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
|
// =================================================================================================
// This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. This
// project loosely follows the Google C++ styleguide and uses a tab-size of two spaces and a max-
// width of 100 characters per line.
//
// Author(s):
// Cedric Nugteren <www.cedricnugteren.nl>
//
// This is part 2 of 3 of the GEMM kernel. See part 1 for more information.
//
// =================================================================================================
// Enables loading of this file using the C++ pre-processor's #include (C++11 standard raw string
// literal). Comment-out this line for syntax-highlighting when developing.
R"(
// =================================================================================================
// The vectorised multiply-add function
INLINE_FUNC realM MultiplyAddVector(realM cvec, const realM avec, const real bval) {
#if USE_VECTOR_MAD == 1
cvec += avec * bval;
#else
#if VWM == 1
MultiplyAdd(cvec, avec, bval);
#elif VWM == 2
MultiplyAdd(cvec.x , avec.x, bval);
MultiplyAdd(cvec.y , avec.y, bval);
#elif VWM == 4
MultiplyAdd(cvec.x , avec.x, bval);
MultiplyAdd(cvec.y , avec.y, bval);
MultiplyAdd(cvec.z , avec.z, bval);
MultiplyAdd(cvec.w , avec.w, bval);
#elif VWM == 8
MultiplyAdd(cvec.s0, avec.s0, bval);
MultiplyAdd(cvec.s1, avec.s1, bval);
MultiplyAdd(cvec.s2, avec.s2, bval);
MultiplyAdd(cvec.s3, avec.s3, bval);
MultiplyAdd(cvec.s4, avec.s4, bval);
MultiplyAdd(cvec.s5, avec.s5, bval);
MultiplyAdd(cvec.s6, avec.s6, bval);
MultiplyAdd(cvec.s7, avec.s7, bval);
#elif VWM == 16
MultiplyAdd(cvec.s0, avec.s0, bval);
MultiplyAdd(cvec.s1, avec.s1, bval);
MultiplyAdd(cvec.s2, avec.s2, bval);
MultiplyAdd(cvec.s3, avec.s3, bval);
MultiplyAdd(cvec.s4, avec.s4, bval);
MultiplyAdd(cvec.s5, avec.s5, bval);
MultiplyAdd(cvec.s6, avec.s6, bval);
MultiplyAdd(cvec.s7, avec.s7, bval);
MultiplyAdd(cvec.s8, avec.s8, bval);
MultiplyAdd(cvec.s9, avec.s9, bval);
MultiplyAdd(cvec.sA, avec.sA, bval);
MultiplyAdd(cvec.sB, avec.sB, bval);
MultiplyAdd(cvec.sC, avec.sC, bval);
MultiplyAdd(cvec.sD, avec.sD, bval);
MultiplyAdd(cvec.sE, avec.sE, bval);
MultiplyAdd(cvec.sF, avec.sF, bval);
#endif
#endif
return cvec;
}
// Performs the actual computation: Cpm += Apm * Bpm
INLINE_FUNC void MultiplyAccumulate(realM cpm[NWI*MWI/VWM], realM apm[MWI/VWM], realN bpm[NWI/VWN]) {
#pragma unroll
for (int _ni = 0; _ni < NWI/VWN; _ni += 1) {
#pragma unroll
for (int _mi = 0; _mi < MWI/VWM; _mi += 1) {
const realM aval = apm[_mi];
#if VWN == 1
cpm[(_ni*VWN + 0)*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 0)*(MWI/VWM) + _mi], aval, bpm[_ni]);
#elif VWN == 2
cpm[(_ni*VWN + 0)*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 0)*(MWI/VWM) + _mi], aval, bpm[_ni].x);
cpm[(_ni*VWN + 1)*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 1)*(MWI/VWM) + _mi], aval, bpm[_ni].y);
#elif VWN == 4
cpm[(_ni*VWN + 0)*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 0)*(MWI/VWM) + _mi], aval, bpm[_ni].x);
cpm[(_ni*VWN + 1)*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 1)*(MWI/VWM) + _mi], aval, bpm[_ni].y);
cpm[(_ni*VWN + 2)*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 2)*(MWI/VWM) + _mi], aval, bpm[_ni].z);
cpm[(_ni*VWN + 3)*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 3)*(MWI/VWM) + _mi], aval, bpm[_ni].w);
#elif VWN == 8
