Optimize the performance of dot by using universal intrinsics in X86/ARM

Qiyu8 · Qiyu8 · commit 1b1a757f5f38 · 2020-09-28T20:36:53.000+08:00
diff --git a/kernel/generic/dot.c b/kernel/generic/dot.c
@@ -47,27 +47,59 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
 
 	if ( (inc_x == 1) && (inc_y == 1) )
 	{
-
 		int n1 = n & -4;
-
-		while(i < n1)
+#if V_SIMD && !defined(DSDOT)
+		const int vstep = v_nlanes_f32;
+        const int unrollx4 = n & (-vstep * 4);
+        const int unrollx  = n &  -vstep;
+		v_f32 vsum0 = v_zero_f32();
+        v_f32 vsum1 = v_zero_f32();
+        v_f32 vsum2 = v_zero_f32();
+        v_f32 vsum3 = v_zero_f32();
+		while(i < unrollx4)
+        {
+            vsum0 = v_muladd_f32(
+                v_loadu_f32(x + i),           v_loadu_f32(y + i),           vsum0
+            );
+            vsum1 = v_muladd_f32(
+                v_loadu_f32(x + i + vstep),   v_loadu_f32(y + i + vstep),   vsum1
+            );
+            vsum2 = v_muladd_f32(
+                v_loadu_f32(x + i + vstep*2), v_loadu_f32(y + i + vstep*2), vsum2
+            );
+            vsum3 = v_muladd_f32(
+                v_loadu_f32(x + i + vstep*3), v_loadu_f32(y + i + vstep*3), vsum3
+            );
+            i += vstep*4;
+        }
+        vsum0 = v_add_f32(
+            v_add_f32(vsum0, vsum1), v_add_f32(vsum2 , vsum3)
+        );
+		while(i < unrollx)
+        {
+            vsum0 = v_muladd_f32(
+                v_loadu_f32(x + i), v_loadu_f32(y + i), vsum0
+            );
+            i += vstep;
+        }
+        dot = v_sum_f32(vsum0);
+#elif defined(DSDOT)
+		for (; i < n1; i += 4)
 		{
-
-#if defined(DSDOT)
 			dot += (double) y[i] * (double) x[i]
 			    + (double) y[i+1] * (double) x[i+1]
 			    + (double) y[i+2] * (double) x[i+2]
 			    + (double) y[i+3] * (double) x[i+3] ;
+		}
 #else
+		for (; i < n1; i += 4)
+		{
 			dot += y[i] * x[i]
 			    + y[i+1] * x[i+1]
 			    + y[i+2] * x[i+2]
 			    + y[i+3] * x[i+3] ;
-#endif
-			i+=4 ;
-
 		}
-
+#endif
 		while(i < n)
 		{
 
diff --git a/kernel/simd/intrin.h b/kernel/simd/intrin.h
@@ -51,6 +51,11 @@ extern "C" {
 #include <immintrin.h>
 #endif
 
+/** NEON **/
+#ifdef HAVE_NEON
+#include <arm_neon.h>
+#endif
+
 // distribute
 #if defined(HAVE_AVX512VL) || defined(HAVE_AVX512BF16)
 #include "intrin_avx512.h"
@@ -60,6 +65,10 @@ extern "C" {
 #include "intrin_sse.h"
 #endif
 
+#ifdef HAVE_NEON
+#include "intrin_neon.h"
+#endif
+
 #ifndef V_SIMD
     #define V_SIMD 0
     #define V_SIMD_F64 0
diff --git a/kernel/simd/intrin_avx.h b/kernel/simd/intrin_avx.h
@@ -1,13 +1,13 @@
 #define V_SIMD 256
 #define V_SIMD_F64 1
-/*
-Data Type
-*/
+/***************************
+ * Data Type
+ ***************************/
 typedef __m256  v_f32;
 #define v_nlanes_f32 8
-/*
-arithmetic
-*/
+/***************************
+ * Arithmetic
+ ***************************/
 #define v_add_f32 _mm256_add_ps
 #define v_mul_f32 _mm256_mul_ps
 
