1+ // Copyright (C) 2018-2023 Intel Corporation
2+ // SPDX-License-Identifier: Apache-2.0
3+ //
4+
5+ #pragma once
6+
7+ #include < cstddef>
8+ #include < cstdint>
9+
10+ namespace ov {
11+ namespace intel_cpu {
12+ /* *
13+ * @brief Computes the length in bytes for the packed matrix B buffer(SGEMM).
14+ *
15+ * @param N Supplies the number of columns of matrix B.
16+ * @param K Supplies the number of rows of matrix B.
17+ * @return bytes of the packing buffer
18+ */
19+ size_t mlas_sgemm_pack_get_size (const int64_t N, const int64_t K);
20+
21+ /* *
22+ * @brief Packs the contents of matrix B
23+ *
24+ * @param transb T for transpose B, N for none-tranpose B
25+ * @param N Supplies the number of columns of matrix B and matrix C.
26+ * @param K Supplies the number of columns of matrix A and the number
27+ of rows of matrix B.
28+ * @param ldb Supplies the first dimension of matrix B.
29+ * @param src Supplies the address of matrix B
30+ * @param dst Supplies pointer to prePacked B buffer
31+ */
32+ void mlas_sgemm_pack (const char * transb,
33+ const int64_t N,
34+ const int64_t K,
35+ const int64_t ldb,
36+ const float * src,
37+ float * dst);
38+
39+ /* *
40+ * @brief SGEMM with planar B matrix
41+ *
42+ * @param transa T for transpose A, N for none-tranpose A.
43+ * @param transb T for transpose B, N for none-tranpose B.
44+ * @param M Supplies the number of rows of matrix A and matrix C.
45+ * @param N Supplies the number of columns of matrix B and matrix C.
46+ * @param K Supplies the number of columns of matrix A and the number
47+ of rows of matrix B.
48+ * @param alpha Supplies the scalar alpha multiplier (see SGEMM definition)
49+ * @param A Supplies the address of matrix A
50+ * @param lda Supplies the first dimension of matrix A.
51+ * @param B Supplies the address of matrix B
52+ * @param ldb Supplies the first dimension of matrix B.
53+ * @param beta Supplies the scalar beta multiplier (see SGEMM definition)
54+ * @param C Supplies the address of matrix C
55+ * @param ldc Supplies the first dimension of matrix C.
56+ * @param thread_num 0 for all threads, otherwise use thread_num
57+ */
58+ void mlas_sgemm (const char * transa,
59+ const char * transb,
60+ const int64_t M,
61+ const int64_t N,
62+ const int64_t K,
63+ const float alpha,
64+ const float * A,
65+ const int64_t lda,
66+ const float * B,
67+ const int64_t ldb,
68+ const float beta,
69+ float * C,
70+ const int64_t ldc,
71+ size_t thread_num = 0 );
72+
73+ /* *
74+ * @brief SGEMM with B matrix prepacked
75+ *
76+ * @param transa T for transpose A, N for none-tranpose A.
77+ * @param transb T for transpose B, N for none-tranpose B.
78+ * @param M Supplies the number of rows of matrix A and matrix C.
79+ * @param N Supplies the number of columns of matrix B and matrix C.
80+ * @param K Supplies the number of columns of matrix A and the number
81+ of rows of matrix B.
82+ * @param alpha Supplies the scalar alpha multiplier (see SGEMM definition)
83+ * @param A Supplies the address of matrix A
84+ * @param lda Supplies the first dimension of matrix A.
85+ * @param B Supplies the address of matrix B
86+ * @param ldb Supplies the first dimension of matrix B.
87+ * @param beta Supplies the scalar beta multiplier (see SGEMM definition)
88+ * @param C Supplies the address of matrix C
89+ * @param ldc Supplies the first dimension of matrix C.
90+ * @param bias Supplies the address of by-channel bias
91+ * @param thread_num 0 for all threads, otherwise use thread_num
92+ */
93+ void mlas_sgemm_compute (const char * transa,
94+ const char * transb,
95+ const int64_t M,
96+ const int64_t N,
97+ const int64_t K,
98+ const float alpha,
99+ const float * A,
100+ const int64_t lda,
101+ const float * B,
102+ const int64_t ldb,
103+ const float beta,
104+ float * C,
105+ const int64_t ldc,
106+ const float * bias = nullptr ,
107+ size_t thread_num = 0 );
108+ } // namespace intel_cpu
109+ } // namespace ov
0 commit comments