root / ETSN / MySteps_3.py @ 280
Historique | Voir | Annoter | Télécharger (6,89 ko)
1 |
#!/usr/bin/env python3
|
---|---|
2 |
|
3 |
import numpy as np |
4 |
import pyopencl as cl |
5 |
|
6 |
# piling 16 arithmetical functions
|
7 |
def MySillyFunction(x): |
8 |
return(np.power(np.sqrt(np.log(np.exp(np.arctanh(np.tanh(np.arcsinh(np.sinh(np.arccosh(np.cosh(np.arctan(np.tan(np.arcsin(np.sin(np.arccos(np.cos(x))))))))))))))),2)) |
9 |
|
10 |
# Native Operation under Numpy (for prototyping & tests
|
11 |
def NativeAddition(a_np,b_np): |
12 |
return(a_np+b_np)
|
13 |
|
14 |
# Native Operation with MySillyFunction under Numpy (for prototyping & tests
|
15 |
def NativeSillyAddition(a_np,b_np): |
16 |
return(MySillyFunction(a_np)+MySillyFunction(b_np))
|
17 |
|
18 |
# CUDA complete operation
|
19 |
def CUDAAddition(a_np,b_np): |
20 |
import pycuda.autoinit |
21 |
import pycuda.driver as drv |
22 |
import numpy |
23 |
|
24 |
from pycuda.compiler import SourceModule |
25 |
mod = SourceModule("""
|
26 |
__global__ void sum(float *dest, float *a, float *b)
|
27 |
{
|
28 |
const int i = threadIdx.x;
|
29 |
dest[i] = a[i] + b[i];
|
30 |
}
|
31 |
""")
|
32 |
|
33 |
sum = mod.get_function("sum")
|
34 |
|
35 |
res_np = numpy.zeros_like(a_np) |
36 |
sum(drv.Out(res_np), drv.In(a_np), drv.In(b_np),
|
37 |
block=(a_np.size,1,1), grid=(1,1)) |
38 |
return(res_np)
|
39 |
|
40 |
# OpenCL complete operation
|
41 |
def OpenCLAddition(a_np,b_np): |
42 |
|
43 |
# Context creation
|
44 |
ctx = cl.create_some_context() |
45 |
# Every process is stored in a queue
|
46 |
queue = cl.CommandQueue(ctx) |
47 |
|
48 |
TimeIn=time.time() |
49 |
# Copy from Host to Device using pointers
|
50 |
mf = cl.mem_flags |
51 |
a_g = cl.Buffer(ctx, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=a_np) |
52 |
b_g = cl.Buffer(ctx, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=b_np) |
53 |
Elapsed=time.time()-TimeIn |
54 |
print("Copy from Host 2 Device : %.3f" % Elapsed)
|
55 |
|
56 |
TimeIn=time.time() |
57 |
# Definition of kernel under OpenCL
|
58 |
prg = cl.Program(ctx, """
|
59 |
__kernel void sum(
|
60 |
__global const float *a_g, __global const float *b_g, __global float *res_g)
|
61 |
{
|
62 |
int gid = get_global_id(0);
|
63 |
res_g[gid] = a_g[gid] + b_g[gid];
|
64 |
}
|
65 |
""").build()
|
66 |
Elapsed=time.time()-TimeIn |
67 |
print("Building kernels : %.3f" % Elapsed)
|
68 |
|
69 |
TimeIn=time.time() |
70 |
# Memory allocation on Device for result
|
71 |
res_g = cl.Buffer(ctx, mf.WRITE_ONLY, a_np.nbytes) |
72 |
Elapsed=time.time()-TimeIn |
73 |
print("Allocation on Device for results : %.3f" % Elapsed)
|
74 |
|
75 |
TimeIn=time.time() |
76 |
# Synthesis of function "sum" inside Kernel Sources
|
77 |
knl = prg.sum # Use this Kernel object for repeated calls
