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
|
// =================================================================================================
// 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 3 of 4 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"(
// =================================================================================================
// A common interface for subgroup functions
#if USE_SUBGROUP_SHUFFLING == 1
INLINE_FUNC int clblast_get_sub_group_local_id() {
// Intel extension
#if SUBGROUP_SHUFFLING_INTEL == 1
return get_sub_group_local_id();
// Nvidia inline PTX
#elif SUBGROUP_SHUFFLING_NVIDIA_PRE_VOLTA == 1 || SUBGROUP_SHUFFLING_NVIDIA_POST_VOLTA == 1
int ret;
asm volatile("mov.u32 %0, %%laneid;" : "=r"(ret) );
return ret;
#endif
}
INLINE_FUNC realN clblast_sub_group_shuffle(realN reg, int src) {
// Intel extension
#if SUBGROUP_SHUFFLING_INTEL == 1
return intel_sub_group_shuffle(reg, src);
// Nvidia inline PTX
// Volta and later requires .sync shuffle instructions with an extra mask arg
#elif SUBGROUP_SHUFFLING_NVIDIA_PRE_VOLTA == 1 || SUBGROUP_SHUFFLING_NVIDIA_POST_VOLTA == 1
realN ret;
#if SUBGROUP_SHUFFLING_NVIDIA_POST_VOLTA == 1
asm volatile("shfl.sync.idx.b32 %0, %1, %2, 0x1f, 0xffffffff;" : "=f"(ret): "f"(reg), "r"(src));
#else
asm volatile("shfl.idx.b32 %0, %1, %2, 0x1f;" : "=f"(ret): "f"(reg), "r"(src));
#endif
return ret;
#endif
}
#endif
// Main body of the matrix-multiplication algorithm. It calls various (inlined) functions.
INLINE_FUNC void XgemmBody(const int kSizeM, const int kSizeN, const int kSizeK,
const __global realM* restrict agm, const __global realN* restrict bgm,
__global realM* cgm, const real alpha, const real beta
#if SA == 1 && SB == 1
, LOCAL_PTR realM* alm, LOCAL_PTR realN* blm
#elif SA == 1
, LOCAL_PTR realM* alm
#elif SB == 1
, LOCAL_PTR realN* blm
#endif
) {
// Allocates workitem-private memory (registers)
#if GEMMK == 0
#pragma promote_to_registers
realM apm[MWI/VWM]; // MWI * 1
#pragma promote_to_registers
realN bpm[NWI/VWN]; // 1 * NWI
#elif GEMMK == 1
#if USE_SUBGROUP_SHUFFLING == 1
#pragma promote_to_registers
realN apm[KREG/VWN]; // KREG (subgroup shuffling in NWI dimension)
#else
#pragma promote_to_registers
realN apm[NWI*(KREG/VWN)]; // NWI * KREG
#endif
#pragma promote_to_registers
realM bpm[KREG*(MWI/VWM)]; // KREG * MWI
#endif
#pragma promote_to_registers
realM cpm[NWI*(MWI/VWM)]; // NWI * MWI
#if GEMMK == 1
const __global real* restrict a_ptr = (const __global real* restrict) &agm[0];
const __global real* restrict b_ptr = (const __global real* restrict) &bgm[0];
