]
+  --enable-libnpp          enable NVIDIA Performance Primitives-based code [no]
   --enable-libopencore-amrnb enable AMR-NB de/encoding via libopencore-amrnb [no]
   --enable-libopencore-amrwb enable AMR-WB decoding via libopencore-amrwb [no]
   --enable-libopencv       enable video filtering via libopencv [no]
@@ -1222,6 +1223,7 @@ EXTERNAL_LIBRARY_LIST="
     libkvazaar
     libmfx
     libmp3lame
+    libnpp
     libopencore_amrnb
     libopencore_amrwb
     libopencv
@@ -2359,6 +2361,7 @@ interlace_filter_deps="gpl"
 ocv_filter_deps="libopencv"
 resample_filter_deps="avresample"
 scale_filter_deps="swscale"
+scale_npp_filter_deps="cuda libnpp"
 scale_vaapi_filter_deps="vaapi VAProcPipelineParameterBuffer"
 
 # examples
@@ -4024,6 +4027,7 @@ die_license_disabled gpl x11grab
 die_license_disabled nonfree cuda
 die_license_disabled nonfree libfaac
 die_license_disabled nonfree libfdk_aac
+die_license_disabled nonfree libnpp
 die_license_disabled nonfree nvenc
 die_license_disabled nonfree openssl
 
@@ -4520,6 +4524,7 @@ enabled libilbc           && require libilbc ilbc.h WebRtcIlbcfix_InitDecode -li
 enabled libkvazaar        && require_pkg_config "kvazaar >= 0.8.1" kvazaar.h kvz_api_get
 enabled libmfx            && require_pkg_config libmfx "mfx/mfxvideo.h" MFXInit
 enabled libmp3lame        && require "libmp3lame >= 3.98.3" lame/lame.h lame_set_VBR_quality -lmp3lame
+enabled libnpp            && require libnpp npp.h nppGetLibVersion -lnppi -lnppc
 enabled libopencore_amrnb && require libopencore_amrnb opencore-amrnb/interf_dec.h Decoder_Interface_init -lopencore-amrnb
 enabled libopencore_amrwb && require libopencore_amrwb opencore-amrwb/dec_if.h D_IF_init -lopencore-amrwb
 enabled libopencv         && require_pkg_config opencv opencv/cv.h cvCreateImageHeader
diff --git a/doc/filters.texi b/doc/filters.texi
index 722c8b03b6..23e4286da8 100644
--- a/doc/filters.texi
+++ b/doc/filters.texi
@@ -2181,6 +2181,44 @@ scale="trunc(3/2*iw/hsub)*hsub:trunc(3/2*ih/vsub)*vsub"
 scale=w='min(500\, iw*3/2):h=-1'
 @end example
 
+@section scale_npp
+
+Use the NVIDIA Performance Primitives (libnpp) to perform scaling and/or pixel
+format conversion on CUDA video frames. Setting the output width and height
+works in the same way as for the @var{scale} filter.
+
+The following additional options are accepted:
+@table @option
+@item format
+The pixel format of the output CUDA frames. If set to the string "same" (the
+default), the input format will be kept. Note that automatic format negotiation
+and conversion is not yet supported for hardware frames
+
+@item interp_algo
+The interpolation algorithm used for resizing. One of the following:
+@table @option
+@item nn
+Nearest neighbour.
+
+@item linear
+@item cubic
+@item cubic2p_bspline
+2-parameter cubic (B=1, C=0)
+
+@item cubic2p_catmullrom
+2-parameter cubic (B=0, C=1/2)
+
+@item cubic2p_b05c03
+2-parameter cubic (B=1/2, C=3/10)
+
+@item super
+Supersampling
+
+@item lanczos
+@end table
+
+@end table
+
 @section select
 Select frames to pass in output.
