FEAT: In dgemm sse2 and fallback, use a 4x4 kernel

bluss · Dec 5, 2018 · 28e7007 · 28e7007
1 parent fd6294a
commit 28e7007
Showing 1 changed file with 6 additions and 6 deletions.
diff --git a/src/dgemm_kernel.rs b/src/dgemm_kernel.rs
@@ -92,8 +92,8 @@ impl GemmKernel for KernelAvx {
 
 impl GemmKernel for KernelSse2 {
     type Elem = T;
-    const MR: usize = MR;
-    const NR: usize = NR;
+    const MR: usize = 4;
+    const NR: usize = 4;
 
     #[inline(always)]
     fn align_to() -> usize { 16 }
@@ -130,8 +130,8 @@ impl GemmKernel for KernelSse2 {
 
 impl GemmKernel for KernelFallback {
     type Elem = T;
-    const MR: usize = MR;
-    const NR: usize = NR;
+    const MR: usize = 4;
+    const NR: usize = 4;
 
     #[inline(always)]
     fn align_to() -> usize { 0 }
@@ -752,7 +752,7 @@ unsafe fn kernel_fallback_impl(k: usize, alpha: T, a: *const T, b: *const T,
 
     // Compute matrix multiplication into ab[i][j]
     unroll_by!(4 => k, {
-        loop_m!(i, loop_n!(j, ab[i][j] += at(a, i) * at(b, j)));
+        loop4!(i, loop4!(j, ab[i][j] += at(a, i) * at(b, j)));
 
         a = a.offset(MR as isize);
         b = b.offset(NR as isize);
@@ -763,7 +763,7 @@ unsafe fn kernel_fallback_impl(k: usize, alpha: T, a: *const T, b: *const T,
     }
 
     // set C = α A B + β C
-    loop_n!(j, loop_m!(i, *c![i, j] = alpha * ab[i][j]));
+    loop4!(j, loop4!(i, *c![i, j] = alpha * ab[i][j]));
 }
 
 #[inline(always)]