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
|
// =================================================================================================
// 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 file contains the common kernels shared among different BLAS functions. This file contains
// a kernel to transpose matrices. This is a 'fast' version with restrictions, see the
// 'padtranspose.opencl' file for a general transpose kernel.
//
// =================================================================================================
// 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"(
// =================================================================================================
// Data-widths
#if TRA_WPT == 1
typedef real realT;
#elif TRA_WPT == 2
typedef real2 realT;
#elif TRA_WPT == 4
typedef real4 realT;
#elif TRA_WPT == 8
typedef real8 realT;
#elif TRA_WPT == 16
typedef real16 realT;
#endif
// =================================================================================================
// Transposes and copies a matrix. Requires both matrices to be of the same dimensions and without
// offset. A more general version is available in 'padtranspose.opencl'.
#if RELAX_WORKGROUP_SIZE == 1
__kernel
#else
__kernel __attribute__((reqd_work_group_size(TRA_DIM, TRA_DIM, 1)))
#endif
void TransposeMatrixFast(const int ld,
__global const realT* restrict src,
__global realT* dest,
const real_arg arg_alpha) {
const real alpha = GetRealArg(arg_alpha);
// Sets the group identifiers. They might be 'shuffled' around to distribute work in a different
// way over workgroups, breaking memory-bank dependencies.
const int gid0 = get_group_id(0);
#if TRA_SHUFFLE == 1
const int gid1 = (get_group_id(0) + get_group_id(1)) % get_num_groups(0);
#else
const int gid1 = get_group_id(1);
#endif
// Local memory to store a tile of the matrix (for coalescing)
__local realT tile[TRA_WPT*TRA_DIM][TRA_DIM + TRA_PAD];
// Loops over the work per thread
#pragma unroll
for (int _w_one = 0; _w_one < TRA_WPT; _w_one += 1) {
// Computes the identifiers for the source matrix. Note that the local and global dimensions
// do not correspond to each other!
const int id_one = gid1 * TRA_DIM + get_local_id(0);
const int id_two = (gid0 * TRA_DIM + get_local_id(1))*TRA_WPT + _w_one;
// Loads data into the local memory
realT value = src[id_two*(ld/TRA_WPT) + id_one];
tile[get_local_id(0)*TRA_WPT + _w_one][get_local_id(1)] = value;
}
// Synchronizes all threads in a workgroup
barrier(CLK_LOCAL_MEM_FENCE);
// Loads transposed data from the local memory
#pragma promote_to_registers
realT vpm[TRA_WPT];
#pragma unroll
for (int _w_one = 0; _w_one < TRA_WPT; _w_one += 1) {
vpm[_w_one] = tile[get_local_id(1)*TRA_WPT + _w_one][get_local_id(0)];
}
// Performs the register-level transpose of the vectorized data
#pragma promote_to_registers
realT results[TRA_WPT];
#if TRA_WPT == 1
results[0] = vpm[0];
