Simplify avx2 chunkset a bit

Put length 16 in the length checking ladder and take care of it there since it's also a simple case to handle. We kind of went out of our way to pretend 128 bit vectors didn't exist when using avx2 but this can be handled in a single instruction. Strangely the intrinsic uses vector register operands but the instruction itself assumes a memory operand for the source. This also means we don't have to handle this case in our "GET_CHUNK_MAG" function.
zlib-ng · Oct 12, 2024 · b52e703 · b52e703
1 parent dae668d
commit b52e703
Show file tree

Hide file tree

Showing 2 changed files with 14 additions and 7 deletions.
diff --git a/arch/x86/chunkset_avx2.c b/arch/x86/chunkset_avx2.c
@@ -14,6 +14,7 @@ typedef __m256i chunk_t;
 #define HAVE_CHUNKMEMSET_2
 #define HAVE_CHUNKMEMSET_4
 #define HAVE_CHUNKMEMSET_8
+#define HAVE_CHUNKMEMSET_16
 #define HAVE_CHUNK_MAG
 
 /* Populate don't cares so that this is a direct lookup (with some indirection into the permute table), because dist can
@@ -68,6 +69,10 @@ static inline void chunkmemset_8(uint8_t *from, chunk_t *chunk) {
     *chunk = _mm256_set1_epi64x(tmp);
 }
 
+static inline void chunkmemset_16(uint8_t *from, chunk_t *chunk) {
+    *chunk = _mm256_broadcastsi128_si256(_mm_loadu_si128((__m128i*)from));
+}
+
 static inline void loadchunk(uint8_t const *s, chunk_t *chunk) {
     *chunk = _mm256_loadu_si256((__m256i *)s);
 }
@@ -99,10 +104,7 @@ static inline chunk_t GET_CHUNK_MAG(uint8_t *buf, uint32_t *chunk_rem, uint32_t
         perm_vec = _mm256_add_epi8(perm_vec, permute_xform);
         ret_vec = _mm256_inserti128_si256(_mm256_castsi128_si256(ret_vec0), ret_vec0, 1);
         ret_vec = _mm256_shuffle_epi8(ret_vec, perm_vec);
-    } else if (dist == 16) {
-        __m128i ret_vec0 = _mm_loadu_si128((__m128i*)buf);
-        return _mm256_inserti128_si256(_mm256_castsi128_si256(ret_vec0), ret_vec0, 1);
-    } else {
+    }  else {
         __m128i ret_vec0 = _mm_loadu_si128((__m128i*)buf);
         __m128i ret_vec1 = _mm_loadu_si128((__m128i*)(buf + 16));
         /* Take advantage of the fact that only the latter half of the 256 bit vector will actually differ */

diff --git a/chunkset_tpl.h b/chunkset_tpl.h
@@ -130,11 +130,16 @@ Z_INTERNAL uint8_t* CHUNKMEMSET(uint8_t *out, unsigned dist, unsigned len) {
 #ifdef HAVE_CHUNKMEMSET_8
     if (dist == 8) {
         chunkmemset_8(from, &chunk_load);
-    } else if (dist == sizeof(chunk_t)) {
-        loadchunk(from, &chunk_load);
     } else
 #endif
-    {
+#ifdef HAVE_CHUNKMEMSET_16
+    if (dist == 16) {
+        chunkmemset_16(from, &chunk_load);
+    } else
+#endif
+    if (dist == sizeof(chunk_t)) {
+        loadchunk(from, &chunk_load);
+    } else {
         chunk_load = GET_CHUNK_MAG(from, &chunk_mod, dist);
     }