Go to most recent revision | Details | Last modification | View Log | RSS feed
Rev | Author | Line No. | Line |
---|---|---|---|
4358 | Serge | 1 | /************************************************************************** |
2 | * |
||
3 | * Copyright 2009 VMware, Inc. |
||
4 | * All Rights Reserved. |
||
5 | * |
||
6 | * Permission is hereby granted, free of charge, to any person obtaining a |
||
7 | * copy of this software and associated documentation files (the |
||
8 | * "Software"), to deal in the Software without restriction, including |
||
9 | * without limitation the rights to use, copy, modify, merge, publish, |
||
10 | * distribute, sub license, and/or sell copies of the Software, and to |
||
11 | * permit persons to whom the Software is furnished to do so, subject to |
||
12 | * the following conditions: |
||
13 | * |
||
14 | * The above copyright notice and this permission notice (including the |
||
15 | * next paragraph) shall be included in all copies or substantial portions |
||
16 | * of the Software. |
||
17 | * |
||
18 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS |
||
19 | * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF |
||
20 | * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. |
||
21 | * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR |
||
22 | * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, |
||
23 | * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE |
||
24 | * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. |
||
25 | * |
||
26 | **************************************************************************/ |
||
27 | |||
28 | |||
29 | /** |
||
30 | * @file |
||
31 | * Unit tests for type conversion. |
||
32 | * |
||
33 | * @author Jose Fonseca |
||
34 | */ |
||
35 | |||
36 | |||
37 | #include "util/u_pointer.h" |
||
38 | #include "gallivm/lp_bld_init.h" |
||
39 | #include "gallivm/lp_bld_type.h" |
||
40 | #include "gallivm/lp_bld_const.h" |
||
41 | #include "gallivm/lp_bld_conv.h" |
||
42 | #include "gallivm/lp_bld_debug.h" |
||
43 | #include "lp_test.h" |
||
44 | |||
45 | |||
46 | typedef void (*conv_test_ptr_t)(const void *src, const void *dst); |
||
47 | |||
48 | |||
49 | void |
||
50 | write_tsv_header(FILE *fp) |
||
51 | { |
||
52 | fprintf(fp, |
||
53 | "result\t" |
||
54 | "cycles_per_channel\t" |
||
55 | "src_type\t" |
||
56 | "dst_type\n"); |
||
57 | |||
58 | fflush(fp); |
||
59 | } |
||
60 | |||
61 | |||
62 | static void |
||
63 | write_tsv_row(FILE *fp, |
||
64 | struct lp_type src_type, |
||
65 | struct lp_type dst_type, |
||
66 | double cycles, |
||
67 | boolean success) |
||
68 | { |
||
69 | fprintf(fp, "%s\t", success ? "pass" : "fail"); |
||
70 | |||
71 | fprintf(fp, "%.1f\t", cycles / MAX2(src_type.length, dst_type.length)); |
||
72 | |||
73 | dump_type(fp, src_type); |
||
74 | fprintf(fp, "\t"); |
||
75 | |||
76 | dump_type(fp, dst_type); |
||
77 | fprintf(fp, "\n"); |
||
78 | |||
79 | fflush(fp); |
||
80 | } |
||
81 | |||
82 | |||
83 | static void |
||
84 | dump_conv_types(FILE *fp, |
||
85 | struct lp_type src_type, |
||
86 | struct lp_type dst_type) |
||
87 | { |
||
88 | fprintf(fp, "src_type="); |
||
89 | dump_type(fp, src_type); |
