--- plugin_ssim.c	2006/11/08 06:55:27	1.6
+++ plugin_ssim.c	2008/11/27 20:34:53	1.12
@@ -23,7 +23,7 @@
  *
  ****************************************************************************/
 
-#include <malloc.h>
+#include <stdlib.h>
 #include <stdio.h>
 #include <math.h>
 #include "../portab.h"
@@ -39,10 +39,20 @@
 
 /*dev 1.0 gaussian weighting. the weight for the pixel x,y is w(x)*w(y)*/
 static float mask8[8] = {
-	0.0069815, 0.1402264, 1.0361408, 2.8165226, 
-	2.8165226, 1.0361408, 0.1402264, 0.0069815
+	0.0069815f, 0.1402264f, 1.0361408f, 2.8165226f, 
+	2.8165226f, 1.0361408f, 0.1402264f, 0.0069815f
 };
 
+/* integer version. Norm: coeffs sums up to 4096.
+  Define USE_INT_GAUSSIAN to use it as replacement to float version */
+
+/* #define USE_INT_GAUSSIAN */
+static const uint16_t imask8[8] = {
+  4, 72, 530, 1442, 1442, 530, 72, 4
+};
+#define GACCUM(X)  ( ((X)+(1<<11)) >> 12 )
+
+
 struct framestat_t{
 	int type;
 	int quant;
@@ -238,7 +248,22 @@
 		sum *=mask8[i];
 		mean += sum;
 	}
-	return (int) mean + 0.5;
+	return (int) (mean + 0.5);
+}
+
+int lum_8x8_gaussian_int(uint8_t* ptr, int stride){
+	uint32_t mean;
+	int i,j;
+	mean = 0;
+	for(i=0;i<8;i++){
+		uint32_t sum = 0;
+		for(j=0;j<8;j++)
+			sum += ptr[i*stride + j]*imask8[j];
+		
+		sum = GACCUM(sum) * imask8[i];
+		mean += sum;
+	}
+	return (int)GACCUM(mean);
 }
 
 /*calculate the difference between two blocks next to each other on a row*/
@@ -279,9 +304,43 @@
 	ptrc += str;
 	}
 
-	*pdevo = (int) (devo - ((lumo*lumo + 32) >> 6)) + 0.5;
-	*pdevc = (int) (devc - ((lumc*lumc + 32) >> 6)) + 0.5;
-	*pcorr = (int) (corr - ((lumo*lumc + 32) >> 6)) + 0.5;
+	*pdevo = (int) ((devo - ((lumo*lumo + 32) >> 6)) + 0.5);
+	*pdevc = (int) ((devc - ((lumc*lumc + 32) >> 6)) + 0.5);
+	*pcorr = (int) ((corr - ((lumo*lumc + 32) >> 6)) + 0.5);
+};
+
+void consim_gaussian_int(uint8_t* ptro, uint8_t* ptrc, int stride, int lumo, int lumc, int* pdevo, int* pdevc, int* pcorr)
+{
+	unsigned int valo, valc,i,j,str;
+	uint32_t  devo=0, devc=0, corr=0;
+	str = stride - 8;
+	for(i=0;i< 8;i++){
+		uint32_t sumo = 0;
+		uint32_t sumc = 0;
+		uint32_t sumcorr = 0;
+		for(j=0;j< 8;j++){
+			valo = *ptro;
+			valc = *ptrc;
+			sumo += valo*valo*imask8[j];
+			sumc += valc*valc*imask8[j];
+			sumcorr += valo*valc*imask8[j];
+			ptro++;
+			ptrc++;
+		}
+
+	devo += GACCUM(sumo)*imask8[i];
+	devc += GACCUM(sumc)*imask8[i];
+	corr += GACCUM(sumcorr)*imask8[i];
+	ptro += str;
+	ptrc += str;
+	}
+
+        devo = GACCUM(devo);
+        devc = GACCUM(devc);
+        corr = GACCUM(corr);
+	*pdevo = (int) ((devo - ((lumo*lumo + 32) >> 6)) + 0.5);
+	*pdevc = (int) ((devc - ((lumc*lumc + 32) >> 6)) + 0.5);
+	*pcorr = (int) ((corr - ((lumo*lumc + 32) >> 6)) + 0.5);
 };
 
