root / ETSN / MySteps_5.py @ 295
Historique | Voir | Annoter | Télécharger (8,37 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 |
const int i = blockIdx.x;
|
30 |
dest[i] = a[i] + b[i];
|
31 |
}
|
32 |
""")
|
33 |
|
34 |
# sum = mod.get_function("sum")
|
35 |
sum = mod.get_function("sum")
|
36 |
|
37 |
res_np = numpy.zeros_like(a_np) |
38 |
sum(drv.Out(res_np), drv.In(a_np), drv.In(b_np),
|
39 |
block=(1,1,1), grid=(a_np.size,1)) |
40 |
return(res_np)
|
41 |
|
42 |
# CUDA Silly complete operation
|
43 |
def CUDASillyAddition(a_np,b_np): |
44 |
import pycuda.autoinit |
45 |
import pycuda.driver as drv |
46 |
import numpy |
47 |
|
48 |
from pycuda.compiler import SourceModule |
49 |
TimeIn=time.time() |
50 |
mod = SourceModule("""
|
51 |
__device__ float MySillyFunction(float x)
|
52 |
{
|
53 |
return(pow(sqrt(log(exp(atanh(tanh(asinh(sinh(acosh(cosh(atan(tan(asin(sin(acos(cos(x))))))))))))))),2));
|
54 |
}
|
55 |
|
56 |
__global__ void sillysum(float *dest, float *a, float *b)
|
57 |
{
|
58 |
const int i = blockIdx.x;
|
59 |
dest[i] = MySillyFunction(a[i]) + MySillyFunction(b[i]);
|
60 |
}
|
61 |
""")
|
62 |
Elapsed=time.time()-TimeIn |
63 |
print("Definition of kernel : %.3f" % Elapsed)
|
64 |
|
65 |
TimeIn=time.time() |
66 |
# sum = mod.get_function("sum")
|
67 |
sillysum = mod.get_function("sillysum")
|
68 |
Elapsed=time.time()-TimeIn |
69 |
print("Synthesis of kernel : %.3f" % Elapsed)
|
70 |
|
71 |
TimeIn=time.time() |
72 |
res_np = numpy.zeros_like(a_np) |
73 |
Elapsed=time.time()-TimeIn |
74 |
print("Allocation on Host for results : %.3f" % Elapsed)
|
75 |
|
76 |
TimeIn=time.time() |
77 |
sillysum(drv.Out(res_np), drv.In(a_np), drv.In(b_np), |
78 |
block=(1,1,1), grid=(a_np.size,1)) |
79 |
Elapsed=time.time()-TimeIn |
80 |
print("Execution of kernel : %.3f" % Elapsed)
|
81 |
return(res_np)
|
82 |
|
83 |
# OpenCL complete operation
|
84 |
def OpenCLAddition(a_np,b_np): |
85 |
|
86 |
# Context creation
|
87 |
ctx = cl.create_some_context() |
88 |
# Every process is stored in a queue
|
89 |
queue = cl.CommandQueue(ctx) |
90 |
|
91 |
TimeIn=time.time() |
92 |
# Copy from Host to Device using pointers
|
93 |
mf = cl.mem_flags |
94 |
a_g = cl.Buffer(ctx, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=a_np) |
95 |
b_g = cl.Buffer(ctx, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=b_np) |
96 |
Elapsed=time.time()-TimeIn |
97 |
print("Copy from Host 2 Device : %.3f" % Elapsed)
|
98 |
|
99 |
TimeIn=time.time() |
100 |
# Definition of kernel under OpenCL
|
101 |
prg = cl.Program(ctx, """
|
102 |
__kernel void sum(
|
103 |
__global const float *a_g, __global const float *b_g, __global float *res_g)
|
104 |
{
|
105 |
int gid = get_global_id(0);
|
106 |
res_g[gid] = a_g[gid] + b_g[gid];
|
107 |
}
|
108 |
""").build()
|
109 |
Elapsed=time.time()-TimeIn |
110 |
print("Building kernels : %.3f" % Elapsed)
|
111 |
|
112 |
TimeIn=time.time() |
113 |
# Memory allocation on Device for result
|
114 |
res_g = cl.Buffer(ctx, mf.WRITE_ONLY, a_np.nbytes) |
115 |