const int tid_x = get_local_id(0) + MDIMC * GetGroupID0();
const int tid_y = get_local_id(1) + NDIMC * GetGroupID1();
#endif
// Combined thread identifier (volatile to disable caching)
#if SA == 1 || SB == 1
volatile int tid = get_local_id(0) + MDIMC*get_local_id(1);
#endif
// Initializes the accumulation registers
#pragma unroll
for (int _mi = 0; _mi < MWI/VWM; _mi += 1) {
#pragma unroll
for (int _ni = 0; _ni < NWI; _ni += 1) {
cpm[_ni * (MWI/VWM) + _mi] = InitAccRegisters();
}
}
// Loops over all workgroup tiles
for (int kwg = 0; kwg < kSizeK; kwg += KWG * KREG) {
// Loads data: off-chip --> local (matrix A)
#if SA == 1
GlobalToLocalA(agm, alm, kSizeM, tid, kwg);
#endif
// Loads data: off-chip --> local (matrix B)
#if SB == 1
GlobalToLocalB(bgm, blm, kSizeN, tid, kwg);
#endif
#if SA == 1 || SB == 1
barrier(CLK_LOCAL_MEM_FENCE);
#endif
// Loops over all workitem tiles, unrolled by a factor KWI
for (int pwi = 0; pwi < KWG * KREG; pwi += KWI * KREG) {
#pragma unroll
for (int _pit = 0; _pit < KWI*KREG; _pit += KREG) {
#if SA == 0 || SB == 0
int idk = kwg + pwi + _pit;
#endif
#if SA == 1 || SB == 1
int kg = pwi + _pit;
#endif
// Loads matrix A (kernel 0) or matrix B (kernel 1)
#pragma unroll
for (int _mi = 0; _mi < MWI/VWM; _mi += 1) {
// Loads data: local --> private (matrix A)
#if GEMMK == 0 && SA == 1
apm[_mi] = LocalToPrivateA(alm, _mi, kg);
// Loads data: off-chip --> private (matrix A)
#elif GEMMK == 0 && SA == 0
apm[_mi] = GlobalToPrivateA(agm, _mi, kSizeM, idk, kwg);
// Loads data: 2D global --> 2D private (matrix B)
#elif GEMMK == 1
#pragma unroll
for (int _ki = 0; _ki < KREG; _ki += 1) {
bpm[_ki * (MWI/VWM) + _mi] = GlobalToPrivateB2D(b_ptr, tid_x, _mi, kSizeN, idk, _ki);
}
#endif
}
// Loads matrix B (kernel 0) or matrix A (kernel 1)
#if GEMMK == 0
#pragma unroll
for (int _ni = 0; _ni < NWI/VWN; _ni += 1) {
// Loads data: local --> private (matrix B)
#if SB == 1
bpm[_ni] = LocalToPrivateB(blm, _ni, kg);
// Loads data: off-chip --> private (matrix B)
#else
bpm[_ni] = GlobalToPrivateB(bgm, _ni, kSizeN, idk);
#endif
}
#elif GEMMK == 1
// Loads data: 2D global --> 2D private (matrix A). Partly, shuffled later among subgroups
#if USE_SUBGROUP_SHUFFLING == 1
const int _ni = clblast_get_sub_group_local_id();
#pragma unroll
for (int _ki = 0; _ki < KREG/VWN; _ki += 1) {
apm[_ki] = GlobalToPrivateA2D(a_ptr, tid_y, _ni, kSizeK, idk, _ki);
}
// Loads data: 2D global --> 2D private (matrix A)
#else
#pragma unroll
for (int _ni = 0; _ni < NWI; _ni += 1) {
#pragma unroll
for (int _ki = 0; _ki < KREG/VWN; _ki += 1) {
apm[_ni * (KREG/VWN) + _ki] = GlobalToPrivateA2D(a_ptr, tid_y, _ni, kSizeK, idk, _ki);