@@ -20,10 +20,22 @@ arithmetic
     { return v_add_f32(v_mul_f32(a, b), c); }
 #endif // !HAVE_FMA3
 
-/*
-memory
-*/
+// Horizontal add: Calculates the sum of all vector elements.
+BLAS_FINLINE float v_sum_f32(__m256 a)
+{
+    __m256 sum_halves = _mm256_hadd_ps(a, a);
+    sum_halves = _mm256_hadd_ps(sum_halves, sum_halves);
+    __m128 lo = _mm256_castps256_ps128(sum_halves);
+    __m128 hi = _mm256_extractf128_ps(sum_halves, 1);
+    __m128 sum = _mm_add_ps(lo, hi);
+    return _mm_cvtss_f32(sum);
+}
+
+/***************************
+ * memory
+ ***************************/
 // unaligned load
 #define v_loadu_f32 _mm256_loadu_ps
 #define v_storeu_f32 _mm256_storeu_ps
-#define v_setall_f32(VAL) _mm256_set1_ps(VAL)
+#define v_setall_f32(VAL) _mm256_set1_ps(VAL)
+#define v_zero_f32 _mm256_setzero_ps
diff --git a/kernel/simd/intrin_avx512.h b/kernel/simd/intrin_avx512.h
@@ -1,21 +1,35 @@
 #define V_SIMD 512
 #define V_SIMD_F64 1
-/*
-Data Type
-*/
+/***************************
+ * Data Type
+ ***************************/
 typedef __m512  v_f32;
 #define v_nlanes_f32 16
-/*
-arithmetic
-*/
+/***************************
+ * Arithmetic
+ ***************************/
 #define v_add_f32 _mm512_add_ps
 #define v_mul_f32 _mm512_mul_ps
 // multiply and add, a*b + c
 #define v_muladd_f32 _mm512_fmadd_ps
-/*
-memory
-*/
+
+BLAS_FINLINE float v_sum_f32(v_f32 a)
+{
+    __m512 h64 = _mm512_shuffle_f32x4(a, a, _MM_SHUFFLE(3, 2, 3, 2));
+    __m512 sum32 = _mm512_add_ps(a, h64);
+    __m512 h32 = _mm512_shuffle_f32x4(sum32, sum32, _MM_SHUFFLE(1, 0, 3, 2));
+    __m512 sum16 = _mm512_add_ps(sum32, h32);
+    __m512 h16 = _mm512_permute_ps(sum16, _MM_SHUFFLE(1, 0, 3, 2));
+    __m512 sum8 = _mm512_add_ps(sum16, h16);
+    __m512 h4 = _mm512_permute_ps(sum8, _MM_SHUFFLE(2, 3, 0, 1));
+    __m512 sum4 = _mm512_add_ps(sum8, h4);
+    return _mm_cvtss_f32(_mm512_castps512_ps128(sum4));
+}
+/***************************
+ * memory
+ ***************************/
 // unaligned load
 #define v_loadu_f32(PTR) _mm512_loadu_ps((const __m512*)(PTR))
 #define v_storeu_f32 _mm512_storeu_ps
 #define v_setall_f32(VAL) _mm512_set1_ps(VAL)
+#define v_zero_f32 _mm512_setzero_ps
diff --git a/kernel/simd/intrin_neon.h b/kernel/simd/intrin_neon.h
@@ -0,0 +1,42 @@
+#define V_SIMD 128
+#ifdef __aarch64__
+    #define V_SIMD_F64 1
+#else
+    #define V_SIMD_F64 0
+#endif
+/***************************
+ * Data Type
+ ***************************/
+typedef float32x4_t v_f32;
+#define v_nlanes_f32 4
+/***************************
+ * Arithmetic
+ ***************************/
+#define v_add_f32 vaddq_f32
+#define v_mul_f32 vmulq_f32