Elapsed=time.time()-TimeIn |
116 |
print("Allocation on Device for results : %.3f" % Elapsed)
|
117 |
|
118 |
TimeIn=time.time() |
119 |
# Synthesis of function "sum" inside Kernel Sources
|
120 |
knl = prg.sum # Use this Kernel object for repeated calls
|
121 |
Elapsed=time.time()-TimeIn |
122 |
print("Synthesis of kernel : %.3f" % Elapsed)
|
123 |
|
124 |
TimeIn=time.time() |
125 |
# Call of kernel previously defined
|
126 |
knl(queue, a_np.shape, None, a_g, b_g, res_g)
|
127 |
Elapsed=time.time()-TimeIn |
128 |
print("Execution of kernel : %.3f" % Elapsed)
|
129 |
|
130 |
TimeIn=time.time() |
131 |
# Creation of vector for result with same size as input vectors
|
132 |
res_np = np.empty_like(a_np) |
133 |
Elapsed=time.time()-TimeIn |
134 |
print("Allocation on Host for results: %.3f" % Elapsed)
|
135 |
|
136 |
TimeIn=time.time() |
137 |
# Copy from Device to Host
|
138 |
cl.enqueue_copy(queue, res_np, res_g) |
139 |
Elapsed=time.time()-TimeIn |
140 |
print("Copy from Device 2 Host : %.3f" % Elapsed)
|
141 |
|
142 |
# Liberation of memory
|
143 |
a_g.release() |
144 |
b_g.release() |
145 |
res_g.release() |
146 |
|
147 |
return(res_np)
|
148 |
|
149 |
# OpenCL complete operation
|
150 |
def OpenCLSillyAddition(a_np,b_np): |
151 |
|
152 |
# Context creation
|
153 |
ctx = cl.create_some_context() |
154 |
# Every process is stored in a queue
|
155 |
queue = cl.CommandQueue(ctx) |
156 |
|
157 |
TimeIn=time.time() |
158 |
# Copy from Host to Device using pointers
|
159 |
mf = cl.mem_flags |
160 |
a_g = cl.Buffer(ctx, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=a_np) |
161 |
b_g = cl.Buffer(ctx, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=b_np) |
162 |
Elapsed=time.time()-TimeIn |
163 |
print("Copy from Host 2 Device : %.3f" % Elapsed)
|
164 |
|
165 |
TimeIn=time.time() |
166 |
# Definition of kernel under OpenCL
|
167 |
prg = cl.Program(ctx, """
|
168 |
|
169 |
float MySillyFunction(float x)
|
170 |
{
|
171 |
return(pow(sqrt(log(exp(atanh(tanh(asinh(sinh(acosh(cosh(atan(tan(asin(sin(acos(cos(x))))))))))))))),2));
|
172 |
}
|
173 |
|
174 |
__kernel void sillysum(
|
175 |
__global const float *a_g, __global const float *b_g, __global float *res_g)
|
176 |
{
|
177 |
int gid = get_global_id(0);
|
178 |
res_g[gid] = MySillyFunction(a_g[gid]) + MySillyFunction(b_g[gid]);
|
179 |
}
|
180 |
|
181 |
__kernel void sum(
|
182 |
__global const float *a_g, __global const float *b_g, __global float *res_g)
|
183 |
{
|
184 |
int gid = get_global_id(0);
|
185 |
res_g[gid] = a_g[gid] + b_g[gid];
|
186 |
}
|
187 |
""").build()
|
188 |
Elapsed=time.time()-TimeIn |
189 |
print("Building kernels : %.3f" % Elapsed)
|
190 |
|
191 |
TimeIn=time.time() |
192 |
# Memory allocation on Device for result
|
193 |
res_g = cl.Buffer(ctx, mf.WRITE_ONLY, a_np.nbytes) |
194 |
Elapsed=time.time()-TimeIn |
195 |
print("Allocation on Device for results : %.3f" % Elapsed)
|
196 |
|
197 |
TimeIn=time.time() |
198 |
# Synthesis of function "sillysum" inside Kernel Sources
|
199 |
