From 85edeac1ae18442aa7dc43074a2e4ddc31b6a731 Mon Sep 17 00:00:00 2001
From: "sgunderson@bigfoot.com" <>
Date: Sun, 3 Feb 2008 17:14:40 +0100
Subject: [PATCH] Add SSE the other way as well, requiring SSE3.

---
 qscale.c | 60 ++++++++++++++++++++++++++++++++++++++++++++------------
 1 file changed, 48 insertions(+), 12 deletions(-)

diff --git a/qscale.c b/qscale.c
index d5536bc..6d529bc 100644
--- a/qscale.c
+++ b/qscale.c
@@ -38,7 +38,7 @@ void hscale(float *pix, unsigned char *npix, unsigned w, unsigned h, unsigned nw
 	int size_coeffs = 8;
 	float *coeffs = (float *)malloc(size_coeffs * sizeof(float));
 	int num_coeffs = 0;
-	int x, y, sx;
+	int x, y;
 	double sf = (double)w / (double)nw;
 	double support = (w > nw) ? (3.0 * sf) : (3.0 / sf);
 
@@ -46,6 +46,7 @@ void hscale(float *pix, unsigned char *npix, unsigned w, unsigned h, unsigned nw
 	for (x = 0; x < nw; ++x) {
 		int start = ceil(x * sf - support);
 		int end = floor(x * sf + support);
+		int sx;
 		double sum = 0.0;
 
 		if (start < 0) {
@@ -55,6 +56,19 @@ void hscale(float *pix, unsigned char *npix, unsigned w, unsigned h, unsigned nw
 			end = w - 1;
 		}
 
+		// round up so we get a multiple of four for the SSE code
+		int num = (end - start + 1);
+		if (num % 4 != 0) {
+			// prefer aligning it if possible
+			if (start % 4 != 0 && start % 4 <= num % 4) {
+				num += start % 4;
+				start -= start % 4;
+			}
+			if (num % 4 != 0) {
+				end += 4 - (num % 4);
+			}
+		}
+
 		pd[x].start = start;
 		pd[x].end = end;
 		pd[x].startcoeff = num_coeffs;
@@ -79,16 +93,35 @@ void hscale(float *pix, unsigned char *npix, unsigned w, unsigned h, unsigned nw
 	for (y = 0; y < h; ++y) {
 		float *sptr = pix + y*w;
 		unsigned char *dptr = npix + y*dstride;
-		unsigned char ch = 0;
+		unsigned char ch;
 		for (x = 0; x < nw; ++x) {
-			float acc = 0.0;
-			float *cf = &coeffs[pd[x].startcoeff];
-			unsigned sx;
-			
-			for (sx = pd[x].start; sx <= pd[x].end; ++sx) {
-				acc += sptr[sx] * *cf++;
-			}
+			float acc;
+			static const float low = 0.0, high = 255.0;
+			asm (
+				"pxor %0, %0               \n"
+				"xor %%eax, %%eax          \n"
+				".lbl2:                    \n"
+				"movups (%2,%%eax),%%xmm1  \n"
+				"movups (%1,%%eax),%%xmm2  \n"
+				"mulps %%xmm2,%%xmm1       \n"
+				"addps %%xmm1,%0           \n"
+				"addl $16,%%eax            \n"
+				"dec %3                    \n"
+				"jnz .lbl2                 \n"
+				"haddps %0,%0              \n"
+				"haddps %0,%0              \n"
+				"maxss %4,%0               \n"
+				"minss %5,%0               \n"
+				: "=x" (acc)
+				: "r" (&coeffs[pd[x].startcoeff]),
+				  "r" (&sptr[pd[x].start]),
+				  "r" ((pd[x].end - pd[x].start + 1)/4),
+				  "m" (low),
+				  "m" (high)
+				: "eax", "xmm1", "xmm2"
+			);
 
+#if 0
 			if (acc < 0.0)
 				ch = 0;
 			else if (acc > 255.0)
@@ -96,7 +129,10 @@ void hscale(float *pix, unsigned char *npix, unsigned w, unsigned h, unsigned nw
 			else
 				ch = (unsigned char)acc;
 			*dptr++ = ch;
+#endif
+			*dptr++ = (unsigned char)acc;
 		}
+		ch = dptr[-1];
 		for ( ; x < dstride; ++x) {
 			*dptr++ = ch;
 		}
@@ -240,9 +276,9 @@ void vscale(unsigned char *pix, float *npix, unsigned w, unsigned h, unsigned nh
 				"addps %%xmm6, %%xmm3    \n"
 
 				/* move along, and loop */
-				"addl $4, %2             \n"
-				"addl %3, %0             \n"
-				"decl %1                 \n"
+				"add $4, %2              \n"
+				"add %3, %0              \n"
+				"dec %1                  \n"
 				"jnz .lbl                \n"
 
 				/* store the values */
-- 
2.39.2