#elif TRA_WPT == 2
results[0].x = vpm[0].x; results[0].y = vpm[1].x;
results[1].x = vpm[0].y; results[1].y = vpm[1].y;
#elif TRA_WPT == 4
results[0].x = vpm[0].x; results[0].y = vpm[1].x; results[0].z = vpm[2].x; results[0].w = vpm[3].x;
results[1].x = vpm[0].y; results[1].y = vpm[1].y; results[1].z = vpm[2].y; results[1].w = vpm[3].y;
results[2].x = vpm[0].z; results[2].y = vpm[1].z; results[2].z = vpm[2].z; results[2].w = vpm[3].z;
results[3].x = vpm[0].w; results[3].y = vpm[1].w; results[3].z = vpm[2].w; results[3].w = vpm[3].w;
#elif TRA_WPT == 8
results[0].s0 = vpm[0].s0; results[0].s1 = vpm[1].s0; results[0].s2 = vpm[2].s0; results[0].s3 = vpm[3].s0; results[0].s4 = vpm[4].s0; results[0].s5 = vpm[5].s0; results[0].s6 = vpm[6].s0; results[0].s7 = vpm[7].s0;
results[1].s0 = vpm[0].s1; results[1].s1 = vpm[1].s1; results[1].s2 = vpm[2].s1; results[1].s3 = vpm[3].s1; results[1].s4 = vpm[4].s1; results[1].s5 = vpm[5].s1; results[1].s6 = vpm[6].s1; results[1].s7 = vpm[7].s1;
results[2].s0 = vpm[0].s2; results[2].s1 = vpm[1].s2; results[2].s2 = vpm[2].s2; results[2].s3 = vpm[3].s2; results[2].s4 = vpm[4].s2; results[2].s5 = vpm[5].s2; results[2].s6 = vpm[6].s2; results[2].s7 = vpm[7].s2;
results[3].s0 = vpm[0].s3; results[3].s1 = vpm[1].s3; results[3].s2 = vpm[2].s3; results[3].s3 = vpm[3].s3; results[3].s4 = vpm[4].s3; results[3].s5 = vpm[5].s3; results[3].s6 = vpm[6].s3; results[3].s7 = vpm[7].s3;
results[4].s0 = vpm[0].s4; results[4].s1 = vpm[1].s4; results[4].s2 = vpm[2].s4; results[4].s3 = vpm[3].s4; results[4].s4 = vpm[4].s4; results[4].s5 = vpm[5].s4; results[4].s6 = vpm[6].s4; results[4].s7 = vpm[7].s4;
results[5].s0 = vpm[0].s5; results[5].s1 = vpm[1].s5; results[5].s2 = vpm[2].s5; results[5].s3 = vpm[3].s5; results[5].s4 = vpm[4].s5; results[5].s5 = vpm[5].s5; results[5].s6 = vpm[6].s5; results[5].s7 = vpm[7].s5;
results[6].s0 = vpm[0].s6; results[6].s1 = vpm[1].s6; results[6].s2 = vpm[2].s6; results[6].s3 = vpm[3].s6; results[6].s4 = vpm[4].s6; results[6].s5 = vpm[5].s6; results[6].s6 = vpm[6].s6; results[6].s7 = vpm[7].s6;
results[7].s0 = vpm[0].s7; results[7].s1 = vpm[1].s7; results[7].s2 = vpm[2].s7; results[7].s3 = vpm[3].s7; results[7].s4 = vpm[4].s7; results[7].s5 = vpm[5].s7; results[7].s6 = vpm[6].s7; results[7].s7 = vpm[7].s7;
#elif TRA_WPT == 16
results[ 0].s0 = vpm[0].s0; results[ 0].s1 = vpm[1].s0; results[ 0].s2 = vpm[2].s0; results[ 0].s3 = vpm[3].s0; results[ 0].s4 = vpm[4].s0; results[ 0].s5 = vpm[5].s0; results[ 0].s6 = vpm[6].s0; results[ 0].s7 = vpm[7].s0; results[ 0].s8 = vpm[8].s0; results[ 0].s9 = vpm[9].s0; results[ 0].sA = vpm[10].s0; results[ 0].sB = vpm[11].s0; results[ 0].sC = vpm[12].s0; results[ 0].sD = vpm[13].s0; results[ 0].sE = vpm[14].s0; results[ 0].sF = vpm[15].s0;
