1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
|
#include "Halide.h"
#include <iostream>
#include <limits>
#include <math.h>
#include <stdio.h>
using namespace Halide;
namespace {
static int num_errors = 0;
template<typename T, typename std::enable_if<std::is_floating_point<T>::value>::type * = nullptr>
bool is_equal(T a, T b) {
if (std::isnan(a) && std::isnan(b)) {
return true;
} else {
return a == b;
}
}
template<typename T, typename std::enable_if<!std::is_floating_point<T>::value>::type * = nullptr>
bool is_equal(T a, T b) {
return a == b;
}
template<typename value_t>
bool relatively_equal(value_t a, value_t b, Target target) {
if (is_equal(a, b)) {
return true;
} else if (!std::numeric_limits<value_t>::is_integer) {
double da = (double)a, db = (double)b;
double relative_error;
// This test seems a bit high.
if (fabs(db - da) < .0001) {
return true;
}
if (fabs(da) > fabs(db)) {
relative_error = fabs((db - da) / da);
} else {
relative_error = fabs((db - da) / db);
}
if (relative_error < .00000125) {
return true;
}
// For HLSL, try again with a lower error threshold, as it might be using
// fast but approximated trigonometric functions:
if (target.supports_device_api(DeviceAPI::D3D12Compute) ||
target.supports_device_api(DeviceAPI::WebGPU)) {
// this threshold value has been empirically determined since there
// is no clear documentation on the precision of these algorithms
const double threshold = .001023;
if (relative_error < threshold) {
std::cout << "relatively_equal: relaxed threshold for (" << a << ", " << b << ") "
<< "with relative error " << relative_error
<< " (shader fast trig)\n";
return true;
}
}
std::cerr
<< "relatively_equal failed for (" << a << ", " << b
<< ") with relative error " << relative_error << "\n";
} else {
std::cerr << "relatively_equal failed for (" << (double)a << ", " << (double)b << ")\n";
}
return false;
}
float absd(float a, float b) {
return a < b ? b - a : a - b;
}
double absd(double a, double b) {
return a < b ? b - a : a - b;
}
uint8_t absd(int8_t a, int8_t b) {
return a < b ? b - a : a - b;
}
uint16_t absd(int16_t a, int16_t b) {
return a < b ? b - a : a - b;
}
uint32_t absd(int32_t a, int32_t b) {
return a < b ? b - a : a - b;
}
uint8_t absd(uint8_t a, uint8_t b) {
return a < b ? b - a : a - b;
}
uint16_t absd(uint16_t a, uint16_t b) {
return a < b ? b - a : a - b;
}
uint32_t absd(uint32_t a, uint32_t b) {
return a < b ? b - a : a - b;
}
template<typename T>
struct TestArgs {
Buffer<T> data;
TestArgs(int steps, T start, T end)
: data(steps) {
for (int i = 0; i < steps; i++) {
data(i) = (T)((double)start + i * ((double)end - start) / steps);
}
}
TestArgs(int steps,
T start_x, T end_x,
T start_y, T end_y)
: data(2, steps) {
for (int i = 0; i < steps; i++) {
data(0, i) = (T)((double)start_x + i * ((double)end_x - start_x) / steps);
data(1, i) = (T)((double)start_y + i * ((double)end_y - start_y) / steps);
}
}
};
// Using macros to expand name as both a C function and an Expr fragment.
// It may well be possible to do this without macros, but that is left
// for another day.
// Version for a one argument function.
