1/**************************************************************************
2 *
3 * Copyright 2011 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#include <limits.h>
30#include <stdio.h>
31#include <stdlib.h>
32
33#include "util/u_pointer.h"
34#include "util/u_memory.h"
35#include "util/u_math.h"
36#include "util/u_cpu_detect.h"
37
38#include "gallivm/lp_bld.h"
39#include "gallivm/lp_bld_debug.h"
40#include "gallivm/lp_bld_init.h"
41#include "gallivm/lp_bld_arit.h"
42
43#include "lp_test.h"
44
45
46void
47write_tsv_header(FILE *fp)
48{
49   fprintf(fp,
50           "result\t"
51           "format\n");
52
53   fflush(fp);
54}
55
56
57typedef void (*unary_func_t)(float *out, const float *in);
58
59
60/**
61 * Describe a test case of one unary function.
62 */
63struct unary_test_t
64{
65   /*
66    * Test name -- name of the mathematical function under test.
67    */
68
69   const char *name;
70
71   LLVMValueRef
72   (*builder)(struct lp_build_context *bld, LLVMValueRef a);
73
74   /*
75    * Reference (pure-C) function.
76    */
77   float
78   (*ref)(float a);
79
80   /*
81    * Test values.
82    */
83   const float *values;
84   unsigned num_values;
85
86   /*
87    * Required precision in bits.
88    */
89   double precision;
90};
91
92
93static float negf(float x)
94{
95   return -x;
96}
97
98
99static float sgnf(float x)
100{
101   if (x > 0.0f) {
102      return 1.0f;
103   }
104   if (x < 0.0f) {
105      return -1.0f;
106   }
107   return 0.0f;
108}
109
110
111const float sgn_values[] = {
112   -INFINITY,
113   -60,
114   -4,
115   -2,
116   -1,
117   -1e-007,
118   0,
119   1e-007,
120   0.01,
121   0.1,
122   0.9,
123   0.99,
124   1,
125   2,
126   4,
127   60,
128   INFINITY,
129   NAN
130};
131
132
133const float exp2_values[] = {
134   -INFINITY,
135   -60,
136   -4,
137   -2,
138   -1,
139   -1e-007,
140   0,
141   1e-007,
142   0.01,
143   0.1,
144   0.9,
145   0.99,
146   1,
147   2,
148   4,
149   60,
150   INFINITY,
151   NAN
152};
153
154
155const float log2_values[] = {
156#if 0
157   /*
158    * Smallest denormalized number; meant just for experimentation, but not
159    * validation.
160    */
161   1.4012984643248171e-45,
162#endif
163   -INFINITY,
164   0,
165   1e-007,
166   0.1,
167   0.5,
168   0.99,
169   1,
170   1.01,
171   1.1,
172   1.9,
173   1.99,
174   2,
175   4,
176   100000,
177   1e+018,
178   INFINITY,
179   NAN
180};
181
182
183static float rcpf(float x)
184{
185   return 1.0/x;
186}
187
188
189const float rcp_values[] = {
190   -0.0, 0.0,
191   -1.0, 1.0,
192   -1e-007, 1e-007,
193   -4.0, 4.0,
194   -1e+035, -100000,
195   100000, 1e+035,
196   5.88e-39f, // denormal
197#if (__STDC_VERSION__ >= 199901L)
198   INFINITY, -INFINITY,
199#endif
200};
201
202
203static float rsqrtf(float x)
204{
205   return 1.0/(float)sqrt(x);
206}
207
208
209const float rsqrt_values[] = {
210   // http://msdn.microsoft.com/en-us/library/windows/desktop/bb147346.aspx
211   0.0, // must yield infinity
212   1.0, // must yield 1.0
213   1e-007, 4.0,
214   100000, 1e+035,
215   5.88e-39f, // denormal
216#if (__STDC_VERSION__ >= 199901L)
217   INFINITY,
218#endif
219};
220
221
222const float sincos_values[] = {
223   -INFINITY,
224   -5*M_PI/4,
225   -4*M_PI/4,
226   -4*M_PI/4,
227   -3*M_PI/4,
228   -2*M_PI/4,
229   -1*M_PI/4,
230   1*M_PI/4,
231   2*M_PI/4,
232   3*M_PI/4,
233   4*M_PI/4,
