modify chunk size

src/simd.rs

@@ -1,11 +1,11 @@
 use std::simd::Simd;
 
 pub fn sum_f32_vector_simd(a: &[f32]) -> f32 {
-    let chunk_size = 4;
+    let chunk_size = 16;
     let chunks = a.len() / chunk_size;
-    let mut sum = Simd::<f32, 4>::splat(0.0);
+    let mut sum = Simd::<f32, 16>::splat(0.0);
     for i in 0..chunks {
-        let a = Simd::<f32, 4>::from_slice(&a[i * chunk_size..]);
+        let a = Simd::<f32, 16>::from_slice(&a[i * chunk_size..]);
         sum += a;
     }
     let mut result = sum.as_array().iter().sum::<f32>();
@@ -21,11 +21,11 @@ pub fn add_f32_vector_simd(a: &[f32], b: &[f32]) -> Vec<f32> {
     if a.is_empty() {
         return result;
     }
-    let chunk_size = 4;
+    let chunk_size = 16;
     let chunks = a.len() / chunk_size;
     for i in 0..chunks {
-        let a = Simd::<f32, 4>::from_slice(&a[i * chunk_size..]);
-        let b = Simd::<f32, 4>::from_slice(&b[i * chunk_size..]);
+        let a = Simd::<f32, 16>::from_slice(&a[i * chunk_size..]);
+        let b = Simd::<f32, 16>::from_slice(&b[i * chunk_size..]);
         let c = a + b;
         result.extend_from_slice(c.as_array());
     }
@@ -41,11 +41,11 @@ pub fn sub_f32_vector_simd(a: &[f32], b: &[f32]) -> Vec<f32> {
     if a.is_empty() {
         return result;
     }
-    let chunk_size = 4;
+    let chunk_size = 16;
     let chunks = a.len() / chunk_size;
     for i in 0..chunks {
-        let a = Simd::<f32, 4>::from_slice(&a[i * chunk_size..]);
-        let b = Simd::<f32, 4>::from_slice(&b[i * chunk_size..]);
+        let a = Simd::<f32, 16>::from_slice(&a[i * chunk_size..]);
+        let b = Simd::<f32, 16>::from_slice(&b[i * chunk_size..]);
         let c = a - b;
         result.extend_from_slice(c.as_array());
     }
@@ -61,11 +61,11 @@ pub fn mul_f32_tensor_simd(a: &[f32], b: &[f32]) -> Vec<f32> {
     if a.is_empty() {
         return result;
     }
-    let chunk_size = 4;
+    let chunk_size = 16;
     let chunks = a.len() / chunk_size;
     for i in 0..chunks {
-        let a = Simd::<f32, 4>::from_slice(&a[i * chunk_size..]);
-        let b = Simd::<f32, 4>::from_slice(&b[i * chunk_size..]);
+        let a = Simd::<f32, 16>::from_slice(&a[i * chunk_size..]);
+        let b = Simd::<f32, 16>::from_slice(&b[i * chunk_size..]);
         let c = a * b;
         result.extend_from_slice(c.as_array());
     }
@@ -81,11 +81,11 @@ pub fn div_f32_tensor_simd(a: &[f32], b: &[f32]) -> Vec<f32> {
     if a.is_empty() {
         return result;
     }
-    let chunk_size = 4;
+    let chunk_size = 16;
     let chunks = a.len() / chunk_size;
     for i in 0..chunks {
-        let a = Simd::<f32, 4>::from_slice(&a[i * chunk_size..]);
-        let b = Simd::<f32, 4>::from_slice(&b[i * chunk_size..]);
+        let a = Simd::<f32, 16>::from_slice(&a[i * chunk_size..]);
+        let b = Simd::<f32, 16>::from_slice(&b[i * chunk_size..]);
         let c = a / b;
         result.extend_from_slice(c.as_array());
     }
@@ -110,7 +110,7 @@ pub fn matmul_f32_tensor_simd(
     let n = b_shape[1];
     let k = a_shape[1];
 
-    let chunk_size = 4;
+    let chunk_size = 16;
     let chunks = k / chunk_size;
 
     let mut result = vec![0.0; m * n];
@@ -121,10 +121,10 @@ pub fn matmul_f32_tensor_simd(
         }
         for i in 0..m {
             let arow = &a[i * k..(i + 1) * k];
-            let mut sum = Simd::<f32, 4>::splat(0.0);
+            let mut sum = Simd::<f32, 16>::splat(0.0);
             for chunk in 0..chunks {
-                let a = Simd::<f32, 4>::from_slice(&arow[chunk * chunk_size..]);
-                let b = Simd::<f32, 4>::from_slice(&bcol[chunk * chunk_size..]);
+                let a = Simd::<f32, 16>::from_slice(&arow[chunk * chunk_size..]);
+                let b = Simd::<f32, 16>::from_slice(&bcol[chunk * chunk_size..]);
                 sum += a * b;
             }
             let mut res = sum.as_array().iter().sum::<f32>();