avfilter/scale_cuda: add CUDA scale filter

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

Changelog

@@ -10,6 +10,7 @@ version <next>:
 - config.log and other configuration files moved into ffbuild/ directory
 - update cuvid/nvenc headers to Video Codec SDK 8.0.14
 - afir audio filter
+- scale_cuda CUDA based video scale filter
 
 version 3.3:
 - CrystalHD decoder moved to new decode API

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

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);

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, \

libavfilter/vf_scale_cuda.cu

+/*
+ * 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);
+    }
+}
+
+}