From: sgunderson@bigfoot.com <>
Date: Mon, 4 Feb 2008 01:11:09 +0000 (+0100)
Subject: Make the SSE code optional.
X-Git-Url: https://git.sesse.net/?p=qscale;a=commitdiff_plain;h=f9fa8c079cf510206fa4f3cf37f0bf7208c0ca85

Make the SSE code optional.
---

diff --git a/qscale.c b/qscale.c
index ca7c7dd..d89c8b9 100644
--- a/qscale.c
+++ b/qscale.c
@@ -23,8 +23,20 @@
 #include <stdlib.h>
 #include "jpeglib.h"
 
+/* The number of pixels to process at a time when scaling vertically. */
 #define CACHE_LINE_FACTOR 16
 
+/* Whether to use SSE for horizontal scaling or not (requires SSE3). */
+#define USE_HORIZONTAL_SSE 1
+
+/* Whether to use SSE for vertical scaling or not (requires only SSE1). */
+#define USE_VERTICAL_SSE 1
+
+#if USE_VERTICAL_SSE
+#undef CACHE_LINE_FACTOR
+#define CACHE_LINE_FACTOR 16
+#endif
+
 #ifndef M_PI
 #define M_PI 3.14159265358979323846264
 #endif
@@ -84,6 +96,7 @@ void hscale(float *pix, unsigned char *npix, unsigned w, unsigned h, unsigned nw
 			end = w - 1;
 		}
 
+#if USE_HORIZONTAL_SSE
 		/* round up so we get a multiple of four for the SSE code */
 		int num = (end - start + 1);
 		if (num % 4 != 0) {
@@ -96,6 +109,7 @@ void hscale(float *pix, unsigned char *npix, unsigned w, unsigned h, unsigned nw
 				end += 4 - (num % 4);
 			}
 		}
+#endif
 
 		pd[x].start = start;
 		pd[x].end = end;
@@ -123,6 +137,7 @@ void hscale(float *pix, unsigned char *npix, unsigned w, unsigned h, unsigned nw
 		unsigned char *dptr = npix + y*dstride;
 		unsigned char ch;
 		for (x = 0; x < nw; ++x) {
+#if USE_HORIZONTAL_SSE
 			int result;
 			float acc;
 			long tmp;
@@ -154,7 +169,16 @@ void hscale(float *pix, unsigned char *npix, unsigned w, unsigned h, unsigned nw
 				: "memory", "xmm1", "xmm2"
 			);
 
-#if 0
+			*dptr++ = (unsigned char)result;
+#else
+			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++;
+			}
+
 			if (acc < 0.0)
 				ch = 0;
 			else if (acc > 255.0)
@@ -163,7 +187,6 @@ void hscale(float *pix, unsigned char *npix, unsigned w, unsigned h, unsigned nw
 				ch = (unsigned char)acc;
 			*dptr++ = ch;
 #endif
-			*dptr++ = (unsigned char)result;
 		}
 		ch = dptr[-1];
 		for ( ; x < dstride; ++x) {
@@ -221,25 +244,7 @@ void vscale(unsigned char *pix, float *npix, unsigned w, unsigned h, unsigned nh
 		unsigned char *sptr = pix + x;
 		float *dptr = npix + x;
 		for (y = 0; y < nh; ++y) {
-#if 0
-			int i;
-			float acc[CACHE_LINE_FACTOR];
-			for (i = 0; i < CACHE_LINE_FACTOR; ++i)
-				acc[i] = 0.0;
-			float *cf = &coeffs[pd[y].startcoeff];
-			unsigned sy;
-		
-			for (sy = pd[y].start; sy <= pd[y].end; ++sy) {
-				for (i = 0; i < CACHE_LINE_FACTOR; ++i) {
-					acc[i] += sptr[sy * w + i] * *cf;
-				}
-				++cf;
-			}
-
-			for (i = 0; i < CACHE_LINE_FACTOR; ++i) {
-				dptr[i] = acc[i];
-			}
-#else
+#if USE_VERTICAL_SSE
 			/*
 			 * xmm0 - xmm3: acc[0..15]
 			 * xmm4: current filter coefficient
@@ -327,6 +332,24 @@ void vscale(unsigned char *pix, float *npix, unsigned w, unsigned h, unsigned nh
 				"r" (dptr)                           /* 4: dstptr base */
 				: "memory", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
 			);
+#else
+			int i;
+			float acc[CACHE_LINE_FACTOR];
+			for (i = 0; i < CACHE_LINE_FACTOR; ++i)
+				acc[i] = 0.0;
+			float *cf = &coeffs[pd[y].startcoeff];
+			unsigned sy;
+		
+			for (sy = pd[y].start; sy <= pd[y].end; ++sy) {
+				for (i = 0; i < CACHE_LINE_FACTOR; ++i) {
+					acc[i] += sptr[sy * w + i] * *cf;
+				}
+				++cf;
+			}
+
+			for (i = 0; i < CACHE_LINE_FACTOR; ++i) {
+				dptr[i] = acc[i];
+			}
 #endif
 			dptr += dstride;
 		}