Implementaion of dasum, sasum with AVX2 & AVX512 intrinsic

kernel/x86_64/KERNEL.HASWELL

@@ -100,3 +100,5 @@ ZTRSMKERNEL_RT	= ../generic/trsm_kernel_RT.c
 CGEMM3MKERNEL    =  cgemm3m_kernel_8x4_haswell.c
 ZGEMM3MKERNEL    =  zgemm3m_kernel_4x4_haswell.c
 
+SASUMKERNEL = sasum.c
+DASUMKERNEL = dasum.c

kernel/x86_64/dasum.c

@@ -0,0 +1,96 @@
+#include "common.h"
+#include <math.h>
+
+#define ABS fabs
+
+#if defined(SKYLAKEX)
+#include "dasum_microk_skylakex-2.c"
+#elif defined(HASWELL)
+#include "dasum_microk_haswell-2.c"
+#endif
+
+#ifndef HAVE_KERNEL_16
+static FLOAT dasum_kernel_16(BLASLONG n, FLOAT *x1)
+{
+
+	BLASLONG i=0;
+	FLOAT *x = x1;
+	FLOAT temp0, temp1, temp2, temp3;
+	FLOAT temp4, temp5, temp6, temp7;
+	FLOAT sum0 = 0.0;
+	FLOAT sum1 = 0.0;
+	FLOAT sum2 = 0.0;
+	FLOAT sum3 = 0.0;
+
+	while ( i< n )
+	{
+
+		temp0 = ABS(x[0]);
+		temp1 = ABS(x[1]);
+		temp2 = ABS(x[2]);
+		temp3 = ABS(x[3]);
+		temp4 = ABS(x[4]);
+		temp5 = ABS(x[5]);
+		temp6 = ABS(x[6]);
+		temp7 = ABS(x[7]);
+
+		sum0 += temp0;
+		sum1 += temp1;
+		sum2 += temp2;
+		sum3 += temp3;
+
+		sum0 += temp4;
+		sum1 += temp5;
+		sum2 += temp6;
+		sum3 += temp7;
+
+		x+=8;
+		i+=8;
+
+	}
+
+	return sum0+sum1+sum2+sum3;
+}
+
+#endif
+
+FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
+{
+	BLASLONG i=0;
+	FLOAT sumf = 0.0;
+	BLASLONG n1;
+
+	if (n <= 0 || inc_x <= 0) return(sumf);
+
+	if ( inc_x == 1 )
+	{
+
+		n1 = n & -16;
+		if ( n1 > 0 )
+		{
+
+			sumf = dasum_kernel_16(n1, x);
+			i=n1;
+		}
+
+		while(i < n)
+		{
+			sumf += ABS(x[i]);
+			i++;
+		}
+
+	}
+	else
+	{
+
+		n *= inc_x;
+		while(i < n)
+		{
+			sumf += ABS(x[i]);
+			i += inc_x;
+		}
+
+	}
+	return(sumf);
+}
+

kernel/x86_64/dasum_microk_haswell-2.c

@@ -0,0 +1,35 @@
+#if (( defined(__GNUC__)  && __GNUC__   > 6 && defined(__AVX2__)) || (defined(__clang__) && __clang_major__ >= 6))
+
+#define HAVE_KERNEL_16 1
+
+#include <immintrin.h>
+
+static FLOAT dasum_kernel_16(BLASLONG n, FLOAT *x1)
+{
+    BLASLONG i = 0;
+    __m256d accum_0, accum_1, accum_2, accum_3;
+
+    accum_0 = _mm256_setzero_pd();
+    accum_1 = _mm256_setzero_pd();
+    accum_2 = _mm256_setzero_pd();
+    accum_3 = _mm256_setzero_pd();
+
+     __m256i abs_mask = _mm256_set1_epi64x(0x7fffffffffffffff);
+    for (; i < n; i += 16) {
+        accum_0 += (__m256d)_mm256_and_si256(_mm256_loadu_si256(&x1[i+ 0]), abs_mask);
+        accum_1 += (__m256d)_mm256_and_si256(_mm256_loadu_si256(&x1[i+ 4]), abs_mask);
+        accum_2 += (__m256d)_mm256_and_si256(_mm256_loadu_si256(&x1[i+ 8]), abs_mask);
+        accum_3 += (__m256d)_mm256_and_si256(_mm256_loadu_si256(&x1[i+12]), abs_mask);
+    }
+
+    accum_0 = accum_0 + accum_1 + accum_2 + accum_3;
+
+    __m128d half_accum0;
+    half_accum0 = _mm_add_pd(_mm256_extractf128_pd(accum_0, 0), _mm256_extractf128_pd(accum_0, 1));
+
+    half_accum0 = _mm_hadd_pd(half_accum0, half_accum0);
+
+    return half_accum0[0];
+
+}
+#endif

kernel/x86_64/dasum_microk_skylakex-2.c

@@ -0,0 +1,27 @@
+/* need a new enough GCC for avx512 support */
+#if (( defined(__GNUC__)  && __GNUC__   > 6 && defined(__AVX2__)) || (defined(__clang__) && __clang_major__ >= 6))
+
+#if defined(__AVX512CD__)
+#define HAVE_KERNEL_16 1
+
+#include <immintrin.h>
+
+static FLOAT dasum_kernel_16(BLASLONG n, FLOAT *x1)
+{
+    BLASLONG i = 0;
+
+    __m512d accum_0, accum_1;
+
+    accum_0 = _mm512_setzero_pd();
+    accum_1 = _mm512_setzero_pd();
+
+    for (; i < n; i += 16) {
+        accum_0 += _mm512_abs_pd(_mm512_loadu_pd(&x1[i+ 0]));
+        accum_1 += _mm512_abs_pd(_mm512_loadu_pd(&x1[i+ 8]));
+    }
+
+    accum_0 += accum_1;
+    return _mm512_reduce_add_pd(accum_0);
+}
+#endif
+#endif

