Merge pull request #2803 from xiegengxin/AVX2-asum

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 +1,82 @@
+#include "common.h"
+
+#ifndef ABS_K
+#define ABS_K(a) ((a) > 0 ? (a) : (-(a)))
+#endif
+
+#if defined(SKYLAKEX)
+#include "dasum_microk_skylakex-2.c"
+#elif defined(HASWELL)
+#include "dasum_microk_haswell-2.c"
+#endif
+
+#ifndef HAVE_DASUM_KERNEL
+static FLOAT dasum_kernel(BLASLONG n, FLOAT *x1)
+{
+
+    BLASLONG i=0;
+    BLASLONG n_8 = n & -8;
+    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;
+    FLOAT sum4 = 0.0;
+    
+    while (i < n_8) {
+        temp0 = ABS_K(x[0]);
+        temp1 = ABS_K(x[1]);
+        temp2 = ABS_K(x[2]);
+        temp3 = ABS_K(x[3]);
+        temp4 = ABS_K(x[4]);
+        temp5 = ABS_K(x[5]);
+        temp6 = ABS_K(x[6]);
+        temp7 = ABS_K(x[7]);
+        
+        sum0 += temp0;
+        sum1 += temp1;
+        sum2 += temp2;
+        sum3 += temp3;
+        
+        sum0 += temp4;
+        sum1 += temp5;
+        sum2 += temp6;
+        sum3 += temp7;
+        
+        x+=8;
+        i+=8;
+    }
+
+     while (i < n) {
+        sum4 += ABS_K(x1[i]);
+        i++;
+     }
+
+    return sum0+sum1+sum2+sum3+sum4;
+}
+
+#endif
+
+FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
+{
+    BLASLONG i=0;
+    FLOAT sumf = 0.0;
+
+    if (n <= 0 || inc_x <= 0) return(sumf);
+
+    if ( inc_x == 1 ) {
+        sumf = dasum_kernel(n, x);
+    } 
+    else {
+        n *= inc_x;
+       
+        while(i < n) {
+            sumf += ABS_K(x[i]);
+            i += inc_x;
+        }
+    }
+    return(sumf);
+}
+

kernel/x86_64/dasum_microk_haswell-2.c

@@ -0,0 +1,86 @@
+#if (( defined(__GNUC__)  && __GNUC__   > 6 ) || (defined(__clang__) && __clang_major__ >= 6)) && defined(__AVX2__)
+
+#define HAVE_DASUM_KERNEL
+
+#include <immintrin.h>
+#include <stdint.h>
+
+#ifndef ABS_K
+#define ABS_K(a) ((a) > 0 ? (a) : (-(a)))
+#endif
+
+static FLOAT dasum_kernel(BLASLONG n, FLOAT *x1)
+{
+    BLASLONG i = 0;
+    FLOAT sumf = 0.0;
+
+    if (n >= 256) {
+        BLASLONG align_256 = ((32 - ((uintptr_t)x1 & (uintptr_t)0x1f)) >> 3) & 0x3;
+
+        for (i = 0; i < align_256; i++) {
+            sumf += ABS_K(x1[i]);
+        }
+
+        n -= align_256;
+        x1 += align_256;
+    }
+
+    BLASLONG tail_index_SSE = n&(~7);
+    BLASLONG tail_index_AVX2 = n&(~255);
+
+    if (n >= 256) {
+        __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 = 0; i < tail_index_AVX2; i += 16) {
+            accum_0 += (__m256d)_mm256_and_si256(_mm256_load_si256(&x1[i+ 0]), abs_mask);
+            accum_1 += (__m256d)_mm256_and_si256(_mm256_load_si256(&x1[i+ 4]), abs_mask);
+            accum_2 += (__m256d)_mm256_and_si256(_mm256_load_si256(&x1[i+ 8]), abs_mask);
+            accum_3 += (__m256d)_mm256_and_si256(_mm256_load_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);
+
+        sumf += half_accum0[0];
+    }
+    
+    if (n >= 8) {
+        __m128d accum_20, accum_21, accum_22, accum_23;
+        accum_20 = _mm_setzero_pd();  
+        accum_21 = _mm_setzero_pd(); 
+        accum_22 = _mm_setzero_pd(); 
+        accum_23 = _mm_setzero_pd(); 
+
+        __m128i abs_mask2 = _mm_set1_epi64x(0x7fffffffffffffff);
+        for (i = tail_index_AVX2; i < tail_index_SSE; i += 8) {
+            accum_20 += (__m128d)_mm_and_si128(_mm_loadu_si128(&x1[i + 0]), abs_mask2);
+            accum_21 += (__m128d)_mm_and_si128(_mm_loadu_si128(&x1[i + 2]), abs_mask2);
+            accum_22 += (__m128d)_mm_and_si128(_mm_loadu_si128(&x1[i + 4]), abs_mask2);
+            accum_23 += (__m128d)_mm_and_si128(_mm_loadu_si128(&x1[i + 6]), abs_mask2);
+        }
+
+        accum_20 = accum_20 + accum_21 + accum_22 + accum_23;
+        __m128d half_accum20;
+        half_accum20 = _mm_hadd_pd(accum_20, accum_20);
+
+        sumf += half_accum20[0];
+    }
+    
+    for (i = tail_index_SSE; i < n; ++i) {
+        sumf += ABS_K(x1[i]);
+    }
+
+    return sumf;
+
+}
+#endif