knl = prg.sillysum # Use this Kernel object for repeated calls
|
200 |
Elapsed=time.time()-TimeIn |
201 |
print("Synthesis of kernel : %.3f" % Elapsed)
|
202 |
|
203 |
TimeIn=time.time() |
204 |
# Call of kernel previously defined
|
205 |
CallCL=knl(queue, a_np.shape, None, a_g, b_g, res_g)
|
206 |
#
|
207 |
CallCL.wait() |
208 |
Elapsed=time.time()-TimeIn |
209 |
print("Execution of kernel : %.3f" % Elapsed)
|
210 |
|
211 |
TimeIn=time.time() |
212 |
# Creation of vector for result with same size as input vectors
|
213 |
res_np = np.empty_like(a_np) |
214 |
Elapsed=time.time()-TimeIn |
215 |
print("Allocation on Host for results: %.3f" % Elapsed)
|
216 |
|
217 |
TimeIn=time.time() |
218 |
# Copy from Device to Host
|
219 |
cl.enqueue_copy(queue, res_np, res_g) |
220 |
Elapsed=time.time()-TimeIn |
221 |
print("Copy from Device 2 Host : %.3f" % Elapsed)
|
222 |
|
223 |
# Liberation of memory
|
224 |
a_g.release() |
225 |
b_g.release() |
226 |
res_g.release() |
227 |
|
228 |
return(res_np)
|
229 |
|
230 |
import sys |
231 |
import time |
232 |
|
233 |
if __name__=='__main__': |
234 |
|
235 |
# Size of input vectors definition based on stdin
|
236 |
import sys |
237 |
try:
|
238 |
SIZE=int(sys.argv[1]) |
239 |
print("Size of vectors set to %i" % SIZE)
|
240 |
except:
|
241 |
SIZE=50000
|
242 |
print("Size of vectors set to default size %i" % SIZE)
|
243 |
|
244 |
a_np = np.random.rand(SIZE).astype(np.float32) |
245 |
b_np = np.random.rand(SIZE).astype(np.float32) |
246 |
|
247 |
# Native Implementation
|
248 |
TimeIn=time.time() |
249 |
# res_np=NativeAddition(a_np,b_np)
|
250 |
res_np=NativeSillyAddition(a_np,b_np) |
251 |
NativeElapsed=time.time()-TimeIn |
252 |
NativeRate=int(SIZE/NativeElapsed)
|
253 |
print("NativeRate: %i" % NativeRate)
|
254 |
|
255 |
# OpenCL Implementation
|
256 |
TimeIn=time.time() |
257 |
# res_cl=OpenCLAddition(a_np,b_np)
|
258 |
res_cl=OpenCLSillyAddition(a_np,b_np) |
259 |
OpenCLElapsed=time.time()-TimeIn |
260 |
OpenCLRate=int(SIZE/OpenCLElapsed)
|
261 |
print("OpenCLRate: %i" % OpenCLRate)
|
262 |
|
263 |
# CUDA Implementation
|
264 |
TimeIn=time.time() |
265 |
# res_cuda=CUDAAddition(a_np,b_np)
|
266 |
res_cuda=CUDASillyAddition(a_np,b_np) |
267 |
CUDAElapsed=time.time()-TimeIn |
268 |
CUDARate=int(SIZE/CUDAElapsed)
|
269 |
print("CUDARate: %i" % CUDARate)
|
270 |
|
271 |
print("OpenCLvsNative ratio: %f" % (OpenCLRate/NativeRate))
|
272 |
print("CUDAvsNative ratio: %f" % (CUDARate/NativeRate))
|
273 |
|
274 |
# Check on OpenCL with Numpy:
|
275 |
print(res_cl - res_np) |
276 |
print(np.linalg.norm(res_cl - res_np)) |
277 |
try:
|
278 |
assert np.allclose(res_np, res_cl)
|
279 |
except:
|
280 |
print("Results between Native & OpenCL seem to be too different!")
|
281 |
|
282 |
# Check on CUDA with Numpy:
|
283 |
print(res_cuda - res_np) |
284 |
print(np.linalg.norm(res_cuda - res_np)) |
285 |
try:
|
286 |
assert np.allclose(res_np, res_cuda)
|
287 |
except:
|
288 |
print("Results between Native & CUDA seem to be too different!")
|
289 |
|
290 |
|