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
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
|
/* ************************************************************************
* Copyright 2013 Advanced Micro Devices, Inc.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* ************************************************************************/
static const char *red_sum_kernel = "
#ifdef DOUBLE_PRECISION
#ifdef cl_khr_fp64
#pragma OPENCL EXTENSION cl_khr_fp64 : enable
#else
#pragma OPENCL EXTENSION cl_amd_fp64 : enable
#endif
#endif
__kernel void %PREFIXred_sum_kernel( __global %TYPE *_X, __global %TYPE *_res,
uint N, uint offx, uint offRes )
{
__global %TYPE *X = _X + offx;
__global %TYPE *res = _res + offRes;
%TYPE redVal = (%TYPE) 0.0;
int gOffset;
for( gOffset=(get_global_id(0) * %V); (gOffset + %V - 1)<N; gOffset+=( get_global_size(0) * %V ) )
{
%TYPE%V reg1;
reg1 = %VLOAD( 0, (X + gOffset) );
redVal += %REDUCE_SUM( reg1 );
}
// Loop for the last thread to handle the tail part of the vector
// Using the same gOffset used above
for( ; gOffset<N; gOffset++ )
{
redVal += X[gOffset];
}
// Note: this has to be called outside any if-conditions- because REDUCTION uses barrier
%REDUCTION_BY_SUM( redVal );
if( (get_local_id(0)) == 0 ) {
res[0] = redVal;
}
}
\n";
static const char *red_max_kernel = "
#ifdef DOUBLE_PRECISION
#ifdef cl_khr_fp64
#pragma OPENCL EXTENSION cl_khr_fp64 : enable
#else
#pragma OPENCL EXTENSION cl_amd_fp64 : enable
#endif
#define MAX 0x1.fffffffffffffp1023 // Max in case of d/z (values from khronos site)
#else
#define MAX 0x1.fffffep127f // Max in case of s/c
#endif
__kernel void %PREFIXred_max_kernel( __global %TYPE *_X, __global %TYPE *_res,
uint N, uint offx, uint offRes )
{
__global %TYPE *X = _X + offx;
__global %TYPE *res = _res + offRes;
%TYPE redVal = (%TYPE) - MAX;
int gOffset;
for( gOffset=(get_global_id(0) * %V); (gOffset + %V - 1)<N; gOffset+=( get_global_size(0) * %V ) )
{
%TYPE%V reg1;
reg1 = %VLOAD( 0, (X + gOffset) );
%TYPE scalarMax = %REDUCE_MAX( reg1 );
redVal = fmax( redVal, scalarMax );
}
// Loop for the last thread to handle the tail part of the vector
// Using the same gOffset used above
for( ; gOffset<N; gOffset++ )
{
redVal = fmax( redVal, X[gOffset] );
}
// Note: this has to be called outside any if-conditions- because REDUCTION uses barrier
%REDUCTION_BY_MAX( redVal );
if( (get_local_id(0)) == 0 ) {
res[0] = redVal;
}
}
\n";
static const char *red_min_kernel = "
#ifdef DOUBLE_PRECISION
#ifdef cl_khr_fp64
#pragma OPENCL EXTENSION cl_khr_fp64 : enable
#else
#pragma OPENCL EXTENSION cl_amd_fp64 : enable
#endif
#define MAX 0x1.fffffffffffffp1023 // Max in case of d/z (values from khronos site)
#else
#define MAX 0x1.fffffep127f // Max in case of s/c
#endif
__kernel void %PREFIXred_min_kernel( __global %TYPE *_X, __global %TYPE *_res,
uint N, uint offx, uint offRes )
{
__global %TYPE *X = _X + offx;
__global %TYPE *res = _res + offRes;
%TYPE redVal = (%TYPE) MAX;
int gOffset;
for( gOffset=(get_global_id(0) * %V); (gOffset + %V - 1)<N; gOffset+=( get_global_size(0) * %V ) )
{
%TYPE%V reg1;
reg1 = %VLOAD( 0, (X + gOffset) );
%TYPE scalarMin = %REDUCE_MIN( reg1 );
redVal = fmin( redVal, scalarMin );
}
// Loop for the last thread to handle the tail part of the vector
// Using the same gOffset used above
for( ; gOffset<N; gOffset++ )
{
redVal = fmin( redVal, X[gOffset] );
}
// Note: this has to be called outside any if-conditions- because REDUCTION uses barrier
%REDUCTION_BY_MIN( redVal );
if( (get_local_id(0)) == 0 ) {
res[0] = redVal;
}
}
\n";
static const char *red_with_index_kernel = "
/******************************************************
* Implementations available for REDUCTION_BY_MAX
0 - ATOMIC_FLI
1 - REG_FLI,
2 - ATOMIC_FHI,
3 - REG_FHI
Implementation available for REDUCE_MAX
0 - FHI
1 - FLI
***************************************************/
#ifdef DOUBLE_PRECISION
#ifdef cl_khr_fp64
#pragma OPENCL EXTENSION cl_khr_fp64 : enable
#else
#pragma OPENCL EXTENSION cl_amd_fp64 : enable
#endif
#define MIN 0x1.0p-1022 // Min in case of d/z (values from khronos site)
#else
#define MIN 0x1.0p-126f // Min in case od s/c
#endif
