Do rounding.

diff --git a/qscale.c b/qscale.c
index 7943f90bc58f5dc8aa372ea5c17c9ee52d3f031e..f3f0fa95ed89cbd585244ac1c3ff8999ca691e5a 100644 (file)
--- a/qscale.c
+++ b/qscale.c
@@ -101,25 +101,28 @@ 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) {
+                       int result;
                        float acc;
                        int tmp;
                        static const float low = 0.0, high = 255.0;
                        asm (
-                               "pxor %0, %0               \n"
-                               "xor %1, %1                \n"
+                               "pxor %1, %1               \n"
+                               "xor %2, %2                \n"
                                ".lbl2:                    \n"
-                               "movups (%3,%1),%%xmm1     \n"
-                               "movups (%2,%1),%%xmm2     \n"
+                               "movups (%4,%2),%%xmm1     \n"
+                               "movups (%3,%2),%%xmm2     \n"
                                "mulps %%xmm2,%%xmm1       \n"
-                               "addps %%xmm1,%0           \n"
-                               "addl $16,%1              \n"
-                               "dec %4                    \n"
+                               "addps %%xmm1,%1           \n"
+                               "addl $16,%2               \n"
+                               "dec %5                    \n"
                                "jnz .lbl2                 \n"
-                               "haddps %0,%0              \n"
-                               "haddps %0,%0              \n"
-                               "maxss %5,%0               \n"
-                               "minss %6,%0               \n"
-                               : "=&x" (acc),
+                               "haddps %1,%1              \n"
+                               "haddps %1,%1              \n"
+                               "maxss %6,%1               \n"
+                               "minss %7,%1               \n"
+                               "cvtss2si %1,%0            \n"
+                               : "=r" (result),
+                                 "=&x" (acc),
                                  "=&r" (tmp)
                                : "r" (&coeffs[pd[x].startcoeff]),
                                  "r" (&sptr[pd[x].start]),
@@ -138,7 +141,7 @@ void hscale(float *pix, unsigned char *npix, unsigned w, unsigned h, unsigned nw
                                ch = (unsigned char)acc;
                        *dptr++ = ch;
 #endif
-                       *dptr++ = (unsigned char)acc;
+                       *dptr++ = (unsigned char)result;
                }
                ch = dptr[-1];
                for ( ; x < dstride; ++x) {