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
|
// =================================================================================================
// 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 "internal/routines/xsyrk.h"
#include <string>
#include <vector>
namespace clblast {
// =================================================================================================
// Specific implementations to get the memory-type based on a template argument
template <> const Precision Xsyrk<float>::precision_ = Precision::kSingle;
template <> const Precision Xsyrk<double>::precision_ = Precision::kDouble;
template <> const Precision Xsyrk<float2>::precision_ = Precision::kComplexSingle;
template <> const Precision Xsyrk<double2>::precision_ = Precision::kComplexDouble;
// =================================================================================================
// Constructor: forwards to base class constructor
template <typename T>
Xsyrk<T>::Xsyrk(CommandQueue &queue, Event &event):
Routine(queue, event, {"Copy", "Pad", "Transpose", "PadTranspose", "Xgemm"}, precision_) {
}
// =================================================================================================
// The main routine
template <typename T>
StatusCode 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 &a_buffer, const size_t a_offset, const size_t a_ld,
const T beta,
const Buffer &c_buffer, const size_t c_offset, const size_t c_ld) {
// Makes sure all dimensions are larger than zero
if ((n == 0) || (k == 0) ) { return StatusCode::kInvalidDimension; }
// Computes whether or not the matrices are transposed in memory. This is based on their layout
// (row or column-major) and whether or not they are requested to be pre-transposed.
auto a_rotated = (layout == Layout::kColMajor && a_transpose != Transpose::kNo) ||
(layout == Layout::kRowMajor && a_transpose == Transpose::kNo);
auto c_rotated = (layout == Layout::kRowMajor);
// In case of complex data-types, the transpose can also become a conjugate transpose
auto a_conjugate = (a_transpose == Transpose::kConjugate);
// Computes the first and second dimensions of the A matrix taking the layout into account
auto a_one = (a_rotated) ? k : n;
auto a_two = (a_rotated) ? n : k;
// 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
auto status = TestMatrixA(a_one, a_two, a_buffer, a_offset, a_ld, sizeof(T));
if (ErrorIn(status)) { return status; }
status = TestMatrixC(n, n, c_buffer, c_offset, c_ld, sizeof(T));
if (ErrorIn(status)) { return status; }
// Calculates the ceiled versions of n and k
auto n_ceiled = Ceil(n, db_["NWG"]);
auto k_ceiled = Ceil(k, db_["KWG"]);
// Decides which kernel to run: the upper-triangular or lower-triangular version
auto kernel_name = (triangle == Triangle::kUpper) ? "XgemmUpper" : "XgemmLower";
// Allocates space on the device for padded and/or transposed input and output matrices.
try {
auto temp_a = Buffer(context_, CL_MEM_READ_WRITE, k_ceiled*n_ceiled*sizeof(T));
auto temp_c = Buffer(context_, CL_MEM_READ_WRITE, n_ceiled*n_ceiled*sizeof(T));
// Loads the program from the database
auto& program = GetProgramFromCache();
// Runs the pre-processing kernel. This transposes the matrices A and B, but also pads zeros to
// fill them up until they reach a certain multiple of size (kernel parameter dependent).
status = PadCopyTransposeMatrix(a_one, a_two, a_ld, a_offset, a_buffer,
n_ceiled, k_ceiled, n_ceiled, 0, temp_a,
a_rotated, a_conjugate, true, false, false, program);
if (ErrorIn(status)) { return status; }
status = PadCopyTransposeMatrix(n, n, c_ld, c_offset, c_buffer,
n_ceiled, n_ceiled, n_ceiled, 0, temp_c,
c_rotated, false, true, false, false, program);
if (ErrorIn(status)) { return status; }
// Retrieves the XgemmUpper or XgemmLower kernel from the compiled binary
try {
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, alpha);
kernel.SetArgument(3, beta);
kernel.SetArgument(4, temp_a());
kernel.SetArgument(5, temp_a());
kernel.SetArgument(6, temp_c());
// 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
status = RunKernel(kernel, global, local);
if (ErrorIn(status)) { return status; }
// Runs the post-processing kernel
auto upper = (triangle == Triangle::kUpper);
auto lower = (triangle == Triangle::kLower);
status = PadCopyTransposeMatrix(n_ceiled, n_ceiled, n_ceiled, 0, temp_c,
n, n, c_ld, c_offset, c_buffer,
c_rotated, false, false, upper, lower, program);
if (ErrorIn(status)) { return status; }
// Successfully finished the computation
return StatusCode::kSuccess;
} catch (...) { return StatusCode::kInvalidKernel; }
} catch (...) { return StatusCode::kTempBufferAllocFailure; }
}
// =================================================================================================
// Compiles the templated class
template class Xsyrk<float>;
template class Xsyrk<double>;
template class Xsyrk<float2>;
template class Xsyrk<double2>;
// =================================================================================================
} // namespace clblast
|