|
78 |
Elapsed=time.time()-TimeIn |
79 |
print("Synthesis of kernel : %.3f" % Elapsed)
|
80 |
|
81 |
TimeIn=time.time() |
82 |
# Call of kernel previously defined
|
83 |
knl(queue, a_np.shape, None, a_g, b_g, res_g)
|
84 |
Elapsed=time.time()-TimeIn |
85 |
print("Execution of kernel : %.3f" % Elapsed)
|
86 |
|
87 |
TimeIn=time.time() |
88 |
# Creation of vector for result with same size as input vectors
|
89 |
res_np = np.empty_like(a_np) |
90 |
Elapsed=time.time()-TimeIn |
91 |
print("Allocation on Host for results: %.3f" % Elapsed)
|
92 |
|
93 |
TimeIn=time.time() |
94 |
# Copy from Device to Host
|
95 |
cl.enqueue_copy(queue, res_np, res_g) |
96 |
Elapsed=time.time()-TimeIn |
97 |
print("Copy from Device 2 Host : %.3f" % Elapsed)
|
98 |
|
99 |
# Liberation of memory
|
100 |
a_g.release() |
101 |
b_g.release() |
102 |
res_g.release() |
103 |
|
104 |
return(res_np)
|
105 |
|
106 |
# OpenCL complete operation
|
107 |
def OpenCLSillyAddition(a_np,b_np): |
108 |
|
109 |
# Context creation
|
110 |
ctx = cl.create_some_context() |
111 |
# Every process is stored in a queue
|
112 |
queue = cl.CommandQueue(ctx) |
113 |
|
114 |
TimeIn=time.time() |
115 |
# Copy from Host to Device using pointers
|
116 |
mf = cl.mem_flags |
117 |
a_g = cl.Buffer(ctx, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=a_np) |
118 |
b_g = cl.Buffer(ctx, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=b_np) |
119 |
Elapsed=time.time()-TimeIn |
120 |
print("Copy from Host 2 Device : %.3f" % Elapsed)
|
121 |
|
122 |
TimeIn=time.time() |
123 |
# Definition of kernel under OpenCL
|
124 |
prg = cl.Program(ctx, """
|
125 |
|
126 |
float MySillyFunction(float x)
|
127 |
{
|
128 |
return(pow(sqrt(log(exp(atanh(tanh(asinh(sinh(acosh(cosh(atan(tan(asin(sin(acos(cos(x))))))))))))))),2));
|
129 |
}
|
130 |
|
131 |
__kernel void sillysum(
|
132 |
__global const float *a_g, __global const float *b_g, __global float *res_g)
|
133 |
{
|
134 |
int gid = get_global_id(0);
|
135 |
res_g[gid] = MySillyFunction(a_g[gid]) + MySillyFunction(b_g[gid]);
|
136 |
}
|
137 |
|
138 |
__kernel void sum(
|
139 |
__global const float *a_g, __global const float *b_g, __global float *res_g)
|
140 |
{
|
141 |
int gid = get_global_id(0);
|
142 |
res_g[gid] = a_g[gid] + b_g[gid];
|
143 |
}
|
144 |
""").build()
|
145 |
Elapsed=time.time()-TimeIn |
146 |
print("Building kernels : %.3f" % Elapsed)
|
147 |
|
148 |
TimeIn=time.time() |
149 |
# Memory allocation on Device for result
|
150 |
res_g = cl.Buffer(ctx, mf.WRITE_ONLY, a_np.nbytes) |
151 |
Elapsed=time.time()-TimeIn |
152 |
print("Allocation on Device for results : %.3f" % Elapsed)
|
153 |
|
154 |
TimeIn=time.time() |
155 |
# Synthesis of function "sillysum" inside Kernel Sources
|
156 |
