| 1 |
/* { dg-do run } */ |
| 2 |
/* { dg-options "-mips3d forbid_cpu=octeon.* (REQUIRES_STDLIB)" } */ |
| 3 |
|
| 4 |
/* Matrix Multiplications */ |
| 5 |
#include <stdlib.h> |
| 6 |
#include <stdio.h> |
| 7 |
|
| 8 |
typedef float v2sf __attribute__((vector_size(8))); |
| 9 |
|
| 10 |
float a[4] = {1.1, 2.2, 3.3, 4.4}; |
| 11 |
float b[4][4] = {{1, 2, 3, 4}, |
| 12 |
{5, 6, 7, 8}, |
| 13 |
{9, 10, 11, 12}, |
| 14 |
{13, 14, 15, 16}}; |
| 15 |
|
| 16 |
float c[4]; /* Result for matrix_multiply1() */ |
| 17 |
float d[4]; /* Result for matrix_multiply2() */ |
| 18 |
float e[4]; /* Result for matrix_multiply3() */ |
| 19 |
float f[4]; /* Result for matrix_multiply4() */ |
| 20 |
|
| 21 |
void matrix_multiply1(); |
| 22 |
NOMIPS16 void matrix_multiply2(); |
| 23 |
NOMIPS16 void matrix_multiply3(); |
| 24 |
NOMIPS16 void matrix_multiply4(); |
| 25 |
|
| 26 |
int main () |
| 27 |
{ |
| 28 |
int i; |
| 29 |
|
| 30 |
/* Version 1. Use float calculations */ |
| 31 |
matrix_multiply1(); |
| 32 |
|
| 33 |
/* Version 2. Use paired-single instructions inside the inner loop*/ |
| 34 |
matrix_multiply2(); |
| 35 |
for (i = 0; i < 4; i++) |
| 36 |
if (d[i] != c[i]) |
| 37 |
abort(); |
| 38 |
|
| 39 |
/* Version 3. Use paired-single instructions and unroll the inner loop */ |
| 40 |
matrix_multiply3(); |
| 41 |
for (i = 0; i < 4; i++) |
| 42 |
if (e[i] != c[i]) |
| 43 |
abort(); |
| 44 |
|
| 45 |
/* Version 4. Use paired-single instructions and unroll all loops */ |
| 46 |
matrix_multiply4(); |
| 47 |
for (i = 0; i < 4; i++) |
| 48 |
if (f[i] != c[i]) |
| 49 |
abort(); |
| 50 |
|
| 51 |
printf ("Test Passes\n"); |
| 52 |
exit (0); |
| 53 |
} |
| 54 |
|
| 55 |
void matrix_multiply1() |
| 56 |
{ |
| 57 |
int i, j; |
| 58 |
|
| 59 |
for (i = 0; i < 4; i++) |
| 60 |
{ |
| 61 |
c[i] = 0.0; |
| 62 |
|
| 63 |
for (j = 0; j < 4; j ++) |
| 64 |
c[i] += a[j] * b[j][i]; |
| 65 |
} |
| 66 |
} |
| 67 |
|
| 68 |
NOMIPS16 void matrix_multiply2() |
| 69 |
{ |
| 70 |
int i, j; |
| 71 |
v2sf m1, m2; |
| 72 |
v2sf result, temp; |
| 73 |
|
| 74 |
for (i = 0; i < 4; i++) |
| 75 |
{ |
| 76 |
result = (v2sf) {0.0, 0.0}; |
| 77 |
|
| 78 |
for (j = 0; j < 4; j+=2) |
| 79 |
{ |
| 80 |
/* Load two float values into m1 */ |
| 81 |
m1 = (v2sf) {a[j], a[j+1]}; |
| 82 |
m2 = (v2sf) {b[j][i], b[j+1][i]}; |
| 83 |
|
| 84 |
/* Multiply and add */ |
| 85 |
result += m1 * m2; |
| 86 |
} |
| 87 |
|
| 88 |
