avfilter/scale_cuda: combine separate CUDA sources

Timo Rothenpieler committed 5 years ago

Commit b0e2e931 parent 91a41a34 files · +138 −240Browse files

Showing the changes to libavfilter (4 of 4 files). Show all

Unified · Split

4 files

libavfilter/Makefile +1 −2
libavfilter/vf_scale_cuda.c +2 −13
libavfilter/vf_scale_cuda.cu +135 −1
libavfilter/vf_scale_cuda_bicubic.cu +0 −224

@@ -395,8 +395,7 @@ OBJS-$(CONFIG_ROTATE_FILTER) += vf_rotate.o · 394 unmodified lines
395	395	OBJS-$(CONFIG_SAB_FILTER) += vf_sab.o
396	396	OBJS-$(CONFIG_SCALE_FILTER) += vf_scale.o scale_eval.o
397	397	OBJS-$(CONFIG_SCALE_CUDA_FILTER) += vf_scale_cuda.o scale_eval.o \
398		− vf_scale_cuda.ptx.o vf_scale_cuda_bicubic.ptx.o \
399		− cuda/load_helper.o
	398	+ vf_scale_cuda.ptx.o cuda/load_helper.o
400	399	OBJS-$(CONFIG_SCALE_NPP_FILTER) += vf_scale_npp.o scale_eval.o
401	400	OBJS-$(CONFIG_SCALE_QSV_FILTER) += vf_scale_qsv.o
402	401	OBJS-$(CONFIG_SCALE_VAAPI_FILTER) += vf_scale_vaapi.o scale_eval.o vaapi_vpp.o

libavfilter/vf_scale_cuda.c+2−13

@@ -276,41 +276,29 @@ static av_cold int cudascale_config_props(AVFilterLink *outlink) · 275 unmodified lines
276	276	int w, h;
277	277	int ret;
278	278
279		− const unsigned char *scaler_ptx;
280		− unsigned int scaler_ptx_len;
281	279	const char *function_infix = "";
282	280
283	281	extern const unsigned char ff_vf_scale_cuda_ptx_data[];
284	282	extern const unsigned int ff_vf_scale_cuda_ptx_len;
285		− extern const unsigned char ff_vf_scale_cuda_bicubic_ptx_data[];
286		− extern const unsigned int ff_vf_scale_cuda_bicubic_ptx_len;
287	283
288	284	switch(s->interp_algo) {
289	285	case INTERP_ALGO_NEAREST:
290		− scaler_ptx = ff_vf_scale_cuda_ptx_data;
291		− scaler_ptx_len = ff_vf_scale_cuda_ptx_len;
292	286	function_infix = "_Nearest";
293	287	s->interp_use_linear = 0;
294	288	s->interp_as_integer = 1;
295	289	break;
296	290	case INTERP_ALGO_BILINEAR:
297		− scaler_ptx = ff_vf_scale_cuda_ptx_data;
298		− scaler_ptx_len = ff_vf_scale_cuda_ptx_len;
299	291	function_infix = "_Bilinear";
300	292	s->interp_use_linear = 1;
301	293	s->interp_as_integer = 1;
302	294	break;
303	295	case INTERP_ALGO_DEFAULT:
304	296	case INTERP_ALGO_BICUBIC:
305		− scaler_ptx = ff_vf_scale_cuda_bicubic_ptx_data;
306		− scaler_ptx_len = ff_vf_scale_cuda_bicubic_ptx_len;
307	297	function_infix = "_Bicubic";
308	298	s->interp_use_linear = 0;
309	299	s->interp_as_integer = 0;
310	300	break;
311	301	case INTERP_ALGO_LANCZOS:
312		− scaler_ptx = ff_vf_scale_cuda_bicubic_ptx_data;
313		− scaler_ptx_len = ff_vf_scale_cuda_bicubic_ptx_len;
314	302	function_infix = "_Lanczos";
315	303	s->interp_use_linear = 0;
316	304	s->interp_as_integer = 0;
@@ -327,7 +315,8 @@ static av_cold int cudascale_config_props(AVFilterLink *outlink) · 10 unmodified lines
327	315	if (ret < 0)
328	316	goto fail;
329	317
330		− ret = ff_cuda_load_module(ctx, device_hwctx, &s->cu_module, scaler_ptx, scaler_ptx_len);
	318	+ ret = ff_cuda_load_module(ctx, device_hwctx, &s->cu_module,
	319	+ ff_vf_scale_cuda_ptx_data, ff_vf_scale_cuda_ptx_len);
331	320	if (ret < 0)
332	321	goto fail;
333	322