 
diff --git a/libavfilter/Makefile b/libavfilter/Makefile
index 0f9643d9f5..39e167faab 100644
--- a/libavfilter/Makefile
+++ b/libavfilter/Makefile
@@ -71,6 +71,7 @@ OBJS-$(CONFIG_OVERLAY_FILTER)                += vf_overlay.o
 OBJS-$(CONFIG_PAD_FILTER)                    += vf_pad.o
 OBJS-$(CONFIG_PIXDESCTEST_FILTER)            += vf_pixdesctest.o
 OBJS-$(CONFIG_SCALE_FILTER)                  += vf_scale.o
+OBJS-$(CONFIG_SCALE_NPP_FILTER)              += vf_scale_npp.o
 OBJS-$(CONFIG_SCALE_VAAPI_FILTER)            += vf_scale_vaapi.o
 OBJS-$(CONFIG_SELECT_FILTER)                 += vf_select.o
 OBJS-$(CONFIG_SETDAR_FILTER)                 += vf_aspect.o
diff --git a/libavfilter/allfilters.c b/libavfilter/allfilters.c
index d76db06f2a..c6eeb1f4e4 100644
--- a/libavfilter/allfilters.c
+++ b/libavfilter/allfilters.c
@@ -97,6 +97,7 @@ void avfilter_register_all(void)
     REGISTER_FILTER(PAD,            pad,            vf);
     REGISTER_FILTER(PIXDESCTEST,    pixdesctest,    vf);
     REGISTER_FILTER(SCALE,          scale,          vf);
+    REGISTER_FILTER(SCALE_NPP,      scale_npp,      vf);
     REGISTER_FILTER(SCALE_VAAPI,    scale_vaapi,    vf);
     REGISTER_FILTER(SELECT,         select,         vf);
     REGISTER_FILTER(SETDAR,         setdar,         vf);
diff --git a/libavfilter/version.h b/libavfilter/version.h
index 1ca0f37a6a..c8102dd7d0 100644
--- a/libavfilter/version.h
+++ b/libavfilter/version.h
@@ -30,7 +30,7 @@
 #include "libavutil/version.h"
 
 #define LIBAVFILTER_VERSION_MAJOR  6
-#define LIBAVFILTER_VERSION_MINOR  3
+#define LIBAVFILTER_VERSION_MINOR  4
 #define LIBAVFILTER_VERSION_MICRO  0
 
 #define LIBAVFILTER_VERSION_INT AV_VERSION_INT(LIBAVFILTER_VERSION_MAJOR, \
diff --git a/libavfilter/vf_scale_npp.c b/libavfilter/vf_scale_npp.c
new file mode 100644
index 0000000000..247baf1133
--- /dev/null
+++ b/libavfilter/vf_scale_npp.c
@@ -0,0 +1,660 @@
+/*
+ * 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
+ */
+
+/**
+ * @file
+ * scale video filter
+ */
+
+#include <nppi.h>
+#include <stdio.h>
+#include <string.h>
+
+#include "libavutil/avstring.h"
+#include "libavutil/common.h"
+#include "libavutil/eval.h"
+#include "libavutil/hwcontext.h"
+#include "libavutil/hwcontext_cuda.h"
+#include "libavutil/internal.h"
+#include "libavutil/mathematics.h"
+#include "libavutil/opt.h"
+#include "libavutil/pixdesc.h"
+
+#include "avfilter.h"
+#include "formats.h"
+#include "internal.h"
+#include "video.h"
+
+static const enum AVPixelFormat supported_formats[] = {
+    AV_PIX_FMT_YUV420P,
+    AV_PIX_FMT_NV12,
+    AV_PIX_FMT_YUV444P,
+};
+
+static const enum AVPixelFormat deinterleaved_formats[][2] = {
+    { AV_PIX_FMT_NV12, AV_PIX_FMT_YUV420P },
+};
+
+static const char *const var_names[] = {
+    "PI",
+    "PHI",
+    "E",
+    "in_w",   "iw",
+    "in_h",   "ih",
+    "out_w",  "ow",
+    "out_h",  "oh",
+    "a", "dar",
+    "sar",
+    NULL
+};
+
+enum var_name {
+    VAR_PI,
+    VAR_PHI,
+    VAR_E,
+    VAR_IN_W,   VAR_IW,
+    VAR_IN_H,   VAR_IH,
+    VAR_OUT_W,  VAR_OW,
+    VAR_OUT_H,  VAR_OH,
+    VAR_A, VAR_DAR,
+    VAR_SAR,
+    VARS_NB
+};
+
+enum ScaleStage {
+    STAGE_DEINTERLEAVE,
+    STAGE_RESIZE,
+    STAGE_INTERLEAVE,
+    STAGE_NB,
+};
+
+typedef struct NPPScaleStageContext {
+    int stage_needed;
+    enum AVPixelFormat in_fmt;
+    enum AVPixelFormat out_fmt;
+
+    struct {
+        int width;
+        int height;
+    } planes_in[3], planes_out[3];
+
+    AVBufferRef *frames_ctx;
+    AVFrame     *frame;
+} NPPScaleStageContext;
+
+typedef struct NPPScaleContext {
