1 #include <stdio.h> 2 #include <string.h> 3 4 int main() 5 { 6 static const int cmp[4][4] = { 7 { 110, 134, 158, 182 }, 8 { 390, 478, 566, 654 }, 9 { 670, 822, 974, 1126 }, 10 { 950, 1166, 1382, 1598 } 11 }; 12 int dst[4][4]; 13 int *tmp = &dst[0][0]; 14 15 asm volatile( 16 ".arch armv8-r+sme\n\t" 17 "smstart\n\t" 18 "index z0.b, #0, #1\n\t" 19 "movprfx z1, z0\n\t" 20 "add z1.b, z1.b, #16\n\t" 21 "ptrue p0.b\n\t" 22 "smopa za0.s, p0/m, p0/m, z0.b, z1.b\n\t" 23 "ptrue p0.s, vl4\n\t" 24 "mov w12, #0\n\t" 25 "st1w { za0h.s[w12, #0] }, p0, [%0]\n\t" 26 "add %0, %0, #16\n\t" 27 "st1w { za0h.s[w12, #1] }, p0, [%0]\n\t" 28 "add %0, %0, #16\n\t" 29 "st1w { za0h.s[w12, #2] }, p0, [%0]\n\t" 30 "add %0, %0, #16\n\t" 31 "st1w { za0h.s[w12, #3] }, p0, [%0]\n\t" 32 "smstop" 33 : "+r"(tmp) : : "memory"); 34 35 if (memcmp(cmp, dst, sizeof(dst)) == 0) { 36 return 0; 37 } 38 39 /* See above for correct results. */ 40 for (int i = 0; i < 4; ++i) { 41 for (int j = 0; j < 4; ++j) { 42 printf("%6d", dst[i][j]); 43 } 44 printf("\n"); 45 } 46 return 1; 47 } 48