libavfilter/vf_scale_cuda.cu+135−1

@@ -1,5 +1,5 @@
1	1	/*
2		− * Copyright (c) 2017, NVIDIA CORPORATION. All rights reserved.
	2	+ * This file is part of FFmpeg.
3	3	*
4	4	* Permission is hereby granted, free of charge, to any person obtaining a
5	5	* copy of this software and associated documentation files (the "Software"),
@@ -21,6 +21,55 @@ · 15 unmodified lines
21	21	*/
22	22
23	23	#include "cuda/vector_helpers.cuh"
	24	+#include "vf_scale_cuda.h"
	25	+
	26	+typedef float4 (*coeffs_function_t)(float, float);
	27	+
	28	+__device__ inline float4 lanczos_coeffs(float x, float param)
	29	+{
	30	+ const float pi = 3.141592654f;
	31	+
	32	+ float4 res = make_float4(
	33	+ pi * (x + 1),
	34	+ pi * x,
	35	+ pi * (x - 1),
	36	+ pi * (x - 2));
	37	+
	38	+ res.x = res.x == 0.0f ? 1.0f :
	39	+ __sinf(res.x) * __sinf(res.x / 2.0f) / (res.x * res.x / 2.0f);
	40	+ res.y = res.y == 0.0f ? 1.0f :
	41	+ __sinf(res.y) * __sinf(res.y / 2.0f) / (res.y * res.y / 2.0f);
	42	+ res.z = res.z == 0.0f ? 1.0f :
	43	+ __sinf(res.z) * __sinf(res.z / 2.0f) / (res.z * res.z / 2.0f);
	44	+ res.w = res.w == 0.0f ? 1.0f :
	45	+ __sinf(res.w) * __sinf(res.w / 2.0f) / (res.w * res.w / 2.0f);
	46	+
	47	+ return res / (res.x + res.y + res.z + res.w);
	48	+}
	49	+
	50	+__device__ inline float4 bicubic_coeffs(float x, float param)
	51	+{
	52	+ const float A = param == SCALE_CUDA_PARAM_DEFAULT ? 0.0f : -param;
	53	+
	54	+ float4 res;
	55	+ res.x = ((A * (x + 1) - 5 * A) * (x + 1) + 8 * A) * (x + 1) - 4 * A;
	56	+ res.y = ((A + 2) * x - (A + 3)) * x * x + 1;
	57	+ res.z = ((A + 2) * (1 - x) - (A + 3)) * (1 - x) * (1 - x) + 1;
	58	+ res.w = 1.0f - res.x - res.y - res.z;
	59	+
	60	+ return res;
	61	+}
	62	+
	63	+template<typename V>
	64	+__device__ inline V apply_coeffs(float4 coeffs, V c0, V c1, V c2, V c3)
	65	+{
	66	+ V res = c0 * coeffs.x;
	67	+ res += c1 * coeffs.y;
	68	+ res += c2 * coeffs.z;
	69	+ res += c3 * coeffs.w;
	70	+
	71	+ return res;
	72	+}
24	73
25	74	template<typename T>
26	75	__device__ inline void Subsample_Nearest(cudaTextureObject_t tex,
@@ -76,6 +125,48 @@ __device__ inline void Subsample_Bilinear(cudaTextureObject_t tex, · 49 unmodified lines
76	125	}
77	126	}
78	127
	128	+template<typename T>
	129	+__device__ inline void Subsample_Bicubic(coeffs_function_t coeffs_function,
	130	+ cudaTextureObject_t tex,
	131	+ T *dst,
	132	+ int dst_width, int dst_height, int dst_pitch,
	133	+ int src_width, int src_height,
	134	+ int bit_depth, float param)
	135	+{
	136	+ int xo = blockIdx.x * blockDim.x + threadIdx.x;
	137	+ int yo = blockIdx.y * blockDim.y + threadIdx.y;
	138	+
	139	+ if (yo < dst_height && xo < dst_width)
	140	+ {
