Merge pull request #38 from AdaWorldAPI/claude/transcode-deepnsm-rust-oNa1Z

feat(hpc): LazyLock frozen SIMD dispatch table — detect once, keep CPU choice forever

Author: AdaWorldAPI

src/hpc/byte_scan.rs

@@ -9,15 +9,15 @@
 // ---------------------------------------------------------------------------
 
 #[cfg(target_arch = "x86_64")]
-mod simd_impl {
+pub(crate) mod simd_impl {
     use core::arch::x86_64::*;
 
     /// Find all positions of `needle` in `haystack` using AVX2 (32 bytes/iter).
     ///
     /// # Safety
     /// Caller must ensure AVX2 is available.
     #[target_feature(enable = "avx2")]
-    pub(super) unsafe fn byte_find_all_avx2(haystack: &[u8], needle: u8) -> Vec<usize> {
+    pub(crate) unsafe fn byte_find_all_avx2(haystack: &[u8], needle: u8) -> Vec<usize> {
         let mut result = Vec::new();
         let n = haystack.len();
         let ptr = haystack.as_ptr();
@@ -52,7 +52,7 @@ mod simd_impl {
     /// # Safety
     /// Caller must ensure AVX-512 BW is available.
     #[target_feature(enable = "avx512bw")]
-    pub(super) unsafe fn byte_find_all_avx512(haystack: &[u8], needle: u8) -> Vec<usize> {
+    pub(crate) unsafe fn byte_find_all_avx512(haystack: &[u8], needle: u8) -> Vec<usize> {
         let mut result = Vec::new();
         let n = haystack.len();
         let ptr = haystack.as_ptr();
@@ -84,7 +84,7 @@ mod simd_impl {
     /// # Safety
     /// Caller must ensure AVX2 is available.
     #[target_feature(enable = "avx2")]
-    pub(super) unsafe fn byte_count_avx2(haystack: &[u8], needle: u8) -> usize {
+    pub(crate) unsafe fn byte_count_avx2(haystack: &[u8], needle: u8) -> usize {
         let n = haystack.len();
         let ptr = haystack.as_ptr();
         let needle_v = _mm256_set1_epi8(needle as i8);
@@ -111,7 +111,7 @@ mod simd_impl {
     /// # Safety
     /// Caller must ensure AVX-512 BW is available.
     #[target_feature(enable = "avx512bw")]
-    pub(super) unsafe fn byte_count_avx512(haystack: &[u8], needle: u8) -> usize {
+    pub(crate) unsafe fn byte_count_avx512(haystack: &[u8], needle: u8) -> usize {
         let n = haystack.len();
         let ptr = haystack.as_ptr();
         let needle_v = _mm512_set1_epi8(needle as i8);

src/hpc/distance.rs

@@ -29,7 +29,7 @@ fn sq_dist_f64(a: [f64; 3], b: [f64; 3]) -> f64 {
 // ---------------------------------------------------------------------------
 
 #[cfg(target_arch = "x86_64")]
-mod simd_impl {
+pub(crate) mod simd_impl {
     #[cfg(target_arch = "x86_64")]
     use core::arch::x86_64::*;
 
@@ -39,7 +39,7 @@ mod simd_impl {
     /// # Safety
     /// Caller must ensure AVX2 is available.
     #[target_feature(enable = "avx2")]
-    pub(super) unsafe fn squared_distances_avx2(
+    pub(crate) unsafe fn squared_distances_avx2(
         query: [f32; 3],
         points: &[[f32; 3]],
         out: &mut Vec<f32>,

src/hpc/mod.rs

@@ -21,6 +21,8 @@
 
 // SIMD capability singleton — detect once, all modules share
 pub mod simd_caps;
+// LazyLock frozen SIMD dispatch — function pointers selected once at startup
+pub mod simd_dispatch;
 
 pub mod blas_level1;
 pub mod blas_level2;

src/hpc/nibble.rs

@@ -40,7 +40,7 @@ pub fn nibble_unpack(packed: &[u8], count: usize) -> Vec<u8> {
     out
 }
 
-fn nibble_unpack_scalar(packed: &[u8], count: usize, out: &mut Vec<u8>) {
+pub(crate) fn nibble_unpack_scalar(packed: &[u8], count: usize, out: &mut Vec<u8>) {
     for i in 0..count {
         let byte = packed[i / 2];
         let val = if i & 1 == 0 { byte & 0x0F } else { byte >> 4 };
@@ -54,7 +54,7 @@ fn nibble_unpack_scalar(packed: &[u8], count: usize, out: &mut Vec<u8>) {
 /// Caller must ensure AVX2 is available and `count >= 32`.
 #[cfg(target_arch = "x86_64")]
 #[target_feature(enable = "avx2")]
-unsafe fn nibble_unpack_avx2(packed: &[u8], count: usize, out: &mut Vec<u8>) {
+pub(crate) unsafe fn nibble_unpack_avx2(packed: &[u8], count: usize, out: &mut Vec<u8>) {
     use core::arch::x86_64::*;
 
     let low_mask = _mm_set1_epi8(0x0F);
@@ -252,7 +252,7 @@ pub fn nibble_above_threshold(packed: &[u8], threshold: u8) -> Vec<usize> {
     nibble_above_threshold_scalar(packed, threshold)
 }
 
-fn nibble_above_threshold_scalar(packed: &[u8], threshold: u8) -> Vec<usize> {
+pub(crate) fn nibble_above_threshold_scalar(packed: &[u8], threshold: u8) -> Vec<usize> {
     let mut result = Vec::new();
     let count = packed.len() * 2;
     for i in 0..count {
@@ -272,7 +272,7 @@ fn nibble_above_threshold_scalar(packed: &[u8], threshold: u8) -> Vec<usize> {
 /// Caller must ensure AVX2 is available and `packed.len() >= 16`.
 #[cfg(target_arch = "x86_64")]
 #[target_feature(enable = "avx2")]
-unsafe fn nibble_above_threshold_avx2(packed: &[u8], threshold: u8) -> Vec<usize> {
+pub(crate) unsafe fn nibble_above_threshold_avx2(packed: &[u8], threshold: u8) -> Vec<usize> {
     use core::arch::x86_64::*;
 
     let mut result = Vec::new();