root / ETSN / MySteps_5c.py @ 291
Historique | Voir | Annoter | Télécharger (10,36 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(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(a_np))))))))))))))))+MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(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 |
mod = SourceModule("""
|
| 50 |
__device__ float MySillyFunction(float x)
|
| 51 |
{
|
| 52 |
return(pow(sqrt(log(exp(atanh(tanh(asinh(sinh(acosh(cosh(atan(tan(asin(sin(acos(cos(x))))))))))))))),2));
|
| 53 |
}
|
| 54 |
|
| 55 |
__global__ void sillysum(float *dest, float *a, float *b)
|
| 56 |
{
|
| 57 |
const int i = blockIdx.x;
|
| 58 |
dest[i] = MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(a[i])))))))))))))))) + MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(b[i]))))))))))))))));
|
| 59 |
}
|
| 60 |
|
| 61 |
__global__ void hybridsillysum(float *dest, float *a, float *b)
|
| 62 |
{
|
| 63 |
const int i = threadIdx.x+blockDim.x*blockIdx.x;
|
| 64 |
dest[i] = MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(a[i])))))))))))))))) + MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(b[i]))))))))))))))));
|
| 65 |
}
|
| 66 |
""")
|
| 67 |
|
| 68 |
# sum = mod.get_function("sum")
|
| 69 |
# sillysum = mod.get_function("sillysum")
|
| 70 |
hybridsillysum = mod.get_function("hybridsillysum")
|
| 71 |
|
| 72 |
res_np = numpy.zeros_like(a_np) |
| 73 |
threads=1024
|
| 74 |
blocks=int(a_np.size/threads)
|
| 75 |
# sillysum(drv.Out(res_np), drv.In(a_np), drv.In(b_np),
|
| 76 |
# block=(threads,1,1), grid=(blocks,1))
|
| 77 |
hybridsillysum(drv.Out(res_np), drv.In(a_np), drv.In(b_np), |
| 78 |
block=(threads,1,1), grid=(blocks,1)) |
| 79 |
return(res_np)
|
| 80 |
|
| 81 |
# OpenCL complete operation
|
| 82 |
def OpenCLAddition(a_np,b_np): |
| 83 |
|
| 84 |
# Context creation
|
| 85 |
ctx = cl.create_some_context() |
| 86 |
# Every process is stored in a queue
|
| 87 |
queue = cl.CommandQueue(ctx) |
| 88 |
|
| 89 |
TimeIn=time.time() |
| 90 |
# Copy from Host to Device using pointers
|
| 91 |
mf = cl.mem_flags |
| 92 |
a_g = cl.Buffer(ctx, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=a_np) |
| 93 |
b_g = cl.Buffer(ctx, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=b_np) |
| 94 |
Elapsed=time.time()-TimeIn |
| 95 |
print("Copy from Host 2 Device : %.3f" % Elapsed)
|
| 96 |
|
| 97 |
TimeIn=time.time() |
| 98 |
# Definition of kernel under OpenCL
|
| 99 |
prg = cl.Program(ctx, """
|
| 100 |
__kernel void sum(
|
| 101 |
__global const float *a_g, __global const float *b_g, __global float *res_g)
|
| 102 |
{
|
| 103 |
int gid = get_global_id(0);
|
| 104 |
res_g[gid] = a_g[gid] + b_g[gid];
|
| 105 |
}
|
| 106 |
""").build()
|
| 107 |
Elapsed=time.time()-TimeIn |
| 108 |
print("Building kernels : %.3f" % Elapsed)
|
| 109 |
|
| 110 |
TimeIn=time.time() |
| 111 |
# Memory allocation on Device for result
|
| 112 |
res_g = cl.Buffer(ctx, mf.WRITE_ONLY, a_np.nbytes) |
| 113 |
Elapsed=time.time()-TimeIn |
| 114 |
print("Allocation on Device for results : %.3f" % Elapsed)
|
| 115 |
|
| 116 |
TimeIn=time.time() |
| 117 |
# Synthesis of function "sum" inside Kernel Sources
|
