-
Notifications
You must be signed in to change notification settings - Fork 15
/
Copy pathtest_cublas.cu
192 lines (178 loc) · 6.16 KB
/
test_cublas.cu
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
#include <sys/time.h>
#include <cuda_profiler_api.h>
#include <cublas_v2.h>
#include <cuda.h>
#include <cuda_fp16.h>
#include <cuda_runtime.h>
#include <stdio.h>
int8_t float2int8(float f, float scale)
{
int8_t i = int8_t(f * scale);
if (i < -127)
i = -127;
if (i > 127)
i = 127;
return i;
}
template <typename T, typename S>
void allocate_memory(int m, int n, int k, T **A, T **B, S **C)
{
cudaMallocManaged(A, m * k * sizeof(T));
cudaMallocManaged(B, k * n * sizeof(T));
cudaMallocManaged(C, m * n * sizeof(S));
}
template <typename T, typename S>
void free_memory(T *A, T *B, S *C)
{
cudaFree(A);
cudaFree(B);
cudaFree(C);
}
template <typename T, typename S>
int cublas_gemm_ex(cublasHandle_t handle, cublasOperation_t transA, cublasOperation_t transB,
int m, int n, int k, T *A, T *B, S *C, int lda, int ldb, int ldc,
S *alpha, S *beta, int algo)
{
cudaDataType_t AType, BType, CType, ComputeType;
if (std::is_same<T, float>::value)
{
AType = BType = CType = ComputeType = CUDA_R_32F;
}
else if (std::is_same<T, __half>::value)
{
AType = BType = CType = ComputeType = CUDA_R_16F;
}
else if (std::is_same<T, int8_t>::value)
{
AType = BType = CUDA_R_8I;
CType = ComputeType = CUDA_R_32I;
}
else
{
printf("Not supported data type.");
return -1;
}
cublasStatus_t status;
status = cublasGemmEx(handle,
transA,
transB,
m,
n,
k,
alpha,
A,
AType,
lda,
B,
BType,
ldb,
beta,
C,
CType,
ldc,
ComputeType,
static_cast<cublasGemmAlgo_t>(algo));
if (status == CUBLAS_STATUS_SUCCESS)
return 1;
else
return -1;
}
template <typename T, typename S>
void test_gemm(cublasHandle_t handle, int m, int n, int k, T *A, T *B, S *C,
S *alpha, S *beta, int algo, int iteration)
{
float total_time = 0;
for (int i = 0; i < iteration; ++i)
{
struct timeval start, end;
cudaDeviceSynchronize();
cudaProfilerStart();
gettimeofday(&start, NULL);
int success = cublas_gemm_ex(handle,
CUBLAS_OP_T,
CUBLAS_OP_N,
n,
m,
k,
B,
A,
C,
n,
k,
n,
alpha,
beta,
static_cast<cublasGemmAlgo_t>(algo));
cudaDeviceSynchronize();
gettimeofday(&end, NULL);
cudaProfilerStop();
if (success > 0 && i > 0)
total_time += (end.tv_sec - start.tv_sec) * 1000 + (end.tv_usec - start.tv_usec) * 0.001;
}
if (total_time > 0)
printf("algo %d: %.3f ms\n", algo, total_time / (iteration - 1));
}
int main()
{
int m = 4096, n = 4096, k = 4096;
printf("shape: (%d, %d) x (%d, %d)\n", m, k, k, n);
int start_algo = CUBLAS_GEMM_DEFAULT;
int end_algo = CUBLAS_GEMM_ALGO23;
int start_algo_t_op = CUBLAS_GEMM_DEFAULT_TENSOR_OP;
int end_algo_t_op = CUBLAS_GEMM_ALGO15_TENSOR_OP;
int iteration = 10;
float *fA, *fB, *fC;
__half *hA, *hB, *hC;
int8_t *iA, *iB;
int32_t *iC;
float f_alpha = 1, f_beta = 0;
__half h_alpha = __float2half_rn(1.0), h_beta = __float2half_rn(0.0);
int32_t i_alpha = 1, i_beta = 0;
allocate_memory(m, n, k, &fA, &fB, &fC);
allocate_memory(m, n, k, &hA, &hB, &hC);
allocate_memory(m, n, k, &iA, &iB, &iC);
for (int i = 0; i < m * k; ++i)
{
fA[i] = float(i % 255 - 127) / 127;
hA[i] = __float2half_rn(fA[i]);
iA[i] = float2int8(fA[i], 127);
}
for (int i = 0; i < k * n; ++i)
{
fB[i] = float(i % 255 - 127) / 127;
hB[i] = __float2half_rn(fB[i]);
iB[i] = float2int8(fB[i], 127);
}
cublasHandle_t handle;
cublasCreate(&handle);
cublasSetMathMode(handle, CUBLAS_TENSOR_OP_MATH);
printf(">>>>>>>>>>>>>>>>> test fp32 >>>>>>>>>>>>>>>>>\n");
for (int algo = start_algo; algo <= end_algo; ++algo)
test_gemm(handle, m, n, k, fA, fB, fC, &f_alpha, &f_beta, algo, iteration);
for (int algo = start_algo_t_op; algo <= end_algo_t_op; ++algo)
test_gemm(handle, m, n, k, fA, fB, fC, &f_alpha, &f_beta, algo, iteration);
printf(">>>>>>>>>>>>>>>>> test fp16 >>>>>>>>>>>>>>>>>\n");
for (int algo = start_algo; algo <= end_algo; ++algo)
test_gemm(handle, m, n, k, hA, hB, hC, &h_alpha, &h_beta, algo, iteration);
for (int algo = start_algo_t_op; algo <= end_algo_t_op; ++algo)
test_gemm(handle, m, n, k, hA, hB, hC, &h_alpha, &h_beta, algo, iteration);
printf(">>>>>>>>>>>>>>>>> test int8 >>>>>>>>>>>>>>>>>\n");
for (int algo = start_algo; algo <= end_algo; ++algo)
test_gemm(handle, m, n, k, iA, iB, iC, &i_alpha, &i_beta, algo, iteration);
for (int algo = start_algo_t_op; algo <= end_algo_t_op; ++algo)
test_gemm(handle, m, n, k, iA, iB, iC, &i_alpha, &i_beta, algo, iteration);
printf(">>>>>>>>>>>>>>>>> compare result >>>>>>>>>>>>>>>>>\n");
printf("fp32: ");
for (int i = 0; i < 10; ++i)
printf("%.5f%c", fC[i], " \n"[i == 9]);
printf("fp16: ");
for (int i = 0; i < 10; ++i)
printf("%.5f%c", float(hC[i]), " \n"[i == 9]);
printf("int8: ");
for (int i = 0; i < 10; ++i)
printf("%.5f%c", float(iC[i]) / 127 / 127, " \n"[i == 9]);
free_memory(iA, iB, iC);
free_memory(fA, fB, fC);
free_memory(hA, hB, hC);
return 0;
}