cpm[(_ni*VWN + 0)*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 0)*(MWI/VWM) + _mi], aval, bpm[_ni].s0);
cpm[(_ni*VWN + 1)*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 1)*(MWI/VWM) + _mi], aval, bpm[_ni].s1);
cpm[(_ni*VWN + 2)*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 2)*(MWI/VWM) + _mi], aval, bpm[_ni].s2);
cpm[(_ni*VWN + 3)*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 3)*(MWI/VWM) + _mi], aval, bpm[_ni].s3);
cpm[(_ni*VWN + 4)*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 4)*(MWI/VWM) + _mi], aval, bpm[_ni].s4);
cpm[(_ni*VWN + 5)*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 5)*(MWI/VWM) + _mi], aval, bpm[_ni].s5);
cpm[(_ni*VWN + 6)*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 6)*(MWI/VWM) + _mi], aval, bpm[_ni].s6);
cpm[(_ni*VWN + 7)*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 7)*(MWI/VWM) + _mi], aval, bpm[_ni].s7);
#elif VWN == 16
cpm[(_ni*VWN + 0 )*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 0 )*(MWI/VWM) + _mi], aval, bpm[_ni].s0);
cpm[(_ni*VWN + 1 )*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 1 )*(MWI/VWM) + _mi], aval, bpm[_ni].s1);
cpm[(_ni*VWN + 2 )*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 2 )*(MWI/VWM) + _mi], aval, bpm[_ni].s2);
cpm[(_ni*VWN + 3 )*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 3 )*(MWI/VWM) + _mi], aval, bpm[_ni].s3);
cpm[(_ni*VWN + 4 )*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 4 )*(MWI/VWM) + _mi], aval, bpm[_ni].s4);
cpm[(_ni*VWN + 5 )*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 5 )*(MWI/VWM) + _mi], aval, bpm[_ni].s5);
cpm[(_ni*VWN + 6 )*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 6 )*(MWI/VWM) + _mi], aval, bpm[_ni].s6);
cpm[(_ni*VWN + 7 )*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 7 )*(MWI/VWM) + _mi], aval, bpm[_ni].s7);
cpm[(_ni*VWN + 8 )*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 8 )*(MWI/VWM) + _mi], aval, bpm[_ni].s8);
cpm[(_ni*VWN + 9 )*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 9 )*(MWI/VWM) + _mi], aval, bpm[_ni].s9);
cpm[(_ni*VWN + 10)*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 10)*(MWI/VWM) + _mi], aval, bpm[_ni].sA);
cpm[(_ni*VWN + 11)*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 11)*(MWI/VWM) + _mi], aval, bpm[_ni].sB);
cpm[(_ni*VWN + 12)*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 12)*(MWI/VWM) + _mi], aval, bpm[_ni].sC);
cpm[(_ni*VWN + 13)*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 13)*(MWI/VWM) + _mi], aval, bpm[_ni].sD);
cpm[(_ni*VWN + 14)*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 14)*(MWI/VWM) + _mi], aval, bpm[_ni].sE);
cpm[(_ni*VWN + 15)*(MWI/VWM) + _mi] = MultiplyAddVector(cpm[(_ni*VWN + 15)*(MWI/VWM) + _mi], aval, bpm[_ni].sF);
#endif
}
}
}
// =================================================================================================
// Merges the results in Cpm with the global array in Cgm. This also performs the multiplication
// with the constants: Cgm = alpha*A*B + beta*Cgm = alpha*Cpm + beta*Cgm
INLINE_FUNC void StoreResults(__global realM* cgm, realM cpm[NWI*MWI/VWM], const int kSizeM,
const real alpha, const real beta) {
#pragma unroll
for (int _ni = 0; _ni < NWI; _ni += 1) {
#pragma unroll
for (int _mi = 0; _mi < MWI/VWM; _mi += 1) {
#if STRM == 0
int mg = _mi + get_local_id(0)*(MWI/VWM);
#elif STRM == 1
int mg = get_local_id(0) + _mi*MDIMC;
#endif
#if STRN == 0
int ng = _ni + get_local_id(1)*NWI;
#elif STRN == 1
