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
| #include <stdio.h>
| #include <stdlib.h>
| #include <string.h>
| #include <time.h>
| #include <math.h>
|
| #include "opencl.h"
| #include "mini_blas.h"
|
| #define STR_HELPER(x) #x
| #define STR(x) STR_HELPER(x)
|
| #define BLOCK 8
|
| cl_kernel get_gemm_kernel()
| {
| static int init = 0;
| static cl_kernel gemm_kernel;
| if(!init){
| gemm_kernel = get_kernel("src/gemm.cl", "gemm", "-D BLOCK=" STR(BLOCK) );
| init = 1;
| }
| return gemm_kernel;
| }
|
| void gpu_gemm(int TA, int TB, int M, int N, int K, float ALPHA,
| float *A, int lda,
| float *B, int ldb,
| float BETA,
| float *C, int ldc)
| {
| cl_setup();
| cl_kernel gemm_kernel = get_gemm_kernel();
| cl_context context = cl.context;
| cl_command_queue queue = cl.queue;
|
| size_t size = sizeof(float)*(TA ? lda*K:lda*M);
| cl_mem A_gpu = clCreateBuffer(context,
| CL_MEM_READ_ONLY|CL_MEM_COPY_HOST_PTR,
| size, A, &cl.error);
| check_error(cl);
|
| size = sizeof(float)*(TB ? ldb*N:ldb*K);
| cl_mem B_gpu = clCreateBuffer(context,
| CL_MEM_READ_ONLY|CL_MEM_COPY_HOST_PTR,
| size, B, &cl.error);
| check_error(cl);
|
| size = sizeof(float)*(ldc*M);
| cl_mem C_gpu = clCreateBuffer(context,
| CL_MEM_WRITE_ONLY|CL_MEM_COPY_HOST_PTR,
| size, C, &cl.error);
| check_error(cl);
|
| cl_uint i = 0;
| cl.error = clSetKernelArg(gemm_kernel, i++, sizeof(TA), (void*) &TA);
| cl.error = clSetKernelArg(gemm_kernel, i++, sizeof(TB), (void*) &TB);
| cl.error = clSetKernelArg(gemm_kernel, i++, sizeof(M), (void*) &M);
| cl.error = clSetKernelArg(gemm_kernel, i++, sizeof(N), (void*) &N);
| cl.error = clSetKernelArg(gemm_kernel, i++, sizeof(K), (void*) &K);
| cl.error = clSetKernelArg(gemm_kernel, i++, sizeof(ALPHA), (void*) &ALPHA);
| cl.error = clSetKernelArg(gemm_kernel, i++, sizeof(A_gpu), (void*) &A_gpu);
| cl.error = clSetKernelArg(gemm_kernel, i++, sizeof(lda), (void*) &lda);
| cl.error = clSetKernelArg(gemm_kernel, i++, sizeof(B_gpu), (void*) &B_gpu);
| cl.error = clSetKernelArg(gemm_kernel, i++, sizeof(ldb), (void*) &ldb);
| cl.error = clSetKernelArg(gemm_kernel, i++, sizeof(BETA), (void*) &BETA);
| cl.error = clSetKernelArg(gemm_kernel, i++, sizeof(C_gpu), (void*) &C_gpu);
| cl.error = clSetKernelArg(gemm_kernel, i++, sizeof(ldc), (void*) &ldc);
| check_error(cl);
|
| const size_t global_size[] = {ceil((float)M/BLOCK)*BLOCK, ceil((float)N/BLOCK)*BLOCK};
| const size_t local_size[] = {BLOCK, BLOCK};
| //printf("%zd %zd %zd %zd\n", global_size[0], global_size[1], local_size[0], local_size[1]);
|
| clEnqueueNDRangeKernel(queue, gemm_kernel, 2, 0, global_size, local_size, 0, 0, 0);
| check_error(cl);
| clEnqueueReadBuffer(queue, C_gpu, CL_TRUE, 0, size, C, 0, 0, 0);
| check_error(cl);
|
| clReleaseMemObject(A_gpu);
| clReleaseMemObject(B_gpu);
| clReleaseMemObject(C_gpu);
|
| }
|
| void time_gpu_random_matrix(int TA, int TB, int m, int k, int n)
| {
| float *a;
| if(!TA) a = random_matrix(m,k);
| else a = random_matrix(k,m);
| int lda = (!TA)?k:m;
| float *b;
| if(!TB) b = random_matrix(k,n);
| else b = random_matrix(n,k);
| int ldb = (!TB)?n:k;
|
| float *c = random_matrix(m,n);
| int i;
| clock_t start = clock(), end;
| for(i = 0; i<1000; ++i){
| gpu_gemm(TA,TB,m,n,k,1,a,lda,b,ldb,1,c,n);
| }
| end = clock();
| printf("Matrix Multiplication %dx%d * %dx%d, TA=%d, TB=%d: %lf ms\n",m,k,k,n, TA, TB, (float)(end-start)/CLOCKS_PER_SEC);
| free(a);
| free(b);
| free(c);
| }
|
| void test_gpu_accuracy(int TA, int TB, int m, int k, int n)
| {
| srand(0);
| float *a;
| if(!TA) a = random_matrix(m,k);
| else a = random_matrix(k,m);
| int lda = (!TA)?k:m;
| float *b;
