Skip to content
Navigation Menu
{{ message }}
forked from AnswerDotAI/gpu.cpp
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathhalf.cpp
More file actions
279 lines (242 loc) · 8.8 KB
/
Copy pathhalf.cpp
File metadata and controls
279 lines (242 loc) · 8.8 KB
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
#include <array>
#include <cfloat>
#include <climits>
#include <cmath>
#include <cstdint>
#include <cstdio>
#include "gpu.h"
#include "numeric_types/half.h"
using namespace gpu;
using std::isinf;
using std::isnan;
#define EPSILON 0.01f
#define COLOR_RESET "\033[0m"
#define COLOR_RED "\033[31m"
#define COLOR_GREEN "\033[32m"
int approximatelyEqual(float a, float b, float epsilon) {
return fabsf(a - b) <= epsilon;
}
void printResult(bool passed, const char *message, float input, float output) {
if (passed) {
printf("[" COLOR_GREEN "PASSED" COLOR_RESET "]"
" : %s (in: %.10f, out: %.10f)\n",
message, input, output);
} else {
printf("[" COLOR_RED "FAILED" COLOR_RESET "]"
" : %s (in: %.10f, out: %.10f)\n",
message, input, output);
}
}
void printResult(bool passed, const char *message, float input,
uint16_t output) {
if (passed) {
printf("[" COLOR_GREEN "PASSED" COLOR_RESET "]"
" : %s (input: %.10f, output: 0x%04x)\n",
message, input, output);
} else {
printf("[" COLOR_RED "FAILED" COLOR_RESET "]"
" : %s (input: %.10f, output: 0x%04x)\n",
message, input, output);
}
}
void testRoundTrip(float value) {
// half h = halfFromFloat(value);
half h = half(value);
float result = static_cast<float>(h);
char message[256];
if (isnan(value)) {
sprintf(message, "NaN correctly round tripped");
printResult(isnan(result), message, value, result);
} else if (isinf(value)) {
sprintf(message, "Infinity correctly round tripped");
printResult(isinf(result) &&
((value > 0 && result > 0) || (value < 0 && result < 0)),
message, value, result);
} else {
sprintf(message, "%.10f correctly round tripped", value);
printResult(approximatelyEqual(result, value, EPSILON), message, value,
result);
}
}
void testRoundTrip(uint16_t value) {
half h;
h.data = value;
float f = halfToFloat(h);
half result = halfFromFloat(f);
char message[256];
sprintf(message, "half 0x%04x correctly round tripped", value);
printResult(result.data == value, message, (float)value, result.data);
}
void testRoundTrip(half value) {
float f = static_cast<float>(value);
half result = half(f);
char message[256];
sprintf(message, "half 0x%04x correctly round tripped", value.data);
printResult(result.data == value.data, message, (float)value, result.data);
}
void testSpecialCases() {
half h;
char message[256];
// Zero
h.data = 0x0000;
sprintf(message, "0x0000 correctly converted to 0.0f");
printResult(halfToFloat(h) == 0.0f, message, 0.0f, halfToFloat(h));
// Negative zero
h.data = 0x8000;
sprintf(message, "0x8000 correctly converted to -0.0f");
printResult(halfToFloat(h) == -0.0f, message, -0.0f, halfToFloat(h));
// Infinity
h.data = 0x7c00;
sprintf(message, "0x7c00 correctly converted to positive infinity");
printResult(isinf(halfToFloat(h)) && halfToFloat(h) > 0, message, INFINITY,
halfToFloat(h));
// Negative infinity
h.data = 0xfc00;
sprintf(message, "0xfc00 correctly converted to negative infinity");
printResult(isinf(halfToFloat(h)) && halfToFloat(h) < 0, message, -INFINITY,
halfToFloat(h));
// NaN
h.data = 0x7e00;
sprintf(message, "0x7e00 correctly converted to NaN");
printResult(isnan(halfToFloat(h)), message, NAN, halfToFloat(h));
// Smallest positive normal number
h.data = 0x0400;
sprintf(message, "0x0400 correctly converted to 6.10352e-05f");
printResult(approximatelyEqual(halfToFloat(h), 6.10352e-05f, EPSILON),
message, 6.10352e-05f, halfToFloat(h));
// Largest denormalized number
h.data = 0x03ff;
sprintf(message, "0x03ff correctly converted to 6.09756e-05f");
printResult(approximatelyEqual(halfToFloat(h), 6.09756e-05f, EPSILON),
message, 6.09756e-05f, halfToFloat(h));
// Smallest positive denormalized number
h.data = 0x0001;
sprintf(message, "0x0001 correctly converted to 5.96046e-08f");
printResult(approximatelyEqual(halfToFloat(h), 5.96046e-08f, EPSILON),
message, 5.96046e-08f, halfToFloat(h));
// Zero
h = halfFromFloat(0.0f);
sprintf(message, "0.0f correctly converted to 0x0000");
printResult(h.data == 0x0000, message, 0.0f, h.data);
// Negative zero
