-
Notifications
You must be signed in to change notification settings - Fork 5
/
exercise_dgemv.hip.cpp
207 lines (167 loc) · 5.29 KB
/
exercise_dgemv.hip.cpp
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
/*
* Matrix vector product
*
* Here we will implement a slightly simplified version of the BLAS
* level 2 routine dgemv() which computes the matrix-vector product
*
* y_i := beta*y_i + alpha*A_ij x_j
*
* for an m x n matrix A_mn and vectors x_n and y_m. The data type
* is double, with alpha and beta scalar constants.
*
* The simplification is that we will consider only
*
* y_i := alpha*A_ij x_j
*
* Again we will assume that we are going to address the matrix with
* the flattened one-dimensional index A_ij = a[i*ncol + j] with ncol
* the number of columns n.
*
* An entirely serial kernel is provided below.
*
* Training material developed by Nick Johnson and Kevin Stratford
* Copyright EPCC, The University of Edinburgh, 2010-2023
*/
#include <cassert>
#include <cfloat>
#include <cstdio>
#include <iomanip>
#include <iostream>
#include <string>
#include "hip/hip_runtime.h"
__host__ void myErrorHandler(hipError_t ifail, std::string file, int line,
int fatal);
#define HIP_ASSERT(call) \
{ myErrorHandler((call), __FILE__, __LINE__, 1); }
/* Kernel parameters (start with 1-d) */
#define THREADS_PER_BLOCK 256
#define THREADS_PER_BLOCKX 16
#define THREADS_PER_BLOCKY 16
/* An entirely serial kernel. */
__global__ void myKernel(int mrow, int ncol, double alpha, double *a, double *x,
double *y) {
int tid = blockIdx.x * blockDim.x + threadIdx.x;
if (tid == 0) {
for (int i = 0; i < mrow; i++) {
double sum = 0.0;
for (int j = 0; j < ncol; j++) {
sum += a[i * ncol + j] * x[j];
}
y[i] += alpha * sum;
}
}
return;
}
/* Main routine */
int main(int argc, char *argv[]) {
int mrow = 1024; /* Number of rows */
int ncol = 256; /* Number of columns (start = THREADS_PER_BLOCK) */
double alpha = 2.0;
double *h_x = NULL;
double *h_y = NULL;
double *d_x = NULL;
double *d_y = NULL;
double *h_a = NULL;
double *d_a = NULL;
/* Print device name (just for information) */
int ndevice = 0;
int deviceNum = -1;
hipDeviceProp_t prop;
HIP_ASSERT(hipGetDeviceCount(&ndevice));
if (ndevice == 0) {
std::cout << "No GPU available!" << std::endl;
std::exit(0);
}
HIP_ASSERT(hipGetDevice(&deviceNum));
HIP_ASSERT(hipGetDeviceProperties(&prop, deviceNum));
std::cout << "Device " << deviceNum << " name: " << prop.name << std::endl;
std::cout << "Maximum number of threads per block: "
<< prop.maxThreadsPerBlock << std::endl;
/*
* Establish some data on the host. x = 1; y = 0; A_ij = 1
*/
h_x = new double[ncol];
h_y = new double[mrow];
h_a = new double[mrow * ncol];
assert(h_x);
assert(h_y);
assert(h_a);
for (int i = 0; i < ncol; i++) {
h_x[i] = 1.0;
}
for (int j = 0; j < mrow; j++) {
h_y[j] = 0.0;
}
for (int i = 0; i < mrow; i++) {
for (int j = 0; j < ncol; j++) {
h_a[i * ncol + j] = 1.0 * i * j;
}
}
static __constant__ double data_read_only[3];
double values[3] = {1.0, 2.0, 3.0};
hipMemcpyToSymbol(data_read_only, values, 3*sizeof(double));
/*
* allocate memory on device
*/
HIP_ASSERT(hipMalloc(&d_x, ncol * sizeof(double)));
HIP_ASSERT(hipMalloc(&d_y, mrow * sizeof(double)));
HIP_ASSERT(hipMalloc(&d_a, mrow * ncol * sizeof(double)));
hipMemcpyKind kind = hipMemcpyHostToDevice;
HIP_ASSERT(hipMemcpy(d_x, h_x, ncol * sizeof(double), kind));
HIP_ASSERT(hipMemcpy(d_y, h_y, mrow * sizeof(double), kind));
HIP_ASSERT(hipMemcpy(d_a, h_a, mrow * ncol * sizeof(double), kind));
/* Kernel */
uint nblockx = 1;
dim3 blocks = {nblockx, 1, 1};
dim3 threadsPerBlock = {THREADS_PER_BLOCK, 1, 1};
myKernel<<<blocks, threadsPerBlock>>>(mrow, ncol, alpha, d_a, d_x, d_y);
/* wait for all threads to complete and check for errors */
HIP_ASSERT(hipPeekAtLastError());
HIP_ASSERT(hipDeviceSynchronize());
kind = hipMemcpyDeviceToHost;
HIP_ASSERT(hipMemcpy(h_y, d_y, mrow * sizeof(double), kind));
std::cout << "Results:" << std::endl;
{
int ncorrect = 0;
for (int i = 0; i < mrow; i++) {
double sum = 0.0;
double yi = 0.0;
for (int j = 0; j < ncol; j++) {
sum += h_a[i * ncol + j] * h_x[j];
}
yi = alpha * sum;
if (fabs(yi - h_y[i]) < DBL_EPSILON)
ncorrect += 1;
/* Can be uncommented to debug ... */
/* printf("Row %5d %14.7e %14.7e\n", i, yi, h_y[i]); */
}
std::cout << "No. rows " << mrow << ", and correct rows " << ncorrect
<< std::endl;
}
/* free device buffer */
HIP_ASSERT(hipFree(d_a));
HIP_ASSERT(hipFree(d_y));
HIP_ASSERT(hipFree(d_x));
/* free host buffers */
delete h_x;
delete h_y;
delete h_a;
return 0;
}
/* It is important to check the return code from API calls, so the
* follow function/macro allow this to be done concisely as
*
* HIP_ASSERT(hipRunTimeAPIFunction(...));
*
* Return codes may be asynchronous, and thus misleading! */
__host__ void myErrorHandler(hipError_t ifail, const std::string file, int line,
int fatal) {
if (ifail != hipSuccess) {
std::cerr << "Line " << line << " (" << file
<< "): " << hipGetErrorName(ifail) << ": "
<< hipGetErrorString(ifail) << std::endl;
if (fatal)
std::exit(ifail);
}
return;
}