Add WebAssembly SIMD backend for universal intrinsics

Author: teddygood

kernel/simd/intrin.h

@@ -56,6 +56,11 @@ extern "C" {
 #include <arm_neon.h>
 #endif
 
+/** WASM SIMD **/
+#if defined(ARCH_WASM) && defined(__wasm_simd128__)
+#include <wasm_simd128.h>
+#endif
+
 // distribute
 #if defined(HAVE_AVX512VL) || defined(HAVE_AVX512BF16)
 #include "intrin_avx512.h"
@@ -69,6 +74,10 @@ extern "C" {
 #include "intrin_neon.h"
 #endif
 
+#if defined(ARCH_WASM) && defined(__wasm_simd128__)
+#include "intrin_wasm.h"
+#endif
+
 #ifndef V_SIMD
     #define V_SIMD 0
     #define V_SIMD_F64 0

kernel/simd/intrin_wasm.h

@@ -0,0 +1,46 @@
+#include <wasm_simd128.h>
+
+#define V_SIMD 128
+#define V_SIMD_F64 1
+
+/***************************
+ * Data Type
+ ***************************/
+typedef v128_t v_f32;
+typedef v128_t v_f64;
+#define v_nlanes_f32 4
+#define v_nlanes_f64 2
+
+/***************************
+ * Arithmetic
+ ***************************/
+#define v_add_f32 wasm_f32x4_add
+#define v_add_f64 wasm_f64x2_add
+#define v_sub_f32 wasm_f32x4_sub
+#define v_sub_f64 wasm_f64x2_sub
+#define v_mul_f32 wasm_f32x4_mul
+#define v_mul_f64 wasm_f64x2_mul
+
+BLAS_FINLINE v_f32 v_muladd_f32(v_f32 a, v_f32 b, v_f32 c)
+{ return v_add_f32(v_mul_f32(a, b), c); }
+
+BLAS_FINLINE v_f64 v_muladd_f64(v_f64 a, v_f64 b, v_f64 c)
+{ return v_add_f64(v_mul_f64(a, b), c); }
+
+BLAS_FINLINE v_f32 v_mulsub_f32(v_f32 a, v_f32 b, v_f32 c)
+{ return v_sub_f32(v_mul_f32(a, b), 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); }
+
+/***************************
+ * memory
+ ***************************/
+#define v_loadu_f32(a) wasm_v128_load((const float*)a)
+#define v_loadu_f64(a) wasm_v128_load((const double*)a)
+#define v_storeu_f32(a, v) wasm_v128_store((float*)a, v)
+#define v_storeu_f64(a, v) wasm_v128_store((double*)a, v)
+#define v_setall_f32(VAL) wasm_f32x4_splat(VAL)
+#define v_setall_f64(VAL) wasm_f64x2_splat(VAL)
+#define v_zero_f32() wasm_f32x4_splat(0.0f)
+#define v_zero_f64() wasm_f64x2_splat(0.0)

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
+#if V_SIMD_F64 && (V_SIMD > 256 || defined(ARCH_WASM))
     const int vstep = v_nlanes_f64;
     const int unrollx4 = n & (-vstep * 4);
     const int unrollx = n & -vstep;

kernel/x86_64/srot.c

@@ -13,7 +13,7 @@ static void srot_kernel(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT c, FLOAT s)
 {
     BLASLONG i = 0;
 
-#if V_SIMD && !defined(C_PGI) && (defined(HAVE_FMA3) || V_SIMD > 128)
+#if V_SIMD && !defined(C_PGI) && (defined(HAVE_FMA3) || V_SIMD > 128 || defined(ARCH_WASM))
     const int vstep = v_nlanes_f32;
     const int unrollx4 = n & (-vstep * 4);
     const int unrollx = n & -vstep;