• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 <math.h>
31 #include <stdio.h>
32 #include <stdlib.h>
33 
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_init.h"
40 #include "gallivm/lp_bld_arit.h"
41 
42 #include "lp_test.h"
43 
44 
45 void
write_tsv_header(FILE * fp)46 write_tsv_header(FILE *fp)
47 {
48    fprintf(fp,
49            "result\t"
50            "format\n");
51 
52    fflush(fp);
53 }
54 
55 
56 typedef void (*unary_func_t)(float *out, const float *in);
57 
58 
59 /**
60  * Describe a test case of one unary function.
61  */
62 struct unary_test_t
63 {
64    /*
65     * Test name -- name of the mathematical function under test.
66     */
67 
68    const char *name;
69 
70    LLVMValueRef
71    (*builder)(struct lp_build_context *bld, LLVMValueRef a);
72 
73    /*
74     * Reference (pure-C) function.
75     */
76    float
77    (*ref)(float a);
78 
79    /*
80     * Test values.
81     */
82    const float *values;
83    unsigned num_values;
84 
85    /*
86     * Required precision in bits.
87     */
88    double precision;
89 };
90 
91 
negf(float x)92 static float negf(float x)
93 {
94    return -x;
95 }
96 
97 
sgnf(float x)98 static float sgnf(float x)
99 {
100    if (x > 0.0f) {
101       return 1.0f;
102    }
103    if (x < 0.0f) {
104       return -1.0f;
105    }
106    return 0.0f;
107 }
108 
109 
110 const float sgn_values[] = {
111    -INFINITY,
112    -60,
113    -4,
114    -2,
115    -1,
116    -1e-007,
117    0,
118    1e-007,
119    0.01,
120    0.1,
121    0.9,
122    0.99,
123    1,
124    2,
125    4,
126    60,
127    INFINITY,
128    NAN
129 };
130 
131 
132 const float exp2_values[] = {
133    -INFINITY,
134    -60,
135    -4,
136    -2,
137    -1,
138    -1e-007,
139    0,
140    1e-007,
141    0.01,
142    0.1,
143    0.9,
144    0.99,
145    1,
146    2,
147    4,
148    60,
149    INFINITY,
150    NAN
151 };
152 
153 
154 const float log2_values[] = {
155 #if 0
156    /*
157     * Smallest denormalized number; meant just for experimentation, but not
158     * validation.
159     */
160    1.4012984643248171e-45,
161 #endif
162    -INFINITY,
163    0,
164    1e-007,
165    0.1,
166    0.5,
167    0.99,
168    1,
169    1.01,
170    1.1,
171    1.9,
172    1.99,
173    2,
174    4,
175    100000,
176    1e+018,
177    INFINITY,
178    NAN
179 };
180 
181 
rcpf(float x)182 static float rcpf(float x)
183 {
184    return 1.0/x;
185 }
186 
187 
188 const float rcp_values[] = {
189    -0.0, 0.0,
190    -1.0, 1.0,
191    -1e-007, 1e-007,
192    -4.0, 4.0,
193    -1e+035, -100000,
194    100000, 1e+035,
195    5.88e-39f, // denormal
196    INFINITY, -INFINITY,
197 };
198 
199 
rsqrtf(float x)200 static float rsqrtf(float x)
201 {
202    return 1.0/(float)sqrt(x);
203 }
204 
205 
206 const float rsqrt_values[] = {
207    // http://msdn.microsoft.com/en-us/library/windows/desktop/bb147346.aspx
208    0.0, // must yield infinity
209    1.0, // must yield 1.0
210    1e-007, 4.0,
211    100000, 1e+035,
212    5.88e-39f, // denormal
213    INFINITY,
214 };
215 
216 
217 const float sincos_values[] = {
218    -INFINITY,
219    -5*M_PI/4,
220    -4*M_PI/4,
221    -4*M_PI/4,
222    -3*M_PI/4,
223    -2*M_PI/4,
224    -1*M_PI/4,
225    1*M_PI/4,
226    2*M_PI/4,
227    3*M_PI/4,
228    4*M_PI/4,
229    5*M_PI/4,
230    INFINITY,
231    NAN
232 };
233 
234 const float round_values[] = {
235       -10.0, -1, 0.0, 12.0,
236       -1.49, -0.25, 1.25, 2.51,
