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_internal.h"
25
26#include "libswscale/swscale.h"
27#include "libswscale/swscale_internal.h"
28
29#include "checkasm.h"
30
31#define randomize_buffers(buf, size)      \
32    do {                                  \
33        int j;                            \
34        for (j = 0; j < size; j+=4)       \
35            AV_WN32(buf + j, rnd());      \
36    } while (0)
37
38// This reference function is the same approximate algorithm employed by the
39// SIMD functions
40static void ref_function(const int16_t *filter, int filterSize,
41                                                 const int16_t **src, uint8_t *dest, int dstW,
42                                                 const uint8_t *dither, int offset)
43{
44    int i, d;
45    d = ((filterSize - 1) * 8 + dither[0]) >> 4;
46    for ( i = 0; i < dstW; i++) {
47        int16_t val = d;
48        int j;
49        union {
50            int val;
51            int16_t v[2];
52        } t;
53        for (j = 0; j < filterSize; j++){
54            t.val = (int)src[j][i + offset] * (int)filter[j];
55            val += t.v[1];
56        }
57        dest[i]= av_clip_uint8(val>>3);
58    }
59}
60
61static void check_yuv2yuvX(void)
62{
63    struct SwsContext *ctx;
64    int fsi, osi, isi, i, j;
65    int dstW;
66#define LARGEST_FILTER 16
67#define FILTER_SIZES 4
68    static const int filter_sizes[FILTER_SIZES] = {1, 4, 8, 16};
69#define LARGEST_INPUT_SIZE 512
70#define INPUT_SIZES 6
71    static const int input_sizes[INPUT_SIZES] = {8, 24, 128, 144, 256, 512};
72
73    declare_func_emms(AV_CPU_FLAG_MMX, void, const int16_t *filter,
74                      int filterSize, const int16_t **src, uint8_t *dest,
75                      int dstW, const uint8_t *dither, int offset);
76
77    const int16_t **src;
78    LOCAL_ALIGNED_16(int16_t, src_pixels, [LARGEST_FILTER * LARGEST_INPUT_SIZE]);
79    LOCAL_ALIGNED_16(int16_t, filter_coeff, [LARGEST_FILTER]);
80    LOCAL_ALIGNED_16(uint8_t, dst0, [LARGEST_INPUT_SIZE]);
81    LOCAL_ALIGNED_16(uint8_t, dst1, [LARGEST_INPUT_SIZE]);
82    LOCAL_ALIGNED_16(uint8_t, dither, [LARGEST_INPUT_SIZE]);
83    union VFilterData{
84        const int16_t *src;
85        uint16_t coeff[8];
86    } *vFilterData;
87    uint8_t d_val = rnd();
88    memset(dither, d_val, LARGEST_INPUT_SIZE);
89    randomize_buffers((uint8_t*)src_pixels, LARGEST_FILTER * LARGEST_INPUT_SIZE * sizeof(int16_t));
90    randomize_buffers((uint8_t*)filter_coeff, LARGEST_FILTER * sizeof(int16_t));
91    ctx = sws_alloc_context();
92    if (sws_init_context(ctx, NULL, NULL) < 0)
93        fail();
94
95    ff_sws_init_scale(ctx);
96    for(isi = 0; isi < INPUT_SIZES; ++isi){
97        dstW = input_sizes[isi];
98        for(osi = 0; osi < 64; osi += 16){
99            for(fsi = 0; fsi < FILTER_SIZES; ++fsi){
100                src = av_malloc(sizeof(int16_t*) * filter_sizes[fsi]);
101                vFilterData = av_malloc((filter_sizes[fsi] + 2) * sizeof(union VFilterData));
102                memset(vFilterData, 0, (filter_sizes[fsi] + 2) * sizeof(union VFilterData));
103                for(i = 0; i < filter_sizes[fsi]; ++i){
104                    src[i] = &src_pixels[i * LARGEST_INPUT_SIZE];
105                    vFilterData[i].src = src[i];
106                    for(j = 0; j < 4; ++j)
107                        vFilterData[i].coeff[j + 4] = filter_coeff[i];
108                }
109                if (check_func(ctx->yuv2planeX, "yuv2yuvX_%d_%d_%d", filter_sizes[fsi], osi, dstW)){
110                    memset(dst0, 0, LARGEST_INPUT_SIZE * sizeof(dst0[0]));
111                    memset(dst1, 0, LARGEST_INPUT_SIZE * sizeof(dst1[0]));
112
113                    // The reference function is not the scalar function selected when mmx
114                    // is deactivated as the SIMD functions do not give the same result as
115                    // the scalar ones due to rounding. The SIMD functions are activated by
116                    // the flag SWS_ACCURATE_RND
117                    ref_function(&filter_coeff[0], filter_sizes[fsi], src, dst0, dstW - osi, dither, osi);
118                    // There's no point in calling new for the reference function
119                    if(ctx->use_mmx_vfilter){
120                        call_new((const int16_t*)vFilterData, filter_sizes[fsi], src, dst1, dstW - osi, dither, osi);
121                        if (memcmp(dst0, dst1, LARGEST_INPUT_SIZE * sizeof(dst0[0])))
122                            fail();
123                        if(dstW == LARGEST_INPUT_SIZE)
124                            bench_new((const int16_t*)vFilterData, filter_sizes[fsi], src, dst1, dstW - osi, dither, osi);
125                    }
126                }
127                av_freep(&src);
128                av_freep(&vFilterData);
129            }
130        }
131    }
132    sws_freeContext(ctx);
133#undef FILTER_SIZES