	141	+ float hscale = (float)src_width / (float)dst_width;
	142	+ float vscale = (float)src_height / (float)dst_height;
	143	+ float xi = (xo + 0.5f) * hscale - 0.5f;
	144	+ float yi = (yo + 0.5f) * vscale - 0.5f;
	145	+ float px = floor(xi);
	146	+ float py = floor(yi);
	147	+ float fx = xi - px;
	148	+ float fy = yi - py;
	149	+
	150	+ float factor = bit_depth > 8 ? 0xFFFF : 0xFF;
	151	+
	152	+ float4 coeffsX = coeffs_function(fx, param);
	153	+ float4 coeffsY = coeffs_function(fy, param);
	154	+
	155	+#define PIX(x, y) tex2D<floatT>(tex, (x), (y))
	156	+
	157	+ dst[yo * dst_pitch + xo] = from_floatN<T, floatT>(
	158	+ apply_coeffs<floatT>(coeffsY,
	159	+ apply_coeffs<floatT>(coeffsX, PIX(px - 1, py - 1), PIX(px, py - 1), PIX(px + 1, py - 1), PIX(px + 2, py - 1)),
	160	+ apply_coeffs<floatT>(coeffsX, PIX(px - 1, py ), PIX(px, py ), PIX(px + 1, py ), PIX(px + 2, py )),
	161	+ apply_coeffs<floatT>(coeffsX, PIX(px - 1, py + 1), PIX(px, py + 1), PIX(px + 1, py + 1), PIX(px + 2, py + 1)),
	162	+ apply_coeffs<floatT>(coeffsX, PIX(px - 1, py + 2), PIX(px, py + 2), PIX(px + 1, py + 2), PIX(px + 2, py + 2))
	163	+ ) * factor
	164	+ );
	165	+
	166	+#undef PIX
	167	+ }
	168	+}
	169	+
79	170	extern "C" {
80	171
81	172	#define NEAREST_KERNEL(T) \
@@ -120,4 +211,47 @@ BILINEAR_KERNEL(ushort) · 38 unmodified lines
120	211	BILINEAR_KERNEL(ushort2)
121	212	BILINEAR_KERNEL(ushort4)
122	213
	214	+#define BICUBIC_KERNEL(T) \
	215	+ __global__ void Subsample_Bicubic_ ## T(cudaTextureObject_t src_tex, \
	216	+ T *dst, \
	217	+ int dst_width, int dst_height, int dst_pitch, \
	218	+ int src_width, int src_height, \
	219	+ int bit_depth, float param) \
	220	+ { \
	221	+ Subsample_Bicubic<T>(&bicubic_coeffs, src_tex, dst, \
	222	+ dst_width, dst_height, dst_pitch, \
	223	+ src_width, src_height, \
	224	+ bit_depth, param); \
	225	+ }
	226	+
	227	+BICUBIC_KERNEL(uchar)
	228	+BICUBIC_KERNEL(uchar2)
	229	+BICUBIC_KERNEL(uchar4)
	230	+
	231	+BICUBIC_KERNEL(ushort)
	232	+BICUBIC_KERNEL(ushort2)
	233	+BICUBIC_KERNEL(ushort4)
	234	+
	235	+
	236	+#define LANCZOS_KERNEL(T) \
	237	+ __global__ void Subsample_Lanczos_ ## T(cudaTextureObject_t src_tex, \
	238	+ T *dst, \
	239	+ int dst_width, int dst_height, int dst_pitch, \
	240	+ int src_width, int src_height, \
	241	+ int bit_depth, float param) \
	242	+ { \
	243	+ Subsample_Bicubic<T>(&lanczos_coeffs, src_tex, dst, \
	244	+ dst_width, dst_height, dst_pitch, \
	245	+ src_width, src_height, \
	246	+ bit_depth, param); \
	247	+ }
	248	+
	249	+LANCZOS_KERNEL(uchar)
	250	+LANCZOS_KERNEL(uchar2)
	251	+LANCZOS_KERNEL(uchar4)
	252	+
	253	+LANCZOS_KERNEL(ushort)
	254	+LANCZOS_KERNEL(ushort2)
	255	+LANCZOS_KERNEL(ushort4)
	256	+
123	257	}