234   5*M_PI/4,
235   INFINITY,
236   NAN
237};
238
239const float round_values[] = {
240      -10.0, -1, 0.0, 12.0,
241      -1.49, -0.25, 1.25, 2.51,
242      -0.99, -0.01, 0.01, 0.99,
243      -1.5, -0.5, 0.5, 1.5,
244      1.401298464324817e-45f, // smallest denormal
245      -1.401298464324817e-45f,
246      1.62981451e-08f,
247      -1.62981451e-08f,
248      1.62981451e15f, // large number not representable as 32bit int
249      -1.62981451e15f,
250      FLT_EPSILON,
251      -FLT_EPSILON,
252      1.0f - 0.5f*FLT_EPSILON,
253      -1.0f + FLT_EPSILON,
254      FLT_MAX,
255      -FLT_MAX
256};
257
258static float fractf(float x)
259{
260   x -= floorf(x);
261   if (x >= 1.0f) {
262      // clamp to the largest number smaller than one
263      x = 1.0f - 0.5f*FLT_EPSILON;
264   }
265   return x;
266}
267
268
269const float fract_values[] = {
270   // http://en.wikipedia.org/wiki/IEEE_754-1985#Examples
271   0.0f,
272   -0.0f,
273   1.0f,
274   -1.0f,
275   0.5f,
276   -0.5f,
277   1.401298464324817e-45f, // smallest denormal
278   -1.401298464324817e-45f,
279   5.88e-39f, // middle denormal
280   1.18e-38f, // largest denormal
281   -1.18e-38f,
282   -1.62981451e-08f,
283   FLT_EPSILON,
284   -FLT_EPSILON,
285   1.0f - 0.5f*FLT_EPSILON,
286   -1.0f + FLT_EPSILON,
287   FLT_MAX,
288   -FLT_MAX
289};
290
291
292/*
293 * Unary test cases.
294 */
295
296#ifdef _MSC_VER
297#define WRAP(func) \
298static float \
299wrap_ ## func(float x) \
300{ \
301   return func(x); \
302}
303WRAP(expf)
304WRAP(logf)
305WRAP(sinf)
306WRAP(cosf)
307WRAP(floorf)
308WRAP(ceilf)
309#define expf wrap_expf
310#define logf wrap_logf
311#define sinf wrap_sinf
312#define cosf wrap_cosf
313#define floorf wrap_floorf
314#define ceilf wrap_ceilf
315#endif
316
317static const struct unary_test_t
318unary_tests[] = {
319   {"abs", &lp_build_abs, &fabsf, sgn_values, ARRAY_SIZE(sgn_values), 20.0 },
320   {"neg", &lp_build_negate, &negf, sgn_values, ARRAY_SIZE(sgn_values), 20.0 },
321   {"sgn", &lp_build_sgn, &sgnf, sgn_values, ARRAY_SIZE(sgn_values), 20.0 },
322   {"exp2", &lp_build_exp2, &exp2f, exp2_values, ARRAY_SIZE(exp2_values), 18.0 },
323   {"log2", &lp_build_log2_safe, &log2f, log2_values, ARRAY_SIZE(log2_values), 20.0 },
324   {"exp", &lp_build_exp, &expf, exp2_values, ARRAY_SIZE(exp2_values), 18.0 },
325   {"log", &lp_build_log_safe, &logf, log2_values, ARRAY_SIZE(log2_values), 20.0 },
326   {"rcp", &lp_build_rcp, &rcpf, rcp_values, ARRAY_SIZE(rcp_values), 20.0 },
327   {"rsqrt", &lp_build_rsqrt, &rsqrtf, rsqrt_values, ARRAY_SIZE(rsqrt_values), 20.0 },
328   {"sin", &lp_build_sin, &sinf, sincos_values, ARRAY_SIZE(sincos_values), 20.0 },
329   {"cos", &lp_build_cos, &cosf, sincos_values, ARRAY_SIZE(sincos_values), 20.0 },
330   {"sgn", &lp_build_sgn, &sgnf, sgn_values, ARRAY_SIZE(sgn_values), 20.0 },
331   {"round", &lp_build_round, &nearbyintf, round_values, ARRAY_SIZE(round_values), 24.0 },
332   {"trunc", &lp_build_trunc, &truncf, round_values, ARRAY_SIZE(round_values), 24.0 },
333   {"floor", &lp_build_floor, &floorf, round_values, ARRAY_SIZE(round_values), 24.0 },
334   {"ceil", &lp_build_ceil, &ceilf, round_values, ARRAY_SIZE(round_values), 24.0 },
335   {"fract", &lp_build_fract_safe, &fractf, fract_values, ARRAY_SIZE(fract_values), 24.0 },
336};
337
338
339/*
340 * Build LLVM function that exercises the unary operator builder.