kernel/x86_64/dasum_microk_skylakex-2.c

@@ -0,0 +1,80 @@
+/* need a new enough GCC for avx512 support */
+#if (( defined(__GNUC__)  && __GNUC__   > 6 && defined(__AVX512CD__)) || (defined(__clang__) && __clang_major__ >= 9))
+
+#define HAVE_DASUM_KERNEL 1
+
+#include <immintrin.h>
+
+#include <stdint.h>
+
+#ifndef ABS_K
+#define ABS_K(a) ((a) > 0 ? (a) : (-(a)))
+#endif
+
+static FLOAT dasum_kernel(BLASLONG n, FLOAT *x1)
+{
+    BLASLONG i = 0;
+    FLOAT sumf = 0.0;
+
+    if (n >= 256) {
+        BLASLONG align_512 = ((64 - ((uintptr_t)x1 & (uintptr_t)0x3f)) >> 3) & 0x7;
+
+        for (i = 0; i < align_512; i++) {
+            sumf += ABS_K(x1[i]);
+        }
+        
+        n -= align_512;
+        x1 += align_512;
+    }
+
+    BLASLONG tail_index_SSE = n&(~7);
+    BLASLONG tail_index_AVX512 = n&(~255);
+
+    //
+    if ( n >= 256 ) {
+
+        __m512d accum_0, accum_1, accum_2, accum_3;
+        accum_0 = _mm512_setzero_pd();
+        accum_1 = _mm512_setzero_pd();
+        accum_2 = _mm512_setzero_pd();
+        accum_3 = _mm512_setzero_pd();
+        for (i = 0; i < tail_index_AVX512; i += 32) {
+            accum_0 += _mm512_abs_pd(_mm512_load_pd(&x1[i + 0]));
+            accum_1 += _mm512_abs_pd(_mm512_load_pd(&x1[i + 8]));
+            accum_2 += _mm512_abs_pd(_mm512_load_pd(&x1[i +16]));
+            accum_3 += _mm512_abs_pd(_mm512_load_pd(&x1[i +24]));
+        }
+
+        accum_0 = accum_0 + accum_1 + accum_2 + accum_3;
+        sumf += _mm512_reduce_add_pd(accum_0);
+    }
+
+    if (n >= 8) {
+        __m128d accum_20, accum_21, accum_22, accum_23;
+        accum_20 = _mm_setzero_pd();  
+        accum_21 = _mm_setzero_pd(); 
+        accum_22 = _mm_setzero_pd(); 
+        accum_23 = _mm_setzero_pd(); 
+
+        __m128i abs_mask2 = _mm_set1_epi64x(0x7fffffffffffffff);
+        for (i = tail_index_AVX512; i < tail_index_SSE; i += 8) {
+            accum_20 += (__m128d)_mm_and_si128(_mm_loadu_si128(&x1[i + 0]), abs_mask2);
+            accum_21 += (__m128d)_mm_and_si128(_mm_loadu_si128(&x1[i + 2]), abs_mask2);
+            accum_22 += (__m128d)_mm_and_si128(_mm_loadu_si128(&x1[i + 4]), abs_mask2);
+            accum_23 += (__m128d)_mm_and_si128(_mm_loadu_si128(&x1[i + 6]), abs_mask2);
+        }
+
+        accum_20 = accum_20 + accum_21 + accum_22 + accum_23;
+        __m128d half_accum20;
+        half_accum20 = _mm_hadd_pd(accum_20, accum_20);
+
+        sumf += half_accum20[0];
+    }
+
+    for (i = tail_index_SSE; i < n; ++i) {
+        sumf += ABS_K(x1[i]);
+    }
+
+    return sumf;
+}
+#endif

