/***************************************
	Auteur : Pierre Aubert
	Mail : pierre.aubert@lapp.in2p3.fr
	Licence : CeCILL-C
****************************************/

#include "phoenix_intrinsics.h"
#include <string.h>
#include "sgemm_intrinsics.h"
///Compute the Matrix-Matrix product of the x,y matrices
/**	@param[out] matOut : result
 * 	@param matX : left matrix
 * 	@param matY : right matrix
 * 	@param size : size of the square matrices
*/
void sgemm_intrinsics(float* matOut, const float* matX, const float* matY, long unsigned int size){
	memset(matOut, 0, sizeof(float)*size*size);
	long unsigned int vecSize(PLIB_VECTOR_SIZE_FLOAT);
	long unsigned int nbVec(size/vecSize);
	
	for(long unsigned int i(0lu); i < size; ++i){			//Iterate over X rows
		const float * rowX = matX + i*size;				//Get current X row
		float * rowOut = matOut + i*size;					//Get current out row
		for(long unsigned int k(0lu); k < size; ++k){					//Part of dot product
			PRegVecf regX = plib_broadcast_ss(rowX[k]);
			
			const float* rowY = matY + k*size;				//Get current Y row
			for(long unsigned int j(0lu); j < nbVec; ++j){
				PRegVecf regY = plib_load_ps(rowY + vecSize*j);
				
				PRegVecf regRes = plib_load_ps(rowOut + vecSize*j);
				
				regRes = plib_add_ps(regRes, plib_mul_ps(regX, regY));
				
				plib_store_ps(rowOut + vecSize*j, regRes);
			}
		}
	}
}