}
}
#endif
#endif
// Performs the accumulation (Cpm += Apm * Bpm)
#if GEMMK == 0
#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
}
}
#elif GEMMK == 1
#pragma unroll
for (int _ni = 0; _ni < NWI; _ni += 1) {
#pragma unroll
for (int _mi = 0; _mi < MWI/VWM; _mi += 1) {
#pragma unroll
for (int _ki = 0; _ki < KREG/VWN; _ki += 1) {
#if USE_SUBGROUP_SHUFFLING == 1
const realN aval = clblast_sub_group_shuffle(apm[_ki], _ni);
#else
const realN aval = apm[_ni * (KREG/VWN) + _ki];
#endif
#if VWN == 1
cpm[_ni * (MWI/VWM) + _mi] = MultiplyAddVector(cpm[_ni * (MWI/VWM) + _mi], bpm[(VWN * _ki + 0) * (MWI/VWM) + _mi], aval);
#elif VWN == 2
cpm[_ni * (MWI/VWM) + _mi] = MultiplyAddVector(cpm[_ni * (MWI/VWM) + _mi], bpm[(VWN * _ki + 0) * (MWI/VWM) + _mi], aval.x);
cpm[_ni * (MWI/VWM) + _mi] = MultiplyAddVector(cpm[_ni * (MWI/VWM) + _mi], bpm[(VWN * _ki + 1) * (MWI/VWM) + _mi], aval.y);
#elif VWN == 4
cpm[_ni * (MWI/VWM) + _mi] = MultiplyAddVector(cpm[_ni * (MWI/VWM) + _mi], bpm[(VWN * _ki + 0) * (MWI/VWM) + _mi], aval.x);
cpm[_ni * (MWI/VWM) + _mi] = MultiplyAddVector(cpm[_ni * (MWI/VWM) + _mi], bpm[(VWN * _ki + 1) * (MWI/VWM) + _mi], aval.y);
cpm[_ni * (MWI/VWM) + _mi] = MultiplyAddVector(cpm[_ni * (MWI/VWM) + _mi], bpm[(VWN * _ki + 2) * (MWI/VWM) + _mi], aval.z);
cpm[_ni * (MWI/VWM) + _mi] = MultiplyAddVector(cpm[_ni * (MWI/VWM) + _mi], bpm[(VWN * _ki + 3) * (MWI/VWM) + _mi], aval.w);
#elif VWN == 8
cpm[_ni * (MWI/VWM) + _mi] = MultiplyAddVector(cpm[_ni * (MWI/VWM) + _mi], bpm[(VWN * _ki + 0) * (MWI/VWM) + _mi], aval.s0);
cpm[_ni * (MWI/VWM) + _mi] = MultiplyAddVector(cpm[_ni * (MWI/VWM) + _mi], bpm[(VWN * _ki + 1) * (MWI/VWM) + _mi], aval.s1);
cpm[_ni * (MWI/VWM) + _mi] = MultiplyAddVector(cpm[_ni * (MWI/VWM) + _mi], bpm[(VWN * _ki + 2) * (MWI/VWM) + _mi], aval.s2);
cpm[_ni * (MWI/VWM) + _mi] = MultiplyAddVector(cpm[_ni * (MWI/VWM) + _mi], bpm[(VWN * _ki + 3) * (MWI/VWM) + _mi], aval.s3);
cpm[_ni * (MWI/VWM) + _mi] = MultiplyAddVector(cpm[_ni * (MWI/VWM) + _mi], bpm[(VWN * _ki + 4) * (MWI/VWM) + _mi], aval.s4);
cpm[_ni * (MWI/VWM) + _mi] = MultiplyAddVector(cpm[_ni * (MWI/VWM) + _mi], bpm[(VWN * _ki + 5) * (MWI/VWM) + _mi], aval.s5);
cpm[_ni * (MWI/VWM) + _mi] = MultiplyAddVector(cpm[_ni * (MWI/VWM) + _mi], bpm[(VWN * _ki + 6) * (MWI/VWM) + _mi], aval.s6);
cpm[_ni * (MWI/VWM) + _mi] = MultiplyAddVector(cpm[_ni * (MWI/VWM) + _mi], bpm[(VWN * _ki + 7) * (MWI/VWM) + _mi], aval.s7);
#elif VWN == 16
cpm[_ni * (MWI/VWM) + _mi] = MultiplyAddVector(cpm[_ni * (MWI/VWM) + _mi], bpm[(VWN * _ki + 0 ) * (MWI/VWM) + _mi], aval.s0);
