3 * This file is part of FFmpeg.
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.
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.
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.
22 #include "libavutil/common.h"
23 #include "libavutil/intreadwrite.h"
24 #include "libavutil/mem_internal.h"
26 #include "libswscale/swscale.h"
27 #include "libswscale/swscale_internal.h"
31 #define randomize_buffers(buf, size) \
34 for (j = 0; j < size; j+=4) \
35 AV_WN32(buf + j, rnd()); \
38 static void yuv2planeX_8_ref(const int16_t *filter, int filterSize,
39 const int16_t **src, uint8_t *dest, int dstW,
40 const uint8_t *dither, int offset)
42 // This corresponds to the yuv2planeX_8_c function
44 for (i = 0; i < dstW; i++) {
45 int val = dither[(i + offset) & 7] << 12;
47 for (j = 0; j < filterSize; j++)
48 val += src[j][i] * filter[j];
50 dest[i]= av_clip_uint8(val >> 19);
54 static int cmp_off_by_n(const uint8_t *ref, const uint8_t *test, size_t n, int accuracy)
56 for (size_t i = 0; i < n; i++) {
57 if (abs(ref[i] - test[i]) > accuracy)
63 static void print_data(uint8_t *p, size_t len, size_t offset)
66 for (; i < len; i++) {
68 printf("0x%04zx: ", i+offset);
70 printf("0x%02x ", (uint32_t) p[i]);
80 static size_t show_differences(uint8_t *a, uint8_t *b, size_t len)
82 for (size_t i = 0; i < len; i++) {
84 size_t offset_of_mismatch = i;
89 print_data(&a[offset], 32, offset);
90 printf("\ntest b:\n");
91 print_data(&b[offset], 32, offset);
93 return offset_of_mismatch;
99 static void check_yuv2yuv1(int accurate)
101 struct SwsContext *ctx;
105 const int input_sizes[] = {8, 24, 128, 144, 256, 512};
106 const int INPUT_SIZES = sizeof(input_sizes)/sizeof(input_sizes[0]);
107 #define LARGEST_INPUT_SIZE 512
109 const int offsets[] = {0, 3, 8, 11, 16, 19};
110 const int OFFSET_SIZES = sizeof(offsets)/sizeof(offsets[0]);
111 const char *accurate_str = (accurate) ? "accurate" : "approximate";
114 const int16_t *src, uint8_t *dest,
115 int dstW, const uint8_t *dither, int offset);
117 LOCAL_ALIGNED_16(int16_t, src_pixels, [LARGEST_INPUT_SIZE]);
118 LOCAL_ALIGNED_16(uint8_t, dst0, [LARGEST_INPUT_SIZE]);
119 LOCAL_ALIGNED_16(uint8_t, dst1, [LARGEST_INPUT_SIZE]);
120 LOCAL_ALIGNED_8(uint8_t, dither, [8]);
122 randomize_buffers((uint8_t*)dither, 8);
123 randomize_buffers((uint8_t*)src_pixels, LARGEST_INPUT_SIZE * sizeof(int16_t));
124 ctx = sws_alloc_context();
126 ctx->flags |= SWS_ACCURATE_RND;
127 if (sws_init_context(ctx, NULL, NULL) < 0)
130 ff_sws_init_scale(ctx);
131 for (isi = 0; isi < INPUT_SIZES; ++isi) {
132 dstW = input_sizes[isi];
133 for (osi = 0; osi < OFFSET_SIZES; osi++) {
134 offset = offsets[osi];
135 if (check_func(ctx->yuv2plane1, "yuv2yuv1_%d_%d_%s", offset, dstW, accurate_str)){
136 memset(dst0, 0, LARGEST_INPUT_SIZE * sizeof(dst0[0]));
137 memset(dst1, 0, LARGEST_INPUT_SIZE * sizeof(dst1[0]));
139 call_ref(src_pixels, dst0, dstW, dither, offset);
140 call_new(src_pixels, dst1, dstW, dither, offset);
141 if (cmp_off_by_n(dst0, dst1, dstW * sizeof(dst0[0]), accurate ? 0 : 2)) {
143 printf("failed: yuv2yuv1_%d_%di_%s\n", offset, dstW, accurate_str);
144 fail_offset = show_differences(dst0, dst1, LARGEST_INPUT_SIZE * sizeof(dst0[0]));
145 printf("failing values: src: 0x%04x dither: 0x%02x dst-c: %02x dst-asm: %02x\n",
146 (int) src_pixels[fail_offset],
147 (int) dither[(fail_offset + fail_offset) & 7],
