C++20 not sufficient for small tables ~2000 elements clearly faster...

C++20 not sufficient for small tables ~2000 elements clearly faster computation with explicit vectorisation hint such as builtin_assume_aligned and __restrict__

src/cpp20/Sequential/hadamard/CMakeLists.txt

@@ -15,6 +15,12 @@ phoenix_compileAndRunExample(perf_hadamard_seq_vectorize_Ofast "-Ofast -ftree-ve
 
 phoenix_plotPerf("hadamard_seqBase" perf_hadamard_seq_O0 perf_hadamard_seq_O1 perf_hadamard_seq_O2 perf_hadamard_seq_O3 perf_hadamard_seq_Ofast)
 
-phoenix_plotPerf("hadamard_seqVectorize" perf_hadamard_seq_O3 perf_hadamard_seq_vectorize_O3 perf_hadamard_seq_vectorize_Ofast)
+set(progVectorizedSrc hadamard_product_vectorized.cpp main_vectorized.cpp)
 
+phoenix_compileAndRunExample(perf_hadamard_seq_explicit_vectorize_O3 "-O3 -ftree-vectorize -march=native -mtune=native -mavx2" "${CONFIG_HADAMARD}" ${progVectorizedSrc})
+phoenix_compileAndRunExample(perf_hadamard_seq_explicit_vectorize_Ofast "-Ofast -ftree-vectorize -march=native -mtune=native -mavx2" "${CONFIG_HADAMARD}" ${progVectorizedSrc})
+
+
+phoenix_plotPerf("hadamard_seqVectorize" perf_hadamard_seq_O3 perf_hadamard_seq_vectorize_O3 perf_hadamard_seq_vectorize_Ofast
+						perf_hadamard_seq_explicit_vectorize_O3 perf_hadamard_seq_explicit_vectorize_Ofast)
 

src/cpp20/Sequential/hadamard/hadamard_product_vectorized.cpp

+/***************************************
+	Auteur : Pierre Aubert
+	Mail : aubertp7@gmail.com
+	Licence : CeCILL-C
+****************************************/
+
+
+#include "hadamard_product_vectorized.h"
+
+#define FLOAT_VECTOR_ALIGNEMENT 32
+
+///Do the Hadamard product
+/**	@param[out] ptabResult : table of results of tabX*tabY
+ * 	@param ptabX : input table
+ * 	@param ptabY : input table
+ * 	@param nbElement : number of elements in the tables
+*/
+void hadamard_product(float* __restrict__ ptabResult, const float* __restrict__ ptabX, const float* __restrict__ ptabY, long unsigned int nbElement){
+	const float* tabX = (const float*)__builtin_assume_aligned(ptabX, FLOAT_VECTOR_ALIGNEMENT);
+	const float* tabY = (const float*)__builtin_assume_aligned(ptabY, FLOAT_VECTOR_ALIGNEMENT);
+	float* tabResult = (float*)__builtin_assume_aligned(ptabResult, FLOAT_VECTOR_ALIGNEMENT);
+	
+	for(long unsigned int i(0lu); i < nbElement; ++i){
+		tabResult[i] = tabX[i]*tabY[i];
+	}
+}
+
+

src/cpp20/Sequential/hadamard/hadamard_product_vectorized.h

+/***************************************
+	Auteur : Pierre Aubert
+	Mail : aubertp7@gmail.com
+	Licence : CeCILL-C
+****************************************/
+
+#ifndef __HADAMARD_PRODUCT_H__
+#define __HADAMARD_PRODUCT_H__
+
+
+
+void hadamard_product(float* __restrict__ ptabResult, const float* __restrict__ ptabX, const float* __restrict__ ptabY, long unsigned int nbElement);
+
+
+#endif

src/cpp20/Sequential/hadamard/main_vectorized.cpp

+/***************************************
+	Auteur : Pierre Aubert
+	Mail : aubertp7@gmail.com
+	Licence : CeCILL-C
+****************************************/
+
+#include <iostream>
+
+#include "micro_benchmark.h"
+#include "PTensor.h"
+
+#include "hadamard_product_vectorized.h"
+
+///Get the number of nanoseconds per elements
+/**	@param nbElement : number of elements of the tables
+*/
+void evaluateHadamardProduct(size_t nbElement){
+	
+	PTensor<float> tabX(AllocMode::ALIGNED, nbElement);
+	PTensor<float> tabY(AllocMode::ALIGNED, nbElement);
+	PTensor<float> tabRes(AllocMode::ALIGNED, nbElement);
+	
+	for(size_t i(0lu); i < nbElement; ++i){
+		tabX[i] = i*19lu%11;
+		tabY[i] = i*27lu%19;
+	}
+	micro_benchmarkAutoNsPrint("evaluate hadamard vectorized", nbElement, hadamard_product, tabRes.getData(), tabX.getData(), tabY.getData(), nbElement);
+}
+
+int main(int argc, char** argv){
+	return micro_benchmarkParseArg(argc, argv, evaluateHadamardProduct);
+}
+
+