||
90 | |||
91 | fprintf(fp, " dst_type="); |
||
92 | dump_type(fp, dst_type); |
||
93 | |||
94 | fprintf(fp, " ...\n"); |
||
95 | fflush(fp); |
||
96 | } |
||
97 | |||
98 | |||
99 | static LLVMValueRef |
||
100 | add_conv_test(struct gallivm_state *gallivm, |
||
101 | struct lp_type src_type, unsigned num_srcs, |
||
102 | struct lp_type dst_type, unsigned num_dsts) |
||
103 | { |
||
104 | LLVMModuleRef module = gallivm->module; |
||
105 | LLVMContextRef context = gallivm->context; |
||
106 | LLVMBuilderRef builder = gallivm->builder; |
||
107 | LLVMTypeRef args[2]; |
||
108 | LLVMValueRef func; |
||
109 | LLVMValueRef src_ptr; |
||
110 | LLVMValueRef dst_ptr; |
||
111 | LLVMBasicBlockRef block; |
||
112 | LLVMValueRef src[LP_MAX_VECTOR_LENGTH]; |
||
113 | LLVMValueRef dst[LP_MAX_VECTOR_LENGTH]; |
||
114 | unsigned i; |
||
115 | |||
116 | args[0] = LLVMPointerType(lp_build_vec_type(gallivm, src_type), 0); |
||
117 | args[1] = LLVMPointerType(lp_build_vec_type(gallivm, dst_type), 0); |
||
118 | |||
119 | func = LLVMAddFunction(module, "test", |
||
120 | LLVMFunctionType(LLVMVoidTypeInContext(context), |
||
121 | args, 2, 0)); |
||
122 | LLVMSetFunctionCallConv(func, LLVMCCallConv); |
||
123 | src_ptr = LLVMGetParam(func, 0); |
||
124 | dst_ptr = LLVMGetParam(func, 1); |
||
125 | |||
126 | block = LLVMAppendBasicBlockInContext(context, func, "entry"); |
||
127 | LLVMPositionBuilderAtEnd(builder, block); |
||
128 | |||
129 | for(i = 0; i < num_srcs; ++i) { |
||
130 | LLVMValueRef index = LLVMConstInt(LLVMInt32TypeInContext(context), i, 0); |
||
131 | LLVMValueRef ptr = LLVMBuildGEP(builder, src_ptr, &index, 1, ""); |
||
132 | src[i] = LLVMBuildLoad(builder, ptr, ""); |
||
133 | } |
||
134 | |||
135 | lp_build_conv(gallivm, src_type, dst_type, src, num_srcs, dst, num_dsts); |
||
136 | |||
137 | for(i = 0; i < num_dsts; ++i) { |
||
138 | LLVMValueRef index = LLVMConstInt(LLVMInt32TypeInContext(context), i, 0); |
||
139 | LLVMValueRef ptr = LLVMBuildGEP(builder, dst_ptr, &index, 1, ""); |
||
140 | LLVMBuildStore(builder, dst[i], ptr); |
||
141 | } |
||
142 | |||
143 | LLVMBuildRetVoid(builder);; |
||
144 | |||
145 | gallivm_verify_function(gallivm, func); |
||
146 | |||
147 | return func; |
||
148 | } |
||
149 | |||
150 | |||
151 | PIPE_ALIGN_STACK |
||
152 | static boolean |
||
153 | test_one(unsigned verbose, |
||
154 | FILE *fp, |
||
155 | struct lp_type src_type, |
||
156 | struct lp_type dst_type) |
||
157 | { |
||
158 | struct gallivm_state *gallivm; |
||
159 | LLVMValueRef func = NULL; |
||
160 | conv_test_ptr_t conv_test_ptr; |
||
161 | boolean success; |
||
162 | const unsigned n = LP_TEST_NUM_SAMPLES; |
||
163 | int64_t cycles[LP_TEST_NUM_SAMPLES]; |
||
164 | double cycles_avg = 0.0; |
||
165 | unsigned num_srcs; |
||
166 | unsigned num_dsts; |
||
167 | double eps; |
||
168 | unsigned i, j; |
||
169 | |||
170 | if ((src_type.width >= dst_type.width && src_type.length > dst_type.length) || |
||
171 | (src_type.width <= dst_type.width && src_type.length < dst_type.length)) { |
||
172 | return TRUE; |
||
173 | } |
||
174 | |||
175 | /* Known failures |
||
176 | * - fixed point 32 -> float 32 |
||
