dunno

Author: Jeff Hammond

lapack/test_xgemm.c

@@ -2,10 +2,17 @@
 #include <stdlib.h>
 #include <string.h>
 
-#ifdef MKL
-#include "mkl.h"
+#if defined(MKL)
+# include <mkl.h>
+# ifdef MKL_ILP64
+#  error Use the MKL library for 32-bit integers!
+# endif
+#elif defined(ACCELERATE)
+/* The location of cblas.h is not in the system
+ * include path when -framework Accelerate is provided. */
+# include <Accelerate/Accelerate.h>
 #else
-#include "cblas.h"
+# include <cblas.h>
 #endif
 
 #if (__STDC_VERSION__ < 199901L)
@@ -18,6 +25,18 @@
 #define PDEBUG(fmt, ...)
 #endif
 
+void test_sgemm(const int m, const int n, const int k)
+{
+
+    const double alpha = 1.0;
+    const double beta  = 1.0;
+    const int lda = k;
+    const int ldb = n;
+    const int ldc = n;
+    cblas_dgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans,
+                m, n, k, alpha, A, lda, B, ldb, beta, C, ldc);
+}
+
 int main(int argc, char* argv[])
 {
     char type;
@@ -96,6 +115,9 @@ int main(int argc, char* argv[])
         printf("You have chosen an unsupported datatype...\n");
         goto input_error;
     }
+    printf("C+=A*B with m=%d, n=%d, k=%d\n", m, n, k);
+
+    test_xgemm(type, m, n k);
 
     printf("SUCCESS\n");
     return 0;
@@ -112,5 +134,6 @@ int main(int argc, char* argv[])
         goto fail;
 
     fail:
+        printf("FAILURE\n");
         return 1;
 }

simd/Makefile

@@ -9,7 +9,7 @@ CXXFLAGS	=	-g -Wall -O3 -qopenmp -xCORE-AVX2
 #CXXFLAGS	=	-g -Wall -O3 -qopenmp -xCORE-AVX512 -qopt-zmm-usage=high
 ASMFLAGS	=	-fsource-asm -fverbose-asm -fasm-blocks -fcode-asm
 
-all: bfloat16 bfloat16.s
+all: bfloat16 bfloat16.s time_conversion
 
 bfloat16: bfloat16.cc
 	${CXX} ${CXXFLAGS} $< -o $@

simd/time_conversion.cc

@@ -21,14 +21,9 @@ int main(int argc, char* argv[])
 
     uint16_t * ibf16 = new uint16_t[n];
     uint16_t * if16  = new uint16_t[n];
-    float    * f32   = new float[n];
     uint16_t * obf16 = new uint16_t[n];
     uint16_t * of16  = new uint16_t[n];
 
-    for (int i=0; i<n; i++) {
-        f32[i] = (float)i;
-    }
-
     for (int i=0; i<n; ++i) {
         uint16_t j = i % 16384;
         ibf16[i] = j;
@@ -37,13 +32,21 @@ int main(int argc, char* argv[])
         of16[i]  = j;
     }
 
+#if defined(__AVX512F__) && defined(__AVX512BW__)
     for (int i=0; i<n; i+=8) {
         __m128i a = _mm_load_si128((__m128i*)&ibf16[i]);
         __m256i b = _mm256_cvtepu16_epi32(a);
         __m256i c = _mm256_slli_epi32(b,16);
 
         _mm_store_si128((__m128i*)&obf16[i],c);
     }
+#else
+    PRAGMA_SIMD
+    for (int i=0; i<n; i++) {
+        uint32_t u32 = ((uint32_t)bf[i]) << 16;
+        float    f32 = *(float*)&u32;
+    }
+#endif
 
     for (int i=0; i<n; i++) {
         std::cout << std::setw(10) << i << ": "