#define fun_1(type_ret, type, name, c_name) \
void test_##type##_##name(Buffer<type> in) { \
Type type_of_type = type_of<type>(); \
Target target = get_jit_target_from_environment(); \
if (!target.supports_type(type_of_type)) { \
return; \
} \
if (target.has_feature(Target::Vulkan) && (type_of_type.is_float() && type_of_type.bits() > 32)) { \
return; \
} \
Func test_##name("test_" #name); \
Var x("x"), xi("xi"); \
test_##name(x) = name(in(x)); \
if (target.has_gpu_feature()) { \
test_##name.gpu_tile(x, xi, 16).vectorize(xi, 2); \
} else if (target.has_feature(Target::HVX)) { \
test_##name.hexagon(); \
} \
Buffer<type_ret> result = test_##name.realize({in.extent(0)}, target); \
for (int i = 0; i < in.extent(0); i++) { \
type_ret c_result = c_name(in(i)); \
if (!relatively_equal(c_result, result(i), target)) { \
fprintf(stderr, "For " #name "(%.20f) == %.20f from C and %.20f from %s.\n", \
(double)in(i), (double)c_result, (double)result(i), \
target.to_string().c_str()); \
num_errors++; \
} \
} \
}
// Version for a two argument function
#define fun_2(type_ret, type, name, c_name) \
void test_##type##_##name(Buffer<type> in) { \
Type type_of_type = type_of<type>(); \
Target target = get_jit_target_from_environment(); \
if (!target.supports_type(type_of_type)) { \
return; \
} \
if (target.has_feature(Target::Vulkan) && (type_of_type.is_float() && type_of_type.bits() > 32)) { \
return; \
} \
Func test_##name("test_" #name); \
Var x("x"), xi("xi"); \
test_##name(x) = name(in(0, x), in(1, x)); \
if (target.has_gpu_feature()) { \
test_##name.gpu_tile(x, xi, 16).vectorize(xi, 2); \
} else if (target.has_feature(Target::HVX)) { \
test_##name.hexagon(); \
} \
Buffer<type_ret> result = test_##name.realize({in.height()}, target); \
for (int i = 0; i < in.height(); i++) { \
type_ret c_result = c_name(in(0, i), in(1, i)); \
if (!relatively_equal(c_result, result(i), target)) { \
fprintf(stderr, "For " #name "(%.20f, %.20f) == %.20f from C and %.20f from %s.\n", \
(double)in(0, i), (double)in(1, i), (double)c_result, (double)result(i), \
target.to_string().c_str()); \
num_errors++; \
} \
} \
}
// clang-format off
#define fun_1_float_types(name) \
fun_1(float, float, name, name) \
fun_1(double, double, name, name)
#define fun_2_float_types(name) \
fun_2(float, float, name, name) \
fun_2(double, double, name, name)
fun_1_float_types(sqrt)
fun_1_float_types(sin)
fun_1_float_types(cos)
fun_1_float_types(exp)
fun_1_float_types(log)
fun_1_float_types(floor)
fun_1_float_types(ceil)
fun_1_float_types(trunc)
fun_1_float_types(asin)
fun_1_float_types(acos)
fun_1_float_types(tan)
fun_1_float_types(atan)
fun_1_float_types(sinh)
fun_1_float_types(cosh)
fun_1_float_types(tanh)
fun_1_float_types(asinh)
fun_1_float_types(acosh)
fun_1_float_types(atanh)
fun_1_float_types(round)
fun_2_float_types(pow)
fun_2_float_types(atan2)
fun_1(float, float, abs, fabsf)
fun_1(double, double, abs, fabs)
fun_1(uint8_t, int8_t, abs, abs)
fun_1(uint16_t, int16_t, abs, abs)
fun_1(uint32_t, int32_t, abs, abs)
fun_2_float_types(absd)
fun_2(uint8_t, int8_t, absd, absd)
fun_2(uint16_t, int16_t, absd, absd)
fun_2(uint32_t, int32_t, absd, absd)
fun_2(uint8_t, uint8_t, absd, absd)
fun_2(uint16_t, uint16_t, absd, absd)
fun_2(uint32_t, uint32_t, absd, absd)
// clang-format on
// Note this test is more oriented toward making sure the paths
// through to math functions all work on a given target rather
// than usefully testing the accuracy of mathematical operations.
// As such little effort has been put into the domains tested,
// other than making sure they are valid for each function.