177 | * - float 32 -> signed normalised integer 32 |
||
178 | */ |
||
179 | if ((src_type.floating && !dst_type.floating && dst_type.sign && dst_type.norm && src_type.width == dst_type.width) || |
||
180 | (!src_type.floating && dst_type.floating && src_type.fixed && src_type.width == dst_type.width)) { |
||
181 | return TRUE; |
||
182 | } |
||
183 | |||
184 | /* Known failures |
||
185 | * - fixed point 32 -> float 32 |
||
186 | * - float 32 -> signed normalised integer 32 |
||
187 | */ |
||
188 | if ((src_type.floating && !dst_type.floating && dst_type.sign && dst_type.norm && src_type.width == dst_type.width) || |
||
189 | (!src_type.floating && dst_type.floating && src_type.fixed && src_type.width == dst_type.width)) { |
||
190 | return TRUE; |
||
191 | } |
||
192 | |||
193 | if(verbose >= 1) |
||
194 | dump_conv_types(stderr, src_type, dst_type); |
||
195 | |||
196 | if (src_type.length > dst_type.length) { |
||
197 | num_srcs = 1; |
||
198 | num_dsts = src_type.length/dst_type.length; |
||
199 | } |
||
200 | else if (src_type.length < dst_type.length) { |
||
201 | num_dsts = 1; |
||
202 | num_srcs = dst_type.length/src_type.length; |
||
203 | } |
||
204 | else { |
||
205 | num_dsts = 1; |
||
206 | num_srcs = 1; |
||
207 | } |
||
208 | |||
209 | /* We must not loose or gain channels. Only precision */ |
||
210 | assert(src_type.length * num_srcs == dst_type.length * num_dsts); |
||
211 | |||
212 | eps = MAX2(lp_const_eps(src_type), lp_const_eps(dst_type)); |
||
213 | |||
214 | gallivm = gallivm_create(); |
||
215 | |||
216 | func = add_conv_test(gallivm, src_type, num_srcs, dst_type, num_dsts); |
||
217 | |||
218 | gallivm_compile_module(gallivm); |
||
219 | |||
220 | conv_test_ptr = (conv_test_ptr_t)gallivm_jit_function(gallivm, func); |
||
221 | |||
222 | success = TRUE; |
||
223 | for(i = 0; i < n && success; ++i) { |
||
224 | unsigned src_stride = src_type.length*src_type.width/8; |
||
225 | unsigned dst_stride = dst_type.length*dst_type.width/8; |
||
226 | PIPE_ALIGN_VAR(LP_MIN_VECTOR_ALIGN) uint8_t src[LP_MAX_VECTOR_LENGTH*LP_MAX_VECTOR_LENGTH]; |
||
227 | PIPE_ALIGN_VAR(LP_MIN_VECTOR_ALIGN) uint8_t dst[LP_MAX_VECTOR_LENGTH*LP_MAX_VECTOR_LENGTH]; |
||
228 | double fref[LP_MAX_VECTOR_LENGTH*LP_MAX_VECTOR_LENGTH]; |
||
229 | uint8_t ref[LP_MAX_VECTOR_LENGTH*LP_MAX_VECTOR_LENGTH]; |
||
230 | int64_t start_counter = 0; |
||
231 | int64_t end_counter = 0; |
||
232 | |||
233 | for(j = 0; j < num_srcs; ++j) { |
||
234 | random_vec(src_type, src + j*src_stride); |
||
235 | read_vec(src_type, src + j*src_stride, fref + j*src_type.length); |
||
236 | } |
||
237 | |||
238 | for(j = 0; j < num_dsts; ++j) { |
||
239 | write_vec(dst_type, ref + j*dst_stride, fref + j*dst_type.length); |
||
240 | } |
||
241 | |||
242 | start_counter = rdtsc(); |
||
243 | conv_test_ptr(src, dst); |
||
244 | end_counter = rdtsc(); |
||
245 | |||
246 | cycles[i] = end_counter - start_counter; |
||
247 | |||
248 | for(j = 0; j < num_dsts; ++j) { |
||
249 | if(!compare_vec_with_eps(dst_type, dst + j*dst_stride, ref + j*dst_stride, eps)) |
||
250 | success = FALSE; |
||
251 | } |
||
252 | |||
253 | if (!success || verbose >= 3) { |
||