knl = prg.sillysum # Use this Kernel object for repeated calls
|
157 |
Elapsed=time.time()-TimeIn |
158 |
print("Synthesis of kernel : %.3f" % Elapsed)
|
159 |
|
160 |
TimeIn=time.time() |
161 |
# Call of kernel previously defined
|
162 |
CallCL=knl(queue, a_np.shape, None, a_g, b_g, res_g)
|
163 |
#
|
164 |
CallCL.wait() |
165 |
Elapsed=time.time()-TimeIn |
166 |
print("Execution of kernel : %.3f" % Elapsed)
|
167 |
|
168 |
TimeIn=time.time() |
169 |
# Creation of vector for result with same size as input vectors
|
170 |
res_np = np.empty_like(a_np) |
171 |
Elapsed=time.time()-TimeIn |
172 |
print("Allocation on Host for results: %.3f" % Elapsed)
|
173 |
|
174 |
TimeIn=time.time() |
175 |
# Copy from Device to Host
|
176 |
cl.enqueue_copy(queue, res_np, res_g) |
177 |
Elapsed=time.time()-TimeIn |
178 |
print("Copy from Device 2 Host : %.3f" % Elapsed)
|
179 |
|
180 |
# Liberation of memory
|
181 |
a_g.release() |
182 |
b_g.release() |
183 |
res_g.release() |
184 |
|
185 |
return(res_np)
|
186 |
|
187 |
import sys |
188 |
import time |
189 |
|
190 |
if __name__=='__main__': |
191 |
|
192 |
# Size of input vectors definition based on stdin
|
193 |
import sys |
194 |
try:
|
195 |
SIZE=int(sys.argv[1]) |
196 |
print("Size of vectors set to %i" % SIZE)
|
197 |
except:
|
198 |
SIZE=50000
|
199 |
print("Size of vectors set to default size %i" % SIZE)
|
200 |
|
201 |
a_np = np.random.rand(SIZE).astype(np.float32) |
202 |
b_np = np.random.rand(SIZE).astype(np.float32) |
203 |
|
204 |
# Native Implementation
|
205 |
TimeIn=time.time() |
206 |
# res_np=NativeSillyAddition(a_np,b_np)
|
207 |
res_np=NativeAddition(a_np,b_np) |
208 |
NativeElapsed=time.time()-TimeIn |
209 |
NativeRate=int(SIZE/NativeElapsed)
|
210 |
print("NativeRate: %i" % NativeRate)
|
211 |
|
212 |
# OpenCL Implementation
|
213 |
TimeIn=time.time() |
214 |
# res_cl=OpenCLSillyAddition(a_np,b_np)
|
215 |
res_cl=OpenCLAddition(a_np,b_np) |
216 |
OpenCLElapsed=time.time()-TimeIn |
217 |
OpenCLRate=int(SIZE/OpenCLElapsed)
|
218 |
print("OpenCLRate: %i" % OpenCLRate)
|
219 |
|
220 |
# CUDA Implementation
|
221 |
TimeIn=time.time() |
222 |
res_cuda=CUDAAddition(a_np,b_np) |
223 |
CUDAElapsed=time.time()-TimeIn |
224 |
CUDARate=int(SIZE/CUDAElapsed)
|
225 |
print("CUDARate: %i" % CUDARate)
|
226 |
|
227 |
print("OpenCLvsNative ratio: %f" % (OpenCLRate/NativeRate))
|
228 |
print("CUDAvsNative ratio: %f" % (CUDARate/NativeRate))
|
229 |
|
230 |
# Check on CPU with Numpy:
|
231 |
print(res_cl - res_np) |
232 |
print(np.linalg.norm(res_cl - res_np)) |
233 |
assert np.allclose(res_np, res_cl)
|
234 |
|
235 |
# Check on CPU with Numpy:
|
236 |
print(res_cuda - res_np) |
237 |
print(np.linalg.norm(res_cuda - res_np)) |
238 |
assert np.allclose(res_np, res_cuda)
|