| 118 |
knl = prg.sum # Use this Kernel object for repeated calls
|
| 119 |
Elapsed=time.time()-TimeIn |
| 120 |
print("Synthesis of kernel : %.3f" % Elapsed)
|
| 121 |
|
| 122 |
TimeIn=time.time() |
| 123 |
# Call of kernel previously defined
|
| 124 |
knl(queue, a_np.shape, None, a_g, b_g, res_g)
|
| 125 |
Elapsed=time.time()-TimeIn |
| 126 |
print("Execution of kernel : %.3f" % Elapsed)
|
| 127 |
|
| 128 |
TimeIn=time.time() |
| 129 |
# Creation of vector for result with same size as input vectors
|
| 130 |
res_np = np.empty_like(a_np) |
| 131 |
Elapsed=time.time()-TimeIn |
| 132 |
print("Allocation on Host for results: %.3f" % Elapsed)
|
| 133 |
|
| 134 |
TimeIn=time.time() |
| 135 |
# Copy from Device to Host
|
| 136 |
cl.enqueue_copy(queue, res_np, res_g) |
| 137 |
Elapsed=time.time()-TimeIn |
| 138 |
print("Copy from Device 2 Host : %.3f" % Elapsed)
|
| 139 |
|
| 140 |
# Liberation of memory
|
| 141 |
a_g.release() |
| 142 |
b_g.release() |
| 143 |
res_g.release() |
| 144 |
|
| 145 |
return(res_np)
|
| 146 |
|
| 147 |
# OpenCL complete operation
|
| 148 |
def OpenCLSillyAddition(a_np,b_np): |
| 149 |
|
| 150 |
# Context creation
|
| 151 |
ctx = cl.create_some_context() |
| 152 |
# Every process is stored in a queue
|
| 153 |
queue = cl.CommandQueue(ctx) |
| 154 |
|
| 155 |
TimeIn=time.time() |
| 156 |
# Copy from Host to Device using pointers
|
| 157 |
mf = cl.mem_flags |
| 158 |
a_g = cl.Buffer(ctx, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=a_np) |
| 159 |
b_g = cl.Buffer(ctx, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=b_np) |
| 160 |
Elapsed=time.time()-TimeIn |
| 161 |
print("Copy from Host 2 Device : %.3f" % Elapsed)
|
| 162 |
|
| 163 |
TimeIn=time.time() |
| 164 |
# Definition of kernel under OpenCL
|
| 165 |
prg = cl.Program(ctx, """
|
| 166 |
|
| 167 |
float MySillyFunction(float x)
|
| 168 |
{
|
| 169 |
return(pow(sqrt(log(exp(atanh(tanh(asinh(sinh(acosh(cosh(atan(tan(asin(sin(acos(cos(x))))))))))))))),2));
|
| 170 |
}
|
| 171 |
|
| 172 |
__kernel void sillysum(
|
| 173 |
__global const float *a_g, __global const float *b_g, __global float *res_g)
|
| 174 |
{
|
| 175 |
int gid = get_global_id(0);
|
| 176 |
res_g[gid] = MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(a_g[gid])))))))))))))))) + MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(MySillyFunction(b_g[gid]))))))))))))))));
|
| 177 |
}
|
| 178 |
|
| 179 |
__kernel void sum(
|
| 180 |
__global const float *a_g, __global const float *b_g, __global float *res_g)
|
| 181 |
{
|
| 182 |
int gid = get_global_id(0);
|
| 183 |
res_g[gid] = a_g[gid] + b_g[gid];
|
| 184 |
}
|
| 185 |
""").build()
|
| 186 |
Elapsed=time.time()-TimeIn |
| 187 |
print("Building kernels : %.3f" % Elapsed)
|
| 188 |
|
| 189 |
TimeIn=time.time() |
| 190 |
# Memory allocation on Device for result
|
| 191 |
res_g = cl.Buffer(ctx, mf.WRITE_ONLY, a_np.nbytes) |
| 192 |
Elapsed=time.time()-TimeIn |
| 193 |
print("Allocation on Device for results : %.3f" % Elapsed)
|
| 194 |
|
| 195 |
TimeIn=time.time() |
| 196 |
# Synthesis of function "sillysum" inside Kernel Sources
|
| 197 |
knl = prg.sillysum # Use this Kernel object for repeated calls