+    const AVClass *class;
+
+    NPPScaleStageContext stages[STAGE_NB];
+    AVFrame *tmp_frame;
+    int passthrough;
+
+    int shift_width, shift_height;
+
+    /**
+     * New dimensions. Special values are:
+     *   0 = original width/height
+     *  -1 = keep original aspect
+     */
+    int w, h;
+
+    /**
+     * 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
+    char *format_str;
+
+    int interp_algo;
+} NPPScaleContext;
+
+static int nppscale_init(AVFilterContext *ctx)
+{
+    NPPScaleContext *s = ctx->priv;
+    int i;
+
+    if (!strcmp(s->format_str, "same")) {
+        s->format = AV_PIX_FMT_NONE;
+    } else {
+        s->format = av_get_pix_fmt(s->format_str);
+        if (s->format == AV_PIX_FMT_NONE) {
+            av_log(ctx, AV_LOG_ERROR, "Unrecognized pixel format: %s\n", s->format_str);
+            return AVERROR(EINVAL);
+        }
+    }
+
+    for (i = 0; i < FF_ARRAY_ELEMS(s->stages); i++) {
+        s->stages[i].frame = av_frame_alloc();
+        if (!s->stages[i].frame)
+            return AVERROR(ENOMEM);
+    }
+    s->tmp_frame = av_frame_alloc();
+    if (!s->tmp_frame)
+        return AVERROR(ENOMEM);
+
+    return 0;
+}
+
+static void nppscale_uninit(AVFilterContext *ctx)
+{
+    NPPScaleContext *s = ctx->priv;
+    int i;
+
+    for (i = 0; i < FF_ARRAY_ELEMS(s->stages); i++) {
+        av_frame_free(&s->stages[i].frame);
+        av_buffer_unref(&s->stages[i].frames_ctx);
+    }
+    av_frame_free(&s->tmp_frame);
+}
+
+static int nppscale_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);
+
+    ff_set_common_formats(ctx, pix_fmts);
+
+    return 0;
+}
+
+static int init_stage(NPPScaleStageContext *stage, 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(stage->in_fmt,  &in_sw,  &in_sh);
+    av_pix_fmt_get_chroma_sub_sample(stage->out_fmt, &out_sw, &out_sh);
+    if (!stage->planes_out[0].width) {
+        stage->planes_out[0].width  = stage->planes_in[0].width;
+        stage->planes_out[0].height = stage->planes_in[0].height;
+    }
+
+    for (i = 1; i < FF_ARRAY_ELEMS(stage->planes_in); i++) {
+        stage->planes_in[i].width   = stage->planes_in[0].width   >> in_sw;
+        stage->planes_in[i].height  = stage->planes_in[0].height  >> in_sh;
+        stage->planes_out[i].width  = stage->planes_out[0].width  >> out_sw;
+        stage->planes_out[i].height = stage->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 = stage->out_fmt;
+    out_ctx->width     = FFALIGN(stage->planes_out[0].width,  32);
+    out_ctx->height    = FFALIGN(stage->planes_out[0].height, 32);
+
+    ret = av_hwframe_ctx_init(out_ref);
+    if (ret < 0)
+        goto fail;
+
+    av_frame_unref(stage->frame);
+    ret = av_hwframe_get_buffer(out_ref, stage->frame, 0);
+    if (ret < 0)
+        goto fail;
+
+    stage->frame->width  = stage->planes_out[0].width;
+    stage->frame->height = stage->planes_out[0].height;
+
+    av_buffer_unref(&stage->frames_ctx);
+    stage->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 enum AVPixelFormat get_deinterleaved_format(enum AVPixelFormat fmt)
+{
+    const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(fmt);
+    int i, planes;
+
+    planes = av_pix_fmt_count_planes(fmt);
+    if (planes == desc->nb_components)
+        return fmt;
+    for (i = 0; i < FF_ARRAY_ELEMS(deinterleaved_formats); i++)
+        if (deinterleaved_formats[i][0] == fmt)
+            return deinterleaved_formats[i][1];
+    return AV_PIX_FMT_NONE;
+}
+
+static int init_processing_chain(AVFilterContext *ctx, int in_width, int in_height,
