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
|
#define A(i,j) a[ (j)*lda + (i) ] #define B(i,j) b[ (j)*ldb + (i) ] #define C(i,j) c[ (j)*ldc + (i) ]
#define mc 256 #define kc 128
#define min( i, j ) ( (i)<(j) ? (i): (j) )
void AddDot4x4( int, double *, int, double *, int, double *, int ); void PackMatrixA( int, double *, int, double * ); void PackMatrixB( int, double *, int, double * ); void InnerKernel( int, int, int, double *, int, double *, int, double *, int, int );
void MY_MMult( int m, int n, int k, double *a, int lda, double *b, int ldb, double *c, int ldc ) { int i, p, pb, ib;
for ( p=0; p<k; p+=kc ){ pb = min( k-p, kc ); for ( i=0; i<m; i+=mc ){ ib = min( m-i, mc ); InnerKernel( ib, n, pb, &A( i,p ), lda, &B(p, 0 ), ldb, &C( i,0 ), ldc, i==0 ); } } }
void InnerKernel( int m, int n, int k, double *a, int lda, double *b, int ldb, double *c, int ldc, int first_time ) { int i, j; double packedA[ m * k ], packedB[ k*n ];
for ( j=0; j<n; j+=4 ){ PackMatrixB( k, &B( 0, j ), ldb, &packedB[ j*k ] ); for ( i=0; i<m; i+=4 ){
if ( j == 0 ) PackMatrixA( k, &A( i, 0 ), lda, &packedA[ i*k ] ); AddDot4x4( k, &packedA[ i*k ], 4, &packedB[ j*k ], k, &C( i,j ), ldc ); } } }
void PackMatrixA( int k, double *a, int lda, double *a_to ) { int j;
for( j=0; j<k; j++){ double *a_ij_pntr = &A( 0, j );
*a_to = *a_ij_pntr; *(a_to+1) = *(a_ij_pntr+1); *(a_to+2) = *(a_ij_pntr+2); *(a_to+3) = *(a_ij_pntr+3);
a_to += 4; } }
void PackMatrixB( int k, double *b, int ldb, double *b_to ) { int i; double *b_i0_pntr = &B( 0, 0 ), *b_i1_pntr = &B( 0, 1 ), *b_i2_pntr = &B( 0, 2 ), *b_i3_pntr = &B( 0, 3 );
for( i=0; i<k; i++){ *b_to++ = *b_i0_pntr++; *b_to++ = *b_i1_pntr++; *b_to++ = *b_i2_pntr++; *b_to++ = *b_i3_pntr++; } }
#include <mmintrin.h> #include <xmmintrin.h> #include <pmmintrin.h> #include <emmintrin.h>
typedef union { __m128d v; double d[2]; } v2df_t;
void AddDot4x4( int k, double *a, int lda, double *b, int ldb, double *c, int ldc ) {
int p; v2df_t c_00_c_10_vreg, c_01_c_11_vreg, c_02_c_12_vreg, c_03_c_13_vreg, c_20_c_30_vreg, c_21_c_31_vreg, c_22_c_32_vreg, c_23_c_33_vreg, a_0p_a_1p_vreg, a_2p_a_3p_vreg, b_p0_vreg, b_p1_vreg, b_p2_vreg, b_p3_vreg;
c_00_c_10_vreg.v = _mm_setzero_pd(); c_01_c_11_vreg.v = _mm_setzero_pd(); c_02_c_12_vreg.v = _mm_setzero_pd(); c_03_c_13_vreg.v = _mm_setzero_pd(); c_20_c_30_vreg.v = _mm_setzero_pd(); c_21_c_31_vreg.v = _mm_setzero_pd(); c_22_c_32_vreg.v = _mm_setzero_pd(); c_23_c_33_vreg.v = _mm_setzero_pd();
for ( p=0; p<k; p++ ){ a_0p_a_1p_vreg.v = _mm_load_pd( (double *) a ); a_2p_a_3p_vreg.v = _mm_load_pd( (double *) ( a+2 ) ); a += 4;
b_p0_vreg.v = _mm_loaddup_pd( (double *) b ); b_p1_vreg.v = _mm_loaddup_pd( (double *) (b+1) ); b_p2_vreg.v = _mm_loaddup_pd( (double *) (b+2) ); b_p3_vreg.v = _mm_loaddup_pd( (double *) (b+3) );
b += 4;
c_00_c_10_vreg.v += a_0p_a_1p_vreg.v * b_p0_vreg.v; c_01_c_11_vreg.v += a_0p_a_1p_vreg.v * b_p1_vreg.v; c_02_c_12_vreg.v += a_0p_a_1p_vreg.v * b_p2_vreg.v; c_03_c_13_vreg.v += a_0p_a_1p_vreg.v * b_p3_vreg.v;
c_20_c_30_vreg.v += a_2p_a_3p_vreg.v * b_p0_vreg.v; c_21_c_31_vreg.v += a_2p_a_3p_vreg.v * b_p1_vreg.v; c_22_c_32_vreg.v += a_2p_a_3p_vreg.v * b_p2_vreg.v; c_23_c_33_vreg.v += a_2p_a_3p_vreg.v * b_p3_vreg.v; }
C( 0, 0 ) += c_00_c_10_vreg.d[0]; C( 0, 1 ) += c_01_c_11_vreg.d[0]; C( 0, 2 ) += c_02_c_12_vreg.d[0]; C( 0, 3 ) += c_03_c_13_vreg.d[0];
C( 1, 0 ) += c_00_c_10_vreg.d[1]; C( 1, 1 ) += c_01_c_11_vreg.d[1]; C( 1, 2 ) += c_02_c_12_vreg.d[1]; C( 1, 3 ) += c_03_c_13_vreg.d[1];
C( 2, 0 ) += c_20_c_30_vreg.d[0]; C( 2, 1 ) += c_21_c_31_vreg.d[0]; C( 2, 2 ) += c_22_c_32_vreg.d[0]; C( 2, 3 ) += c_23_c_33_vreg.d[0];
C( 3, 0 ) += c_20_c_30_vreg.d[1]; C( 3, 1 ) += c_21_c_31_vreg.d[1]; C( 3, 2 ) += c_22_c_32_vreg.d[1]; C( 3, 3 ) += c_23_c_33_vreg.d[1]; }
|