kernel/x86_64/sasum.c

@@ -0,0 +1,104 @@
+#include "common.h"
+#include <math.h>
+
+#if defined(DOUBLE)
+
+#error supports float only
+
+#else
+
+#define ABS fabsf
+
+#endif
+
+#if defined(SKYLAKEX)
+#include "sasum_microk_skylakex-2.c"
+#elif defined(HASWELL)
+#include "sasum_microk_haswell-2.c"
+#endif
+
+#ifndef HAVE_KERNEL_32
+
+static FLOAT sasum_kernel_32(BLASLONG n, FLOAT *x1)
+{
+
+	BLASLONG i=0;
+	FLOAT *x = x1;
+	FLOAT temp0, temp1, temp2, temp3;
+	FLOAT temp4, temp5, temp6, temp7;
+	FLOAT sum0 = 0.0;
+	FLOAT sum1 = 0.0;
+	FLOAT sum2 = 0.0;
+	FLOAT sum3 = 0.0;
+
+	while ( i< n )
+	{
+
+		temp0 = ABS(x[0]);
+		temp1 = ABS(x[1]);
+		temp2 = ABS(x[2]);
+		temp3 = ABS(x[3]);
+		temp4 = ABS(x[4]);
+		temp5 = ABS(x[5]);
+		temp6 = ABS(x[6]);
+		temp7 = ABS(x[7]);
+
+		sum0 += temp0;
+		sum1 += temp1;
+		sum2 += temp2;
+		sum3 += temp3;
+
+		sum0 += temp4;
+		sum1 += temp5;
+		sum2 += temp6;
+		sum3 += temp7;
+
+		x+=8;
+		i+=8;
+
+	}
+
+	return sum0+sum1+sum2+sum3;
+}
+
+#endif
+
+FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
+{
+	BLASLONG i=0;
+	FLOAT sumf = 0.0;
+	BLASLONG n1;
+
+	if (n <= 0 || inc_x <= 0) return(sumf);
+
+	if ( inc_x == 1 )
+	{
+
+		n1 = n & -32;
+		if ( n1 > 0 )
+		{
+
+			sumf = sasum_kernel_32(n1, x);
+			i=n1;
+		}
+
+		while(i < n)
+		{
+			sumf += ABS(x[i]);
+			i++;
+		}
+
+	}
+	else
+	{
+
+		n *= inc_x;
+		while(i < n)
+		{
+			sumf += ABS(x[i]);
+			i += inc_x;
+		}
+
+	}
+	return(sumf);
+}

kernel/x86_64/sasum_microk_haswell-2.c

@@ -0,0 +1,36 @@
+#if (( defined(__GNUC__)  && __GNUC__   > 6 && defined(__AVX2__)) || (defined(__clang__) && __clang_major__ >= 6))
+
+#define HAVE_KERNEL_32 1
+
+#include <immintrin.h>
+
+static FLOAT sasum_kernel_32(BLASLONG n, FLOAT *x1)
+{
+    BLASLONG i = 0;
+    __m256 accum_0, accum_1, accum_2, accum_3;
+
+    accum_0 = _mm256_setzero_ps();
+    accum_1 = _mm256_setzero_ps();
+    accum_2 = _mm256_setzero_ps();
+    accum_3 = _mm256_setzero_ps();
+
+    __m256i abs_mask = _mm256_set1_epi32(0x7fffffff);
+    for (; i < n; i += 32) {
+        accum_0 += (__m256)_mm256_and_si256(_mm256_loadu_si256(&x1[i+ 0]), abs_mask);
+        accum_1 += (__m256)_mm256_and_si256(_mm256_loadu_si256(&x1[i+ 8]), abs_mask);
+        accum_2 += (__m256)_mm256_and_si256(_mm256_loadu_si256(&x1[i+16]), abs_mask);
+        accum_3 += (__m256)_mm256_and_si256(_mm256_loadu_si256(&x1[i+24]), abs_mask);
+    }
+
+    accum_0 = accum_0 + accum_1 + accum_2 + accum_3;
+
+    __m128 half_accum0;
+    half_accum0 = _mm_add_ps(_mm256_extractf128_ps(accum_0, 0), _mm256_extractf128_ps(accum_0, 1));
+
+    half_accum0 = _mm_hadd_ps(half_accum0, half_accum0);
+    half_accum0 = _mm_hadd_ps(half_accum0, half_accum0);
+
+    return half_accum0[0];
+
+}
+#endif

kernel/x86_64/sasum_microk_skylakex-2.c

@@ -0,0 +1,27 @@
+/* need a new enough GCC for avx512 support */
+#if (( defined(__GNUC__)  && __GNUC__   > 6 && defined(__AVX2__)) || (defined(__clang__) && __clang_major__ >= 6))
+
+#if defined(__AVX512CD__)
+#define HAVE_KERNEL_32 1
+
+#include <immintrin.h>
+
+static FLOAT sasum_kernel_32(BLASLONG n, FLOAT *x1)
+{
+    BLASLONG i = 0;
+
+    __m512 accum_0, accum_1;
+
+    accum_0 = _mm512_setzero_ps();
+    accum_1 = _mm512_setzero_ps();
+
+    for (; i < n; i += 32) {
+        accum_0 += _mm512_abs_ps(_mm512_loadu_ps(&x1[i+ 0]));
+        accum_1 += _mm512_abs_ps(_mm512_loadu_ps(&x1[i+ 16]));
+    }
+
+    accum_0 += accum_1;
+    return _mm512_reduce_add_ps(accum_0);
+}
+#endif
+#endif