libavfilter/vf_scale_cuda_bicubic.cu+0−224

@@ -1,224 +0,0 @@
1		−/*
2		− * This file is part of FFmpeg.
3		− *
4		− * Permission is hereby granted, free of charge, to any person obtaining a
5		− * copy of this software and associated documentation files (the "Software"),
6		− * to deal in the Software without restriction, including without limitation
7		− * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8		− * and/or sell copies of the Software, and to permit persons to whom the
9		− * Software is furnished to do so, subject to the following conditions:
10		− *
11		− * The above copyright notice and this permission notice shall be included in
12		− * all copies or substantial portions of the Software.
13		− *
14		− * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15		− * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16		− * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17		− * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
18		− * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
19		− * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
20		− * DEALINGS IN THE SOFTWARE.
21		− */
22		−
23		−#include "cuda/vector_helpers.cuh"
24		−#include "vf_scale_cuda.h"
25		−
26		−typedef float4 (*coeffs_function_t)(float, float);
27		−
28		−__device__ inline float4 lanczos_coeffs(float x, float param)
29		−{
30		− const float pi = 3.141592654f;
31		−
32		− float4 res = make_float4(
33		− pi * (x + 1),
34		− pi * x,
35		− pi * (x - 1),
36		− pi * (x - 2));
37		−
38		− res.x = res.x == 0.0f ? 1.0f :
39		− __sinf(res.x) * __sinf(res.x / 2.0f) / (res.x * res.x / 2.0f);
40		− res.y = res.y == 0.0f ? 1.0f :
41		− __sinf(res.y) * __sinf(res.y / 2.0f) / (res.y * res.y / 2.0f);
42		− res.z = res.z == 0.0f ? 1.0f :
43		− __sinf(res.z) * __sinf(res.z / 2.0f) / (res.z * res.z / 2.0f);
44		− res.w = res.w == 0.0f ? 1.0f :
45		− __sinf(res.w) * __sinf(res.w / 2.0f) / (res.w * res.w / 2.0f);
46		−
47		− return res / (res.x + res.y + res.z + res.w);
48		−}
49		−
50		−__device__ inline float4 bicubic_coeffs(float x, float param)
51		−{
52		− const float A = param == SCALE_CUDA_PARAM_DEFAULT ? 0.0f : -param;
53		−
54		− float4 res;
55		− res.x = ((A * (x + 1) - 5 * A) * (x + 1) + 8 * A) * (x + 1) - 4 * A;
56		− res.y = ((A + 2) * x - (A + 3)) * x * x + 1;
57		− res.z = ((A + 2) * (1 - x) - (A + 3)) * (1 - x) * (1 - x) + 1;
58		− res.w = 1.0f - res.x - res.y - res.z;
59		−
60		− return res;
61		−}
62		−
63		−__device__ inline void derived_fast_coeffs(float4 coeffs, float x, float h0, float h1, float *s)
64		−{
65		− float g0 = coeffs.x + coeffs.y;
66		− float g1 = coeffs.z + coeffs.w;
67		−
68		− *h0 = coeffs.y / g0 - 0.5f;
69		− *h1 = coeffs.w / g1 + 1.5f;
70		− *s = g0 / (g0 + g1);
71		−}
72		−
73		−template<typename V>
74		−__device__ inline V apply_coeffs(float4 coeffs, V c0, V c1, V c2, V c3)
75		−{
76		− V res = c0 * coeffs.x;
77		− res += c1 * coeffs.y;
78		− res += c2 * coeffs.z;
79		− res += c3 * coeffs.w;
80		−
81		− return res;
82		−}
83		−
84		−template<typename T>
85		−__device__ inline void Subsample_Bicubic(coeffs_function_t coeffs_function,
86		− cudaTextureObject_t src_tex,
87		− T *dst,
88		− int dst_width, int dst_height, int dst_pitch,
89		− int src_width, int src_height,
90		− int bit_depth, float param)
91		−{
92		− int xo = blockIdx.x * blockDim.x + threadIdx.x;
93		− int yo = blockIdx.y * blockDim.y + threadIdx.y;
94		−
95		− if (yo < dst_height && xo < dst_width)
96		− {
97		− float hscale = (float)src_width / (float)dst_width;
98		− float vscale = (float)src_height / (float)dst_height;
99		− float xi = (xo + 0.5f) * hscale - 0.5f;
100		− float yi = (yo + 0.5f) * vscale - 0.5f;
101		− float px = floor(xi);
102		− float py = floor(yi);
103		− float fx = xi - px;
104		− float fy = yi - py;
105		−
106		− float factor = bit_depth > 8 ? 0xFFFF : 0xFF;
107		−
108		− float4 coeffsX = coeffs_function(fx, param);
109		− float4 coeffsY = coeffs_function(fy, param);