341 */
342static LLVMValueRef
343build_unary_test_func(struct gallivm_state *gallivm,
344                      const struct unary_test_t *test,
345                      unsigned length,
346                      const char *test_name)
347{
348   struct lp_type type = lp_type_float_vec(32, length * 32);
349   LLVMContextRef context = gallivm->context;
350   LLVMModuleRef module = gallivm->module;
351   LLVMTypeRef vf32t = lp_build_vec_type(gallivm, type);
352   LLVMTypeRef args[2] = { LLVMPointerType(vf32t, 0), LLVMPointerType(vf32t, 0) };
353   LLVMValueRef func = LLVMAddFunction(module, test_name,
354                                       LLVMFunctionType(LLVMVoidTypeInContext(context),
355                                                        args, ARRAY_SIZE(args), 0));
356   LLVMValueRef arg0 = LLVMGetParam(func, 0);
357   LLVMValueRef arg1 = LLVMGetParam(func, 1);
358   LLVMBuilderRef builder = gallivm->builder;
359   LLVMBasicBlockRef block = LLVMAppendBasicBlockInContext(context, func, "entry");
360   LLVMValueRef ret;
361
362   struct lp_build_context bld;
363
364   lp_build_context_init(&bld, gallivm, type);
365
366   LLVMSetFunctionCallConv(func, LLVMCCallConv);
367
368   LLVMPositionBuilderAtEnd(builder, block);
369
370   arg1 = LLVMBuildLoad(builder, arg1, "");
371
372   ret = test->builder(&bld, arg1);
373
374   LLVMBuildStore(builder, ret, arg0);
375
376   LLVMBuildRetVoid(builder);
377
378   gallivm_verify_function(gallivm, func);
379
380   return func;
381}
382
383
384/*
385 * Flush denorms to zero.
386 */
387static float
388flush_denorm_to_zero(float val)
389{
390   /*
391    * If we have a denorm manually set it to (+-)0.
392    * This is because the reference may or may not do the right thing
393    * otherwise because we want the result according to treating all
394    * denormals as zero (FTZ/DAZ). Not using fpclassify because
395    * a) some compilers are stuck at c89 (msvc)
396    * b) not sure it reliably works with non-standard ftz/daz mode
397    * And, right now we only disable denorms with jited code on x86/sse
398    * (albeit this should be classified as a bug) so to get results which
399    * match we must only flush them to zero here in that case too.
400    */
401   union fi fi_val;
402
403   fi_val.f = val;
404
405#if defined(PIPE_ARCH_SSE)
406   if (util_get_cpu_caps()->has_sse) {
407      if ((fi_val.ui & 0x7f800000) == 0) {
408         fi_val.ui &= 0xff800000;
409      }
410   }
411#endif
412
413   return fi_val.f;
414}
415
416/*
417 * Test one LLVM unary arithmetic builder function.
