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
|
// =================================================================================================
// 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 Xgemm routine. The precision is implemented using a template argument.
//
// =================================================================================================
#ifndef CLBLAST_ROUTINES_XGEMM_H_
#define CLBLAST_ROUTINES_XGEMM_H_
#include "routine.hpp"
namespace clblast {
// =================================================================================================
// See comment at top of file for a description of the class
template <typename T>
class Xgemm: public Routine {
public:
// Defines the assumptions of the GEMM kernels
static const bool a_want_rotated_(const size_t gemm_kernel_id) { return gemm_kernel_id == 1; }
static const bool b_want_rotated_(const size_t gemm_kernel_id) { return true; }
static const bool c_want_rotated_(const size_t gemm_kernel_id) { return gemm_kernel_id == 1; }
// Computes the size of the temporary GEMM buffer based on user-arguments
static size_t GetTempSize(const Layout layout, const Transpose a_transpose, const Transpose b_transpose,
const size_t m, const size_t n, const size_t k,
const size_t a_offset, const size_t a_ld,
const size_t b_offset, const size_t b_ld,
const size_t c_offset, const size_t c_ld,
const size_t mwg, const size_t nwg, const size_t kwg,
const size_t gemm_kernel_id) {
// 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;
ProcessArguments(layout, a_transpose, b_transpose, m, 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,
gemm_kernel_id);
// 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.
size_t a_one_i, a_two_i, b_one_i, b_two_i, c_one_i, c_two_i;
CalculateInternalDimensions(m, n, k, mwg, nwg, kwg,
a_one_i, a_two_i, b_one_i, b_two_i, c_one_i, c_two_i,
gemm_kernel_id);
// Determines whether or not temporary matrices are needed
auto a_no_temp = NoTempBuffer(a_one, a_one_i, a_two, a_two_i, a_ld, a_offset, a_do_transpose, a_conjugate);
auto b_no_temp = NoTempBuffer(b_one, b_one_i, b_two, b_two_i, b_ld, b_offset, b_do_transpose, b_conjugate);
auto c_no_temp = NoTempBuffer(c_one, c_one_i, c_two, c_two_i, c_ld, c_offset, c_do_transpose, false);
// Computes the sizes and offsets for (optional) temporary buffers for the 3 matrices
auto b_temp_offset = size_t{0};
auto c_temp_offset = size_t{0};
return ComputeTempSize(a_no_temp, b_no_temp, c_no_temp,
a_one_i*a_two_i, b_one_i*b_two_i, c_one_i*c_two_i,
b_temp_offset, c_temp_offset);
}
// Selects which version of GEMM to run
static bool UseDirectKernel(const size_t m, const size_t n, const size_t k,
const size_t min_indirect_size) {
const auto m_n_k = static_cast<unsigned long long>(m) * static_cast<unsigned long long>(n) *
static_cast<unsigned long long>(k);
const auto min_indirect_size_ll = static_cast<unsigned long long>(min_indirect_size);
const auto min_indirect_size_e3 = min_indirect_size_ll * min_indirect_size_ll * min_indirect_size_ll;
return (m_n_k < min_indirect_size_e3);
}
// Process the user-arguments, computes secondary parameters
static void ProcessArguments(const Layout layout, const Transpose a_transpose, const Transpose b_transpose,
const size_t m, const size_t n, const size_t k,
size_t& a_one, size_t& a_two, size_t& b_one,
size_t& b_two, size_t& c_one, size_t& c_two,
bool& a_do_transpose, bool& b_do_transpose, bool& c_do_transpose,
bool& a_conjugate, bool& b_conjugate,
const size_t gemm_kernel_id) {
// Makes sure all dimensions are larger than zero
if ((m == 0) || (n == 0) || (k == 0)) { throw BLASError(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. Note
// that the Xgemm kernel expects either matrices A and C (in case of row-major) or B (in case of
// col-major) to be transformed, so transposing requirements are not the same as whether or not
// the matrix is actually transposed in memory.