237       -0.99, -0.01, 0.01, 0.99,
238       -1.5, -0.5, 0.5, 1.5,
239       1.401298464324817e-45f, // smallest denormal
240       -1.401298464324817e-45f,
241       1.62981451e-08f,
242       -1.62981451e-08f,
243       1.62981451e15f, // large number not representable as 32bit int
244       -1.62981451e15f,
245       FLT_EPSILON,
246       -FLT_EPSILON,
247       1.0f - 0.5f*FLT_EPSILON,
248       -1.0f + FLT_EPSILON,
249       FLT_MAX,
250       -FLT_MAX
251 };
252 
fractf(float x)253 static float fractf(float x)
254 {
255    x -= floorf(x);
256    if (x >= 1.0f) {
257       // clamp to the largest number smaller than one
258       x = 1.0f - 0.5f*FLT_EPSILON;
259    }
260    return x;
261 }
262 
263 
264 const float fract_values[] = {
265    // http://en.wikipedia.org/wiki/IEEE_754-1985#Examples
266    0.0f,
267    -0.0f,
268    1.0f,
269    -1.0f,
270    0.5f,
271    -0.5f,
272    1.401298464324817e-45f, // smallest denormal
273    -1.401298464324817e-45f,
274    5.88e-39f, // middle denormal
275    1.18e-38f, // largest denormal
276    -1.18e-38f,
277    -1.62981451e-08f,
278    FLT_EPSILON,
279    -FLT_EPSILON,
280    1.0f - 0.5f*FLT_EPSILON,
281    -1.0f + FLT_EPSILON,
282    FLT_MAX,
283    -FLT_MAX
284 };
285 
286 
287 /*
288  * Unary test cases.
289  */
290 
291 #ifdef _MSC_VER
292 #define WRAP(func) \
293 static float \
294 wrap_ ## func(float x) \
295 { \
296    return func(x); \
297 }
298 WRAP(log2f)
299 WRAP(expf)
300 WRAP(logf)
301 WRAP(sinf)
302 WRAP(cosf)
303 WRAP(nearbyintf)
304 WRAP(floorf)
305 WRAP(ceilf)
306 #define log2f wrap_log2f
307 #define expf wrap_expf
308 #define logf wrap_logf
309 #define sinf wrap_sinf
310 #define cosf wrap_cosf
311 #define nearbyintf wrap_nearbyintf
312 #define floorf wrap_floorf
313 #define ceilf wrap_ceilf
314 #endif
315 
316 static const struct unary_test_t
317 unary_tests[] = {
318    {"abs", &lp_build_abs, &fabsf, sgn_values, ARRAY_SIZE(sgn_values), 20.0 },
319    {"neg", &lp_build_negate, &negf, sgn_values, ARRAY_SIZE(sgn_values), 20.0 },
320    {"sgn", &lp_build_sgn, &sgnf, sgn_values, ARRAY_SIZE(sgn_values), 20.0 },
321    {"exp2", &lp_build_exp2, &exp2f, exp2_values, ARRAY_SIZE(exp2_values), 18.0 },
322    {"log2", &lp_build_log2_safe, &log2f, log2_values, ARRAY_SIZE(log2_values), 20.0 },
323    {"exp", &lp_build_exp, &expf, exp2_values, ARRAY_SIZE(exp2_values), 18.0 },
324    {"log", &lp_build_log_safe, &logf, log2_values, ARRAY_SIZE(log2_values), 20.0 },
325    {"rcp", &lp_build_rcp, &rcpf, rcp_values, ARRAY_SIZE(rcp_values), 20.0 },
326    {"rsqrt", &lp_build_rsqrt, &rsqrtf, rsqrt_values, ARRAY_SIZE(rsqrt_values), 20.0 },
327    {"sin", &lp_build_sin, &sinf, sincos_values, ARRAY_SIZE(sincos_values), 20.0 },
328    {"cos", &lp_build_cos, &cosf, sincos_values, ARRAY_SIZE(sincos_values), 20.0 },
329    {"sgn", &lp_build_sgn, &sgnf, sgn_values, ARRAY_SIZE(sgn_values), 20.0 },
330    {"round", &lp_build_round, &nearbyintf, round_values, ARRAY_SIZE(round_values), 24.0 },
331    {"trunc", &lp_build_trunc, &truncf, round_values, ARRAY_SIZE(round_values), 24.0 },
332    {"floor", &lp_build_floor, &floorf, round_values, ARRAY_SIZE(round_values), 24.0 },
333    {"ceil", &lp_build_ceil, &ceilf, round_values, ARRAY_SIZE(round_values), 24.0 },
334    {"fract", &lp_build_fract_safe, &fractf, fract_values, ARRAY_SIZE(fract_values), 24.0 },
335 };
336 
337 
338 /*
339  * Build LLVM function that exercises the unary operator builder.
