Merge pull request #2897 from Qiyu8/usimd-double

Add double precision universal intrinsics for X86/ARM

kernel/arm/sum.c

@@ -43,6 +43,26 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
 	if (inc_x == 1)
 	{
 #if V_SIMD
+#ifdef DOUBLE
+		const int vstep = v_nlanes_f64;
+		const int unrollx2 = n & (-vstep * 2);
+		const int unrollx = n & -vstep;
+		v_f64 vsum0 = v_zero_f64();
+		v_f64 vsum1 = v_zero_f64();
+		while (i < unrollx2)
+		{
+			vsum0 = v_add_f64(vsum0, v_loadu_f64(x));
+			vsum1 = v_add_f64(vsum1, v_loadu_f64(x + vstep));
+			i += vstep * 2;
+		}
+		vsum0 = v_add_f64(vsum0, vsum1);
+		while (i < unrollx)
+		{
+			vsum0 = v_add_f64(vsum0, v_loadu_f64(x + i));
+			i += vstep;
+		}
+		sumf = v_sum_f64(vsum0);
+#else
 		const int vstep = v_nlanes_f32;
 		const int unrollx4 = n & (-vstep * 4);
 		const int unrollx = n & -vstep;
@@ -66,6 +86,7 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
 			i += vstep;
 		}
 		sumf = v_sum_f32(vsum0);
+#endif
 #else
 		int n1 = n & -4;
 		for (; i < n1; i += 4)

kernel/simd/intrin_avx.h

@@ -4,20 +4,27 @@
  * Data Type
  ***************************/
 typedef __m256  v_f32;
+typedef __m256d  v_f64;
 #define v_nlanes_f32 8
+#define v_nlanes_f64 4
 /***************************
  * Arithmetic
  ***************************/
 #define v_add_f32 _mm256_add_ps
+#define v_add_f64 _mm256_add_pd
 #define v_mul_f32 _mm256_mul_ps
+#define v_mul_f64 _mm256_mul_pd
 
 #ifdef HAVE_FMA3
     // multiply and add, a*b + c
     #define v_muladd_f32 _mm256_fmadd_ps
+    #define v_muladd_f64 _mm256_fmadd_pd
 #else
     // multiply and add, a*b + c
     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); }
 #endif // !HAVE_FMA3
 
 // Horizontal add: Calculates the sum of all vector elements.
@@ -31,11 +38,23 @@ BLAS_FINLINE float v_sum_f32(__m256 a)
     return _mm_cvtss_f32(sum);
 }
 
+BLAS_FINLINE double v_sum_f64(__m256d a)
+{
+    __m256d sum_halves = _mm256_hadd_pd(a, a);
+    __m128d lo = _mm256_castpd256_pd128(sum_halves);
+    __m128d hi = _mm256_extractf128_pd(sum_halves, 1);
+    __m128d sum = _mm_add_pd(lo, hi);
+    return _mm_cvtsd_f64(sum);
+}
 /***************************
  * memory
  ***************************/
 // unaligned load
 #define v_loadu_f32 _mm256_loadu_ps
+#define v_loadu_f64 _mm256_loadu_pd
 #define v_storeu_f32 _mm256_storeu_ps
+#define v_storeu_f64 _mm256_storeu_pd
 #define v_setall_f32(VAL) _mm256_set1_ps(VAL)
-#define v_zero_f32 _mm256_setzero_ps
+#define v_setall_f64(VAL) _mm256_set1_pd(VAL)
+#define v_zero_f32 _mm256_setzero_ps
+#define v_zero_f64 _mm256_setzero_pd

kernel/simd/intrin_avx512.h

@@ -4,15 +4,19 @@
  * Data Type
  ***************************/
 typedef __m512  v_f32;
+typedef __m512d  v_f64;
 #define v_nlanes_f32 16
+#define v_nlanes_f64 8
 /***************************
  * Arithmetic
  ***************************/
 #define v_add_f32 _mm512_add_ps
+#define v_add_f64 _mm512_add_pd
 #define v_mul_f32 _mm512_mul_ps
+#define v_mul_f64 _mm512_mul_pd
 // multiply and add, a*b + c
 #define v_muladd_f32 _mm512_fmadd_ps
-
+#define v_muladd_f64 _mm512_fmadd_pd
 BLAS_FINLINE float v_sum_f32(v_f32 a)
 {
     __m512 h64 = _mm512_shuffle_f32x4(a, a, _MM_SHUFFLE(3, 2, 3, 2));
@@ -25,11 +29,26 @@ BLAS_FINLINE float v_sum_f32(v_f32 a)
     __m512 sum4 = _mm512_add_ps(sum8, h4);
     return _mm_cvtss_f32(_mm512_castps512_ps128(sum4));
 }
+
+BLAS_FINLINE double v_sum_f64(v_f64 a)
+{
+    __m512d h64   = _mm512_shuffle_f64x2(a, a, _MM_SHUFFLE(3, 2, 3, 2));
+    __m512d sum32 = _mm512_add_pd(a, h64);
+    __m512d h32   = _mm512_permutex_pd(sum32, _MM_SHUFFLE(1, 0, 3, 2));
+    __m512d sum16 = _mm512_add_pd(sum32, h32);
+    __m512d h16   = _mm512_permute_pd(sum16, _MM_SHUFFLE(2, 3, 0, 1));
+    __m512d sum8  = _mm512_add_pd(sum16, h16);
+    return _mm_cvtsd_f64(_mm512_castpd512_pd128(sum8));
+}
 /***************************
  * memory
  ***************************/
 // unaligned load
 #define v_loadu_f32(PTR) _mm512_loadu_ps((const __m512*)(PTR))
+#define v_loadu_f64(PTR) _mm512_loadu_pd((const __m512*)(PTR))
 #define v_storeu_f32 _mm512_storeu_ps
+#define v_storeu_f64 _mm512_storeu_pd
 #define v_setall_f32(VAL) _mm512_set1_ps(VAL)
+#define v_setall_f64(VAL) _mm512_set1_pd(VAL)
 #define v_zero_f32 _mm512_setzero_ps
+#define v_zero_f64 _mm512_setzero_pd

