Merge pull request #82 from xianyi/develop

rebase

Author: martin-frbg

Makefile.arm64

@@ -66,6 +66,11 @@ FCOMMON_OPT += -march=armv8.1-a -mtune=thunderx2t99
 endif
 endif
 
+ifeq ($(CORE), VORTEX)
+CCOMMON_OPT += -march=armv8.3-a
+FCOMMON_OPT += -march=armv8.3-a
+endif
+
 ifeq ($(GCCVERSIONGTEQ9), 1)
 ifeq ($(CORE), TSV110)
 CCOMMON_OPT += -march=armv8.2-a -mtune=tsv110

TargetList.txt

@@ -98,6 +98,7 @@ THUNDERX
 THUNDERX2T99
 TSV110
 THUNDERX3T110
+VORTEX
 
 9.System Z:
 ZARCH_GENERIC

benchmark/asum.c

@@ -128,8 +128,13 @@ int main(int argc, char *argv[]){
   int to   = 200;
   int step =   1;
 
+#if defined(__WIN32__) || defined(__WIN64__) || !defined(_POSIX_TIMERS)
   struct timeval start, stop;
   double time1,timeg;
+#else
+  struct timespec start = { 0, 0 }, stop = { 0, 0 };
+  double time1, timeg;
+#endif
 
   argc--;argv++;
 
@@ -160,26 +165,30 @@ int main(int argc, char *argv[]){
 
    fprintf(stderr, " %6d : ", (int)m);
 
-
    for (l=0; l<loops; l++)
    {
 
    	for(i = 0; i < m * COMPSIZE * abs(inc_x); i++){
 			x[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;
    	}
-
+#if defined(__WIN32__) || defined(__WIN64__) || !defined(_POSIX_TIMERS)
     	gettimeofday( &start, (struct timezone *)0);
-
+#else
+        clock_gettime(CLOCK_REALTIME, &start);
+#endif
     	result = ASUM (&m, x, &inc_x);
-
-    	gettimeofday( &stop, (struct timezone *)0);
-
-    	time1 = (double)(stop.tv_sec - start.tv_sec) + (double)((stop.tv_usec - start.tv_usec)) * 1.e-6;
+#if defined(__WIN32__) || defined(__WIN64__) || !defined(_POSIX_TIMERS)
+    	clock_gettime(CLOCK_REALTIME, &stop);
+   	time1 = (double)(stop.tv_sec - start.tv_sec) + (double)((stop.tv_usec - start.tv_usec)) * 1.e-6;
+#else
+  	gettimeofday( &stop, (struct timezone *)0);
+        time1 = (double)(stop.tv_sec - start.tv_sec) + (double)((stop.tv_nsec - start.tv_nsec)) / 1.e9;
+#endif
 
 	timeg += time1;
 
     }
-
+if (loops >1)
     timeg /= loops;
 
 #ifdef COMPLEX

cpuid_arm64.c

@@ -26,6 +26,11 @@
   *****************************************************************************/
 
 #include <string.h>
+#ifdef OS_DARWIN
+#include <sys/sysctl.h>
+int32_t value;
+size_t length=sizeof(value);
+#endif
 
 #define CPU_UNKNOWN     	0
 #define CPU_ARMV8       	1
@@ -45,6 +50,8 @@
 #define CPU_TSV110        9
 // Ampere
 #define CPU_EMAG8180	 10
+// Apple
+#define CPU_VORTEX       13
 
 static char *cpuname[] = {
   "UNKNOWN",
@@ -59,7 +66,8 @@ static char *cpuname[] = {
   "TSV110",
   "EMAG8180",
   "NEOVERSEN1",
-  "THUNDERX3T110"
+  "THUNDERX3T110",
+  "VORTEX"	
 };
 
 static char *cpuname_lower[] = {
@@ -75,7 +83,8 @@ static char *cpuname_lower[] = {
   "tsv110",
   "emag8180",
   "neoversen1",
-  "thunderx3t110"
+  "thunderx3t110",
+  "vortex"	
 };
 
 int get_feature(char *search)
@@ -198,6 +207,10 @@ int detect(void)
 
 	}
 #else
+#ifdef DARWIN
+	sysctlbyname("hw.cpufamily",&value,&length,NULL,0);
+	if (value ==131287967) return CPU_VORTEX;
+#endif
 	return CPU_ARMV8;	
 #endif
 
@@ -247,7 +260,10 @@ int n=0;
 
 	printf("#define NUM_CORES %d\n",n);
 #endif
-
+#ifdef DARWIN
+	sysctlbyname("hw.physicalcpu_max",&value,&length,NULL,0);
+	printf("#define NUM_CORES %d\n",value);
+#endif	
 }
 
 
@@ -398,6 +414,19 @@ void get_cpuconfig(void)
 			printf("#define DTB_DEFAULT_ENTRIES  64       \n");
 			printf("#define DTB_SIZE             4096     \n");
 			break;
+#ifdef DARWIN
+		case CPU_VORTEX:
+			printf("#define VORTEX			      \n");
+			sysctlbyname("hw.l1icachesize",&value,&length,NULL,0);
+			printf("#define L1_CODE_SIZE	     %d       \n",value);
+			sysctlbyname("hw.cachelinesize",&value,&length,NULL,0);
+			printf("#define L1_CODE_LINESIZE     %d       \n",value);
+			sysctlbyname("hw.l1dcachesize",&value,&length,NULL,0);
+			printf("#define L1_DATA_SIZE	     %d       \n",value);
+			sysctlbyname("hw.l2dcachesize",&value,&length,NULL,0);
+			printf("#define L2_DATA_SIZE	     %d       \n",value);
+			break;
+#endif			
 	}
 	get_cpucount();
 }

kernel/arm64/KERNEL.VORTEX

@@ -0,0 +1 @@
+include $(KERNELDIR)/KERNEL.ARMV8

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