kernel/x86_64/sasum.c

@@ -0,0 +1,90 @@
+#include "common.h"
+
+#if defined(DOUBLE)
+#error supports float only
+#else
+#ifndef ABS_K
+#define ABS_K(a) ((a) > 0 ? (a) : (-(a)))
+#endif
+
+#endif
+
+#if defined(SKYLAKEX)
+#include "sasum_microk_skylakex-2.c"
+#elif defined(HASWELL)
+#include "sasum_microk_haswell-2.c"
+#endif
+
+#ifndef HAVE_SASUM_KERNEL
+
+static FLOAT sasum_kernel(BLASLONG n, FLOAT *x1)
+{
+
+    BLASLONG i=0;
+    BLASLONG n_8 = n & -8;
+    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;
+    FLOAT sum4 = 0.0;
+
+    while (i < n_8) {
+
+        temp0 = ABS_K(x[0]);
+        temp1 = ABS_K(x[1]);
+        temp2 = ABS_K(x[2]);
+        temp3 = ABS_K(x[3]);
+        temp4 = ABS_K(x[4]);
+        temp5 = ABS_K(x[5]);
+        temp6 = ABS_K(x[6]);
+        temp7 = ABS_K(x[7]);
+
+        sum0 += temp0;
+        sum1 += temp1;
+        sum2 += temp2;
+        sum3 += temp3;
+
+        sum0 += temp4;
+        sum1 += temp5;
+        sum2 += temp6;
+        sum3 += temp7;
+
+        x+=8;
+        i+=8;
+
+    }
+
+    while (i < n) {
+        sum4 += ABS_K(x1[i]);
+        i++;
+    }
+
+    return sum0+sum1+sum2+sum3+sum4;
+}
+
+#endif
+
+FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
+{
+    BLASLONG i=0;
+    FLOAT sumf = 0.0;
+
+    if (n <= 0 || inc_x <= 0) return(sumf);
+
+    if ( inc_x == 1 ) {
+        sumf = sasum_kernel(n, x);
+    }
+    else {
+
+        n *= inc_x;
+        while(i < n) {
+            sumf += ABS_K(x[i]);
+            i += inc_x;
+        }
+
+    }
+    return(sumf);
+}

kernel/x86_64/sasum_microk_haswell-2.c

@@ -0,0 +1,82 @@
+#if (( defined(__GNUC__)  && __GNUC__   > 6 ) || (defined(__clang__) && __clang_major__ >= 6)) && defined(__AVX2__)
+
+#define HAVE_SASUM_KERNEL 1
+
+#include <immintrin.h>
+#include <stdint.h>
+
+#ifndef ABS_K
+#define ABS_K(a) ((a) > 0 ? (a) : (-(a)))
+#endif
+
+static FLOAT sasum_kernel(BLASLONG n, FLOAT *x1)
+{
+    BLASLONG i = 0;
+    FLOAT sumf = 0.0;
+
+    if (n >= 256) { 
+        BLASLONG align_256 = ((32 - ((uintptr_t)x1 & (uintptr_t)0x1f)) >> 2) & 0x7;
+
+        for (i = 0; i < align_256; i++) {
+            sumf += ABS_K(x1[i]);
+        }
+
+        n -= align_256;
+        x1 += align_256;
+    }
+
+    BLASLONG tail_index_SSE = n&(~7);
+    BLASLONG tail_index_AVX2 = n&(~255);
+
+    if (n >= 256) {
+        __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 = 0; i < tail_index_AVX2; i += 32) {
+            accum_0 += (__m256)_mm256_and_si256(_mm256_load_si256(&x1[i+ 0]), abs_mask);
+            accum_1 += (__m256)_mm256_and_si256(_mm256_load_si256(&x1[i+ 8]), abs_mask);
+            accum_2 += (__m256)_mm256_and_si256(_mm256_load_si256(&x1[i+16]), abs_mask);
+            accum_3 += (__m256)_mm256_and_si256(_mm256_load_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);
+
+        sumf += half_accum0[0];
+        
+    }
+    
+    if (n >= 8) {
+        __m128 accum_20, accum_21;
+        accum_20 = _mm_setzero_ps();
+        accum_21 = _mm_setzero_ps();
+
+        __m128i abs_mask2 = _mm_set1_epi32(0x7fffffff);
+        for (i = tail_index_AVX2; i < tail_index_SSE; i += 8) {
+            accum_20 += (__m128)_mm_and_si128(_mm_loadu_si128(&x1[i + 0]), abs_mask2);
+            accum_21 += (__m128)_mm_and_si128(_mm_loadu_si128(&x1[i + 4]), abs_mask2);
+        }
+        
+        accum_20 += accum_21;
+        accum_20 = _mm_hadd_ps(accum_20, accum_20);
+        accum_20 = _mm_hadd_ps(accum_20, accum_20);
+        
+        sumf += accum_20[0];
+    }
+
+    for (i = tail_index_SSE; i < n; ++i) {
+        sumf += ABS_K(x1[i]);
+    }
+
+    return sumf;
+}
+#endif

