xref: /aosp_15_r20/external/mesa3d/src/gallium/tests/unit/translate_test.c (revision 6104692788411f58d303aa86923a9ff6ecaded22)
1 /**************************************************************************
2  *
3  * Copyright © 2010 Luca Barbieri
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the "Software"),
7  * to deal in the Software without restriction, including without limitation
8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9  * and/or sell copies of the Software, and to permit persons to whom the
10  * Software is furnished to do so, subject to the following conditions:
11  *
12  * The above copyright notice and this permission notice (including the next
13  * paragraph) shall be included in all copies or substantial portions of the
14  * Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
21  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
22  * DEALINGS IN THE SOFTWARE.
23  *
24  **************************************************************************/
25 
26 #include <stdio.h>
27 #include "translate/translate.h"
28 #include "util/u_memory.h"
29 #include "util/format/u_format.h"
30 #include "util/half_float.h"
31 #include "util/u_cpu_detect.h"
32 
33 /* don't use this for serious use */
rand_double()34 static double rand_double()
35 {
36    const double rm = (double)RAND_MAX + 1;
37    double div = 1;
38    double v = 0;
39    unsigned i;
40    for(i = 0; i < 4; ++i)
41    {
42       div *= rm;
43       v += (double)rand() / div;
44    }
45    return v;
46 }
47 
48 char cpu_caps_override_env[128];
49 
main(int argc,char ** argv)50 int main(int argc, char** argv)
51 {
52    struct translate *(*create_fn)(const struct translate_key *key) = 0;
53 
54    struct translate_key key;
55    unsigned output_format;
56    unsigned input_format;
57    unsigned buffer_size = 4096;
58    unsigned char* buffer[5];
59    unsigned char* byte_buffer;
60    float* float_buffer;
61    double* double_buffer;
62    uint16_t *half_buffer;
63    unsigned * elts;
64    unsigned count = 4;
65    unsigned i, j, k;
66    unsigned passed = 0;
67    unsigned total = 0;
68    const float error = 0.03125;
69 
70    create_fn = 0;
71 
72    if (argc <= 1 ||
73        !strcmp(argv[1], "default") )
74       create_fn = translate_create;
75    else if (!strcmp(argv[1], "generic"))
76       create_fn = translate_generic_create;
77    else if (!strcmp(argv[1], "x86"))
78       create_fn = translate_sse2_create;
79    else
80    {
81       const char *translate_options[] = {
82          "nosse", "sse", "sse2", "sse3", "ssse3", "sse4.1", "avx",
83          NULL
84       };
85       const char **option;
86       for (option = translate_options; *option; ++option)
87       {
88          if (!strcmp(argv[1], *option))
89          {
90             create_fn = translate_sse2_create;
91             break;
92          }
93       }
94       if (create_fn) {
95          snprintf(cpu_caps_override_env, sizeof(cpu_caps_override_env), "GALLIUM_OVERRIDE_CPU_CAPS=%s", argv[1]);
96          putenv(cpu_caps_override_env);
97       }
98    }
99 
100    if (!create_fn)
101    {
102       printf("Usage: ./translate_test [default|generic|x86|nosse|sse|sse2|sse3|ssse3|sse4.1|avx]\n");
103       return 2;
104    }
105 
106    for (i = 1; i < ARRAY_SIZE(buffer); ++i)
107       buffer[i] = align_malloc(buffer_size, 4096);
108 
109    byte_buffer = align_malloc(buffer_size, 4096);
110    float_buffer = align_malloc(buffer_size, 4096);
111    double_buffer = align_malloc(buffer_size, 4096);
112    half_buffer = align_malloc(buffer_size, 4096);
113 
114    elts = align_malloc(count * sizeof *elts, 4096);
115 
116    key.nr_elements = 1;
117    key.element[0].input_buffer = 0;
118    key.element[0].input_offset = 0;
119    key.element[0].output_offset = 0;
120    key.element[0].type = TRANSLATE_ELEMENT_NORMAL;
121    key.element[0].instance_divisor = 0;
122 
123    srand(4359025);
124 
125    /* avoid negative values that work badly when converted to unsigned format*/
126    for (i = 0; i < buffer_size; ++i)
127       byte_buffer[i] = rand() & 0x7f7f7f7f;
128 
129    for (i = 0; i < buffer_size / sizeof(float); ++i)
130       float_buffer[i] = (float)rand_double();
131 
132    for (i = 0; i < buffer_size / sizeof(double); ++i)
133       double_buffer[i] = rand_double();
134 
135    for (i = 0; i < buffer_size / sizeof(double); ++i)
136       half_buffer[i] = _mesa_float_to_half((float) rand_double());
137 
138    for (i = 0; i < count; ++i)
139       elts[i] = i;
140 
141    for (output_format = 1; output_format < PIPE_FORMAT_COUNT; ++output_format)
142    {
143       const struct util_format_description* output_format_desc = util_format_description(output_format);
144       const struct util_format_pack_description* output_format_pack = util_format_pack_description(output_format);
145       util_format_fetch_rgba_func_ptr fetch_rgba =
146          util_format_fetch_rgba_func(output_format);
147       unsigned output_format_size;
148       unsigned output_normalized = 0;
149 
150       if (!fetch_rgba
151           || !output_format_pack->pack_rgba_float
152           || output_format_desc->colorspace != UTIL_FORMAT_COLORSPACE_RGB
153           || output_format_desc->layout != UTIL_FORMAT_LAYOUT_PLAIN
154           || !translate_is_output_format_supported(output_format))
155          continue;
156 
157       for(i = 0; i < output_format_desc->nr_channels; ++i)
158       {
159          if(output_format_desc->channel[i].type != UTIL_FORMAT_TYPE_FLOAT)
160             output_normalized |= (1 << output_format_desc->channel[i].normalized);
161       }
162 
163       output_format_size = util_format_get_stride(output_format, 1);
164 
165       for (input_format = 1; input_format < PIPE_FORMAT_COUNT; ++input_format)
166       {
167          const struct util_format_description* input_format_desc = util_format_description(input_format);