const auto a_rotated = (layout == Layout::kColMajor && a_transpose != Transpose::kNo) ||
(layout == Layout::kRowMajor && a_transpose == Transpose::kNo);
const auto b_rotated = (layout == Layout::kColMajor && b_transpose != Transpose::kNo) ||
(layout == Layout::kRowMajor && b_transpose == Transpose::kNo);
const auto c_rotated = (layout == Layout::kRowMajor);
a_do_transpose = a_rotated != a_want_rotated_(gemm_kernel_id);
b_do_transpose = b_rotated != b_want_rotated_(gemm_kernel_id);
c_do_transpose = c_rotated != c_want_rotated_(gemm_kernel_id);
// In case of complex data-types, the transpose can also become a conjugate transpose
a_conjugate = (a_transpose == Transpose::kConjugate);
b_conjugate = (b_transpose == Transpose::kConjugate);
// Computes the first and second dimensions of the 3 matrices taking into account whether the
// matrices are rotated or not
a_one = (a_rotated) ? k : m;
a_two = (a_rotated) ? m : k;
b_one = (b_rotated) ? n : k;
b_two = (b_rotated) ? k : n;
c_one = (c_rotated) ? n : m;
c_two = (c_rotated) ? m : n;
}
// Computes the sizes and offsets for (optional) temporary buffers for the 3 matrices
static size_t ComputeTempSize(const bool a_no_temp, const bool b_no_temp, const bool c_no_temp,
const size_t a_size, const size_t b_size, const size_t c_size,
size_t &b_temp_offset, size_t &c_temp_offset) {
auto temp_size = size_t{0};
if (!a_no_temp) { temp_size += a_size; }
if (!b_no_temp) { b_temp_offset = temp_size; temp_size += b_size; }
if (!c_no_temp) { c_temp_offset = temp_size; temp_size += c_size; }
return temp_size;
}
// Determines whether or not temporary matrices are needed
static bool NoTempBuffer(const size_t one, const size_t one_i, const size_t two, const size_t two_i,
const size_t ld, const size_t offset,
const bool do_transpose, const bool conjugate) {
return one == one_i && two == two_i && ld == one && offset == 0 && !do_transpose && !conjugate;
}
// 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.
static void CalculateInternalDimensions(const size_t m, const size_t n, const size_t k,
const size_t mwg, const size_t nwg, const size_t kwg,
size_t& a_one_i, size_t& a_two_i, size_t& b_one_i,
size_t& b_two_i, size_t& c_one_i, size_t& c_two_i,
const size_t gemm_kernel_id) {
const auto m_ceiled = Ceil(m, mwg);
const auto n_ceiled = Ceil(n, nwg);
const auto k_ceiled = Ceil(k, kwg);
a_one_i = (a_want_rotated_(gemm_kernel_id)) ? k_ceiled : m_ceiled;
a_two_i = (a_want_rotated_(gemm_kernel_id)) ? m_ceiled : k_ceiled;
b_one_i = (b_want_rotated_(gemm_kernel_id)) ? n_ceiled : k_ceiled;
b_two_i = (b_want_rotated_(gemm_kernel_id)) ? k_ceiled : n_ceiled;
c_one_i = (c_want_rotated_(gemm_kernel_id)) ? n_ceiled : m_ceiled;
c_two_i = (c_want_rotated_(gemm_kernel_id)) ? m_ceiled : n_ceiled;
}
// Constructor
Xgemm(Queue &queue, EventPointer event, const std::string &name = "GEMM");
// Templated-precision implementation of the routine
void DoGemm(const Layout layout, const Transpose a_transpose, const Transpose b_transpose,
const size_t m, 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,
const Buffer<T> &temp_buffer = Buffer<T>(0), const bool temp_buffer_provided = false);
// Indirect version of GEMM (with pre and post-processing kernels)
void GemmIndirect(const size_t m, 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,
const bool a_do_transpose, const bool b_do_transpose, const bool c_do_transpose,
const bool a_conjugate, const bool b_conjugate,
const size_t a_one, const size_t a_two,
const size_t b_one, const size_t b_two,
const size_t c_one, const size_t c_two,
const Buffer<T> &temp_buffer, const bool temp_buffer_provided);
// Direct version of GEMM (no pre and post-processing kernels)
void GemmDirect(const size_t m, 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,
const bool a_do_transpose, const bool b_do_transpose, const bool c_do_transpose,
const bool a_conjugate, const bool b_conjugate);
};
// =================================================================================================
} // namespace clblast
// CLBLAST_ROUTINES_XGEMM_H_
#endif
|