+                                 int out_width, int out_height)
+{
+    NPPScaleContext *s = ctx->priv;
+
+    AVHWFramesContext *in_frames_ctx;
+
+    enum AVPixelFormat in_format;
+    enum AVPixelFormat out_format;
+    enum AVPixelFormat in_deinterleaved_format;
+    enum AVPixelFormat out_deinterleaved_format;
+
+    int i, ret, last_stage = -1;
+
+    /* 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);
+    }
+
+    in_deinterleaved_format  = get_deinterleaved_format(in_format);
+    out_deinterleaved_format = get_deinterleaved_format(out_format);
+    if (in_deinterleaved_format  == AV_PIX_FMT_NONE ||
+        out_deinterleaved_format == AV_PIX_FMT_NONE)
+        return AVERROR_BUG;
+
+    /* figure out which stages need to be done */
+    if (in_width != out_width || in_height != out_height ||
+        in_deinterleaved_format != out_deinterleaved_format)
+        s->stages[STAGE_RESIZE].stage_needed = 1;
+
+    if (!s->stages[STAGE_RESIZE].stage_needed && in_format == out_format)
+        s->passthrough = 1;
+
+    if (!s->passthrough) {
+        if (in_format != in_deinterleaved_format)
+            s->stages[STAGE_DEINTERLEAVE].stage_needed = 1;
+        if (out_format != out_deinterleaved_format)
+            s->stages[STAGE_INTERLEAVE].stage_needed = 1;
+    }
+
+    s->stages[STAGE_DEINTERLEAVE].in_fmt              = in_format;
+    s->stages[STAGE_DEINTERLEAVE].out_fmt             = in_deinterleaved_format;
+    s->stages[STAGE_DEINTERLEAVE].planes_in[0].width  = in_width;
+    s->stages[STAGE_DEINTERLEAVE].planes_in[0].height = in_height;
+
+    s->stages[STAGE_RESIZE].in_fmt               = in_deinterleaved_format;
+    s->stages[STAGE_RESIZE].out_fmt              = out_deinterleaved_format;
+    s->stages[STAGE_RESIZE].planes_in[0].width   = in_width;
+    s->stages[STAGE_RESIZE].planes_in[0].height  = in_height;
+    s->stages[STAGE_RESIZE].planes_out[0].width  = out_width;
+    s->stages[STAGE_RESIZE].planes_out[0].height = out_height;
+
+    s->stages[STAGE_INTERLEAVE].in_fmt              = out_deinterleaved_format;
+    s->stages[STAGE_INTERLEAVE].out_fmt             = out_format;
+    s->stages[STAGE_INTERLEAVE].planes_in[0].width  = out_width;
+    s->stages[STAGE_INTERLEAVE].planes_in[0].height = out_height;
+
+    /* init the hardware contexts */
+    for (i = 0; i < FF_ARRAY_ELEMS(s->stages); i++) {
+        if (!s->stages[i].stage_needed)
+            continue;
+
+        ret = init_stage(&s->stages[i], in_frames_ctx->device_ref);
+        if (ret < 0)
+            return ret;
+
+        last_stage = i;
+    }
+
+    if (last_stage < 0)
+        return AVERROR_BUG;
+    ctx->outputs[0]->hw_frames_ctx = av_buffer_ref(s->stages[last_stage].frames_ctx);
+    if (!ctx->outputs[0]->hw_frames_ctx)
+        return AVERROR(ENOMEM);
+
+    return 0;
+}
+
+static int nppscale_config_props(AVFilterLink *outlink)
+{
+    AVFilterContext *ctx = outlink->src;
+    AVFilterLink *inlink = outlink->src->inputs[0];
+    NPPScaleContext  *s = ctx->priv;
+    int64_t w, h;
+    double var_values[VARS_NB], res;
+    char *expr;
+    int ret;
+
+    var_values[VAR_PI]    = M_PI;
+    var_values[VAR_PHI]   = M_PHI;
+    var_values[VAR_E]     = M_E;
+    var_values[VAR_IN_W]  = var_values[VAR_IW] = inlink->w;
+    var_values[VAR_IN_H]  = var_values[VAR_IH] = inlink->h;
+    var_values[VAR_OUT_W] = var_values[VAR_OW] = NAN;
+    var_values[VAR_OUT_H] = var_values[VAR_OH] = NAN;
+    var_values[VAR_A]     = (double) inlink->w / inlink->h;
+    var_values[VAR_SAR]   = inlink->sample_aspect_ratio.num ?