| if(!TB) b = random_matrix(k,n);
| else b = random_matrix(n,k);
| int ldb = (!TB)?n:k;
|
| float *c = random_matrix(m,n);
| float *c_gpu = random_matrix(m,n);
| memset(c, 0, m*n*sizeof(float));
| memset(c_gpu, 0, m*n*sizeof(float));
| int i;
| //pm(m,k,b);
| gpu_gemm(TA,TB,m,n,k,1,a,lda,b,ldb,1,c_gpu,n);
| //pm(m, n, c_gpu);
| cpu_gemm(TA,TB,m,n,k,1,a,lda,b,ldb,1,c,n);
| //pm(m, n, c);
| double sse = 0;
| for(i = 0; i < m*n; ++i) {
| //printf("%f %f\n", c[i], c_gpu[i]);
| sse += pow(c[i]-c_gpu[i], 2);
| }
| printf("Matrix Multiplication %dx%d * %dx%d, TA=%d, TB=%d: %g MSE\n",m,k,k,n, TA, TB, sse/(m*n));
| free(a);
| free(b);
| free(c);
| }
|
| void test_gpu_blas()
| {
| test_gpu_accuracy(0,0,17,10,10);
| test_gpu_accuracy(1,0,17,10,10);
| test_gpu_accuracy(0,1,17,10,10);
| test_gpu_accuracy(1,1,17,10,10);
|
| test_gpu_accuracy(0,0,1000,10,100);
| test_gpu_accuracy(1,0,1000,10,100);
| test_gpu_accuracy(0,1,1000,10,100);
| test_gpu_accuracy(1,1,1000,10,100);
|
| time_gpu_random_matrix(0,0,1000,1000,100);
| time_random_matrix(0,0,1000,1000,100);
|
| time_gpu_random_matrix(0,1,1000,1000,100);
| time_random_matrix(0,1,1000,1000,100);
|
| time_gpu_random_matrix(1,0,1000,1000,100);
| time_random_matrix(1,0,1000,1000,100);
|
| time_gpu_random_matrix(1,1,1000,1000,100);
| time_random_matrix(1,1,1000,1000,100);
|
| }
|
| /*
| cl_kernel get_gemm_kernel_slow()
| {
| static int init = 0;
| static cl_kernel gemm_kernel;
| if(!init){
| gemm_kernel = get_kernel("src/gemm.cl", "gemm_slow");
| init = 1;
| }
| return gemm_kernel;
| }
|
| void gpu_gemm_slow(int TA, int TB, int M, int N, int K, float ALPHA,
| float *A, int lda,
| float *B, int ldb,
| float BETA,
| float *C, int ldc)
| {
| cl_setup();
| cl_kernel gemm_kernel = get_gemm_kernel_slow();
| cl_context context = cl.context;
| cl_command_queue queue = cl.queue;
|
| size_t size = sizeof(float)*(TA ? lda*K:lda*M);
| cl_mem A_gpu = clCreateBuffer(context,
| CL_MEM_READ_ONLY|CL_MEM_COPY_HOST_PTR,
| size, A, &cl.error);
| check_error(cl);
|
| size = sizeof(float)*(TB ? ldb*N:ldb*K);
| cl_mem B_gpu = clCreateBuffer(context,
| CL_MEM_READ_ONLY|CL_MEM_COPY_HOST_PTR,
| size, B, &cl.error);
| check_error(cl);
|
| size = sizeof(float)*(ldc*M);
| cl_mem C_gpu = clCreateBuffer(context,
| CL_MEM_READ_ONLY|CL_MEM_COPY_HOST_PTR,
| size, C, &cl.error);
| check_error(cl);
|
| cl_uint i = 0;
| cl.error = clSetKernelArg(gemm_kernel, i++, sizeof(TA), (void*) &TA);
| cl.error = clSetKernelArg(gemm_kernel, i++, sizeof(TB), (void*) &TB);
| cl.error = clSetKernelArg(gemm_kernel, i++, sizeof(M), (void*) &M);
| cl.error = clSetKernelArg(gemm_kernel, i++, sizeof(N), (void*) &N);
| cl.error = clSetKernelArg(gemm_kernel, i++, sizeof(K), (void*) &K);
| cl.error = clSetKernelArg(gemm_kernel, i++, sizeof(ALPHA), (void*) &ALPHA);
| cl.error = clSetKernelArg(gemm_kernel, i++, sizeof(A_gpu), (void*) &A_gpu);
| cl.error = clSetKernelArg(gemm_kernel, i++, sizeof(lda), (void*) &lda);
| cl.error = clSetKernelArg(gemm_kernel, i++, sizeof(B_gpu), (void*) &B_gpu);
| cl.error = clSetKernelArg(gemm_kernel, i++, sizeof(ldb), (void*) &ldb);
| cl.error = clSetKernelArg(gemm_kernel, i++, sizeof(BETA), (void*) &BETA);
| cl.error = clSetKernelArg(gemm_kernel, i++, sizeof(C_gpu), (void*) &C_gpu);
| cl.error = clSetKernelArg(gemm_kernel, i++, sizeof(ldc), (void*) &ldc);
| check_error(cl);
|
| const size_t global_size[] = {M, N};
|
| clEnqueueNDRangeKernel(queue, gemm_kernel, 2, 0, global_size, 0, 0, 0, 0);
| clEnqueueReadBuffer(queue, C_gpu, CL_TRUE, 0, size, C, 0, 0, 0);
|
| clReleaseMemObject(A_gpu);
| clReleaseMemObject(B_gpu);
| clReleaseMemObject(C_gpu);
|
| }
| */
|
|