254 | if(verbose < 1) |
||
255 | dump_conv_types(stderr, src_type, dst_type); |
||
256 | if (success) { |
||
257 | fprintf(stderr, "PASS\n"); |
||
258 | } |
||
259 | else { |
||
260 | fprintf(stderr, "MISMATCH\n"); |
||
261 | } |
||
262 | |||
263 | for(j = 0; j < num_srcs; ++j) { |
||
264 | fprintf(stderr, " Src%u: ", j); |
||
265 | dump_vec(stderr, src_type, src + j*src_stride); |
||
266 | fprintf(stderr, "\n"); |
||
267 | } |
||
268 | |||
269 | #if 1 |
||
270 | fprintf(stderr, " Ref: "); |
||
271 | for(j = 0; j < src_type.length*num_srcs; ++j) |
||
272 | fprintf(stderr, " %f", fref[j]); |
||
273 | fprintf(stderr, "\n"); |
||
274 | #endif |
||
275 | |||
276 | for(j = 0; j < num_dsts; ++j) { |
||
277 | fprintf(stderr, " Dst%u: ", j); |
||
278 | dump_vec(stderr, dst_type, dst + j*dst_stride); |
||
279 | fprintf(stderr, "\n"); |
||
280 | |||
281 | fprintf(stderr, " Ref%u: ", j); |
||
282 | dump_vec(stderr, dst_type, ref + j*dst_stride); |
||
283 | fprintf(stderr, "\n"); |
||
284 | } |
||
285 | } |
||
286 | } |
||
287 | |||
288 | /* |
||
289 | * Unfortunately the output of cycle counter is not very reliable as it comes |
||
290 | * -- sometimes we get outliers (due IRQs perhaps?) which are |
||
291 | * better removed to avoid random or biased data. |
||
292 | */ |
||
293 | { |
||
294 | double sum = 0.0, sum2 = 0.0; |
||
295 | double avg, std; |
||
296 | unsigned m; |
||
297 | |||
298 | for(i = 0; i < n; ++i) { |
||
299 | sum += cycles[i]; |
||
300 | sum2 += cycles[i]*cycles[i]; |
||
301 | } |
||
302 | |||
303 | avg = sum/n; |
||
304 | std = sqrtf((sum2 - n*avg*avg)/n); |
||
305 | |||
306 | m = 0; |
||
307 | sum = 0.0; |
||
308 | for(i = 0; i < n; ++i) { |
||
309 | if(fabs(cycles[i] - avg) <= 4.0*std) { |
||
310 | sum += cycles[i]; |
||
311 | ++m; |
||
312 | } |
||
313 | } |
||
314 | |||
315 | cycles_avg = sum/m; |
||
316 | |||
317 | } |
||
318 | |||
319 | if(fp) |
||
320 | write_tsv_row(fp, src_type, dst_type, cycles_avg, success); |
||
321 | |||
322 | gallivm_free_function(gallivm, func, conv_test_ptr); |
||
323 | |||
324 | gallivm_destroy(gallivm); |
||
325 | |||
326 | return success; |
||
327 | } |
||
328 | |||
329 | |||
330 | const struct lp_type conv_types[] = { |
||
331 | /* float, fixed, sign, norm, width, len */ |
||
332 | |||
333 | /* Float */ |
||
334 | { TRUE, FALSE, TRUE, TRUE, 32, 4 }, |
||
335 | { TRUE, FALSE, TRUE, FALSE, 32, 4 }, |
||
336 | { TRUE, FALSE, FALSE, TRUE, 32, 4 }, |
||
337 | { TRUE, FALSE, FALSE, FALSE, 32, 4 }, |
||
338 | |||
339 | { TRUE, FALSE, TRUE, TRUE, 32, 8 }, |
||
340 | { TRUE, FALSE, TRUE, FALSE, 32, 8 }, |
||
341 | { TRUE, FALSE, FALSE, TRUE, 32, 8 }, |
||
342 | { TRUE, FALSE, FALSE, FALSE, 32, 8 }, |
||
343 | |||
344 | /* Fixed */ |
||
345 | { FALSE, TRUE, TRUE, TRUE, 32, 4 }, |
||
346 | { FALSE, TRUE, TRUE, FALSE, 32, 4 }, |
||
347 | { FALSE, TRUE, FALSE, TRUE, 32, 4 }, |
||
348 | { FALSE, TRUE, FALSE, FALSE, 32, 4 }, |
||
349 | |||
350 | { FALSE, TRUE, TRUE, TRUE, 32, 8 }, |
||
351 | { FALSE, TRUE, TRUE, FALSE, 32, 8 }, |
||
352 | { FALSE, TRUE, FALSE, TRUE, 32, 8 }, |
||
353 | { FALSE, TRUE, FALSE, FALSE, 32, 8 }, |
||
354 | |||
355 | /* Integer */ |
||