340  */
341 static LLVMValueRef
build_unary_test_func(struct gallivm_state * gallivm,const struct unary_test_t * test,unsigned length,const char * test_name)342 build_unary_test_func(struct gallivm_state *gallivm,
343                       const struct unary_test_t *test,
344                       unsigned length,
345                       const char *test_name)
346 {
347    struct lp_type type = lp_type_float_vec(32, length * 32);
348    LLVMContextRef context = gallivm->context;
349    LLVMModuleRef module = gallivm->module;
350    LLVMTypeRef vf32t = lp_build_vec_type(gallivm, type);
351    LLVMTypeRef args[2] = { LLVMPointerType(vf32t, 0), LLVMPointerType(vf32t, 0) };
352    LLVMValueRef func = LLVMAddFunction(module, test_name,
353                                        LLVMFunctionType(LLVMVoidTypeInContext(context),
354                                                         args, ARRAY_SIZE(args), 0));
355    LLVMValueRef arg0 = LLVMGetParam(func, 0);
356    LLVMValueRef arg1 = LLVMGetParam(func, 1);
357    LLVMBuilderRef builder = gallivm->builder;
358    LLVMBasicBlockRef block = LLVMAppendBasicBlockInContext(context, func, "entry");
359    LLVMValueRef ret;
360 
361    struct lp_build_context bld;
362 
363    lp_build_context_init(&bld, gallivm, type);
364 
365    LLVMSetFunctionCallConv(func, LLVMCCallConv);
366 
367    LLVMPositionBuilderAtEnd(builder, block);
368 
369    arg1 = LLVMBuildLoad2(builder, vf32t, arg1, "");
370 
371    ret = test->builder(&bld, arg1);
372 
373    LLVMBuildStore(builder, ret, arg0);
374 
375    LLVMBuildRetVoid(builder);
376 
377    gallivm_verify_function(gallivm, func);
378 
379    return func;
380 }
381 
382 
383 /*
384  * Flush denorms to zero.
385  */
386 static float
flush_denorm_to_zero(float val)387 flush_denorm_to_zero(float val)
388 {
389    /*
390     * If we have a denorm manually set it to (+-)0.
391     * This is because the reference may or may not do the right thing
392     * otherwise because we want the result according to treating all
393     * denormals as zero (FTZ/DAZ). Not using fpclassify because
394     * a) some compilers are stuck at c89 (msvc)
395     * b) not sure it reliably works with non-standard ftz/daz mode
396     * And, right now we only disable denorms with jited code on x86/sse
397     * (albeit this should be classified as a bug) so to get results which
398     * match we must only flush them to zero here in that case too.
399     */
400    union fi fi_val;
401 
402    fi_val.f = val;
403 
404 #if DETECT_ARCH_SSE
405    if (util_get_cpu_caps()->has_sse) {
406       if ((fi_val.ui & 0x7f800000) == 0) {
407          fi_val.ui &= 0xff800000;
408       }
409    }
410 #endif
411 
412    return fi_val.f;
413 }
414 
415 /*
416  * Test one LLVM unary arithmetic builder function.
