1 /*
2 *
3 * This file is part of FFmpeg.
4 *
5 * FFmpeg is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation; either version 2 of the License, or
8 * (at your option) any later version.
9 *
10 * FFmpeg is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
14 *
15 * You should have received a copy of the GNU General Public License along
16 * with FFmpeg; if not, write to the Free Software Foundation, Inc.,
17 * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
18 */
19
20 #include <string.h>
21
22 #include "libavutil/common.h"
23 #include "libavutil/intreadwrite.h"
24 #include "libavutil/mem.h"
25 #include "libavutil/mem_internal.h"
26
27 #include "libswscale/swscale.h"
28 #include "libswscale/swscale_internal.h"
29
30 #include "checkasm.h"
31
32 #define randomize_buffers(buf, size) \
33 do { \
34 int j; \
35 for (j = 0; j < size; j+=4) \
36 AV_WN32(buf + j, rnd()); \
37 } while (0)
38
39 // This reference function is the same approximate algorithm employed by the
40 // SIMD functions
ref_function(const int16_t * filter,int filterSize,const int16_t ** src,uint8_t * dest,int dstW,const uint8_t * dither,int offset)41 static void ref_function(const int16_t *filter, int filterSize,
42 const int16_t **src, uint8_t *dest, int dstW,
43 const uint8_t *dither, int offset)
44 {
45 int i, d;
46 d = ((filterSize - 1) * 8 + dither[0]) >> 4;
47 for ( i = 0; i < dstW; i++) {
48 int16_t val = d;
49 int j;
50 union {
51 int val;
52 int16_t v[2];
53 } t;
54 for (j = 0; j < filterSize; j++){
55 t.val = (int)src[j][i + offset] * (int)filter[j];
56 val += t.v[1];
57 }
58 dest[i]= av_clip_uint8(val>>3);
59 }
60 }
61
check_yuv2yuvX(void)62 static void check_yuv2yuvX(void)
63 {
64 struct SwsContext *ctx;
65 int fsi, osi, isi, i, j;
66 int dstW;
67 #define LARGEST_FILTER 16
68 #define FILTER_SIZES 4
69 static const int filter_sizes[FILTER_SIZES] = {1, 4, 8, 16};
70 #define LARGEST_INPUT_SIZE 512
71 #define INPUT_SIZES 6
72 static const int input_sizes[INPUT_SIZES] = {8, 24, 128, 144, 256, 512};
73
74 declare_func_emms(AV_CPU_FLAG_MMX, void, const int16_t *filter,
75 int filterSize, const int16_t **src, uint8_t *dest,
76 int dstW, const uint8_t *dither, int offset);
77
78 const int16_t **src;
79 LOCAL_ALIGNED_8(int16_t, src_pixels, [LARGEST_FILTER * LARGEST_INPUT_SIZE]);
80 LOCAL_ALIGNED_8(int16_t, filter_coeff, [LARGEST_FILTER]);
81 LOCAL_ALIGNED_8(uint8_t, dst0, [LARGEST_INPUT_SIZE]);
82 LOCAL_ALIGNED_8(uint8_t, dst1, [LARGEST_INPUT_SIZE]);
83 LOCAL_ALIGNED_8(uint8_t, dither, [LARGEST_INPUT_SIZE]);
84 union VFilterData{
85 const int16_t *src;
86 uint16_t coeff[8];
87 } *vFilterData;
88 uint8_t d_val = rnd();
89 memset(dither, d_val, LARGEST_INPUT_SIZE);
90 randomize_buffers((uint8_t*)src_pixels, LARGEST_FILTER * LARGEST_INPUT_SIZE * sizeof(int16_t));
91 randomize_buffers((uint8_t*)filter_coeff, LARGEST_FILTER * sizeof(int16_t));
92 ctx = sws_alloc_context();
93 if (sws_init_context(ctx, NULL, NULL) < 0)
94 fail();
95
96 ff_getSwsFunc(ctx);
97 for(isi = 0; isi < INPUT_SIZES; ++isi){
98 dstW = input_sizes[isi];
99 for(osi = 0; osi < 64; osi += 16){
100 for(fsi = 0; fsi < FILTER_SIZES; ++fsi){
101 src = av_malloc(sizeof(int16_t*) * filter_sizes[fsi]);
102 vFilterData = av_malloc((filter_sizes[fsi] + 2) * sizeof(union VFilterData));
103 memset(vFilterData, 0, (filter_sizes[fsi] + 2) * sizeof(union VFilterData));
104 for(i = 0; i < filter_sizes[fsi]; ++i){
105 src[i] = &src_pixels[i * LARGEST_INPUT_SIZE];
106 vFilterData[i].src = src[i];
107 for(j = 0; j < 4; ++j)
108 vFilterData[i].coeff[j + 4] = filter_coeff[i];
109 }
110 if (check_func(ctx->yuv2planeX, "yuv2yuvX_%d_%d_%d", filter_sizes[fsi], osi, dstW)){
111 memset(dst0, 0, LARGEST_INPUT_SIZE * sizeof(dst0[0]));
112 memset(dst1, 0, LARGEST_INPUT_SIZE * sizeof(dst1[0]));
113
114 // The reference function is not the scalar function selected when mmx
115 // is deactivated as the SIMD functions do not give the same result as