#define call_1(type, name, steps, start, end) \
do { \
printf("Testing " #name "(" #type ")\n"); \
TestArgs<type> args(steps, start, end); \
test_##type##_##name(args.data); \
} while (0)
#define call_2(type, name, steps, start1, end1, start2, end2) \
do { \
printf("Testing " #name "(" #type ")\n"); \
TestArgs<type> args(steps, start1, end1, start2, end2); \
test_##type##_##name(args.data); \
} while (0)
#define call_1_float_types(name, steps, start, end) \
do { \
call_1(float, name, steps, start, end); \
call_1(double, name, steps, start, end); \
} while (0)
#define call_2_float_types(name, steps, start1, end1, start2, end2) \
do { \
call_2(float, name, steps, start1, end1, start2, end2); \
call_2(double, name, steps, start1, end1, start2, end2); \
} while (0)
} // namespace
int main(int argc, char **argv) {
printf("host is: %s\n", get_host_target().to_string().c_str());
printf("HL_JIT_TARGET is: %s\n", get_jit_target_from_environment().to_string().c_str());
call_1_float_types(abs, 256, -1000, 1000);
call_1_float_types(sqrt, 256, 0, 1000000);
call_1_float_types(sin, 256, 5 * -3.1415f, 5 * 3.1415f);
call_1_float_types(cos, 256, 5 * -3.1415f, 5 * 3.1415f);
call_1_float_types(tan, 256, 0.49f * -3.1415f, 0.49f * 3.1415f);
call_1_float_types(asin, 256, -1.0, 1.0);
call_1_float_types(acos, 256, -1.0, 1.0);
call_1_float_types(atan, 256, -256, 100);
call_2_float_types(atan2, 256, -20, 20, -2, 2.001f);
call_1_float_types(sinh, 256, 5 * -3.1415f, 5 * 3.1415f);
call_1_float_types(cosh, 256, 0, 1);
call_1_float_types(tanh, 256, 5 * -3.1415f, 5 * 3.1415f);
call_1_float_types(asinh, 256, -10.0, 10.0);
call_1_float_types(acosh, 256, 1.0, 10);
call_1_float_types(atanh, 256, -1.0, 1.0);
call_1_float_types(round, 256, -15, 15);
call_1_float_types(exp, 256, 0, 20);
call_1_float_types(log, 256, 1, 1000000);
call_1_float_types(floor, 256, -25, 25);
call_1_float_types(ceil, 256, -25, 25);
call_1_float_types(trunc, 256, -25, 25);
call_2_float_types(pow, 256, -10.0, 10.0, -4.0f, 4.0f);
const int8_t int8_min = std::numeric_limits<int8_t>::min();
const int16_t int16_min = std::numeric_limits<int16_t>::min();
const int32_t int32_min = std::numeric_limits<int32_t>::min();
const int8_t int8_max = std::numeric_limits<int8_t>::max();
const int16_t int16_max = std::numeric_limits<int16_t>::max();
const int32_t int32_max = std::numeric_limits<int32_t>::max();
const uint8_t uint8_min = std::numeric_limits<uint8_t>::min();
const uint16_t uint16_min = std::numeric_limits<uint16_t>::min();
const uint32_t uint32_min = std::numeric_limits<uint32_t>::min();
const uint8_t uint8_max = std::numeric_limits<uint8_t>::max();
const uint16_t uint16_max = std::numeric_limits<uint16_t>::max();
const uint32_t uint32_max = std::numeric_limits<uint32_t>::max();
call_1_float_types(abs, 256, -25, 25);
call_1(int8_t, abs, 255, -int8_max, int8_max);
call_1(int16_t, abs, 255, -int16_max, int16_max);
call_1(int32_t, abs, 255, -int32_max, int32_max);
call_2_float_types(absd, 256, -25, 25, -25, 25);
call_2(int8_t, absd, 256, int8_min, int8_max, int8_min, int8_max);
call_2(int16_t, absd, 256, int16_min, int16_max, int16_min, int16_max);
call_2(int32_t, absd, 256, int32_min, int32_max, int32_min, int32_max);
call_2(uint8_t, absd, 256, uint8_min, uint8_max, uint8_min, uint8_max);
call_2(uint16_t, absd, 256, uint16_min, uint16_max, uint16_min, uint16_max);
call_2(uint32_t, absd, 256, uint32_min, uint32_max, uint32_min, uint32_max);
// TODO: int64 isn't tested because the testing mechanism relies
// on integer types being representable with doubles.
if (num_errors) {
fprintf(stderr, "Failed with %d total errors\n", num_errors);
exit(1);
}
printf("Success!\n");
return 0;
}
|