results[ 1].s0 = vpm[0].s1; results[ 1].s1 = vpm[1].s1; results[ 1].s2 = vpm[2].s1; results[ 1].s3 = vpm[3].s1; results[ 1].s4 = vpm[4].s1; results[ 1].s5 = vpm[5].s1; results[ 1].s6 = vpm[6].s1; results[ 1].s7 = vpm[7].s1; results[ 1].s8 = vpm[8].s1; results[ 1].s9 = vpm[9].s1; results[ 1].sA = vpm[10].s1; results[ 1].sB = vpm[11].s1; results[ 1].sC = vpm[12].s1; results[ 1].sD = vpm[13].s1; results[ 1].sE = vpm[14].s1; results[ 1].sF = vpm[15].s1;
results[ 2].s0 = vpm[0].s2; results[ 2].s1 = vpm[1].s2; results[ 2].s2 = vpm[2].s2; results[ 2].s3 = vpm[3].s2; results[ 2].s4 = vpm[4].s2; results[ 2].s5 = vpm[5].s2; results[ 2].s6 = vpm[6].s2; results[ 2].s7 = vpm[7].s2; results[ 2].s8 = vpm[8].s2; results[ 2].s9 = vpm[9].s2; results[ 2].sA = vpm[10].s2; results[ 2].sB = vpm[11].s2; results[ 2].sC = vpm[12].s2; results[ 2].sD = vpm[13].s2; results[ 2].sE = vpm[14].s2; results[ 2].sF = vpm[15].s2;
results[ 3].s0 = vpm[0].s3; results[ 3].s1 = vpm[1].s3; results[ 3].s2 = vpm[2].s3; results[ 3].s3 = vpm[3].s3; results[ 3].s4 = vpm[4].s3; results[ 3].s5 = vpm[5].s3; results[ 3].s6 = vpm[6].s3; results[ 3].s7 = vpm[7].s3; results[ 3].s8 = vpm[8].s3; results[ 3].s9 = vpm[9].s3; results[ 3].sA = vpm[10].s3; results[ 3].sB = vpm[11].s3; results[ 3].sC = vpm[12].s3; results[ 3].sD = vpm[13].s3; results[ 3].sE = vpm[14].s3; results[ 3].sF = vpm[15].s3;
results[ 4].s0 = vpm[0].s4; results[ 4].s1 = vpm[1].s4; results[ 4].s2 = vpm[2].s4; results[ 4].s3 = vpm[3].s4; results[ 4].s4 = vpm[4].s4; results[ 4].s5 = vpm[5].s4; results[ 4].s6 = vpm[6].s4; results[ 4].s7 = vpm[7].s4; results[ 4].s8 = vpm[8].s4; results[ 4].s9 = vpm[9].s4; results[ 4].sA = vpm[10].s4; results[ 4].sB = vpm[11].s4; results[ 4].sC = vpm[12].s4; results[ 4].sD = vpm[13].s4; results[ 4].sE = vpm[14].s4; results[ 4].sF = vpm[15].s4;
results[ 5].s0 = vpm[0].s5; results[ 5].s1 = vpm[1].s5; results[ 5].s2 = vpm[2].s5; results[ 5].s3 = vpm[3].s5; results[ 5].s4 = vpm[4].s5; results[ 5].s5 = vpm[5].s5; results[ 5].s6 = vpm[6].s5; results[ 5].s7 = vpm[7].s5; results[ 5].s8 = vpm[8].s5; results[ 5].s9 = vpm[9].s5; results[ 5].sA = vpm[10].s5; results[ 5].sB = vpm[11].s5; results[ 5].sC = vpm[12].s5; results[ 5].sD = vpm[13].s5; results[ 5].sE = vpm[14].s5; results[ 5].sF = vpm[15].s5;
results[ 6].s0 = vpm[0].s6; results[ 6].s1 = vpm[1].s6; results[ 6].s2 = vpm[2].s6; results[ 6].s3 = vpm[3].s6; results[ 6].s4 = vpm[4].s6; results[ 6].s5 = vpm[5].s6; results[ 6].s6 = vpm[6].s6; results[ 6].s7 = vpm[7].s6; results[ 6].s8 = vpm[8].s6; results[ 6].s9 = vpm[9].s6; results[ 6].sA = vpm[10].s6; results[ 6].sB = vpm[11].s6; results[ 6].sC = vpm[12].s6; results[ 6].sD = vpm[13].s6; results[ 6].sE = vpm[14].s6; results[ 6].sF = vpm[15].s6;
results[ 7].s0 = vpm[0].s7; results[ 7].s1 = vpm[1].s7; results[ 7].s2 = vpm[2].s7; results[ 7].s3 = vpm[3].s7; results[ 7].s4 = vpm[4].s7; results[ 7].s5 = vpm[5].s7; results[ 7].s6 = vpm[6].s7; results[ 7].s7 = vpm[7].s7; results[ 7].s8 = vpm[8].s7; results[ 7].s9 = vpm[9].s7; results[ 7].sA = vpm[10].s7; results[ 7].sB = vpm[11].s7; results[ 7].sC = vpm[12].s7; results[ 7].sD = vpm[13].s7; results[ 7].sE = vpm[14].s7; results[ 7].sF = vpm[15].s7;
