performance improvements

Author: thomasjungblut

recordio/simd/search.c

@@ -60,7 +60,7 @@ int cpu_supports_avx512() {
 }
 
 // Optimized SSE4 implementation for 3-byte pattern search
-// Checks 14 positions per iteration by advancing 1 byte at a time
+// Checks 14 positions per iteration by advancing 13 bytes at a time
 // Pattern bytes are broadcast once outside the loop for better performance
 int find_magic_numbers_sse4(const unsigned char* data, size_t off, size_t len) {
     if (len < 3) return -1;
@@ -75,8 +75,8 @@ int find_magic_numbers_sse4(const unsigned char* data, size_t off, size_t len) {
     __m128i p2 = _mm_set1_epi8(pattern[2]);
 
     // Process 16 bytes per loop, checking 14 positions per iteration
-    // Advance by 1 byte to check every possible starting position
-    for (; i + 16 <= end; i += 1) {
+    // Advance by 14 bytes (16 - 3 + 1) to check all positions without gaps
+    for (; i + 16 <= end; i += 14) {
         __m128i d0 = _mm_loadu_si128((const __m128i*)(data + i));
         __m128i d1 = _mm_loadu_si128((const __m128i*)(data + i + 1));
         __m128i d2 = _mm_loadu_si128((const __m128i*)(data + i + 2));
@@ -107,7 +107,7 @@ int find_magic_numbers_sse4(const unsigned char* data, size_t off, size_t len) {
 }
 
 // Optimized AVX2 implementation for 3-byte pattern search
-// Checks 30 positions per iteration by advancing 1 byte at a time
+// Checks 30 positions per iteration by advancing 29 bytes at a time
 // Pattern bytes are broadcast once outside the loop for better performance
 int find_magic_numbers_avx2(const unsigned char* data, size_t off, size_t len) {
     if (len < 3) return -1;
@@ -122,8 +122,8 @@ int find_magic_numbers_avx2(const unsigned char* data, size_t off, size_t len) {
     __m256i p2 = _mm256_set1_epi8(pattern[2]);
 
     // Process 32 bytes per loop, checking 30 positions per iteration
-    // Advance by 1 byte to check every possible starting position
-    for (; i + 32 <= end; i += 1) {
+    // Advance by 30 bytes (32 - 3 + 1) to check all positions without gaps
+    for (; i + 32 <= end; i += 30) {
         __m256i d0 = _mm256_loadu_si256((const __m256i*)(data + i));
         __m256i d1 = _mm256_loadu_si256((const __m256i*)(data + i + 1));
         __m256i d2 = _mm256_loadu_si256((const __m256i*)(data + i + 2));
@@ -154,7 +154,7 @@ int find_magic_numbers_avx2(const unsigned char* data, size_t off, size_t len) {
 }
 
 // Optimized AVX512 implementation for 3-byte pattern search
-// Checks 62 positions per iteration by advancing 1 byte at a time
+// Checks 62 positions per iteration by advancing 61 bytes at a time
 // Pattern bytes are broadcast once outside the loop for better performance
 int find_magic_numbers_avx512(const unsigned char* data, size_t off, size_t len) {
     if (len < 3) return -1;
@@ -169,8 +169,8 @@ int find_magic_numbers_avx512(const unsigned char* data, size_t off, size_t len)
     __m512i p2 = _mm512_set1_epi8(pattern[2]);
 
     // Process 64 bytes per loop, checking 62 positions per iteration
-    // Advance by 1 byte to check every possible starting position
-    for (; i + 64 <= end; i += 1) {
+    // Advance by 62 bytes (64 - 3 + 1) to check all positions without gaps
+    for (; i + 64 <= end; i += 62) {
         __m512i d0 = _mm512_loadu_si512((const void*)(data + i));
         __m512i d1 = _mm512_loadu_si512((const void*)(data + i + 1));
         __m512i d2 = _mm512_loadu_si512((const void*)(data + i + 2));