+        (double) inlink->sample_aspect_ratio.num / inlink->sample_aspect_ratio.den : 1;
+    var_values[VAR_DAR]   = var_values[VAR_A] * var_values[VAR_SAR];
+
+    /* evaluate width and height */
+    av_expr_parse_and_eval(&res, (expr = s->w_expr),
+                           var_names, var_values,
+                           NULL, NULL, NULL, NULL, NULL, 0, ctx);
+    s->w = var_values[VAR_OUT_W] = var_values[VAR_OW] = res;
+    if ((ret = av_expr_parse_and_eval(&res, (expr = s->h_expr),
+                                      var_names, var_values,
+                                      NULL, NULL, NULL, NULL, NULL, 0, ctx)) < 0)
+        goto fail;
+    s->h = var_values[VAR_OUT_H] = var_values[VAR_OH] = res;
+    /* evaluate again the width, as it may depend on the output height */
+    if ((ret = av_expr_parse_and_eval(&res, (expr = s->w_expr),
+                                      var_names, var_values,
+                                      NULL, NULL, NULL, NULL, NULL, 0, ctx)) < 0)
+        goto fail;
+    s->w = res;
+
+    w = s->w;
+    h = s->h;
+
+    /* sanity check params */
+    if (w <  -1 || h <  -1) {
+        av_log(ctx, AV_LOG_ERROR, "Size values less than -1 are not acceptable.\n");
+        return AVERROR(EINVAL);
+    }
+    if (w == -1 && h == -1)
+        s->w = s->h = 0;
+
+    if (!(w = s->w))
+        w = inlink->w;
+    if (!(h = s->h))
+        h = inlink->h;
+    if (w == -1)
+        w = av_rescale(h, inlink->w, inlink->h);
+    if (h == -1)
+        h = av_rescale(w, inlink->h, inlink->w);
+
+    if (w > INT_MAX || h > INT_MAX ||
+        (h * inlink->w) > INT_MAX  ||
+        (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:
+    av_log(NULL, AV_LOG_ERROR,
+           "Error when evaluating the expression '%s'\n", expr);
+    return ret;
+}
+
+static int nppscale_deinterleave(AVFilterContext *ctx, NPPScaleStageContext *stage,
+                                 AVFrame *out, AVFrame *in)
+{
+    AVHWFramesContext *in_frames_ctx = (AVHWFramesContext*)in->hw_frames_ctx->data;
+    NppStatus err;
+
+    switch (in_frames_ctx->sw_format) {
+    case AV_PIX_FMT_NV12:
+        err = nppiYCbCr420_8u_P2P3R(in->data[0], in->linesize[0],
+                                    in->data[1], in->linesize[1],
+                                    out->data, out->linesize,
+                                    (NppiSize){ in->width, in->height });
+        break;
+    default:
+        return AVERROR_BUG;
+    }
+    if (err != NPP_SUCCESS) {
+        av_log(ctx, AV_LOG_ERROR, "NPP deinterleave error: %d\n", err);
+        return AVERROR_UNKNOWN;
+    }
+
+    return 0;
+}
+
+static int nppscale_resize(AVFilterContext *ctx, NPPScaleStageContext *stage,
+                           AVFrame *out, AVFrame *in)
+{
+    NPPScaleContext *s = ctx->priv;
+    NppStatus err;
+    int i;
+
+    for (i = 0; i < FF_ARRAY_ELEMS(in->data) && in->data[i]; i++) {
+        int iw = stage->planes_in[i].width;
+        int ih = stage->planes_in[i].height;
+        int ow = stage->planes_out[i].width;
+        int oh = stage->planes_out[i].height;
+
+        err = nppiResizeSqrPixel_8u_C1R(in->data[i], (NppiSize){ iw, ih },