417  */
418 static bool
test_unary(unsigned verbose,FILE * fp,const struct unary_test_t * test,unsigned length)419 test_unary(unsigned verbose, FILE *fp, const struct unary_test_t *test, unsigned length)
420 {
421    char test_name[128];
422    snprintf(test_name, sizeof test_name, "%s.v%u", test->name, length);
423    lp_context_ref context;
424    struct gallivm_state *gallivm;
425    LLVMValueRef test_func;
426    unary_func_t test_func_jit;
427    bool success = true;
428    int i, j;
429    float *in, *out;
430 
431    in = align_malloc(length * 4, length * 4);
432    out = align_malloc(length * 4, length * 4);
433 
434    /* random NaNs or 0s could wreak havoc */
435    for (i = 0; i < length; i++) {
436       in[i] = 1.0;
437    }
438 
439    lp_context_create(&context);
440    gallivm = gallivm_create("test_module", &context, NULL);
441 
442    test_func = build_unary_test_func(gallivm, test, length, test_name);
443 
444    gallivm_compile_module(gallivm);
445 
446    test_func_jit = (unary_func_t) gallivm_jit_function(gallivm, test_func, test_name);
447 
448    gallivm_free_ir(gallivm);
449 
450    for (j = 0; j < (test->num_values + length - 1) / length; j++) {
451       int num_vals = ((j + 1) * length <= test->num_values) ? length :
452                                                               test->num_values % length;
453 
454       for (i = 0; i < num_vals; ++i) {
455          in[i] = test->values[i+j*length];
456       }
457 
458       test_func_jit(out, in);
459       for (i = 0; i < num_vals; ++i) {
460          float testval, ref;
461          double error, precision;
462          bool expected_pass = true;
463          bool pass;
464 
465          testval = flush_denorm_to_zero(in[i]);
466          ref = flush_denorm_to_zero(test->ref(testval));
467 
468          if (util_inf_sign(ref) && util_inf_sign(out[i]) == util_inf_sign(ref)) {
469             error = 0;
470          } else {
471             error = fabs(out[i] - ref);
472          }
473          precision = error ? -log2(error/fabs(ref)) : FLT_MANT_DIG;
474 
475          pass = precision >= test->precision;
476 
477          if (isnan(ref)) {
478             continue;
479          }
480 
481          if (test->ref == &nearbyintf && length == 2 &&
482              !util_get_cpu_caps()->has_neon &&
483              util_get_cpu_caps()->family != CPU_S390X &&
484              !(util_get_cpu_caps()->has_sse4_1 && LLVM_VERSION_MAJOR >= 8) &&
485              ref != roundf(testval)) {
486             /* FIXME: The generic (non SSE) path in lp_build_iround, which is
487              * always taken for length==2 regardless of native round support,
488              * does not round to even. */
489             expected_pass = false;
490          }
491 
492          if (test->ref == &expf && util_inf_sign(testval) == -1) {
493             /* Some older 64-bit MSVCRT versions return -inf instead of 0
494             * for expf(-inf). As detecting the VC runtime version is
495             * non-trivial, just ignore the test result. */
496 #if defined(_MSC_VER) && defined(_WIN64)
497             expected_pass = pass;
498 #endif
499          }
500 
501          if (pass != expected_pass || verbose) {
502             printf("%s(%.9g): ref = %.9g, out = %.9g, precision = %f bits, %s%s\n",
503                   test_name, in[i], ref, out[i], precision,
504                   pass ? "PASS" : "FAIL",
505                   !expected_pass ? (pass ? " (unexpected)" : " (expected)" ): "");
506             fflush(stdout);
507          }
508 
509          if (pass != expected_pass) {
510             success = false;
511          }
512       }
513    }
514 
515    gallivm_destroy(gallivm);
516    lp_context_destroy(&context);
517 
518    align_free(in);
519    align_free(out);
520 
521    return success;
522 }
523 
524 
525 bool
test_all(unsigned verbose,FILE * fp)526 test_all(unsigned verbose, FILE *fp)
527 {
528    bool success = true;
529    int i;
530 
531    for (i = 0; i < ARRAY_SIZE(unary_tests); ++i) {
532       unsigned max_length = lp_native_vector_width / 32;
533       unsigned length;
534       for (length = 1; length <= max_length; length *= 2) {
535          if (!test_unary(verbose, fp, &unary_tests[i], length)) {
536             success = false;
537          }
538       }
539    }
540 
541    return success;
542 }
543 
544 
545 bool
test_some(unsigned verbose,FILE * fp,unsigned long n)546 test_some(unsigned verbose, FILE *fp,
547           unsigned long n)
548 {
549    /*
550     * Not randomly generated test cases, so test all.
551     */
552 
553    return test_all(verbose, fp);
554 }
555 
556 
557 bool
test_single(unsigned verbose,FILE * fp)558 test_single(unsigned verbose, FILE *fp)
559 {
560    return true;
561 }
562