116 // the scalar ones due to rounding. The SIMD functions are activated by
117 // the flag SWS_ACCURATE_RND
118 ref_function(&filter_coeff[0], filter_sizes[fsi], src, dst0, dstW - osi, dither, osi);
119 // There's no point in calling new for the reference function
120 if(ctx->use_mmx_vfilter){
121 call_new((const int16_t*)vFilterData, filter_sizes[fsi], src, dst1, dstW - osi, dither, osi);
122 if (memcmp(dst0, dst1, LARGEST_INPUT_SIZE * sizeof(dst0[0])))
123 fail();
124 if(dstW == LARGEST_INPUT_SIZE)
125 bench_new((const int16_t*)vFilterData, filter_sizes[fsi], src, dst1, dstW - osi, dither, osi);
126 }
127 }
128 av_freep(&src);
129 av_freep(&vFilterData);
130 }
131 }
132 }
133 sws_freeContext(ctx);
134 #undef FILTER_SIZES
135 }
136
137 #undef SRC_PIXELS
138 #define SRC_PIXELS 128
139
check_hscale(void)140 static void check_hscale(void)
141 {
142 #define MAX_FILTER_WIDTH 40
143 #define FILTER_SIZES 5
144 static const int filter_sizes[FILTER_SIZES] = { 4, 8, 16, 32, 40 };
145
146 #define HSCALE_PAIRS 2
147 static const int hscale_pairs[HSCALE_PAIRS][2] = {
148 { 8, 14 },
149 { 8, 18 },
150 };
151
152 int i, j, fsi, hpi, width;
153 struct SwsContext *ctx;
154
155 // padded
156 LOCAL_ALIGNED_32(uint8_t, src, [FFALIGN(SRC_PIXELS + MAX_FILTER_WIDTH - 1, 4)]);
157 LOCAL_ALIGNED_32(uint32_t, dst0, [SRC_PIXELS]);
158 LOCAL_ALIGNED_32(uint32_t, dst1, [SRC_PIXELS]);
159
160 // padded
161 LOCAL_ALIGNED_32(int16_t, filter, [SRC_PIXELS * MAX_FILTER_WIDTH + MAX_FILTER_WIDTH]);
162 LOCAL_ALIGNED_32(int32_t, filterPos, [SRC_PIXELS]);
163
164 // The dst parameter here is either int16_t or int32_t but we use void* to
165 // just cover both cases.
166 declare_func_emms(AV_CPU_FLAG_MMX, void, void *c, void *dst, int dstW,
167 const uint8_t *src, const int16_t *filter,
168 const int32_t *filterPos, int filterSize);
169
170 ctx = sws_alloc_context();
171 if (sws_init_context(ctx, NULL, NULL) < 0)
172 fail();
173
174 randomize_buffers(src, SRC_PIXELS + MAX_FILTER_WIDTH - 1);
175
176 for (hpi = 0; hpi < HSCALE_PAIRS; hpi++) {
177 for (fsi = 0; fsi < FILTER_SIZES; fsi++) {
178 width = filter_sizes[fsi];
179
180 ctx->srcBpc = hscale_pairs[hpi][0];
181 ctx->dstBpc = hscale_pairs[hpi][1];
182 ctx->hLumFilterSize = ctx->hChrFilterSize = width;
183
184 for (i = 0; i < SRC_PIXELS; i++) {
185 filterPos[i] = i;
186
187 // These filter cofficients are chosen to try break two corner
188 // cases, namely:
189 //
190 // - Negative filter coefficients. The filters output signed
191 // values, and it should be possible to end up with negative
192 // output values.
193 //
194 // - Positive clipping. The hscale filter function has clipping
195 // at (1<<15) - 1
196 //
197 // The coefficients sum to the 1.0 point for the hscale
198 // functions (1 << 14).
199
200 for (j = 0; j < width; j++) {
201 filter[i * width + j] = -((1 << 14) / (width - 1));
202 }
203 filter[i * width + (rnd() % width)] = ((1 << 15) - 1);
204 }
205
206 for (i = 0; i < MAX_FILTER_WIDTH; i++) {
207 // These values should be unused in SIMD implementations but
208 // may still be read, random coefficients here should help show
209 // issues where they are used in error.
210
211 filter[SRC_PIXELS * width + i] = rnd();
212 }
213 ff_getSwsFunc(ctx);
214
215 if (check_func(ctx->hcScale, "hscale_%d_to_%d_width%d", ctx->srcBpc, ctx->dstBpc + 1, width)) {
216 memset(dst0, 0, SRC_PIXELS * sizeof(dst0[0]));
217 memset(dst1, 0, SRC_PIXELS * sizeof(dst1[0]));
218
219 call_ref(NULL, dst0, SRC_PIXELS, src, filter, filterPos, width);
220 call_new(NULL, dst1, SRC_PIXELS, src, filter, filterPos, width);
221 if (memcmp(dst0, dst1, SRC_PIXELS * sizeof(dst0[0])))
222 fail();
223 bench_new(NULL, dst0, SRC_PIXELS, src, filter, filterPos, width);
224 }
225 }
226 }
227 sws_freeContext(ctx);
228 }
229
checkasm_check_sw_scale(void)230 void checkasm_check_sw_scale(void)
231 {
232 check_hscale();
233 report("hscale");
234 check_yuv2yuvX();
235 report("yuv2yuvX");
236 }
237