+                                        in->linesize[i], (NppiRect){ 0, 0, iw, ih },
+                                        out->data[i], out->linesize[i],
+                                        (NppiRect){ 0, 0, ow, oh },
+                                        (double)ow / iw, (double)oh / ih,
+                                        0.0, 0.0, s->interp_algo);
+        if (err != NPP_SUCCESS) {
+            av_log(ctx, AV_LOG_ERROR, "NPP resize error: %d\n", err);
+            return AVERROR_UNKNOWN;
+        }
+    }
+
+    return 0;
+}
+
+static int nppscale_interleave(AVFilterContext *ctx, NPPScaleStageContext *stage,
+                               AVFrame *out, AVFrame *in)
+{
+    AVHWFramesContext *out_frames_ctx = (AVHWFramesContext*)out->hw_frames_ctx->data;
+    NppStatus err;
+
+    switch (out_frames_ctx->sw_format) {
+    case AV_PIX_FMT_NV12:
+        err = nppiYCbCr420_8u_P3P2R((const uint8_t**)in->data,
+                                    in->linesize,
+                                    out->data[0], out->linesize[0],
+                                    out->data[1], out->linesize[1],
+                                    (NppiSize){ in->width, in->height });
+        break;
+    default:
+        return AVERROR_BUG;
+    }
+    if (err != NPP_SUCCESS) {
+        av_log(ctx, AV_LOG_ERROR, "NPP deinterleave error: %d\n", err);
+        return AVERROR_UNKNOWN;
+    }
+
+    return 0;
+}
+
+static int (*const nppscale_process[])(AVFilterContext *ctx, NPPScaleStageContext *stage,
+                                       AVFrame *out, AVFrame *in) = {
+    [STAGE_DEINTERLEAVE] = nppscale_deinterleave,
+    [STAGE_RESIZE]       = nppscale_resize,
+    [STAGE_INTERLEAVE]   = nppscale_interleave,
+};
+
+static int nppscale_scale(AVFilterContext *ctx, AVFrame *out, AVFrame *in)
+{
+    NPPScaleContext *s = ctx->priv;
+    AVFrame *src = in;
+    int i, ret, last_stage = -1;
+
+    for (i = 0; i < FF_ARRAY_ELEMS(s->stages); i++) {
+        if (!s->stages[i].stage_needed)
+            continue;
+
+        ret = nppscale_process[i](ctx, &s->stages[i], s->stages[i].frame, src);
+        if (ret < 0)
+            return ret;
+
+        src        = s->stages[i].frame;
+        last_stage = i;
+    }
+
+    if (last_stage < 0)
+        return AVERROR_BUG;
+    ret = av_hwframe_get_buffer(src->hw_frames_ctx, s->tmp_frame, 0);
+    if (ret < 0)
+        return ret;
+
+    av_frame_move_ref(out, src);
+    av_frame_move_ref(src, s->tmp_frame);
+
+    ret = av_frame_copy_props(out, in);
+    if (ret < 0)
+        return ret;
+
+    return 0;
+}
+
+static int nppscale_filter_frame(AVFilterLink *link, AVFrame *in)
+{
+    AVFilterContext              *ctx = link->dst;
+    NPPScaleContext                *s = ctx->priv;
+    AVFilterLink             *outlink = ctx->outputs[0];
+    AVHWFramesContext     *frames_ctx = (AVHWFramesContext*)outlink->hw_frames_ctx->data;
+    AVCUDADeviceContext *device_hwctx = frames_ctx->device_ctx->hwctx;
+
+    AVFrame *out = NULL;
+    CUresult err;
+    CUcontext dummy;
+    int ret = 0;
+
+    if (s->passthrough)
+        return ff_filter_frame(outlink, in);
+
+    out = av_frame_alloc();
+    if (!out) {
+        ret = AVERROR(ENOMEM);
