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
|
// =================================================================================================
// 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 all the interfaces to common kernels, such as copying, padding, and
// transposing a matrix. These functions are templated and thus header-only. This file also contains
// other common functions to routines, such as a function to launch a kernel.
//
// =================================================================================================
#ifndef CLBLAST_ROUTINES_COMMON_H_
#define CLBLAST_ROUTINES_COMMON_H_
#include <string>
#include <vector>
#include "utilities/utilities.hpp"
#include "utilities/compile.hpp"
#include "database/database.hpp"
namespace clblast {
// =================================================================================================
// Enqueues a kernel, waits for completion, and checks for errors
void RunKernel(Kernel &kernel, Queue &queue, const Device &device,
std::vector<size_t> global, const std::vector<size_t> &local,
EventPointer event, const std::vector<Event> &waitForEvents = {});
// =================================================================================================
// Sets all elements of a matrix to a constant value
template <typename T>
void FillMatrix(Queue &queue, const Device &device,
const Program &program, const Databases &,
EventPointer event, const std::vector<Event> &waitForEvents,
const size_t m, const size_t n, const size_t ld, const size_t offset,
const Buffer<T> &dest,
const T constant_value);
// Sets all elements of a vector to a constant value
template <typename T>
void FillVector(Queue &queue, const Device &device,
const Program &program, const Databases &,
EventPointer event, const std::vector<Event> &waitForEvents,
const size_t n, const size_t inc, const size_t offset,
const Buffer<T> &dest,
const T constant_value);
// =================================================================================================
// Copies or transposes a matrix and optionally pads/unpads it with zeros. This method is also able
// to write to symmetric and triangular matrices through optional arguments.
template <typename T>
void PadCopyTransposeMatrix(Queue &queue, const Device &device,
const Databases &db,
EventPointer event, const std::vector<Event> &waitForEvents,
const size_t src_one, const size_t src_two,
const size_t src_ld, const size_t src_offset,
const Buffer<T> &src,
const size_t dest_one, const size_t dest_two,
const size_t dest_ld, const size_t dest_offset,
const Buffer<T> &dest,
const T alpha,
const Program &program, const bool do_pad,
const bool do_transpose, const bool do_conjugate,
const bool upper = false, const bool lower = false,
const bool diagonal_imag_zero = false) {
// Determines whether or not the fast-version could potentially be used
auto use_fast_kernel = (src_offset == 0) && (dest_offset == 0) && (do_conjugate == false) &&
(src_one == dest_one) && (src_two == dest_two) && (src_ld == dest_ld) &&
(upper == false) && (lower == false) && (diagonal_imag_zero == false);
// Determines the right kernel
auto kernel_name = std::string{};
if (do_transpose) {
if (use_fast_kernel &&
IsMultiple(src_ld, db["TRA_WPT"]) &&
IsMultiple(src_one, db["TRA_WPT"]*db["TRA_DIM"]) &&
IsMultiple(src_two, db["TRA_WPT"]*db["TRA_DIM"])) {
kernel_name = "TransposeMatrixFast";
}
else {
use_fast_kernel = false;
kernel_name = (do_pad) ? "TransposePadMatrix" : "TransposeMatrix";
}
}
else {
if (use_fast_kernel &&
IsMultiple(src_ld, db["COPY_VW"]) &&
IsMultiple(src_one, db["COPY_VW"]*db["COPY_DIMX"]) &&
IsMultiple(src_two, db["COPY_WPT"]*db["COPY_DIMY"])) {
kernel_name = "CopyMatrixFast";
}
else {
use_fast_kernel = false;
kernel_name = (do_pad) ? "CopyPadMatrix" : "CopyMatrix";
}
}
// Retrieves the kernel from the compiled binary
auto kernel = Kernel(program, kernel_name);
// Sets the kernel arguments
if (use_fast_kernel) {
kernel.SetArgument(0, static_cast<int>(src_ld));
kernel.SetArgument(1, src());
kernel.SetArgument(2, dest());
kernel.SetArgument(3, GetRealArg(alpha));
}
else {
kernel.SetArgument(0, static_cast<int>(src_one));
kernel.SetArgument(1, static_cast<int>(src_two));
kernel.SetArgument(2, static_cast<int>(src_ld));
kernel.SetArgument(3, static_cast<int>(src_offset));
kernel.SetArgument(4, src());
kernel.SetArgument(5, static_cast<int>(dest_one));
kernel.SetArgument(6, static_cast<int>(dest_two));
kernel.SetArgument(7, static_cast<int>(dest_ld));
kernel.SetArgument(8, static_cast<int>(dest_offset));
kernel.SetArgument(9, dest());
kernel.SetArgument(10, GetRealArg(alpha));
if (do_pad) {
kernel.SetArgument(11, static_cast<int>(do_conjugate));
}
else {
kernel.SetArgument(11, static_cast<int>(upper));
kernel.SetArgument(12, static_cast<int>(lower));
kernel.SetArgument(13, static_cast<int>(diagonal_imag_zero));
}
}
// Launches the kernel and returns the error code. Uses global and local thread sizes based on
// parameters in the database.