134}
135
136#undef SRC_PIXELS
137#define SRC_PIXELS 512
138
139static void check_hscale(void)
140{
141#define MAX_FILTER_WIDTH 40
142#define FILTER_SIZES 6
143    static const int filter_sizes[FILTER_SIZES] = { 4, 8, 12, 16, 32, 40 };
144
145#define HSCALE_PAIRS 2
146    static const int hscale_pairs[HSCALE_PAIRS][2] = {
147        { 8, 14 },
148        { 8, 18 },
149    };
150
151#define LARGEST_INPUT_SIZE 512
152#define INPUT_SIZES 6
153    static const int input_sizes[INPUT_SIZES] = {8, 24, 128, 144, 256, 512};
154
155    int i, j, fsi, hpi, width, dstWi;
156    struct SwsContext *ctx;
157
158    // padded
159    LOCAL_ALIGNED_32(uint8_t, src, [FFALIGN(SRC_PIXELS + MAX_FILTER_WIDTH - 1, 4)]);
160    LOCAL_ALIGNED_32(uint32_t, dst0, [SRC_PIXELS]);
161    LOCAL_ALIGNED_32(uint32_t, dst1, [SRC_PIXELS]);
162
163    // padded
164    LOCAL_ALIGNED_32(int16_t, filter, [SRC_PIXELS * MAX_FILTER_WIDTH + MAX_FILTER_WIDTH]);
165    LOCAL_ALIGNED_32(int32_t, filterPos, [SRC_PIXELS]);
166    LOCAL_ALIGNED_32(int16_t, filterAvx2, [SRC_PIXELS * MAX_FILTER_WIDTH + MAX_FILTER_WIDTH]);
167    LOCAL_ALIGNED_32(int32_t, filterPosAvx, [SRC_PIXELS]);
168
169    // The dst parameter here is either int16_t or int32_t but we use void* to
170    // just cover both cases.
171    declare_func_emms(AV_CPU_FLAG_MMX, void, void *c, void *dst, int dstW,
172                      const uint8_t *src, const int16_t *filter,
173                      const int32_t *filterPos, int filterSize);
174
175    int cpu_flags = av_get_cpu_flags();
176
177    ctx = sws_alloc_context();
178    if (sws_init_context(ctx, NULL, NULL) < 0)
179        fail();
180
181    randomize_buffers(src, SRC_PIXELS + MAX_FILTER_WIDTH - 1);
182
183    for (hpi = 0; hpi < HSCALE_PAIRS; hpi++) {
184        for (fsi = 0; fsi < FILTER_SIZES; fsi++) {
185            for (dstWi = 0; dstWi < INPUT_SIZES; dstWi++) {
186                width = filter_sizes[fsi];
187
188                ctx->srcBpc = hscale_pairs[hpi][0];
189                ctx->dstBpc = hscale_pairs[hpi][1];
190                ctx->hLumFilterSize = ctx->hChrFilterSize = width;
191
192                for (i = 0; i < SRC_PIXELS; i++) {
193                    filterPos[i] = i;
194                    filterPosAvx[i] = i;
195
196                    // These filter cofficients are chosen to try break two corner
197                    // cases, namely:
198                    //
199                    // - Negative filter coefficients. The filters output signed
200                    //   values, and it should be possible to end up with negative
201                    //   output values.
202                    //
203                    // - Positive clipping. The hscale filter function has clipping
204                    //   at (1<<15) - 1
205                    //
206                    // The coefficients sum to the 1.0 point for the hscale
207                    // functions (1 << 14).
208
209                    for (j = 0; j < width; j++) {
210                        filter[i * width + j] = -((1 << 14) / (width - 1));
211                    }
212                    filter[i * width + (rnd() % width)] = ((1 << 15) - 1);
213                }
214
215                for (i = 0; i < MAX_FILTER_WIDTH; i++) {
216                    // These values should be unused in SIMD implementations but
217                    // may still be read, random coefficients here should help show
218                    // issues where they are used in error.
219
220                    filter[SRC_PIXELS * width + i] = rnd();
221                }
222                ctx->dstW = ctx->chrDstW = input_sizes[dstWi];
223                ff_sws_init_scale(ctx);
224                memcpy(filterAvx2, filter, sizeof(uint16_t) * (SRC_PIXELS * MAX_FILTER_WIDTH + MAX_FILTER_WIDTH));
225                if ((cpu_flags & AV_CPU_FLAG_AVX2) && !(cpu_flags & AV_CPU_FLAG_SLOW_GATHER))
226                    ff_shuffle_filter_coefficients(ctx, filterPosAvx, width, filterAvx2, SRC_PIXELS);
227
228                if (check_func(ctx->hcScale, "hscale_%d_to_%d__fs_%d_dstW_%d", ctx->srcBpc, ctx->dstBpc + 1, width, ctx->dstW)) {
229                    memset(dst0, 0, SRC_PIXELS * sizeof(dst0[0]));
230                    memset(dst1, 0, SRC_PIXELS * sizeof(dst1[0]));
231
232                    call_ref(NULL, dst0, ctx->dstW, src, filter, filterPos, width);
233                    call_new(NULL, dst1, ctx->dstW, src, filterAvx2, filterPosAvx, width);
234                    if (memcmp(dst0, dst1, ctx->dstW * sizeof(dst0[0])))
235                        fail();
236                    bench_new(NULL, dst0, ctx->dstW, src, filter, filterPosAvx, width);
237                }
238            }
239        }
240    }
241    sws_freeContext(ctx);
242}
243
244void checkasm_check_sw_scale(void)
245{
246    check_hscale();
247    report("hscale");
248    check_yuv2yuvX();
249    report("yuv2yuvX");
250}
251