int ng = _ni%VWN + get_local_id(1)*VWN + (_ni/VWN)*VWN*NDIMC;
#endif
int idm = mg + GetGroupID0() * (MWG/VWM);
int idn = ng + GetGroupID1() * NWG;
int index = idn*(kSizeM/VWM) + idm;
realM result;
realM xval = cpm[_ni * (MWI/VWM) + _mi];
// The final multiplication with alpha (in case beta == 0)
if (IsZero(beta)) {
#if VWM == 1
Multiply(result, alpha, xval);
#elif VWM == 2
Multiply(result.x, alpha, xval.x);
Multiply(result.y, alpha, xval.y);
#elif VWM == 4
Multiply(result.x, alpha, xval.x);
Multiply(result.y, alpha, xval.y);
Multiply(result.z, alpha, xval.z);
Multiply(result.w, alpha, xval.w);
#elif VWM == 8
Multiply(result.s0, alpha, xval.s0);
Multiply(result.s1, alpha, xval.s1);
Multiply(result.s2, alpha, xval.s2);
Multiply(result.s3, alpha, xval.s3);
Multiply(result.s4, alpha, xval.s4);
Multiply(result.s5, alpha, xval.s5);
Multiply(result.s6, alpha, xval.s6);
Multiply(result.s7, alpha, xval.s7);
#elif VWM == 16
Multiply(result.s0, alpha, xval.s0);
Multiply(result.s1, alpha, xval.s1);
Multiply(result.s2, alpha, xval.s2);
Multiply(result.s3, alpha, xval.s3);
Multiply(result.s4, alpha, xval.s4);
Multiply(result.s5, alpha, xval.s5);
Multiply(result.s6, alpha, xval.s6);
Multiply(result.s7, alpha, xval.s7);
Multiply(result.s8, alpha, xval.s8);
Multiply(result.s9, alpha, xval.s9);
Multiply(result.sA, alpha, xval.sA);
Multiply(result.sB, alpha, xval.sB);
Multiply(result.sC, alpha, xval.sC);
Multiply(result.sD, alpha, xval.sD);
Multiply(result.sE, alpha, xval.sE);
Multiply(result.sF, alpha, xval.sF);
#endif
}
// The final multiplication with alpha and the addition with beta*C
else {
realM yval = cgm[index];
#if VWM == 1
AXPBY(result, alpha, xval, beta, yval);
#elif VWM == 2
AXPBY(result.x, alpha, xval.x, beta, yval.x);
AXPBY(result.y, alpha, xval.y, beta, yval.y);
#elif VWM == 4
AXPBY(result.x, alpha, xval.x, beta, yval.x);
AXPBY(result.y, alpha, xval.y, beta, yval.y);
AXPBY(result.z, alpha, xval.z, beta, yval.z);
AXPBY(result.w, alpha, xval.w, beta, yval.w);
#elif VWM == 8
AXPBY(result.s0, alpha, xval.s0, beta, yval.s0);
AXPBY(result.s1, alpha, xval.s1, beta, yval.s1);
AXPBY(result.s2, alpha, xval.s2, beta, yval.s2);
AXPBY(result.s3, alpha, xval.s3, beta, yval.s3);
AXPBY(result.s4, alpha, xval.s4, beta, yval.s4);
AXPBY(result.s5, alpha, xval.s5, beta, yval.s5);
AXPBY(result.s6, alpha, xval.s6, beta, yval.s6);
AXPBY(result.s7, alpha, xval.s7, beta, yval.s7);
#elif VWM == 16
AXPBY(result.s0, alpha, xval.s0, beta, yval.s0);
AXPBY(result.s1, alpha, xval.s1, beta, yval.s1);
AXPBY(result.s2, alpha, xval.s2, beta, yval.s2);
AXPBY(result.s3, alpha, xval.s3, beta, yval.s3);
AXPBY(result.s4, alpha, xval.s4, beta, yval.s4);
AXPBY(result.s5, alpha, xval.s5, beta, yval.s5);
AXPBY(result.s6, alpha, xval.s6, beta, yval.s6);
AXPBY(result.s7, alpha, xval.s7, beta, yval.s7);
AXPBY(result.s8, alpha, xval.s8, beta, yval.s8);
AXPBY(result.s9, alpha, xval.s9, beta, yval.s9);
AXPBY(result.sA, alpha, xval.sA, beta, yval.sA);
AXPBY(result.sB, alpha, xval.sB, beta, yval.sB);
AXPBY(result.sC, alpha, xval.sC, beta, yval.sC);
AXPBY(result.sD, alpha, xval.sD, beta, yval.sD);
AXPBY(result.sE, alpha, xval.sE, beta, yval.sE);
AXPBY(result.sF, alpha, xval.sF, beta, yval.sF);
#endif
}
cgm[index] = result;
}
}
}
// =================================================================================================
// End of the C++11 raw string literal
)"
// =================================================================================================
|