if (do_transpose) {
if (use_fast_kernel) {
const auto global = std::vector<size_t>{
dest_one / db["TRA_WPT"],
dest_two / db["TRA_WPT"]
};
const auto local = std::vector<size_t>{db["TRA_DIM"], db["TRA_DIM"]};
RunKernel(kernel, queue, device, global, local, event, waitForEvents);
}
else {
const auto global = std::vector<size_t>{
Ceil(CeilDiv(dest_one, db["PADTRA_WPT"]), db["PADTRA_TILE"]),
Ceil(CeilDiv(dest_two, db["PADTRA_WPT"]), db["PADTRA_TILE"])
};
const auto local = std::vector<size_t>{db["PADTRA_TILE"], db["PADTRA_TILE"]};
RunKernel(kernel, queue, device, global, local, event, waitForEvents);
}
}
else {
if (use_fast_kernel) {
const auto global = std::vector<size_t>{
dest_one / db["COPY_VW"],
dest_two / db["COPY_WPT"]
};
const auto local = std::vector<size_t>{db["COPY_DIMX"], db["COPY_DIMY"]};
RunKernel(kernel, queue, device, global, local, event, waitForEvents);
}
else {
const auto global = std::vector<size_t>{
Ceil(CeilDiv(dest_one, db["PAD_WPTX"]), db["PAD_DIMX"]),
Ceil(CeilDiv(dest_two, db["PAD_WPTY"]), db["PAD_DIMY"])
};
const auto local = std::vector<size_t>{db["PAD_DIMX"], db["PAD_DIMY"]};
RunKernel(kernel, queue, device, global, local, event, waitForEvents);
}
}
}
// Batched version of the above
template <typename T>
void PadCopyTransposeMatrixBatched(Queue &queue, const Device &device,
const Databases &db,
EventPointer event, const std::vector<Event> &waitForEvents,
const size_t src_one, const size_t src_two,
const size_t src_ld, const Buffer<int> &src_offsets,
const Buffer<T> &src,
const size_t dest_one, const size_t dest_two,
const size_t dest_ld, const Buffer<int> &dest_offsets,
const Buffer<T> &dest,
const Program &program, const bool do_pad,
const bool do_transpose, const bool do_conjugate,
const size_t batch_count) {
// Determines the right kernel
auto kernel_name = std::string{};
if (do_transpose) {
kernel_name = (do_pad) ? "TransposePadMatrixBatched" : "TransposeMatrixBatched";
}
else {
kernel_name = (do_pad) ? "CopyPadMatrixBatched" : "CopyMatrixBatched";
}
// Retrieves the kernel from the compiled binary
auto kernel = Kernel(program, kernel_name);
// Sets the kernel arguments
kernel.SetArgument(0, static_cast<int>(src_one));
kernel.SetArgument(1, static_cast<int>(src_two));
kernel.SetArgument(2, static_cast<int>(src_ld));
kernel.SetArgument(3, src_offsets());
kernel.SetArgument(4, src());
kernel.SetArgument(5, static_cast<int>(dest_one));
kernel.SetArgument(6, static_cast<int>(dest_two));
kernel.SetArgument(7, static_cast<int>(dest_ld));
kernel.SetArgument(8, dest_offsets());
kernel.SetArgument(9, dest());
if (do_pad) {
kernel.SetArgument(10, static_cast<int>(do_conjugate));
}
// Launches the kernel and returns the error code. Uses global and local thread sizes based on
// parameters in the database.
if (do_transpose) {
const auto global = std::vector<size_t>{
Ceil(CeilDiv(dest_one, db["PADTRA_WPT"]), db["PADTRA_TILE"]),
Ceil(CeilDiv(dest_two, db["PADTRA_WPT"]), db["PADTRA_TILE"]),
batch_count
};
const auto local = std::vector<size_t>{db["PADTRA_TILE"], db["PADTRA_TILE"], 1};
RunKernel(kernel, queue, device, global, local, event, waitForEvents);
}
else {
const auto global = std::vector<size_t>{
Ceil(CeilDiv(dest_one, db["PAD_WPTX"]), db["PAD_DIMX"]),
Ceil(CeilDiv(dest_two, db["PAD_WPTY"]), db["PAD_DIMY"]),
batch_count
};
const auto local = std::vector<size_t>{db["PAD_DIMX"], db["PAD_DIMY"], 1};
RunKernel(kernel, queue, device, global, local, event, waitForEvents);
}
}
// =================================================================================================
} // namespace clblast
// CLBLAST_ROUTINES_COMMON_H_
#endif
|
