Refine WebAssembly SIMD backend scope

Author: teddygood

kernel/simd/intrin_wasm.h

@@ -33,6 +33,23 @@ BLAS_FINLINE v_f32 v_mulsub_f32(v_f32 a, v_f32 b, v_f32 c)
 BLAS_FINLINE v_f64 v_mulsub_f64(v_f64 a, v_f64 b, v_f64 c)
 { return v_sub_f64(v_mul_f64(a, b), c); }
 
+/***************************
+ * reduction
+ ***************************/
+BLAS_FINLINE float v_sum_f32(v_f32 a)
+{
+    return wasm_f32x4_extract_lane(a, 0)
+         + wasm_f32x4_extract_lane(a, 1)
+         + wasm_f32x4_extract_lane(a, 2)
+         + wasm_f32x4_extract_lane(a, 3);
+}
+
+BLAS_FINLINE double v_sum_f64(v_f64 a)
+{
+    return wasm_f64x2_extract_lane(a, 0)
+         + wasm_f64x2_extract_lane(a, 1);
+}
+
 /***************************
  * memory
  ***************************/

kernel/x86_64/drot.c

@@ -12,7 +12,7 @@
 static void drot_kernel(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT c, FLOAT s)
 {
     BLASLONG i = 0;
-#if V_SIMD_F64 && (V_SIMD > 256 || defined(ARCH_WASM))
+#if V_SIMD_F64 && V_SIMD > 256
     const int vstep = v_nlanes_f64;
     const int unrollx4 = n & (-vstep * 4);
     const int unrollx = n & -vstep;