+        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);
+
+    err = cuCtxPushCurrent(device_hwctx->cuda_ctx);
+    if (err != CUDA_SUCCESS) {
+        ret = AVERROR_UNKNOWN;
+        goto fail;
+    }
+
+    ret = nppscale_scale(ctx, out, in);
+
+    cuCtxPopCurrent(&dummy);
+    if (ret < 0)
+        goto fail;
+
+    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(NPPScaleContext, x)
+#define FLAGS 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 },
+    { "format", "Output pixel format", OFFSET(format_str), AV_OPT_TYPE_STRING, { .str = "same" }, .flags = FLAGS },
+
+    { "interp_algo", "Interpolation algorithm used for resizing", OFFSET(interp_algo), AV_OPT_TYPE_INT, { .i64 = NPPI_INTER_CUBIC }, 0, INT_MAX, FLAGS, "interp_algo" },
+        { "nn",                 "nearest neighbour",                 0, AV_OPT_TYPE_CONST, { .i64 = NPPI_INTER_NN                 }, 0, 0, FLAGS, "interp_algo" },
+        { "linear",             "linear",                            0, AV_OPT_TYPE_CONST, { .i64 = NPPI_INTER_LINEAR             }, 0, 0, FLAGS, "interp_algo" },
+        { "cubic",              "cubic",                             0, AV_OPT_TYPE_CONST, { .i64 = NPPI_INTER_CUBIC              }, 0, 0, FLAGS, "interp_algo" },
+        { "cubic2p_bspline",    "2-parameter cubic (B=1, C=0)",      0, AV_OPT_TYPE_CONST, { .i64 = NPPI_INTER_CUBIC2P_BSPLINE    }, 0, 0, FLAGS, "interp_algo" },
+        { "cubic2p_catmullrom", "2-parameter cubic (B=0, C=1/2)",    0, AV_OPT_TYPE_CONST, { .i64 = NPPI_INTER_CUBIC2P_CATMULLROM }, 0, 0, FLAGS, "interp_algo" },
+        { "cubic2p_b05c03",     "2-parameter cubic (B=1/2, C=3/10)", 0, AV_OPT_TYPE_CONST, { .i64 = NPPI_INTER_CUBIC2P_B05C03     }, 0, 0, FLAGS, "interp_algo" },
+        { "super",              "supersampling",                     0, AV_OPT_TYPE_CONST, { .i64 = NPPI_INTER_SUPER              }, 0, 0, FLAGS, "interp_algo" },
+        { "lanczos",            "Lanczos",                           0, AV_OPT_TYPE_CONST, { .i64 = NPPI_INTER_LANCZOS            }, 0, 0, FLAGS, "interp_algo" },
+    { NULL },
+};
+
+static const AVClass nppscale_class = {
+    .class_name = "nppscale",
+    .item_name  = av_default_item_name,
+    .option     = options,
+    .version    = LIBAVUTIL_VERSION_INT,
+};
+
+static const AVFilterPad nppscale_inputs[] = {
+    {
+        .name        = "default",
+        .type        = AVMEDIA_TYPE_VIDEO,
+        .filter_frame = nppscale_filter_frame,
+    },
+    { NULL }
+};
+
+static const AVFilterPad nppscale_outputs[] = {
+    {
+        .name         = "default",
+        .type         = AVMEDIA_TYPE_VIDEO,
+        .config_props = nppscale_config_props,
+    },
+    { NULL }
+};
+
+AVFilter ff_vf_scale_npp = {
+    .name      = "scale_npp",
+    .description = NULL_IF_CONFIG_SMALL("NVIDIA Performance Primitives video "
+                                        "scaling and format conversion"),
+
+    .init          = nppscale_init,
+    .uninit        = nppscale_uninit,
+    .query_formats = nppscale_query_formats,
+
+    .priv_size = sizeof(NPPScaleContext),
+    .priv_class = &nppscale_class,
+
+    .inputs    = nppscale_inputs,
+    .outputs   = nppscale_outputs,
+};