148 (int) dst0[fail_offset],
149 (int) dst1[fail_offset]);
151 if(dstW == LARGEST_INPUT_SIZE)
152 bench_new(src_pixels, dst1, dstW, dither, offset);
156 sws_freeContext(ctx);
159 static void check_yuv2yuvX(int accurate)
161 struct SwsContext *ctx;
162 int fsi, osi, isi, i, j;
164 #define LARGEST_FILTER 16
165 // ff_yuv2planeX_8_sse2 can't handle odd filter sizes
166 const int filter_sizes[] = {2, 4, 8, 16};
167 const int FILTER_SIZES = sizeof(filter_sizes)/sizeof(filter_sizes[0]);
168 #define LARGEST_INPUT_SIZE 512
169 static const int input_sizes[] = {8, 24, 128, 144, 256, 512};
170 const int INPUT_SIZES = sizeof(input_sizes)/sizeof(input_sizes[0]);
171 const char *accurate_str = (accurate) ? "accurate" : "approximate";
173 declare_func_emms(AV_CPU_FLAG_MMX, void, const int16_t *filter,
174 int filterSize, const int16_t **src, uint8_t *dest,
175 int dstW, const uint8_t *dither, int offset);
178 LOCAL_ALIGNED_16(int16_t, src_pixels, [LARGEST_FILTER * LARGEST_INPUT_SIZE]);
179 LOCAL_ALIGNED_16(int16_t, filter_coeff, [LARGEST_FILTER]);
180 LOCAL_ALIGNED_16(uint8_t, dst0, [LARGEST_INPUT_SIZE]);
181 LOCAL_ALIGNED_16(uint8_t, dst1, [LARGEST_INPUT_SIZE]);
182 LOCAL_ALIGNED_16(uint8_t, dither, [LARGEST_INPUT_SIZE]);
187 uint8_t d_val = rnd();
188 memset(dither, d_val, LARGEST_INPUT_SIZE);
189 randomize_buffers((uint8_t*)src_pixels, LARGEST_FILTER * LARGEST_INPUT_SIZE * sizeof(int16_t));
190 ctx = sws_alloc_context();
192 ctx->flags |= SWS_ACCURATE_RND;
193 if (sws_init_context(ctx, NULL, NULL) < 0)
196 ff_sws_init_scale(ctx);
197 for(isi = 0; isi < INPUT_SIZES; ++isi){
198 dstW = input_sizes[isi];
199 for(osi = 0; osi < 64; osi += 16){
202 for (fsi = 0; fsi < FILTER_SIZES; ++fsi) {
203 // Generate filter coefficients for the given filter size,
204 // with some properties:
205 // - The coefficients add up to the intended sum (4096, 1<<12)
206 // - The coefficients contain negative values
207 // - The filter intermediates don't overflow for worst case
208 // inputs (all positive coefficients are coupled with
209 // input_max and all negative coefficients with input_min,
211 // Produce a filter with all coefficients set to
212 // -((1<<12)/(filter_size-1)) except for one (randomly chosen)
213 // which is set to ((1<<13)-1).
214 for (i = 0; i < filter_sizes[fsi]; ++i)
215 filter_coeff[i] = -((1 << 12) / (filter_sizes[fsi] - 1));
216 filter_coeff[rnd() % filter_sizes[fsi]] = (1 << 13) - 1;
218 src = av_malloc(sizeof(int16_t*) * filter_sizes[fsi]);
219 vFilterData = av_malloc((filter_sizes[fsi] + 2) * sizeof(union VFilterData));
220 memset(vFilterData, 0, (filter_sizes[fsi] + 2) * sizeof(union VFilterData));
221 for (i = 0; i < filter_sizes[fsi]; ++i) {
222 src[i] = &src_pixels[i * LARGEST_INPUT_SIZE];
223 vFilterData[i].src = src[i] - osi;
224 for(j = 0; j < 4; ++j)
225 vFilterData[i].coeff[j + 4] = filter_coeff[i];
227 if (check_func(ctx->yuv2planeX, "yuv2yuvX_%d_%d_%d_%s", filter_sizes[fsi], osi, dstW, accurate_str)){
228 // use vFilterData for the mmx function
229 const int16_t *filter = ctx->use_mmx_vfilter ? (const int16_t*)vFilterData : &filter_coeff[0];
230 memset(dst0, 0, LARGEST_INPUT_SIZE * sizeof(dst0[0]));
231 memset(dst1, 0, LARGEST_INPUT_SIZE * sizeof(dst1[0]));
233 // We can't use call_ref here, because we don't know if use_mmx_vfilter was set for that
234 // function or not, so we can't pass it the parameters correctly.