 /*calculate contrast and correlation of the two blocks*/
@@ -309,10 +368,10 @@
 
 /*calculate the final ssim value*/
 static float calc_ssim(float meano, float meanc, float devo, float devc, float corr){
-	static const float c1 = (0.01*255)*(0.01*255);
-	static const float c2 = (0.03*255)*(0.03*255);
+	static const float c1 = (0.01f*255)*(0.01f*255);
+	static const float c2 = (0.03f*255)*(0.03f*255);
 	/*printf("meano: %f meanc: %f devo: %f devc: %f corr: %f\n",meano,meanc,devo,devc,corr);*/
-	return ((2.0*meano*meanc + c1)*(corr/32.0 + c2))/((meano*meano + meanc*meanc + c1)*(devc/64.0 + devo/64.0 + c2));
+	return ((2.0f*meano*meanc + c1)*(corr/32.0f + c2))/((meano*meano + meanc*meanc + c1)*(devc/64.0f + devo/64.0f + c2));
 }
 
 static void ssim_after(xvid_plg_data_t* data, ssim_data_t* ssim){
@@ -407,13 +466,10 @@
 static int ssim_create(xvid_plg_create_t* create, void** handle){
 	ssim_data_t* ssim;
 	plg_ssim_param_t* param;
-	int cpu_flags;
 	param = (plg_ssim_param_t*) malloc(sizeof(plg_ssim_param_t));
 	*param = *((plg_ssim_param_t*) create->param);
 	ssim = (ssim_data_t*) malloc(sizeof(ssim_data_t));
 
-	cpu_flags = check_cpu_features();
-
 	ssim->func8x8 = lum_8x8_c;
 	ssim->func2x8 = lum_2x8_c;
 	ssim->consim = consim_c;
@@ -422,23 +478,36 @@
 
 	ssim->grid = param->acc;
 
-#if defined(ARCH_IS_IA32)
-	if((cpu_flags & XVID_CPU_MMX) && (param->acc > 0)){
-		ssim->func8x8 = lum_8x8_mmx;
-		ssim->consim = consim_mmx;
-	}
-	if((cpu_flags & XVID_CPU_SSE2) && (param->acc > 0)){
-		ssim->consim = consim_sse2;
+#if defined(ARCH_IS_IA32) || defined(ARCH_IS_X86_64)
+	{
+        int cpu_flags = (param->cpu_flags & XVID_CPU_FORCE) ? param->cpu_flags : check_cpu_features();
+
+		if((cpu_flags & XVID_CPU_MMX) && (param->acc > 0)){
+			ssim->func8x8 = lum_8x8_mmx;
+			ssim->consim = consim_mmx;
+		}
+		if((cpu_flags & XVID_CPU_SSE2) && (param->acc > 0)){
+			ssim->consim = consim_sse2;
+		}
 	}
 #endif
 
 	/*gaussian weigthing not implemented*/
+#if !defined(USE_INT_GAUSSIAN)
 	if(ssim->grid == 0){
 		ssim->grid = 1;
 		ssim->func8x8 = lum_8x8_gaussian;
 		ssim->func2x8 = NULL;
 		ssim->consim = consim_gaussian;
 	}
+#else
+	if(ssim->grid == 0){
+		ssim->grid = 1;
+		ssim->func8x8 = lum_8x8_gaussian_int;
+		ssim->func2x8 = NULL;
+		ssim->consim = consim_gaussian_int;
+	}
+#endif
 	if(ssim->grid > 4) ssim->grid = 4;
 
 	ssim->ssim_sum = 0.0;