|
| 198 |
Elapsed=time.time()-TimeIn |
| 199 |
print("Synthesis of kernel : %.3f" % Elapsed)
|
| 200 |
|
| 201 |
TimeIn=time.time() |
| 202 |
# Call of kernel previously defined
|
| 203 |
CallCL=knl(queue, a_np.shape, None, a_g, b_g, res_g)
|
| 204 |
#
|
| 205 |
CallCL.wait() |
| 206 |
Elapsed=time.time()-TimeIn |
| 207 |
print("Execution of kernel : %.3f" % Elapsed)
|
| 208 |
|
| 209 |
TimeIn=time.time() |
| 210 |
# Creation of vector for result with same size as input vectors
|
| 211 |
res_np = np.empty_like(a_np) |
| 212 |
Elapsed=time.time()-TimeIn |
| 213 |
print("Allocation on Host for results: %.3f" % Elapsed)
|
| 214 |
|
| 215 |
TimeIn=time.time() |
| 216 |
# Copy from Device to Host
|
| 217 |
cl.enqueue_copy(queue, res_np, res_g) |
| 218 |
Elapsed=time.time()-TimeIn |
| 219 |
print("Copy from Device 2 Host : %.3f" % Elapsed)
|
| 220 |
|
| 221 |
# Liberation of memory
|
| 222 |
a_g.release() |
| 223 |
b_g.release() |
| 224 |
res_g.release() |
| 225 |
|
| 226 |
return(res_np)
|
| 227 |
|
| 228 |
import sys |
| 229 |
import time |
| 230 |
|
| 231 |
if __name__=='__main__': |
| 232 |
|
| 233 |
# Size of input vectors definition based on stdin
|
| 234 |
import sys |
| 235 |
try:
|
| 236 |
SIZE=int(sys.argv[1]) |
| 237 |
print("Size of vectors set to %i" % SIZE)
|
| 238 |
except:
|
| 239 |
SIZE=50000
|
| 240 |
print("Size of vectors set to default size %i" % SIZE)
|
| 241 |
|
| 242 |
a_np = np.random.rand(SIZE).astype(np.float32) |
| 243 |
b_np = np.random.rand(SIZE).astype(np.float32) |
| 244 |
|
| 245 |
# Native Implementation
|
| 246 |
TimeIn=time.time() |
| 247 |
# res_np=NativeAddition(a_np,b_np)
|
| 248 |
res_np=NativeSillyAddition(a_np,b_np) |
| 249 |
NativeElapsed=time.time()-TimeIn |
| 250 |
NativeRate=int(SIZE/NativeElapsed)
|
| 251 |
print("NativeRate: %i" % NativeRate)
|
| 252 |
|
| 253 |
# OpenCL Implementation
|
| 254 |
TimeIn=time.time() |
| 255 |
# res_cl=OpenCLAddition(a_np,b_np)
|
| 256 |
res_cl=OpenCLSillyAddition(a_np,b_np) |
| 257 |
OpenCLElapsed=time.time()-TimeIn |
| 258 |
OpenCLRate=int(SIZE/OpenCLElapsed)
|
| 259 |
print("OpenCLRate: %i" % OpenCLRate)
|
| 260 |
|
| 261 |
# CUDA Implementation
|
| 262 |
TimeIn=time.time() |
| 263 |
# res_cuda=CUDAAddition(a_np,b_np)
|
| 264 |
res_cuda=CUDASillyAddition(a_np,b_np) |
| 265 |
CUDAElapsed=time.time()-TimeIn |
| 266 |
CUDARate=int(SIZE/CUDAElapsed)
|
| 267 |
print("CUDARate: %i" % CUDARate)
|
| 268 |
|
| 269 |
print("OpenCLvsNative ratio: %f" % (OpenCLRate/NativeRate))
|
| 270 |
print("CUDAvsNative ratio: %f" % (CUDARate/NativeRate))
|
| 271 |
|
| 272 |
# Check on CPU with Numpy:
|
| 273 |
print(res_cl - res_np) |
| 274 |
print(np.linalg.norm(res_cl - res_np)) |
| 275 |
try:
|
| 276 |
assert np.allclose(res_np, res_cl)
|
| 277 |
except:
|
| 278 |
print("Results between Native & OpenCL seem to be too different!")
|
| 279 |
|
| 280 |
# Check on CPU with Numpy:
|
| 281 |
print(res_cuda - res_np) |
| 282 |
print(np.linalg.norm(res_cuda - res_np)) |
| 283 |
try:
|
| 284 |
assert np.allclose(res_np, res_cuda)
|
| 285 |
except:
|
| 286 |
print("Results between Native & CUDA seem to be too different!")
|
| 287 |
|