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
|
// =================================================================================================
// 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 implements the Xsyrk class (see the header for information about the class).
//
// =================================================================================================
#include "routines/level3/xsyrk.hpp"
#include "routines/level3/xgemm.hpp"
#include <string>
#include <vector>
namespace clblast {
// =================================================================================================
// Constructor: forwards to base class constructor
template <typename T>
Xsyrk<T>::Xsyrk(Queue &queue, EventPointer event, const std::string &name):
Routine(queue, event, name, {"Copy","Pad","Transpose","Padtranspose","Xgemm"}, PrecisionValue<T>(), {}, {
#include "../../kernels/level3/level3.opencl"
#include "../../kernels/level3/copy_fast.opencl"
#include "../../kernels/level3/copy_pad.opencl"
#include "../../kernels/level3/transpose_fast.opencl"
#include "../../kernels/level3/transpose_pad.opencl"
, // separated in multiple parts to prevent C1091 in MSVC 2013
#include "../../kernels/level3/xgemm_part1.opencl"
#include "../../kernels/level3/xgemm_part2.opencl"
#include "../../kernels/level3/xgemm_part3.opencl"
#include "../../kernels/level3/xgemm_part4.opencl"
}) {
}
// =================================================================================================
// The main routine
template <typename T>
void Xsyrk<T>::DoSyrk(const Layout layout, const Triangle triangle, const Transpose a_transpose,
const size_t n, const size_t k,
const T alpha,
const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld,
const T beta,
const Buffer<T> &c_buffer, const size_t c_offset, const size_t c_ld) {
const auto b_transpose = (a_transpose != Transpose::kNo) ? Transpose::kNo : Transpose::kYes;
const auto b_buffer = a_buffer;
const auto b_offset = a_offset;
const auto b_ld = a_ld;
SyrkAB(layout, triangle, a_transpose, b_transpose, n, k, alpha,
a_buffer, a_offset, a_ld, b_buffer, b_offset, b_ld, beta, c_buffer, c_offset, c_ld, event_);
}
template <typename T>
void Xsyrk<T>::SyrkAB(const Layout layout, const Triangle triangle, const Transpose a_transpose, const Transpose b_transpose,
const size_t n, const size_t k,
const T alpha,
const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld,
const Buffer<T> &b_buffer, const size_t b_offset, const size_t b_ld,
const T beta,
const Buffer<T> &c_buffer, const size_t c_offset, const size_t c_ld,
EventPointer final_event) {
// Computes the transpose/conjugate options and sets the a/b/c sizes based on that
bool a_do_transpose, b_do_transpose, c_do_transpose, a_conjugate, b_conjugate;
size_t a_one, a_two, b_one, b_two, c_one, c_two;
Xgemm<T>::ProcessArguments(layout, a_transpose, b_transpose, n, n, k,
a_one, a_two, b_one, b_two, c_one, c_two,
a_do_transpose, b_do_transpose, c_do_transpose, a_conjugate, b_conjugate,
db_["GEMMK"]);
// Tests the two matrices (A, C) for validity, first from a perspective of the OpenCL buffers and
// their sizes, and then from a perspective of parameter values (e.g. n, k). Tests whether the
// OpenCL buffers are valid and non-zero and whether the OpenCL buffers have sufficient storage
// space. Also tests that the leading dimensions of:
// matrix A cannot be less than N when rotated, or less than K when not-rotated
// matrix C cannot be less than N
TestMatrixA(a_one, a_two, a_buffer, a_offset, a_ld);
TestMatrixB(b_one, b_two, b_buffer, b_offset, b_ld);
TestMatrixC(c_one, c_two, c_buffer, c_offset, c_ld);
// Calculates the ceiled versions of n and k
auto n_ceiled = Ceil(Ceil(n, db_["MWG"]), db_["NWG"]);
auto k_ceiled = Ceil(k, db_["KWG"] * db_["KREG"]);
// Computes the first and second "internal" (ceiled) dimensions of the 3 matrices taking into account
// whether the matrices need to be rotated or not for the kernel.