/* Reduction add at the end */ |
| 89 |
temp = __builtin_mips_addr_ps (result, result); |
| 90 |
d[i] = __builtin_mips_cvt_s_pl (temp); |
| 91 |
} |
| 92 |
} |
| 93 |
|
| 94 |
NOMIPS16 void matrix_multiply3() |
| 95 |
{ |
| 96 |
int i; |
| 97 |
v2sf m1, m2, n1, n2; |
| 98 |
v2sf result, temp; |
| 99 |
|
| 100 |
m1 = (v2sf) {a[0], a[1]}; |
| 101 |
m2 = (v2sf) {a[2], a[3]}; |
| 102 |
|
| 103 |
for (i = 0; i < 4; i++) |
| 104 |
{ |
| 105 |
n1 = (v2sf) {b[0][i], b[1][i]}; |
| 106 |
n2 = (v2sf) {b[2][i], b[3][i]}; |
| 107 |
|
| 108 |
/* Multiply and add */ |
| 109 |
result = m1 * n1 + m2 * n2; |
| 110 |
|
| 111 |
/* Reduction add at the end */ |
| 112 |
temp = __builtin_mips_addr_ps (result, result); |
| 113 |
e[i] = __builtin_mips_cvt_s_pl (temp); |
| 114 |
} |
| 115 |
} |
| 116 |
|
| 117 |
NOMIPS16 void matrix_multiply4() |
| 118 |
{ |
| 119 |
v2sf m1, m2; |
| 120 |
v2sf n1, n2, n3, n4, n5, n6, n7, n8; |
| 121 |
v2sf temp1, temp2, temp3, temp4; |
| 122 |
v2sf result1, result2; |
| 123 |
|
| 124 |
/* Load a[0] a[1] values into m1 |
| 125 |
Load a[2] a[3] values into m2 */ |
| 126 |
m1 = (v2sf) {a[0], a[1]}; |
| 127 |
m2 = (v2sf) {a[2], a[3]}; |
| 128 |
|
| 129 |
/* Load b[0][0] b[1][0] values into n1 |
| 130 |
Load b[2][0] b[3][0] values into n2 |
| 131 |
Load b[0][1] b[1][1] values into n3 |
| 132 |
Load b[2][1] b[3][1] values into n4 |
| 133 |
Load b[0][2] b[1][2] values into n5 |
| 134 |
Load b[2][2] b[3][2] values into n6 |
| 135 |
Load b[0][3] b[1][3] values into n7 |
| 136 |
Load b[2][3] b[3][3] values into n8 */ |
| 137 |
n1 = (v2sf) {b[0][0], b[1][0]}; |
| 138 |
n2 = (v2sf) {b[2][0], b[3][0]}; |
| 139 |
n3 = (v2sf) {b[0][1], b[1][1]}; |
| 140 |
n4 = (v2sf) {b[2][1], b[3][1]}; |
| 141 |
n5 = (v2sf) {b[0][2], b[1][2]}; |
| 142 |
n6 = (v2sf) {b[2][2], b[3][2]}; |
| 143 |
n7 = (v2sf) {b[0][3], b[1][3]}; |
| 144 |
n8 = (v2sf) {b[2][3], b[3][3]}; |
| 145 |
|
| 146 |
temp1 = m1 * n1 + m2 * n2; |
| 147 |
temp2 = m1 * n3 + m2 * n4; |
| 148 |
temp3 = m1 * n5 + m2 * n6; |
| 149 |
temp4 = m1 * n7 + m2 * n8; |
| 150 |
|
| 151 |
result1 = __builtin_mips_addr_ps (temp1, temp2); |
| 152 |
result2 = __builtin_mips_addr_ps (temp3, temp4); |
| 153 |
|
| 154 |
f[0] = __builtin_mips_cvt_s_pu (result1); |
| 155 |
f[1] = __builtin_mips_cvt_s_pl (result1); |
| 156 |
f[2] = __builtin_mips_cvt_s_pu (result2); |
| 157 |
f[3] = __builtin_mips_cvt_s_pl (result2); |
| 158 |
} |