kernel/x86_64/sasum_microk_skylakex-2.c

@@ -0,0 +1,73 @@
+/* need a new enough GCC for avx512 support */
+#if (( defined(__GNUC__)  && __GNUC__   > 6 && defined(__AVX512CD__)) || (defined(__clang__) && __clang_major__ >= 9))
+
+#define HAVE_SASUM_KERNEL 1
+
+#ifndef ABS_K
+#define ABS_K(a) ((a) > 0 ? (a) : (-(a)))
+#endif
+
+#include <immintrin.h>
+#include <stdint.h>
+
+static FLOAT sasum_kernel(BLASLONG n, FLOAT *x1)
+{
+    BLASLONG i = 0;
+    FLOAT sumf = 0.0;
+
+    if (n >= 256) {
+        BLASLONG align_512 = ((64 - ((uintptr_t)x1 & (uintptr_t)0x3f)) >> 2) & 0xf;
+
+        for (i = 0; i < align_512; i++) {
+            sumf += ABS_K(x1[i]);
+        }
+        n -= align_512;
+        x1 += align_512;
+    }
+
+    BLASLONG tail_index_SSE = n&(~7);
+    BLASLONG tail_index_AVX512 = n&(~255);
+
+    if (n >= 256) {
+        __m512 accum_0, accum_1, accum_2, accum_3;
+        accum_0 = _mm512_setzero_ps();
+        accum_1 = _mm512_setzero_ps();
+        accum_2 = _mm512_setzero_ps();
+        accum_3 = _mm512_setzero_ps();
+
+        for (i = 0; i < tail_index_AVX512; i += 64) {
+            accum_0 += _mm512_abs_ps(_mm512_load_ps(&x1[i + 0]));
+            accum_1 += _mm512_abs_ps(_mm512_load_ps(&x1[i +16]));
+            accum_2 += _mm512_abs_ps(_mm512_load_ps(&x1[i +32]));
+            accum_3 += _mm512_abs_ps(_mm512_load_ps(&x1[i +48]));
+        }
+
+        accum_0 = accum_0 + accum_1 + accum_2 + accum_3;
+        sumf += _mm512_reduce_add_ps(accum_0);
+    }
+
+    if (n >= 8) {
+        __m128 accum_20, accum_21;
+        accum_20 = _mm_setzero_ps();
+        accum_21 = _mm_setzero_ps();
+
+        __m128i abs_mask2 = _mm_set1_epi32(0x7fffffff);
+        for (i = tail_index_AVX512; i < tail_index_SSE; i += 8) {
+            accum_20 += (__m128)_mm_and_si128(_mm_loadu_si128(&x1[i + 0]), abs_mask2);
+            accum_21 += (__m128)_mm_and_si128(_mm_loadu_si128(&x1[i + 4]), abs_mask2);
+        }
+        
+        accum_20 += accum_21;
+        accum_20 = _mm_hadd_ps(accum_20, accum_20);
+        accum_20 = _mm_hadd_ps(accum_20, accum_20);
+        
+        sumf += accum_20[0];
+    }
+
+    for (i = tail_index_SSE; i < n; i++) {
+        sumf += ABS_K(x1[i]);
+    }
+
+    return sumf;
+}
+#endif