const auto a_one_i = (Xgemm<T>::a_want_rotated_(db_["GEMMK"])) ? k_ceiled : n_ceiled;
const auto a_two_i = (Xgemm<T>::a_want_rotated_(db_["GEMMK"])) ? n_ceiled : k_ceiled;
const auto b_one_i = (!Xgemm<T>::b_want_rotated_(db_["GEMMK"])) ? k_ceiled : n_ceiled;
const auto b_two_i = (!Xgemm<T>::b_want_rotated_(db_["GEMMK"])) ? n_ceiled : k_ceiled;
// Decides which kernel to run: the upper-triangular or lower-triangular version
auto kernel_name = (triangle == Triangle::kUpper) ? "XgemmUpper" : "XgemmLower";
// Determines whether or not temporary matrices are needed
const auto a_no_temp = Xgemm<T>::NoTempBuffer(a_one, a_one_i, a_two, a_two_i, a_ld, a_offset, a_do_transpose, a_conjugate);
const auto b_no_temp = Xgemm<T>::NoTempBuffer(b_one, b_one_i, b_two, b_two_i, b_ld, b_offset, b_do_transpose, b_conjugate);
// Creates the temporary matrices
auto a_temp = (a_no_temp) ? a_buffer : Buffer<T>(context_, a_one_i * a_two_i);
auto b_temp = (b_no_temp) ? b_buffer : Buffer<T>(context_, b_one_i * b_two_i);
auto c_temp = Buffer<T>(context_, n_ceiled*n_ceiled);
// Events of all kernels (including pre/post processing kernels)
auto eventWaitList = std::vector<Event>();
auto emptyEventList = std::vector<Event>();
// Runs the pre-processing kernel for matrix A. This transposes the matrix, but also pads zeros
// to fill it up until it reaches a certain multiple of size (kernel parameter dependent). In
// case nothing has to be done, these kernels can be skipped.
if (!a_no_temp) {
auto eventProcessA = Event();
PadCopyTransposeMatrix(queue_, device_, db_, eventProcessA.pointer(), emptyEventList,
a_one, a_two, a_ld, a_offset, a_buffer,
a_one_i, a_two_i, a_one_i, 0, a_temp,
ConstantOne<T>(), program_,
true, a_do_transpose, false);
eventWaitList.push_back(eventProcessA);
}
if (!b_no_temp) {
auto eventProcessB = Event();
PadCopyTransposeMatrix(queue_, device_, db_, eventProcessB.pointer(), emptyEventList,
b_one, b_two, b_ld, b_offset, b_buffer,
b_one_i, b_two_i, b_one_i, 0, b_temp,
ConstantOne<T>(), program_,
true, b_do_transpose, false);
eventWaitList.push_back(eventProcessB);
}
// Furthermore, also creates a (possibly padded) copy of matrix C, since it is not allowed to
// modify the other triangle.
auto eventProcessC = Event();
PadCopyTransposeMatrix(queue_, device_, db_, eventProcessC.pointer(), emptyEventList,
n, n, c_ld, c_offset, c_buffer,
n_ceiled, n_ceiled, n_ceiled, 0, c_temp,
ConstantOne<T>(), program_,
true, c_do_transpose, false);
eventWaitList.push_back(eventProcessC);
// Retrieves the XgemmUpper or XgemmLower kernel from the compiled binary
auto kernel = Kernel(program_, kernel_name);
// Sets the kernel arguments
kernel.SetArgument(0, static_cast<int>(n_ceiled));
kernel.SetArgument(1, static_cast<int>(k_ceiled));
kernel.SetArgument(2, GetRealArg(alpha));
kernel.SetArgument(3, GetRealArg(beta));
kernel.SetArgument(4, a_temp());
kernel.SetArgument(5, b_temp());
kernel.SetArgument(6, c_temp());
// Computes the global and local thread sizes
auto global = std::vector<size_t>{
(n_ceiled * db_["MDIMC"]) / db_["MWG"],
(n_ceiled * db_["NDIMC"]) / db_["NWG"]
};
auto local = std::vector<size_t>{db_["MDIMC"], db_["NDIMC"]};
// Launches the kernel
auto eventKernel = Event();
RunKernel(kernel, queue_, device_, global, local, eventKernel.pointer(), eventWaitList);
eventWaitList.push_back(eventKernel);
// Runs the post-processing kernel
const auto upper = Xgemm<T>::c_want_rotated_(db_["GEMMK"]) ? (triangle == Triangle::kLower) :
(triangle == Triangle::kUpper);
const auto lower = !upper;
PadCopyTransposeMatrix(queue_, device_, db_, final_event, eventWaitList,
n_ceiled, n_ceiled, n_ceiled, 0, c_temp,
n, n, c_ld, c_offset, c_buffer,
ConstantOne<T>(), program_,
false, c_do_transpose, false, upper, lower, false);
}
// =================================================================================================
// Compiles the templated class
template class Xsyrk<half>;
template class Xsyrk<float>;
template class Xsyrk<double>;
template class Xsyrk<float2>;
template class Xsyrk<double2>;
// =================================================================================================
} // namespace clblast
|
