Implement some vec instructions with NEON

ish-app · Dec 13, 2020 · 89dfac7 · 89dfac7
1 parent dccc553
commit 89dfac7
Showing 1 changed file with 24 additions and 0 deletions.
diff --git a/emu/vec.c b/emu/vec.c
@@ -1,3 +1,6 @@
+#ifdef __ARM_NEON__
+#include <arm_neon.h>
+#endif
 #include <math.h>
 #include <string.h>
 
@@ -91,8 +94,15 @@ void vec_shiftr_q128(NO_CPU, union xmm_reg *amount, union xmm_reg *dst) {
 }
 
 void vec_add_b128(NO_CPU, union xmm_reg *src, union xmm_reg *dst) {
+#ifdef __ARM_NEON__
+    uint8x16_t neon_dst = vld1q_u8(dst->u8);
+    uint8x16_t neon_src = vld1q_u8(src->u8);
+    uint8x16_t neon_res = vaddq_u8(neon_dst, neon_src);
+    vst1q_u8(dst->u8, neon_res);
+#else
     for (unsigned i = 0; i < array_size(src->u8); i++)
         dst->u8[i] += src->u8[i];
+#endif
 }
 void vec_add_d128(NO_CPU, union xmm_reg *src, union xmm_reg *dst) {
     for (unsigned i = 0; i < array_size(src->u32); i++)
@@ -142,9 +152,16 @@ void vec_xor64(NO_CPU, union mm_reg *src, union mm_reg *dst) {
 }
 
 void vec_min_ub128(NO_CPU, union xmm_reg *src, union xmm_reg *dst) {
+#ifdef __ARM_NEON__
+    uint8x16_t neon_dst = vld1q_u8(dst->u8);
+    uint8x16_t neon_src = vld1q_u8(src->u8);
+    uint8x16_t neon_res = vminq_u8(neon_dst, neon_src);
+    vst1q_u8(dst->u8, neon_res);
+#else
     for (unsigned i = 0; i < array_size(src->u8); i++)
         if (src->u8[i] < dst->u8[i])
             dst->u8[i] = src->u8[i];
+#endif
 }
 
 static bool cmpd(double a, double b, int type) {
@@ -253,8 +270,15 @@ void vec_shuffle_d128(NO_CPU, const union xmm_reg *src, union xmm_reg *dst, uint
 }
 
 void vec_compare_eqb128(NO_CPU, const union xmm_reg *src, union xmm_reg *dst) {
+#ifdef __ARM_NEON__
+    uint8x16_t neon_dst = vld1q_u8(dst->u8);
+    uint8x16_t neon_src = vld1q_u8(src->u8);
+    uint8x16_t neon_res = vceqq_u8(neon_dst, neon_src);
+    vst1q_u8(dst->u8, neon_res);
+#else
     for (unsigned i = 0; i < array_size(src->u8); i++)
         dst->u8[i] = dst->u8[i] == src->u8[i] ? ~0 : 0;
+#endif
 }
 void vec_compare_eqd128(NO_CPU, const union xmm_reg *src, union xmm_reg *dst) {
     for (unsigned i = 0; i < array_size(src->u32); i++)