__kernel void %PREFIXred_with_index_kernel( __global %TYPE *_X, __global uint *_res,
uint N, uint offx, uint offRes )
{
__global %TYPE *X = _X + offx;
__global uint *XIndex = (__global uint*)(&X[N]);
__global uint *res = _res + offRes;
%TYPE maxVal = (%TYPE)MIN, val = (%TYPE)MIN;
uint maxIndex = 0, index = 0;
int gOffset;
for( gOffset=(get_global_id(0) * %V); (gOffset + %V - 1)<N; gOffset+=( get_global_size(0) * %V ) )
{
%TYPE%V vReg1 = %VLOAD( 0, (X + gOffset) );
%REDUCE_MAX(vReg1,val,index,1); // Find max within a vector
if(val > maxVal)
{
maxVal = val;
maxIndex = XIndex[(gOffset + index)];
}
}
// Loop for the last thread to handle the tail part of the vector
// Using the same gOffset used above
for( ; gOffset<N; gOffset++ )
{
%TYPE sreg1 = X[gOffset];
if(sreg1 > maxVal)
{
maxVal = sreg1;
maxIndex = XIndex[gOffset];
}
}
// Note: this has to be called outside any if-conditions- because REDUCTION uses barrier
#ifdef REDUCE_MAX_WITH_INDEX_ATOMICS
%REDUCTION_BY_MAX(maxVal,maxIndex,0);
#else
%REDUCTION_BY_MAX(maxVal,maxIndex,1);
#endif
if(get_local_id(0) == 0)
{
res[0] = maxIndex;
}
}
\n";
static const char *red_hypot_kernel = "
#ifdef DOUBLE_PRECISION
#ifdef cl_khr_fp64
#pragma OPENCL EXTENSION cl_khr_fp64 : enable
#else
#pragma OPENCL EXTENSION cl_amd_fp64 : enable
#endif
#endif
__kernel void %PREFIXred_hypot_kernel( __global %TYPE *_X, __global %TYPE *_res,
uint N, uint offx, uint offRes )
{
__global %TYPE *X = _X + offx;
__global %TYPE *res = _res + offRes;
%TYPE redVal = (%TYPE) 0.0;
int gOffset;
for( gOffset=(get_global_id(0) * %V); (gOffset + %V - 1)<N; gOffset+=( get_global_size(0) * %V ) )
{
%TYPE%V reg1;
reg1 = %VLOAD( 0, (X + gOffset) );
%TYPE scalarHypot = %REDUCE_HYPOT( reg1 );
redVal = hypot( redVal, scalarHypot );
}
// Loop for the last thread to handle the tail part of the vector
// Using the same gOffset used above
for( ; gOffset<N; gOffset++ )
{
redVal = hypot( redVal, X[gOffset] );
}
// Note: this has to be called outside any if-conditions- because REDUCTION uses barrier
%REDUCTION_BY_HYPOT( redVal );
if( (get_local_id(0)) == 0 ) {
res[0] = redVal;
}
}
\n";
static const char *red_ssq_kernel = "
#ifdef DOUBLE_PRECISION
#ifdef cl_khr_fp64
#pragma OPENCL EXTENSION cl_khr_fp64 : enable
#else
#pragma OPENCL EXTENSION cl_amd_fp64 : enable
#endif
#define MAX 0x1.fffffffffffffp1023 // Max in case of d/z (values from khronos site)
#else
#define MAX 0x1.fffffep127f // Max in case of s/c
#endif
#define ZERO (%TYPE)0.0
// Since scale & ssq are always of primitive type,
// This kernel will always be called only for float/double
__kernel void %PREFIXred_ssq_kernel( __global %TYPE *_X, __global %TYPE *_res,
uint N, uint offx, uint offRes )
{
__global %TYPE *X = _X + offx;
__global %TYPE *res = _res + offRes;
%TYPE scale = -MAX;
int gOffset;
for( gOffset=(get_global_id(0) * %V); (gOffset + %V - 1)<N; gOffset+=( get_global_size(0) * %V ) )
{
%TYPE%V scale1;
scale1 = %VLOAD( 0, (X + gOffset) );
%TYPE regMax = %REDUCE_MAX( scale1 );
scale = fmax( scale, regMax );
}
for( ; gOffset<N; gOffset++ )
{
%TYPE sReg;
sReg = X[gOffset];
scale = fmax( scale, sReg );
}
%REDUCTION_BY_MAX( scale );
__local %TYPE _scaleOfWG;
if( (get_local_id(0)) == 0 ) {
_scaleOfWG = scale;
}
barrier(CLK_LOCAL_MEM_FENCE);
// At this point we have scale.
// Now we calculate ssq by loading the array again and dividing the
// elements by scale and squaring it.
%TYPE ssq = (%TYPE) 0.0;
%TYPE scaleOfWG = _scaleOfWG;
// If scale was zero, that means the whole array encountered before was filled with zeroes
// Note: scale is a local variable, either all enter or none
if(isnotequal(scaleOfWG, ZERO))
{
for( gOffset=(get_global_id(0) * %V); (gOffset + %V - 1)<N; gOffset+=( get_global_size(0) * %V ) )
{
%TYPE%V scale1, ssq1;
scale1 = %VLOAD( 0, (X + gOffset) );
ssq1 = %VLOAD( 0, (X + gOffset + N) );
%TYPE%V tempSsq = (scale1 / scaleOfWG) * (scale1 / scaleOfWG) * ssq1;
ssq += %REDUCE_SUM( tempSsq );
}
for( ; gOffset<N; gOffset++ )
{
%TYPE scale1, ssq1;
scale1 = X[gOffset];
ssq1 = X[gOffset + N];
ssq += (scale1 / scaleOfWG) * (scale1 / scaleOfWG) * ssq1;
}
%REDUCTION_BY_SUM( ssq );
}
if( (get_local_id(0)) == 0 ) {
res[0] = scaleOfWG * sqrt(ssq);
}
}
\n";
|
