1 // -*- C++ -*-
2 /** @file */
3 #pragma once
4
5 /** \addtogroup SCALARALG
6 * @{
7 */
8
9 /**
10 * @brief Matrix times matrix for scalar architecture and float
11 *
12 * @param[in] pSrcA The source a
13 * @param[in] pSrcB The source b
14 * @param pDst The destination
15 *
16 * @tparam MA Left hand side datatype
17 * @tparam MB Right hand side datatype
18 * @tparam RES Result datatype
19 * @tparam <unnamed> Check if float
20 */
21 template<typename MA,
22 typename MB,
23 typename RES,
24 typename std::enable_if<number_traits<typename traits<MA>::Scalar>::is_float,bool>::type = true>
_dot_m_m(const MA & pSrcA,const MB & pSrcB,RES && pDst,const Scalar * =nullptr)25 __STATIC_INLINE void _dot_m_m(const MA&pSrcA,const MB&pSrcB,
26 RES &&pDst,
27 const Scalar* = nullptr)
28 {
29 using T = typename traits<MA>::Scalar;
30 using Acc = typename number_traits<T>::accumulator;
31 //using Comp = typename number_traits<T>::compute_type;
32 T *pIn1 = pSrcA.ptr(); /* Input data matrix pointer A */
33 T *pIn2 = pSrcB.ptr(); /* Input data matrix pointer B */
34 T *pInA = pSrcA.ptr(); /* Input data matrix pointer A */
35 T *pInB = pSrcB.ptr(); /* Input data matrix pointer B */
36 T *pOut = pDst.ptr(); /* Output data matrix pointer */
37 T *px; /* Temporary output data matrix pointer */
38 Acc sum; /* Accumulator */
39 uint16_t numRowsA = pSrcA.rows(); /* Number of rows of input matrix A */
40 uint16_t numColsB = pSrcB.columns(); /* Number of columns of input matrix B */
41 uint16_t numColsA = pSrcA.columns(); /* Number of columns of input matrix A */
42 uint32_t col, i = 0U, row = numRowsA, colCnt; /* Loop counters */
43
44
45 /* The following loop performs the dot-product of each row in pSrcA with each column in pSrcB */
46 /* row loop */
47 do
48 {
49 /* Output pointer is set to starting address of row being processed */
50 px = pOut + i;
51
52 /* For every row wise process, column loop counter is to be initiated */
53 col = numColsB;
54
55 /* For every row wise process, pIn2 pointer is set to starting address of pSrcB data */
56 pIn2 = pSrcB.ptr();
57
58 /* column loop */
59 do
60 {
61 /* Set the variable sum, that acts as accumulator, to zero */
62 sum = Acc{};
63
64 /* Initialize pointer pIn1 to point to starting address of column being processed */
65 pIn1 = pInA;
66
67
68 /* Loop unrolling: Compute 4 MACs at a time. */
69 colCnt = numColsA >> 2U;
70
71 /* matrix multiplication */
72 while (colCnt > 0U)
73 {
74 /* c(m,p) = a(m,1) * b(1,p) + a(m,2) * b(2,p) + .... + a(m,n) * b(n,p) */
75
76 /* Perform the multiply-accumulates */
77 sum = inner::mac(sum, *pIn1++, *pIn2);
78 pIn2 += pSrcB.stride();
79
80 sum = inner::mac(sum, *pIn1++, *pIn2);
81 pIn2 += pSrcB.stride();
82
83 sum = inner::mac(sum, *pIn1++, *pIn2);
84 pIn2 += pSrcB.stride();
85
86 sum = inner::mac(sum, *pIn1++, *pIn2);
87 pIn2 += pSrcB.stride();
88
89 /* Decrement loop counter */
90 colCnt--;
91 }
92
93 /* Loop unrolling: Compute remaining MACs */
94 colCnt = numColsA % 0x4U;
95
96 while (colCnt > 0U)
97 {
98 /* c(m,p) = a(m,1) * b(1,p) + a(m,2) * b(2,p) + .... + a(m,n) * b(n,p) */
99
100 /* Perform the multiply-accumulates */
101 sum = inner::mac(sum, *pIn1++, *pIn2);
102 pIn2 += pSrcB.stride();
103
104 /* Decrement loop counter */
105 colCnt--;
106 }
107
108 /* Store result in destination buffer */
109 *px++ = inner::from_accumulator(sum);
110
111 /* Decrement column loop counter */
112 col--;
113
114 /* Update pointer pIn2 to point to starting address of next column */
115 pIn2 = pInB + (numColsB - col);
116
117 } while (col > 0U);
118
119 /* Update pointer pInA to point to starting address of next row */
120 i = i + pDst.stride();
121 pInA = pInA + pSrcA.stride();
122
123 /* Decrement row loop counter */
124 row--;
125
126 } while (row > 0U);
127
128
129 }
130
131 /*! @} */
