Only use the ARM NEON 32-way unrolled rANS on AArch64.

NEON alone isn't a sufficient guard as AArch32 also has some limited Neon capabilities. While we could no doubt have a 32-bit alternative, for now this is the simple fix and let aarch32 use the scalar implementation. Doing a 32-bit neon is a complex task and without having access to the hardware it's pretty much impossible. I also wouldn't have high hopes for any significant speed gains over scalar with only half the lanes available. Fixes #81
samtools · Apr 19, 2023 · 479b5be · 479b5be
1 parent 5aecc6e
commit 479b5be
Show file tree

Hide file tree

Showing 3 changed files with 3 additions and 3 deletions.
diff --git a/htscodecs/rANS_static32x16pr.h b/htscodecs/rANS_static32x16pr.h
@@ -146,7 +146,7 @@ unsigned char *rans_uncompress_O1_32x16_avx512(unsigned char *in,
 
 //----------------------------------------------------------------------
 // Arm Neon implementation
-#ifdef __ARM_NEON
+#if defined(__ARM_NEON) && defined(__aarch64__)
 unsigned char *rans_compress_O0_32x16_neon(unsigned char *in,
                                            unsigned int in_size,
                                            unsigned char *out,

diff --git a/htscodecs/rANS_static32x16pr_neon.c b/htscodecs/rANS_static32x16pr_neon.c
@@ -32,7 +32,7 @@
  */
 
 #include "config.h"
-#ifdef __ARM_NEON
+#if defined(__ARM_NEON) && defined(__aarch64__)
 #include <arm_neon.h>
 
 #include <limits.h>

diff --git a/htscodecs/rANS_static4x16pr.c b/htscodecs/rANS_static4x16pr.c
@@ -1006,7 +1006,7 @@ unsigned char *(*rans_dec_func(int do_simd, int order))
     }
 }
 
-#elif defined(__ARM_NEON)
+#elif defined(__ARM_NEON) && defined(__aarch64__)
 
 #if defined(__linux__) || defined(__FreeBSD__)
 #include <sys/auxv.h>