results[ 8].s0 = vpm[0].s8; results[ 8].s1 = vpm[1].s8; results[ 8].s2 = vpm[2].s8; results[ 8].s3 = vpm[3].s8; results[ 8].s4 = vpm[4].s8; results[ 8].s5 = vpm[5].s8; results[ 8].s6 = vpm[6].s8; results[ 8].s7 = vpm[7].s8; results[ 8].s8 = vpm[8].s8; results[ 8].s9 = vpm[9].s8; results[ 8].sA = vpm[10].s8; results[ 8].sB = vpm[11].s8; results[ 8].sC = vpm[12].s8; results[ 8].sD = vpm[13].s8; results[ 8].sE = vpm[14].s8; results[ 8].sF = vpm[15].s8;
results[ 9].s0 = vpm[0].s9; results[ 9].s1 = vpm[1].s9; results[ 9].s2 = vpm[2].s9; results[ 9].s3 = vpm[3].s9; results[ 9].s4 = vpm[4].s9; results[ 9].s5 = vpm[5].s9; results[ 9].s6 = vpm[6].s9; results[ 9].s7 = vpm[7].s9; results[ 9].s8 = vpm[8].s9; results[ 9].s9 = vpm[9].s9; results[ 9].sA = vpm[10].s9; results[ 9].sB = vpm[11].s9; results[ 9].sC = vpm[12].s9; results[ 9].sD = vpm[13].s9; results[ 9].sE = vpm[14].s9; results[ 9].sF = vpm[15].s9;
results[10].s0 = vpm[0].sA; results[10].s1 = vpm[1].sA; results[10].s2 = vpm[2].sA; results[10].s3 = vpm[3].sA; results[10].s4 = vpm[4].sA; results[10].s5 = vpm[5].sA; results[10].s6 = vpm[6].sA; results[10].s7 = vpm[7].sA; results[10].s8 = vpm[8].sA; results[10].s9 = vpm[9].sA; results[10].sA = vpm[10].sA; results[10].sB = vpm[11].sA; results[10].sC = vpm[12].sA; results[10].sD = vpm[13].sA; results[10].sE = vpm[14].sA; results[10].sF = vpm[15].sA;
results[11].s0 = vpm[0].sB; results[11].s1 = vpm[1].sB; results[11].s2 = vpm[2].sB; results[11].s3 = vpm[3].sB; results[11].s4 = vpm[4].sB; results[11].s5 = vpm[5].sB; results[11].s6 = vpm[6].sB; results[11].s7 = vpm[7].sB; results[11].s8 = vpm[8].sB; results[11].s9 = vpm[9].sB; results[11].sA = vpm[10].sB; results[11].sB = vpm[11].sB; results[11].sC = vpm[12].sB; results[11].sD = vpm[13].sB; results[11].sE = vpm[14].sB; results[11].sF = vpm[15].sB;
results[12].s0 = vpm[0].sC; results[12].s1 = vpm[1].sC; results[12].s2 = vpm[2].sC; results[12].s3 = vpm[3].sC; results[12].s4 = vpm[4].sC; results[12].s5 = vpm[5].sC; results[12].s6 = vpm[6].sC; results[12].s7 = vpm[7].sC; results[12].s8 = vpm[8].sC; results[12].s9 = vpm[9].sC; results[12].sA = vpm[10].sC; results[12].sB = vpm[11].sC; results[12].sC = vpm[12].sC; results[12].sD = vpm[13].sC; results[12].sE = vpm[14].sC; results[12].sF = vpm[15].sC;
results[13].s0 = vpm[0].sD; results[13].s1 = vpm[1].sD; results[13].s2 = vpm[2].sD; results[13].s3 = vpm[3].sD; results[13].s4 = vpm[4].sD; results[13].s5 = vpm[5].sD; results[13].s6 = vpm[6].sD; results[13].s7 = vpm[7].sD; results[13].s8 = vpm[8].sD; results[13].s9 = vpm[9].sD; results[13].sA = vpm[10].sD; results[13].sB = vpm[11].sD; results[13].sC = vpm[12].sD; results[13].sD = vpm[13].sD; results[13].sE = vpm[14].sD; results[13].sF = vpm[15].sD;
