Significant optimizations of MCT, DWT, MQ and T1 modules by Peter Wimmer (thanks...

[openjpeg.git] / libopenjpeg / mct.c

diff --git a/libopenjpeg/mct.c b/libopenjpeg/mct.c
index ca21744f3e48ec6d1c50e05edfc61aa6d271be95..870993b06d250b01f4a8163b960dd00934190224 100644 (file)
--- a/libopenjpeg/mct.c
+++ b/libopenjpeg/mct.c
@@ -29,6 +29,10 @@
  * POSSIBILITY OF SUCH DAMAGE.
  */
 
+#ifdef __SSE__
+#include <xmmintrin.h>
+#endif
+
 #include "opj_includes.h"
 
 /* <summary> */
@@ -127,6 +131,44 @@ void mct_decode_real(
                int n)
 {
        int i;
+#ifdef __SSE__
+       __m128 vrv, vgu, vgv, vbu;
+       vrv = _mm_set1_ps(1.402f);
+       vgu = _mm_set1_ps(0.34413f);
+       vgv = _mm_set1_ps(0.71414f);
+       vbu = _mm_set1_ps(1.772f);
+       for (i = 0; i < (n >> 3); ++i) {
+               __m128 vy, vu, vv;
+               __m128 vr, vg, vb;
+
+               vy = _mm_load_ps(c0);
+               vu = _mm_load_ps(c1);
+               vv = _mm_load_ps(c2);
+               vr = _mm_add_ps(vy, _mm_mul_ps(vv, vrv));
+               vg = _mm_sub_ps(_mm_sub_ps(vy, _mm_mul_ps(vu, vgu)), _mm_mul_ps(vv, vgv));
+               vb = _mm_add_ps(vy, _mm_mul_ps(vu, vbu));
+               _mm_store_ps(c0, vr);
+               _mm_store_ps(c1, vg);
+               _mm_store_ps(c2, vb);
+               c0 += 4;
+               c1 += 4;
+               c2 += 4;
+
+               vy = _mm_load_ps(c0);
+               vu = _mm_load_ps(c1);
+               vv = _mm_load_ps(c2);
+               vr = _mm_add_ps(vy, _mm_mul_ps(vv, vrv));
+               vg = _mm_sub_ps(_mm_sub_ps(vy, _mm_mul_ps(vu, vgu)), _mm_mul_ps(vv, vgv));
+               vb = _mm_add_ps(vy, _mm_mul_ps(vu, vbu));
+               _mm_store_ps(c0, vr);
+               _mm_store_ps(c1, vg);
+               _mm_store_ps(c2, vb);
+               c0 += 4;
+               c1 += 4;
+               c2 += 4;
+       }
+       n &= 7;
+#endif
        for(i = 0; i < n; ++i) {
                float y = c0[i];
                float u = c1[i];