Fix compilation of NEON code [ci skip]

althonos · Oct 21, 2024 · 78f7c1e · 78f7c1e
1 parent eec8d98
commit 78f7c1e
Showing 1 changed file with 5 additions and 4 deletions.
diff --git a/src/pyfastani/_sequtils/neon.c b/src/pyfastani/_sequtils/neon.c
@@ -1,15 +1,16 @@
 #include <ctype.h>
-#include <x86intrin.h>
+#include <stddef.h>
+#include <arm_neon.h>
 
 extern void neon_copy_upper(char* dst, const char* src, size_t len) {
     const int8x16_t ascii_a = vdupq_n_s8('a' - 1);
     const int8x16_t ascii_z = vdupq_n_s8('z');
     const int8x16_t offset  = vdupq_n_s8('a' - 'A');
 
     while (len >= sizeof(int8x16_t)) {
-        int8x16_t inp = vld1q_u8((int8_t*) src);
-        int8x16_t greater_than_a = vcgtq_s8(inp, ascii_a);
-        int8x16_t less_equal_z = vcgtq_s8(ascii_z, inp);
+        int8x16_t inp = vld1q_s8((int8_t*) src);
+        int8x16_t greater_than_a = vreinterpretq_s8_u8(vcgtq_s8(inp, ascii_a));
+        int8x16_t less_equal_z = vreinterpretq_s8_u8(vcgtq_s8(ascii_z, inp));
         int8x16_t mask = vandq_s8(greater_than_a, less_equal_z);
         int8x16_t diff = vandq_s8(mask, offset);
         int8x16_t added = vsubq_s8(inp, diff);