235 yuv2planeX_8_ref(&filter_coeff[0], filter_sizes[fsi], src, dst0, dstW - osi, dither, osi);
237 call_new(filter, filter_sizes[fsi], src, dst1, dstW - osi, dither, osi);
238 if (cmp_off_by_n(dst0, dst1, LARGEST_INPUT_SIZE * sizeof(dst0[0]), accurate ? 0 : 2)) {
240 printf("failed: yuv2yuvX_%d_%d_%d_%s\n", filter_sizes[fsi], osi, dstW, accurate_str);
241 show_differences(dst0, dst1, LARGEST_INPUT_SIZE * sizeof(dst0[0]));
243 if(dstW == LARGEST_INPUT_SIZE)
244 bench_new((const int16_t*)vFilterData, filter_sizes[fsi], src, dst1, dstW - osi, dither, osi);
248 av_freep(&vFilterData);
252 sws_freeContext(ctx);
257 #define SRC_PIXELS 512
259 static void check_hscale(void)
261 #define MAX_FILTER_WIDTH 40
262 #define FILTER_SIZES 6
263 static const int filter_sizes[FILTER_SIZES] = { 4, 8, 12, 16, 32, 40 };
265 #define HSCALE_PAIRS 2
266 static const int hscale_pairs[HSCALE_PAIRS][2] = {
271 #define LARGEST_INPUT_SIZE 512
272 #define INPUT_SIZES 6
273 static const int input_sizes[INPUT_SIZES] = {8, 24, 128, 144, 256, 512};
275 int i, j, fsi, hpi, width, dstWi;
276 struct SwsContext *ctx;
279 LOCAL_ALIGNED_32(uint8_t, src, [FFALIGN(SRC_PIXELS + MAX_FILTER_WIDTH - 1, 4)]);
280 LOCAL_ALIGNED_32(uint32_t, dst0, [SRC_PIXELS]);
281 LOCAL_ALIGNED_32(uint32_t, dst1, [SRC_PIXELS]);
284 LOCAL_ALIGNED_32(int16_t, filter, [SRC_PIXELS * MAX_FILTER_WIDTH + MAX_FILTER_WIDTH]);
285 LOCAL_ALIGNED_32(int32_t, filterPos, [SRC_PIXELS]);
286 LOCAL_ALIGNED_32(int16_t, filterAvx2, [SRC_PIXELS * MAX_FILTER_WIDTH + MAX_FILTER_WIDTH]);
287 LOCAL_ALIGNED_32(int32_t, filterPosAvx, [SRC_PIXELS]);
289 // The dst parameter here is either int16_t or int32_t but we use void* to
290 // just cover both cases.
291 declare_func(void, void *c, void *dst, int dstW,
292 const uint8_t *src, const int16_t *filter,
293 const int32_t *filterPos, int filterSize);
295 ctx = sws_alloc_context();
296 if (sws_init_context(ctx, NULL, NULL) < 0)
299 randomize_buffers(src, SRC_PIXELS + MAX_FILTER_WIDTH - 1);
301 for (hpi = 0; hpi < HSCALE_PAIRS; hpi++) {
302 for (fsi = 0; fsi < FILTER_SIZES; fsi++) {
303 for (dstWi = 0; dstWi < INPUT_SIZES; dstWi++) {
304 width = filter_sizes[fsi];
306 ctx->srcBpc = hscale_pairs[hpi][0];
307 ctx->dstBpc = hscale_pairs[hpi][1];
308 ctx->hLumFilterSize = ctx->hChrFilterSize = width;
310 for (i = 0; i < SRC_PIXELS; i++) {
314 // These filter cofficients are chosen to try break two corner
317 // - Negative filter coefficients. The filters output signed
318 // values, and it should be possible to end up with negative
321 // - Positive clipping. The hscale filter function has clipping
324 // The coefficients sum to the 1.0 point for the hscale
325 // functions (1 << 14).
327 for (j = 0; j < width; j++) {
328 filter[i * width + j] = -((1 << 14) / (width - 1));
330 filter[i * width + (rnd() % width)] = ((1 << 15) - 1);
333 for (i = 0; i < MAX_FILTER_WIDTH; i++) {
334 // These values should be unused in SIMD implementations but
335 // may still be read, random coefficients here should help show
336 // issues where they are used in error.
338 filter[SRC_PIXELS * width + i] = rnd();
340 ctx->dstW = ctx->chrDstW = input_sizes[dstWi];
341 ff_sws_init_scale(ctx);
342 memcpy(filterAvx2, filter, sizeof(uint16_t) * (SRC_PIXELS * MAX_FILTER_WIDTH + MAX_FILTER_WIDTH));
343 ff_shuffle_filter_coefficients(ctx, filterPosAvx, width, filterAvx2, ctx->dstW);
345 if (check_func(ctx->hcScale, "hscale_%d_to_%d__fs_%d_dstW_%d", ctx->srcBpc, ctx->dstBpc + 1, width, ctx->dstW)) {
346 memset(dst0, 0, SRC_PIXELS * sizeof(dst0[0]));
347 memset(dst1, 0, SRC_PIXELS * sizeof(dst1[0]));
349 call_ref(NULL, dst0, ctx->dstW, src, filter, filterPos, width);
350 call_new(NULL, dst1, ctx->dstW, src, filterAvx2, filterPosAvx, width);
351 if (memcmp(dst0, dst1, ctx->dstW * sizeof(dst0[0])))
353 bench_new(NULL, dst0, ctx->dstW, src, filter, filterPosAvx, width);
358 sws_freeContext(ctx);
361 void checkasm_check_sw_scale(void)