kernel/simd/intrin_neon.h

@@ -8,12 +8,18 @@
  * Data Type
  ***************************/
 typedef float32x4_t v_f32;
+#if V_SIMD_F64
+    typedef float64x2_t v_f64;
+#endif
 #define v_nlanes_f32 4
+#define v_nlanes_f64 2
 /***************************
  * Arithmetic
  ***************************/
 #define v_add_f32 vaddq_f32
+#define v_add_f64 vaddq_f64
 #define v_mul_f32 vmulq_f32
+#define v_mul_f64 vmulq_f64
 
 // FUSED F32
 #ifdef HAVE_VFPV4 // FMA
@@ -26,12 +32,26 @@ typedef float32x4_t v_f32;
     { return vmlaq_f32(c, a, b); }
 #endif
 
+// FUSED F64
+#if V_SIMD_F64
+    BLAS_FINLINE v_f64 v_muladd_f64(v_f64 a, v_f64 b, v_f64 c)
+    { return vfmaq_f64(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);
 }
+
+#if V_SIMD_F64
+    BLAS_FINLINE double v_sum_f64(float64x2_t a)
+    {
+        return vget_lane_f64(vget_low_f64(a) + vget_high_f64(a), 0);
+    }
+#endif
+
 /***************************
  * memory
  ***************************/
@@ -39,4 +59,10 @@ BLAS_FINLINE float v_sum_f32(float32x4_t a)
 #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)
+#define v_zero_f32() vdupq_n_f32(0.0f)
+#if V_SIMD_F64
+    #define v_loadu_f64(a) vld1q_f64((const double*)a)
+    #define v_storeu_f64 vst1q_f64
+    #define v_setall_f64 vdupq_n_f64
+    #define v_zero_f64() vdupq_n_f64(0.0)
+#endif

kernel/simd/intrin_sse.h

@@ -4,22 +4,30 @@
  * Data Type
  ***************************/
 typedef __m128  v_f32;
+typedef __m128d  v_f64;
 #define v_nlanes_f32 4
+#define v_nlanes_f64 2
 /***************************
  * Arithmetic
  ***************************/
 #define v_add_f32 _mm_add_ps
+#define v_add_f64 _mm_add_pd
 #define v_mul_f32 _mm_mul_ps
+#define v_mul_f64 _mm_mul_pd
 #ifdef HAVE_FMA3
     // multiply and add, a*b + c
     #define v_muladd_f32 _mm_fmadd_ps
+    #define v_muladd_f64 _mm_fmadd_pd
 #elif defined(HAVE_FMA4)
     // multiply and add, a*b + c
     #define v_muladd_f32 _mm_macc_ps
+    #define v_muladd_f64 _mm_macc_pd
 #else
     // multiply and add, a*b + c
     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); }
 #endif // HAVE_FMA3
 
 // Horizontal add: Calculates the sum of all vector elements.
@@ -36,11 +44,24 @@ BLAS_FINLINE float v_sum_f32(__m128 a)
     return _mm_cvtss_f32(t4);
 #endif
 }
+
+BLAS_FINLINE double v_sum_f64(__m128d a)
+{
+#ifdef HAVE_SSE3
+    return _mm_cvtsd_f64(_mm_hadd_pd(a, a));
+#else
+    return _mm_cvtsd_f64(_mm_add_pd(a, _mm_unpackhi_pd(a, a)));
+#endif
+}
 /***************************
  * memory
  ***************************/
 // unaligned load
 #define v_loadu_f32 _mm_loadu_ps
+#define v_loadu_f64 _mm_loadu_pd
 #define v_storeu_f32 _mm_storeu_ps
+#define v_storeu_f64 _mm_storeu_pd
 #define v_setall_f32(VAL) _mm_set1_ps(VAL)
-#define v_zero_f32 _mm_setzero_ps
+#define v_setall_f64(VAL) _mm_set1_pd(VAL)
+#define v_zero_f32 _mm_setzero_ps
+#define v_zero_f64 _mm_setzero_pd

kernel/x86_64/daxpy.c

@@ -53,6 +53,15 @@ static void daxpy_kernel_8(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
 	BLASLONG register i = 0;
 	FLOAT a = *alpha;
 #if V_SIMD
+#ifdef DOUBLE
+	v_f64 __alpha, tmp;
+	__alpha =  v_setall_f64(*alpha);
+	const int vstep = v_nlanes_f64;
+	for (; i < n; i += vstep) {
+		tmp = v_muladd_f64(__alpha, v_loadu_f64( x + i ), v_loadu_f64(y + i));
+		v_storeu_f64(y + i, tmp);
+	}
+#else
 	v_f32 __alpha, tmp;
 	__alpha =  v_setall_f32(*alpha);
 	const int vstep = v_nlanes_f32;
@@ -60,6 +69,7 @@ static void daxpy_kernel_8(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
 		tmp = v_muladd_f32(__alpha, v_loadu_f32( x + i ), v_loadu_f32(y + i));
 		v_storeu_f32(y + i, tmp);
 	}
+#endif
 #else
 	while(i < n)
 	{