cpm[_ni * (MWI/VWM) + _mi] = MultiplyAddVector(cpm[_ni * (MWI/VWM) + _mi], bpm[(VWN * _ki + 1 ) * (MWI/VWM) + _mi], aval.s1);
cpm[_ni * (MWI/VWM) + _mi] = MultiplyAddVector(cpm[_ni * (MWI/VWM) + _mi], bpm[(VWN * _ki + 2 ) * (MWI/VWM) + _mi], aval.s2);
cpm[_ni * (MWI/VWM) + _mi] = MultiplyAddVector(cpm[_ni * (MWI/VWM) + _mi], bpm[(VWN * _ki + 3 ) * (MWI/VWM) + _mi], aval.s3);
cpm[_ni * (MWI/VWM) + _mi] = MultiplyAddVector(cpm[_ni * (MWI/VWM) + _mi], bpm[(VWN * _ki + 4 ) * (MWI/VWM) + _mi], aval.s4);
cpm[_ni * (MWI/VWM) + _mi] = MultiplyAddVector(cpm[_ni * (MWI/VWM) + _mi], bpm[(VWN * _ki + 5 ) * (MWI/VWM) + _mi], aval.s5);
cpm[_ni * (MWI/VWM) + _mi] = MultiplyAddVector(cpm[_ni * (MWI/VWM) + _mi], bpm[(VWN * _ki + 6 ) * (MWI/VWM) + _mi], aval.s6);
cpm[_ni * (MWI/VWM) + _mi] = MultiplyAddVector(cpm[_ni * (MWI/VWM) + _mi], bpm[(VWN * _ki + 7 ) * (MWI/VWM) + _mi], aval.s7);
cpm[_ni * (MWI/VWM) + _mi] = MultiplyAddVector(cpm[_ni * (MWI/VWM) + _mi], bpm[(VWN * _ki + 8 ) * (MWI/VWM) + _mi], aval.s8);
cpm[_ni * (MWI/VWM) + _mi] = MultiplyAddVector(cpm[_ni * (MWI/VWM) + _mi], bpm[(VWN * _ki + 9 ) * (MWI/VWM) + _mi], aval.s9);
cpm[_ni * (MWI/VWM) + _mi] = MultiplyAddVector(cpm[_ni * (MWI/VWM) + _mi], bpm[(VWN * _ki + 10) * (MWI/VWM) + _mi], aval.sA);
cpm[_ni * (MWI/VWM) + _mi] = MultiplyAddVector(cpm[_ni * (MWI/VWM) + _mi], bpm[(VWN * _ki + 11) * (MWI/VWM) + _mi], aval.sB);
cpm[_ni * (MWI/VWM) + _mi] = MultiplyAddVector(cpm[_ni * (MWI/VWM) + _mi], bpm[(VWN * _ki + 12) * (MWI/VWM) + _mi], aval.sC);
cpm[_ni * (MWI/VWM) + _mi] = MultiplyAddVector(cpm[_ni * (MWI/VWM) + _mi], bpm[(VWN * _ki + 13) * (MWI/VWM) + _mi], aval.sD);
cpm[_ni * (MWI/VWM) + _mi] = MultiplyAddVector(cpm[_ni * (MWI/VWM) + _mi], bpm[(VWN * _ki + 14) * (MWI/VWM) + _mi], aval.sE);
cpm[_ni * (MWI/VWM) + _mi] = MultiplyAddVector(cpm[_ni * (MWI/VWM) + _mi], bpm[(VWN * _ki + 15) * (MWI/VWM) + _mi], aval.sF);
#endif
}
}
}
#endif
}
}
#if SA == 1 || SB == 1
barrier(CLK_LOCAL_MEM_FENCE);
#endif
}
#if GLOBAL_MEM_FENCE == 1
barrier(CLK_GLOBAL_MEM_FENCE);
#endif
// Stores an MWG * NWG tile of results and performs the multiplication with alpha and beta
#if GEMMK == 0
const int cld = kSizeM;
#elif GEMMK == 1
const int cld = kSizeN;
#endif
#pragma unroll
for (int _ni = 0; _ni < NWI; _ni += 1) {
#pragma unroll
for (int _mi = 0; _mi < MWI/VWM; _mi += 1) {
StoreResults(cgm, cpm[_ni * (MWI/VWM) + _mi], _mi, _ni, cld, alpha, beta);
}
}
}
// =================================================================================================
// End of the C++11 raw string literal
)"
// =================================================================================================
|
