ELECH473 - SIMD _ movapd surprisingly working

gistfile1.txt

#include "StdAfx.h"
#include "stdio.h"
#include "time.h"
#include <stdlib.h>
 
void blackWhiteASM(char* pixels, int lSize, int t, char* output) {
	int ii=lSize/16; // Set the counter
	unsigned char *threshold = (unsigned char*) malloc(sizeof(unsigned char)*16);
	
	int i;
	for(i=0;i<16;i++) {
		threshold[i] = t;
	}
	_asm {
		mov		esi , pixels; // datain ptr of the line in register
		mov		ecx , ii; // counter
		mov		edi , output; // dataout pointer
		mov		eax , threshold; // treshold value pointer
		movapd	xmm7 , [eax]; // 1st pipe mask
	l1:
		movapd	xmm0 , [esi]; // load first line
		pcmpeqb	xmm0 , xmm7; // compare
		movapd	[edi] , xmm0; // Move result to memory destination
		add		edi , 16; // New result pointer
		add		esi , 16; // New source pointer
		sub		ecx , 1;
	jnz		l1;
		//emms;
	}
}
 
void ASMtest() {
	int counter = 4;
	unsigned char *pixels = (unsigned char*) malloc(sizeof(unsigned char*)*65);
	char *output = (char*) malloc(sizeof(char)*65);
	int i;
	for(int i=0;i<65;i++) {
		pixels[i] = 200;
	}
	unsigned char *threshold = (unsigned char*) malloc(sizeof(unsigned char)*16);
	for(i=0;i<16;i++) {
		threshold[i] = 127;
	}
	_asm {
		mov		esi , pixels; // datain ptr of the line in register
		mov		ecx , counter; // counter
		mov		edi , output; // dataout pointer
		mov		eax , threshold; // treshold value pointer
		movapd	xmm7 , [eax]; // 1st pipe mask
	l1:
		movapd	xmm0 , [esi]; // load first line
		pcmpeqb	xmm0 , xmm7; // compare
		movapd	[edi] , xmm0; // Move result to memory destination
		add		edi , 16; // New result pointer
		add		esi , 16; // New source pointer
		sub		ecx , 1;
	jnz		l1;
		//emms;
	}
}
 
void blackWhiteBuffer(char* pixels, int lSize, int threshold) {
	int i;
	for(i=0;i<lSize;i++) {
		if(pixels[i]<threshold) {
			pixels[i]=0;
		}
		else {
			pixels[i]=255;
		}
	}
}
 
 
int main(int argc, char* argv) {
	time_t startTimeA, endTimeA, startTimeB, endTimeB, startTimeC, endTimeC;
	int const height = 1024;
	int const width = 1024;
	int threshold = 75;
	int i,j;
	unsigned char pixel;
	float dtA, dtB, dtC;
 
	//******
	//ASM testing
	//******
	//ASMtest();
 
	
	//***************
	//Using a buffer
	//***************
	//Start time
	startTimeA = clock();
	FILE* imgA = fopen("C:\\test.raw","rb");
	FILE* imgOutputA = fopen("C:\\outputA.raw","wb");
	//file size
	fseek(imgA , 0 , SEEK_END);
	int lSize = ftell (imgA);
	rewind (imgA);
	// allocate memory to contain the whole file:
	char *pixels = (char*) malloc(sizeof(char)*lSize);
	char *output = (char*) malloc(sizeof(char)*lSize);
	fread(pixels, sizeof(unsigned char), lSize, imgA);
	fclose(imgA);
	//black & white
	blackWhiteBuffer(pixels, lSize, threshold);
	/*for(i=0;i<lSize;i++) {
		if(pixels[i]<threshold) {
			pixels[i]=0;
		}
		else {
			pixels[i]=255;
		}
	}*/
	fwrite(pixels, sizeof(unsigned char), lSize, imgOutputA);
	fclose(imgOutputA);
	//end time
	endTimeA = clock();
	dtA = (endTimeA-startTimeA)/(float)(CLOCKS_PER_SEC);
 
	//**************
	//Streaming
	//**************
	//Start time
	startTimeB = clock();
	FILE* imgB = fopen("C:\\test.raw","rb");
	FILE* imgOutputB = fopen("C:\\outputB.raw","wb");
	//black & white
	for(i=0;i<lSize;i++) {
		pixel = fgetc(imgB);
		if(pixel<threshold) {
			pixel=0;
		}
		else {
			pixel=255;
		}
		fputc(pixel, imgOutputB);
	}
	fclose(imgB);
 
	fwrite(pixels, sizeof(unsigned char), lSize, imgOutputB);
	fclose(imgOutputB);
	//end time
	endTimeB = clock();
	dtB = (endTimeB-startTimeB)/(float)(CLOCKS_PER_SEC);
 
	//**********
	//ASM
	//**********
	//Start time
	startTimeC = clock();
	FILE* imgC = fopen("C:\\test.raw","rb");
	FILE* imgOutputC = fopen("C:\\outputC.raw","wb");
	//file size
	fseek(imgC , 0 , SEEK_END);
	lSize = ftell (imgC);
	rewind (imgC);
	// allocate memory to contain the whole file:
	lSize++;
	char *pixelsC = (char*) malloc(sizeof(char)*lSize);
	char *outputC = (char*) malloc(sizeof(char)*lSize);
	fread(pixelsC, sizeof(unsigned char), lSize, imgC);
	fclose(imgC);
	//black & white
	blackWhiteASM(pixelsC, lSize, threshold, outputC);
 
	fwrite(outputC, sizeof(unsigned char), lSize, imgOutputC);
	fclose(imgOutputC);
	//end time
	endTimeC = clock();
	dtC = (endTimeC-startTimeC)/(float)(CLOCKS_PER_SEC);
 
 
	//*******
	//Display time spent.
	//*******
	printf("Time A (Buffer, C): %f\n",dtA);
	printf("Start time A: %d\n",startTimeA);
	printf("End time A: %d\n",endTimeA);
	printf("Time B (Stream, C): %f\n",dtB);
	printf("Start time B: %d\n",startTimeB);
	printf("End time B: %d\n",endTimeB);
	printf("Time C (Buffer, ASM): %f\n",dtC);
	printf("Start time C: %d\n",startTimeC);
	printf("End time C: %d\n",endTimeC);
 
	getchar();
	return 0;
}