h = halfFromFloat(-0.0f);
sprintf(message, "-0.0f correctly converted to 0x8000");
printResult(h.data == 0x8000, message, -0.0f, h.data);
// Infinity
h = halfFromFloat(INFINITY);
sprintf(message, "positive infinity correctly converted to 0x7c00");
printResult(h.data == 0x7c00, message, INFINITY, h.data);
// Negative infinity
h = halfFromFloat(-INFINITY);
sprintf(message, "negative infinity correctly converted to 0xfc00");
printResult(h.data == 0xfc00, message, -INFINITY, h.data);
// NaN
h = halfFromFloat(NAN);
sprintf(message, "NaN correctly converted to NaN representation");
printResult((h.data & 0x7c00) == 0x7c00 && (h.data & 0x03ff) != 0x0000,
message, NAN, h.data);
// Smallest positive normal number
h = halfFromFloat(6.10352e-05f);
sprintf(message, "6.10352e-05f correctly converted to 0x0400");
printResult(h.data == 0x0400, message, 6.10352e-05f, h.data);
// Largest denormalized number
h = halfFromFloat(6.09756e-05f);
sprintf(message, "6.09756e-05f correctly converted to 0x03ff");
printResult(h.data == 0x03ff, message, 6.09756e-05f, h.data);
// Smallest positive denormalized number
h = halfFromFloat(5.96046448e-08f);
sprintf(message, "5.96046448e-08f correctly converted to 0x0001");
printResult(h.data == 0x0001, message, 5.96046e-08f, h.data);
}
void testContainers() {
{
std::array<half, 4> h = {0.0f, -0.0f, INFINITY, NAN};
testRoundTrip(h[0]);
testRoundTrip(h[1]);
testRoundTrip(h[2]);
testRoundTrip(h[3]);
}
{
Context ctx = createContext();
std::array<half, 8> h = {1.0f, 0.5f, 2.0f, 3.14f, 1.0, 2.0, 3.0, 4.0};
Tensor devH = createTensor(ctx, {h.size()}, kf16, h.data());
std::array<half, 8> h2;
toCPU(ctx, devH, h2.data(), sizeof(h2));
for (int i = 0; i < 8; ++i) {
printResult(h[i].data == h2[i].data, "Container round trip",
static_cast<float>(h[i]), static_cast<float>(h2[i]));
}
}
}
void testWGSL() {
static const char *kGelu = R"(
const GELU_SCALING_FACTOR: f16 = 0.7978845608028654; // sqrt(2.0 / PI)
@group(0) @binding(0) var<storage, read_write> inp: array<{{precision}}>;
@group(0) @binding(1) var<storage, read_write> out: array<{{precision}}>;
@group(0) @binding(1) var<storage, read_write> dummy: array<{{precision}}>;
@compute @workgroup_size({{workgroupSize}})
fn main(
@builtin(global_invocation_id) GlobalInvocationID: vec3<u32>) {
let i: u32 = GlobalInvocationID.x;
if (i < arrayLength(&inp)) {
let x: f16 = inp[i];
out[i] = select(0.5 * x * (1.0 + tanh(GELU_SCALING_FACTOR
* (x + .044715 * x * x * x))), x, x > 10.0);
}
}
)";
Context ctx = createContext(
{}, {},
/*device descriptor, enabling f16 in WGSL*/
{
.requiredFeatureCount = 1,
.requiredFeatures = std::array{WGPUFeatureName_ShaderF16}.data(),
});
static constexpr size_t N = 10000;
std::array<half, N> inputArr, outputArr;
for (int i = 0; i < N; ++i) {
inputArr[i] = half(static_cast<float>(i) / 10.0f); // dummy input data
}
Tensor input = createTensor(ctx, Shape{N}, kf16, inputArr.data());
Tensor output = createTensor(ctx, Shape{N}, kf16);
std::promise<void> promise;
std::future<void> future = promise.get_future();
Kernel op = createKernel(ctx, {kGelu, 256, kf16}, Bindings{input, output},
{cdiv(N, 256), 1, 1});
dispatchKernel(ctx, op, promise);
wait(ctx, future);
toCPU(ctx, output, outputArr.data(), sizeof(outputArr));
for (int i = 0; i < 12; ++i) {
printf(" gelu(%.2f) = %.2f\n", static_cast<float>(inputArr[i]),
static_cast<float>(outputArr[i]));
}
}
int main() {
printf("\nHalf-precision float tests\n==========================\n");
printf("\nRegular values float round trips\n\n");
testRoundTrip(1.0f);
testRoundTrip(0.5f);
testRoundTrip(2.0f);
testRoundTrip(3.14f);
testRoundTrip(-1.0f);
testRoundTrip(-0.5f);
testRoundTrip(-2.0f);
testRoundTrip(-3.14f);
printf("\nEdge Case float round trips\n\n");
testRoundTrip(0.0f);
testRoundTrip(-0.0f);
testRoundTrip(INFINITY);
testRoundTrip(-INFINITY);
testRoundTrip(NAN);
// testRoundTrip(FLT_MAX); // since FLT_MAX is not representable as half it
// is not expected to round-trip correctly testRoundTrip(FLT_MIN);
testRoundTrip(FLT_TRUE_MIN);
printf("\nSpecial half values\n\n");
testSpecialCases();
printf("\nContainers and CPU/GPU round trip\n\n");
testContainers();
printf("\nWGSL f16 extension test\n\n");
testWGSL();
printf("\nTests completed.\n");
return 0;
}
You can’t perform that action at this time.