+
+// FUSED F32
+#ifdef HAVE_VFPV4 // FMA
+    // multiply and add, a*b + c
+    BLAS_FINLINE v_f32 v_muladd_f32(v_f32 a, v_f32 b, v_f32 c)
+    { return vfmaq_f32(c, a, b); }
+#else
+    // multiply and add, a*b + c
+    BLAS_FINLINE v_f32 v_muladd_f32(v_f32 a, v_f32 b, v_f32 c)
+    { return vmlaq_f32(c, a, b); }
+#endif
+
+// Horizontal add: Calculates the sum of all vector elements.
+BLAS_FINLINE float v_sum_f32(float32x4_t a)
+{
+    float32x2_t r = vadd_f32(vget_high_f32(a), vget_low_f32(a));
+    return vget_lane_f32(vpadd_f32(r, r), 0);
+}
+/***************************
+ * memory
+ ***************************/
+// unaligned load
+#define v_loadu_f32(a) vld1q_f32((const float*)a)
+#define v_storeu_f32 vst1q_f32
+#define v_setall_f32(VAL) vdupq_n_f32(VAL)
+#define v_zero_f32() vdupq_n_f32(0.0f)
diff --git a/kernel/simd/intrin_sse.h b/kernel/simd/intrin_sse.h
@@ -1,13 +1,13 @@
 #define V_SIMD 128
 #define V_SIMD_F64 1
-/*
-Data Type
-*/
+/***************************
+ * Data Type
+ ***************************/
 typedef __m128  v_f32;
 #define v_nlanes_f32 4
-/*
-arithmetic
-*/
+/***************************
+ * Arithmetic
+ ***************************/
 #define v_add_f32 _mm_add_ps
 #define v_mul_f32 _mm_mul_ps
 #ifdef HAVE_FMA3
@@ -21,10 +21,26 @@ arithmetic
     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); }
 #endif // HAVE_FMA3
-/*
-memory
-*/
+
+// Horizontal add: Calculates the sum of all vector elements.
+BLAS_FINLINE float v_sum_f32(__m128 a)
+{
+#ifdef HAVE_SSE3
+    __m128 sum_halves = _mm_hadd_ps(a, a);
+    return _mm_cvtss_f32(_mm_hadd_ps(sum_halves, sum_halves));
+#else
+    __m128 t1 = _mm_movehl_ps(a, a);
+    __m128 t2 = _mm_add_ps(a, t1);
+    __m128 t3 = _mm_shuffle_ps(t2, t2, 1);
+    __m128 t4 = _mm_add_ss(t2, t3);
+    return _mm_cvtss_f32(t4);
+#endif
+}
+/***************************
+ * memory
+ ***************************/
 // unaligned load
 #define v_loadu_f32 _mm_loadu_ps
 #define v_storeu_f32 _mm_storeu_ps
-#define v_setall_f32(VAL) _mm_set1_ps(VAL)
+#define v_setall_f32(VAL) _mm_set1_ps(VAL)
+#define v_zero_f32 _mm_setzero_ps
diff --git a/utest/test_dsdot.c b/utest/test_dsdot.c
@@ -47,3 +47,17 @@ CTEST(dsdot,dsdot_n_1)
 	ASSERT_DBL_NEAR_TOL(res2, res1, DOUBLE_EPS);
 
 }
+
+CTEST(dsdot,dsdot_n_2)
+{
+	float x[] = {0.1F, 0.2F, 0.3F, 0.4F, 0.5F, 0.6F, 0.7F, 0.8F};
+	float y[] = {0.1F, 0.2F, 0.3F, 0.4F, 0.5F, 0.6F, 0.7F, 0.8F};
+	blasint incx=1;
+	blasint incy=1;
+	blasint n=8;
+
+	double res1=0.0f, res2= 2.0400000444054616;
+
+	res1=BLASFUNC(dsdot)(&n, &x, &incx, &y, &incy);
+	ASSERT_DBL_NEAR_TOL(res2, res1, DOUBLE_EPS);
+}