168          const struct util_format_pack_description* input_format_pack = util_format_pack_description(input_format);
169          util_format_fetch_rgba_func_ptr fetch_rgba =
170             util_format_fetch_rgba_func(input_format);
171          unsigned input_format_size;
172          struct translate* translate[2];
173          unsigned fail = 0;
174          unsigned used_generic = 0;
175          unsigned input_normalized = 0;
176          bool input_is_float = false;
177 
178          if (!fetch_rgba
179              || !input_format_pack->pack_rgba_float
180              || input_format_desc->colorspace != UTIL_FORMAT_COLORSPACE_RGB
181              || input_format_desc->layout != UTIL_FORMAT_LAYOUT_PLAIN
182              || !translate_is_output_format_supported(input_format))
183             continue;
184 
185          input_format_size = util_format_get_stride(input_format, 1);
186 
187          for(i = 0; i < input_format_desc->nr_channels; ++i)
188          {
189             if(input_format_desc->channel[i].type == UTIL_FORMAT_TYPE_FLOAT)
190             {
191                input_is_float = 1;
192                input_normalized |= 1 << 1;
193             }
194             else
195                input_normalized |= (1 << input_format_desc->channel[i].normalized);
196          }
197 
198          if(((input_normalized | output_normalized) == 3)
199                || ((input_normalized & 1) && (output_normalized & 1)
200                      && input_format_size * output_format_desc->nr_channels > output_format_size * input_format_desc->nr_channels))
201             continue;
202 
203          key.element[0].input_format = input_format;
204          key.element[0].output_format = output_format;
205          key.output_stride = output_format_size;
206          translate[0] = create_fn(&key);
207          if (!translate[0])
208             continue;
209 
210          key.element[0].input_format = output_format;
211          key.element[0].output_format = input_format;
212          key.output_stride = input_format_size;
213          translate[1] = create_fn(&key);
214          if(!translate[1])
215          {
216             used_generic = 1;
217             translate[1] = translate_generic_create(&key);
218             if(!translate[1])
219                continue;
220          }
221 
222          for(i = 1; i < 5; ++i)
223             memset(buffer[i], 0xcd - (0x22 * i), 4096);
224 
225          if(input_is_float && input_format_desc->channel[0].size == 32)
226             buffer[0] = (unsigned char*)float_buffer;
227          else if(input_is_float && input_format_desc->channel[0].size == 64)
228             buffer[0] = (unsigned char*)double_buffer;
229          else if(input_is_float && input_format_desc->channel[0].size == 16)
230             buffer[0] = (unsigned char*)half_buffer;
231          else if(input_is_float)
232             abort();
233          else
234             buffer[0] = byte_buffer;
235 
236          translate[0]->set_buffer(translate[0], 0, buffer[0], input_format_size, count - 1);
237          translate[0]->run_elts(translate[0], elts, count, 0, 0, buffer[1]);
238          translate[1]->set_buffer(translate[1], 0, buffer[1], output_format_size, count - 1);
239          translate[1]->run_elts(translate[1], elts, count, 0, 0, buffer[2]);
240          translate[0]->set_buffer(translate[0], 0, buffer[2], input_format_size, count - 1);
241          translate[0]->run_elts(translate[0], elts, count, 0, 0, buffer[3]);
242          translate[1]->set_buffer(translate[1], 0, buffer[3], output_format_size, count - 1);
243          translate[1]->run_elts(translate[1], elts, count, 0, 0, buffer[4]);
244 
245          for (i = 0; i < count; ++i)
246          {
247             float a[4];
248             float b[4];
249             fetch_rgba(a, buffer[2] + i * input_format_size, 0, 0);
250             fetch_rgba(b, buffer[4] + i * input_format_size, 0, 0);
251 
252             for (j = 0; j < count; ++j)
253             {
254                float d = a[j] - b[j];
255                if (d > error || d < -error)
256                {
257                   fail = 1;
258                   break;
259                }
260             }
261          }
262 
263          printf("%s%s: %s -> %s -> %s -> %s -> %s\n",
264                fail ? "FAIL" : "PASS",
265                used_generic ? "[GENERIC]" : "",
266                input_format_desc->name, output_format_desc->name, input_format_desc->name, output_format_desc->name, input_format_desc->name);
267 
268          if (1)
269          {
270             for (i = 0; i < ARRAY_SIZE(buffer); ++i)
271             {
272                unsigned format_size = (i & 1) ? output_format_size : input_format_size;
273                printf("%c ", (i == 2 || i == 4) ? '*' : ' ');
274                for (j = 0; j < count; ++j)
275                {
276                   for (k = 0; k < format_size; ++k)
277                   {
278                      printf("%02x", buffer[i][j * format_size + k]);
279                   }
280                   printf(" ");
281                }
282                printf("\n");
283             }
284          }
285 
286          if (!fail)
287             ++passed;
288          ++total;
289 
290          if(translate[1])
291             translate[1]->release(translate[1]);
292          translate[0]->release(translate[0]);
293       }
294    }
295 
296    printf("%u/%u tests passed for translate_%s\n", passed, total, argv[1]);
297 
298    for (i = 1; i < ARRAY_SIZE(buffer); ++i)
299       align_free(buffer[i]);
300 
301    align_free(byte_buffer);
302    align_free(float_buffer);
303    align_free(double_buffer);
304    align_free(half_buffer);
305    align_free(elts);
306    return passed != total;
307 }
308