Merge pull request #5685 from teddygood/wasm-intrin-backend-exp

Add a WebAssembly SIMD backend for reusable intrinsics kernels

Author: martin-frbg

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 +1,63 @@
+#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); }
+
+/***************************
+ * 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
+ ***************************/
+#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/wasm/KERNEL.WASM128_GENERIC

@@ -40,7 +40,7 @@ DSUMKERNEL  = ../arm/sum.c
 CSUMKERNEL  = ../arm/zsum.c
 ZSUMKERNEL  = ../arm/zsum.c
 
-SAXPYKERNEL  = ../riscv64/axpy.c
+SAXPYKERNEL  = ../x86_64/saxpy.c
 DAXPYKERNEL  = ../x86_64/daxpy.c
 CAXPYKERNEL  = ../riscv64/zaxpy.c
 ZAXPYKERNEL  = ../riscv64/zaxpy.c

kernel/x86_64/saxpy.c

@@ -43,12 +43,23 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 
 #ifndef HAVE_KERNEL_16
+#include"../simd/intrin.h"
 
 static void saxpy_kernel_16(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
 {
 	BLASLONG register i = 0;
 	FLOAT a = *alpha;
 
+#if V_SIMD
+	v_f32 __alpha, tmp;
+	__alpha = v_setall_f32(*alpha);
+	const int vstep = v_nlanes_f32;
+
+	for (; i < n; i += vstep) {
+		tmp = v_muladd_f32(__alpha, v_loadu_f32(x + i), v_loadu_f32(y + i));
+		v_storeu_f32(y + i, tmp);
+	}
+#else
 	while(i < n)
         {
               y[i]   += a * x[i];
@@ -62,6 +73,7 @@ static void saxpy_kernel_16(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
               i+=8 ;
 
        }
+#endif
 
 }
 
@@ -131,4 +143,3 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x, BLAS
 
 }
 
-

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;