356 | { FALSE, FALSE, TRUE, TRUE, 32, 4 }, |
||
357 | { FALSE, FALSE, TRUE, FALSE, 32, 4 }, |
||
358 | { FALSE, FALSE, FALSE, TRUE, 32, 4 }, |
||
359 | { FALSE, FALSE, FALSE, FALSE, 32, 4 }, |
||
360 | |||
361 | { FALSE, FALSE, TRUE, TRUE, 32, 8 }, |
||
362 | { FALSE, FALSE, TRUE, FALSE, 32, 8 }, |
||
363 | { FALSE, FALSE, FALSE, TRUE, 32, 8 }, |
||
364 | { FALSE, FALSE, FALSE, FALSE, 32, 8 }, |
||
365 | |||
366 | { FALSE, FALSE, TRUE, TRUE, 16, 8 }, |
||
367 | { FALSE, FALSE, TRUE, FALSE, 16, 8 }, |
||
368 | { FALSE, FALSE, FALSE, TRUE, 16, 8 }, |
||
369 | { FALSE, FALSE, FALSE, FALSE, 16, 8 }, |
||
370 | |||
371 | { FALSE, FALSE, TRUE, TRUE, 8, 16 }, |
||
372 | { FALSE, FALSE, TRUE, FALSE, 8, 16 }, |
||
373 | { FALSE, FALSE, FALSE, TRUE, 8, 16 }, |
||
374 | { FALSE, FALSE, FALSE, FALSE, 8, 16 }, |
||
375 | |||
376 | { FALSE, FALSE, TRUE, TRUE, 8, 4 }, |
||
377 | { FALSE, FALSE, TRUE, FALSE, 8, 4 }, |
||
378 | { FALSE, FALSE, FALSE, TRUE, 8, 4 }, |
||
379 | { FALSE, FALSE, FALSE, FALSE, 8, 4 }, |
||
380 | |||
381 | { FALSE, FALSE, FALSE, TRUE, 8, 8 }, |
||
382 | }; |
||
383 | |||
384 | |||
385 | const unsigned num_types = sizeof(conv_types)/sizeof(conv_types[0]); |
||
386 | |||
387 | |||
388 | boolean |
||
389 | test_all(unsigned verbose, FILE *fp) |
||
390 | { |
||
391 | const struct lp_type *src_type; |
||
392 | const struct lp_type *dst_type; |
||
393 | boolean success = TRUE; |
||
394 | int error_count = 0; |
||
395 | |||
396 | for(src_type = conv_types; src_type < &conv_types[num_types]; ++src_type) { |
||
397 | for(dst_type = conv_types; dst_type < &conv_types[num_types]; ++dst_type) { |
||
398 | |||
399 | if(src_type == dst_type) |
||
400 | continue; |
||
401 | |||
402 | if(!test_one(verbose, fp, *src_type, *dst_type)){ |
||
403 | success = FALSE; |
||
404 | ++error_count; |
||
405 | } |
||
406 | } |
||
407 | } |
||
408 | |||
409 | fprintf(stderr, "%d failures\n", error_count); |
||
410 | |||
411 | return success; |
||
412 | } |
||
413 | |||
414 | |||
415 | boolean |
||
416 | test_some(unsigned verbose, FILE *fp, |
||
417 | unsigned long n) |
||
418 | { |
||
419 | const struct lp_type *src_type; |
||
420 | const struct lp_type *dst_type; |
||
421 | unsigned long i; |
||
422 | boolean success = TRUE; |
||
423 | |||
424 | for(i = 0; i < n; ++i) { |
||
425 | src_type = &conv_types[rand() % num_types]; |
||
426 | |||
427 | do { |
||
428 | dst_type = &conv_types[rand() % num_types]; |
||
429 | } while (src_type == dst_type || src_type->norm != dst_type->norm); |
||
430 | |||
431 | if(!test_one(verbose, fp, *src_type, *dst_type)) |
||
432 | success = FALSE; |
||
433 | } |
||
434 | |||
435 | return success; |
||
436 | } |
||
437 | |||
438 | |||
439 | boolean |
||
440 | test_single(unsigned verbose, FILE *fp) |
||
441 | { |
||
442 | /* float, fixed, sign, norm, width, len */ |
||
443 | struct lp_type f32x4_type = |
||
444 | { TRUE, FALSE, TRUE, TRUE, 32, 4 }; |
||
445 | struct lp_type ub8x4_type = |
||
446 | { FALSE, FALSE, FALSE, TRUE, 8, 16 }; |
||
447 | |||
448 | boolean success; |
||
449 | |||
450 | success = test_one(verbose, fp, f32x4_type, ub8x4_type); |
||
451 | |||
452 | return success; |
||
453 | }>>>=>>>>>>>>>>>>>=>>> |