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
|
#!/usr/bin/env python3
# Make a "naive" implementation of a matrix multiplication.
# - Start from the template from last lecture - all global memory, each work item calculates one value of the result independent of the others, and so on.
# - Feel free to make assumptions on the size, for instance that the matrices are square, small enough to fit in GPU memory and so on.
# - Make sure to test your kernel thoroughly enough that you trust it is correct.
# Im asuming that the buffers fit in memory
import numpy as np
import pyopencl as cl
import time
# Source of the kernel itself.
kernel_source = """
__kernel void matrixmult(
const uint shared_dim,
__global const float *a_device,
__global const float *b_device,
__global float *result_device)
{
// get the i'th row of matrix a
int index_a = get_global_id(0) * shared_dim;
// get the start of the i'th column of b.
// Remember we should index this by jumping b's row size (bcols, or get_global_size(1)).
int index_b = get_global_id(1);
int b_jump = get_global_size(1);
// Do the vector dot
float result = 0;
for (int i = 0; i < shared_dim; i++) {
result += a_device[index_a + i] * b_device[index_b];
// Remember we need to move b by it's column size to
// skip to the next row
index_b += b_jump;
}
// Save the result
result_device[get_global_id(0) * get_global_size(1) + get_global_id(1)] = result;
}
"""
# matrix a rows
arows = 500
bcols = 1000
# A columns and b rows
shared = 1000
# Create the context (containing platform and device information) and command queue.
context = cl.create_some_context()
cmd_queue = cl.CommandQueue(context)
# Create the host side data and a empty array to hold the result.
a_host = np.random.random((arows, shared)).astype(np.float32)
b_host = np.random.random((shared, bcols)).astype(np.float32)
result_host = np.empty((arows, bcols)).astype(np.float32)
# If you want to keep the kernel in a seperate file uncomment this line and adjust the filename
#kernel_source = open("kernel.cl").read()
# Create a new program from the kernel and build the source.
prog = cl.Program(context, kernel_source).build()
start = time.time()
# Create a device side read-only memory buffer and copy the data from "hostbuf" into it.
# Create as many
# You can find the other possible mem_flags values at
# https://www.khronos.org/registry/OpenCL/sdk/1.2/docs/man/xhtml/clCreateBuffer.html
mf = cl.mem_flags
a_device = cl.Buffer(context, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=a_host)
b_device = cl.Buffer(context, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=b_host)
result_device = cl.Buffer(context, mf.WRITE_ONLY, result_host.nbytes)
# Execute the "sum" kernel in the program. Parameters are:
#
# Command queue Work group size Kernel param 1
# ↓ Global grid size ↓ Kernel param 0 ↓ Kernel param 2
# ↓ ↓ ↓ ↓ ↓ ↓
prog.matrixmult(cmd_queue, result_host.shape, None, np.uint32(shared), a_device, b_device, result_device)
# Copy the result back from device to host.
cl.enqueue_copy(cmd_queue, result_host, result_device)
end = time.time()
print(f"GPU time {end - start} sec")
start = time.time()
hostcalc = a_host @ b_host
end = time.time()
print(f"CPU time {end - start} sec")
# Check the results in the host array with Numpy.
print("All elements close?", np.allclose(result_host, hostcalc))
|