418 */
419static boolean
420test_unary(unsigned verbose, FILE *fp, const struct unary_test_t *test, unsigned length)
421{
422   char test_name[128];
423   snprintf(test_name, sizeof test_name, "%s.v%u", test->name, length);
424   LLVMContextRef context;
425   struct gallivm_state *gallivm;
426   LLVMValueRef test_func;
427   unary_func_t test_func_jit;
428   boolean success = TRUE;
429   int i, j;
430   float *in, *out;
431
432   in = align_malloc(length * 4, length * 4);
433   out = align_malloc(length * 4, length * 4);
434
435   /* random NaNs or 0s could wreak havoc */
436   for (i = 0; i < length; i++) {
437      in[i] = 1.0;
438   }
439
440   context = LLVMContextCreate();
441   gallivm = gallivm_create("test_module", context, NULL);
442
443   test_func = build_unary_test_func(gallivm, test, length, test_name);
444
445   gallivm_compile_module(gallivm);
446
447   test_func_jit = (unary_func_t) gallivm_jit_function(gallivm, test_func);
448
449   gallivm_free_ir(gallivm);
450
451   for (j = 0; j < (test->num_values + length - 1) / length; j++) {
452      int num_vals = ((j + 1) * length <= test->num_values) ? length :
453                                                              test->num_values % length;
454
455      for (i = 0; i < num_vals; ++i) {
456         in[i] = test->values[i+j*length];
457      }
458
459      test_func_jit(out, in);
460      for (i = 0; i < num_vals; ++i) {
461         float testval, ref;
462         double error, precision;
463         boolean expected_pass = TRUE;
464         bool pass;
465
466         testval = flush_denorm_to_zero(in[i]);
467         ref = flush_denorm_to_zero(test->ref(testval));
468
469         if (util_inf_sign(ref) && util_inf_sign(out[i]) == util_inf_sign(ref)) {
470            error = 0;
471         } else {
472            error = fabs(out[i] - ref);
473         }
474         precision = error ? -log2(error/fabs(ref)) : FLT_MANT_DIG;
475
476         pass = precision >= test->precision;
477
478         if (isnan(ref)) {
479            continue;
480         }
481
482         if (!util_get_cpu_caps()->has_neon &&
483             test->ref == &nearbyintf && length == 2 &&
484             ref != roundf(testval)) {
485            /* FIXME: The generic (non SSE) path in lp_build_iround, which is
486             * always taken for length==2 regardless of native round support,
487             * does not round to even. */
488            expected_pass = FALSE;
489         }
490
491         if (test->ref == &expf && util_inf_sign(testval) == -1) {
492            /* Some older 64-bit MSVCRT versions return -inf instead of 0
493	     * for expf(-inf). As detecting the VC runtime version is
494	     * non-trivial, just ignore the test result. */
495#if defined(_MSC_VER) && defined(_WIN64)
496            expected_pass = pass;
497#endif
498         }
499
500         if (pass != expected_pass || verbose) {
501            printf("%s(%.9g): ref = %.9g, out = %.9g, precision = %f bits, %s%s\n",
502                  test_name, in[i], ref, out[i], precision,
503                  pass ? "PASS" : "FAIL",
504                  !expected_pass ? (pass ? " (unexpected)" : " (expected)" ): "");
505            fflush(stdout);
506         }
507
508         if (pass != expected_pass) {
509            success = FALSE;
510         }
511      }
512   }
513
514   gallivm_destroy(gallivm);
515   LLVMContextDispose(context);
516
517   align_free(in);
518   align_free(out);
519
520   return success;
521}
522
523
524boolean
525test_all(unsigned verbose, FILE *fp)
526{
527   boolean success = TRUE;
528   int i;
529
530   for (i = 0; i < ARRAY_SIZE(unary_tests); ++i) {
531      unsigned max_length = lp_native_vector_width / 32;
532      unsigned length;
533      for (length = 1; length <= max_length; length *= 2) {
534         if (!test_unary(verbose, fp, &unary_tests[i], length)) {
535            success = FALSE;
536         }
537      }
538   }
539
540   return success;
541}
542
543
544boolean
545test_some(unsigned verbose, FILE *fp,
546          unsigned long n)
547{
548   /*
549    * Not randomly generated test cases, so test all.
550    */
551
552   return test_all(verbose, fp);
553}
554
555
556boolean
557test_single(unsigned verbose, FILE *fp)
558{
559   return TRUE;
560}
561