#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
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) {
end += 4 - (num % 4);
}
}
+#endif
pd[x].start = start;
pd[x].end = end;
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;
: "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)
ch = (unsigned char)acc;
*dptr++ = ch;
#endif
- *dptr++ = (unsigned char)result;
}
ch = dptr[-1];
for ( ; x < dstride; ++x) {
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
"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;
}