results[14].s0 = vpm[0].sE; results[14].s1 = vpm[1].sE; results[14].s2 = vpm[2].sE; results[14].s3 = vpm[3].sE; results[14].s4 = vpm[4].sE; results[14].s5 = vpm[5].sE; results[14].s6 = vpm[6].sE; results[14].s7 = vpm[7].sE; results[14].s8 = vpm[8].sE; results[14].s9 = vpm[9].sE; results[14].sA = vpm[10].sE; results[14].sB = vpm[11].sE; results[14].sC = vpm[12].sE; results[14].sD = vpm[13].sE; results[14].sE = vpm[14].sE; results[14].sF = vpm[15].sE;
results[15].s0 = vpm[0].sF; results[15].s1 = vpm[1].sF; results[15].s2 = vpm[2].sF; results[15].s3 = vpm[3].sF; results[15].s4 = vpm[4].sF; results[15].s5 = vpm[5].sF; results[15].s6 = vpm[6].sF; results[15].s7 = vpm[7].sF; results[15].s8 = vpm[8].sF; results[15].s9 = vpm[9].sF; results[15].sA = vpm[10].sF; results[15].sB = vpm[11].sF; results[15].sC = vpm[12].sF; results[15].sD = vpm[13].sF; results[15].sE = vpm[14].sF; results[15].sF = vpm[15].sF;
#endif
// Multiplies by alpha and then stores the results into the destination matrix
#pragma unroll
for (int _w_two = 0; _w_two < TRA_WPT; _w_two += 1) {
realT result;
#if TRA_WPT == 1
Multiply(result, alpha, results[_w_two]);
#elif TRA_WPT == 2
Multiply(result.x, alpha, results[_w_two].x);
Multiply(result.y, alpha, results[_w_two].y);
#elif TRA_WPT == 4
Multiply(result.x, alpha, results[_w_two].x);
Multiply(result.y, alpha, results[_w_two].y);
Multiply(result.z, alpha, results[_w_two].z);
Multiply(result.w, alpha, results[_w_two].w);
#elif TRA_WPT == 8
Multiply(result.s0, alpha, results[_w_two].s0);
Multiply(result.s1, alpha, results[_w_two].s1);
Multiply(result.s2, alpha, results[_w_two].s2);
Multiply(result.s3, alpha, results[_w_two].s3);
Multiply(result.s4, alpha, results[_w_two].s4);
Multiply(result.s5, alpha, results[_w_two].s5);
Multiply(result.s6, alpha, results[_w_two].s6);
Multiply(result.s7, alpha, results[_w_two].s7);
#elif TRA_WPT == 16
Multiply(result.s0, alpha, results[_w_two].s0);
Multiply(result.s1, alpha, results[_w_two].s1);
Multiply(result.s2, alpha, results[_w_two].s2);
Multiply(result.s3, alpha, results[_w_two].s3);
Multiply(result.s4, alpha, results[_w_two].s4);
Multiply(result.s5, alpha, results[_w_two].s5);
Multiply(result.s6, alpha, results[_w_two].s6);
Multiply(result.s7, alpha, results[_w_two].s7);
Multiply(result.s8, alpha, results[_w_two].s8);
Multiply(result.s9, alpha, results[_w_two].s9);
Multiply(result.sA, alpha, results[_w_two].sA);
Multiply(result.sB, alpha, results[_w_two].sB);
Multiply(result.sC, alpha, results[_w_two].sC);
Multiply(result.sD, alpha, results[_w_two].sD);
Multiply(result.sE, alpha, results[_w_two].sE);
Multiply(result.sF, alpha, results[_w_two].sF);
#endif
const int id_one = gid0*TRA_DIM + get_local_id(0);
const int id_two = (gid1*TRA_DIM + get_local_id(1))*TRA_WPT + _w_two;
dest[id_two*(ld/TRA_WPT) + id_one] = result;
}
}
// =================================================================================================
// End of the C++11 raw string literal
)"
// =================================================================================================
|