110		−
111		−#define PIX(x, y) tex2D<floatT>(src_tex, (x), (y))
112		−
113		− dst[yo * dst_pitch + xo] = from_floatN<T, floatT>(
114		− apply_coeffs<floatT>(coeffsY,
115		− apply_coeffs<floatT>(coeffsX, PIX(px - 1, py - 1), PIX(px, py - 1), PIX(px + 1, py - 1), PIX(px + 2, py - 1)),
116		− apply_coeffs<floatT>(coeffsX, PIX(px - 1, py ), PIX(px, py ), PIX(px + 1, py ), PIX(px + 2, py )),
117		− apply_coeffs<floatT>(coeffsX, PIX(px - 1, py + 1), PIX(px, py + 1), PIX(px + 1, py + 1), PIX(px + 2, py + 1)),
118		− apply_coeffs<floatT>(coeffsX, PIX(px - 1, py + 2), PIX(px, py + 2), PIX(px + 1, py + 2), PIX(px + 2, py + 2))
119		− ) * factor
120		− );
121		−
122		−#undef PIX
123		− }
124		−}
125		−
126		−/* This does not yield correct results. Most likely because of low internal precision in tex2D linear interpolation */
127		−template<typename T>
128		−__device__ inline void Subsample_FastBicubic(coeffs_function_t coeffs_function,
129		− cudaTextureObject_t src_tex,
130		− T *dst,
131		− int dst_width, int dst_height, int dst_pitch,
132		− int src_width, int src_height,
133		− int bit_depth, float param)
134		−{
135		− int xo = blockIdx.x * blockDim.x + threadIdx.x;
136		− int yo = blockIdx.y * blockDim.y + threadIdx.y;
137		−
138		− if (yo < dst_height && xo < dst_width)
139		− {
140		− float hscale = (float)src_width / (float)dst_width;
141		− float vscale = (float)src_height / (float)dst_height;
142		− float xi = (xo + 0.5f) * hscale - 0.5f;
143		− float yi = (yo + 0.5f) * vscale - 0.5f;
144		− float px = floor(xi);
145		− float py = floor(yi);
146		− float fx = xi - px;
147		− float fy = yi - py;
148		−
149		− float factor = bit_depth > 8 ? 0xFFFF : 0xFF;
150		−
151		− float4 coeffsX = coeffs_function(fx, param);
152		− float4 coeffsY = coeffs_function(fy, param);
153		−
154		− float h0x, h1x, sx;
155		− float h0y, h1y, sy;
156		− derived_fast_coeffs(coeffsX, fx, &h0x, &h1x, &sx);
157		− derived_fast_coeffs(coeffsY, fy, &h0y, &h1y, &sy);
158		−
159		−#define PIX(x, y) tex2D<floatT>(src_tex, (x), (y))
160		−
161		− floatT pix[4] = {
162		− PIX(px + h0x, py + h0y),
163		− PIX(px + h1x, py + h0y),
164		− PIX(px + h0x, py + h1y),
165		− PIX(px + h1x, py + h1y)
166		− };
167		−
168		−#undef PIX
169		−
170		− dst[yo * dst_pitch + xo] = from_floatN<T, floatT>(
171		− lerp_scalar(
172		− lerp_scalar(pix[3], pix[2], sx),
173		− lerp_scalar(pix[1], pix[0], sx),
174		− sy) * factor
175		− );
176		− }
177		−}
178		−
179		−extern "C" {
180		−
181		−#define BICUBIC_KERNEL(T) \
182		− __global__ void Subsample_Bicubic_ ## T(cudaTextureObject_t src_tex, \
183		− T *dst, \
184		− int dst_width, int dst_height, int dst_pitch, \
185		− int src_width, int src_height, \
186		− int bit_depth, float param) \
187		− { \
188		− Subsample_Bicubic<T>(&bicubic_coeffs, src_tex, dst, \
189		− dst_width, dst_height, dst_pitch, \
190		− src_width, src_height, \
191		− bit_depth, param); \
192		− }
193		−
194		−BICUBIC_KERNEL(uchar)
195		−BICUBIC_KERNEL(uchar2)
196		−BICUBIC_KERNEL(uchar4)
197		−
198		−BICUBIC_KERNEL(ushort)
199		−BICUBIC_KERNEL(ushort2)
200		−BICUBIC_KERNEL(ushort4)
201		−
202		−
203		−#define LANCZOS_KERNEL(T) \
204		− __global__ void Subsample_Lanczos_ ## T(cudaTextureObject_t src_tex, \
205		− T *dst, \
206		− int dst_width, int dst_height, int dst_pitch, \
207		− int src_width, int src_height, \
208		− int bit_depth, float param) \
209		− { \
210		− Subsample_Bicubic<T>(&lanczos_coeffs, src_tex, dst, \
211		− dst_width, dst_height, dst_pitch, \
212		− src_width, src_height, \
213		− bit_depth, param); \
214		− }
215		−
216		−LANCZOS_KERNEL(uchar)
217		−LANCZOS_KERNEL(uchar2)
218		−LANCZOS_KERNEL(uchar4)
219		−
220		−LANCZOS_KERNEL(ushort)
221		−LANCZOS_KERNEL(ushort2)
222		−LANCZOS_KERNEL(ushort4)
223		−
224		−}

avfilter/scale_cuda: combine separate CUDA sources

Keyboard shortcuts

General

Issues and pull requests

Issues, pull requests and Inbox

Repository code