Author: post
Date: 2012-01-31 17:22:10 +0100 (Tue, 31 Jan 2012)
New Revision: 4113
Modified:
trunk/plugins/resample/resample-avx.c
trunk/plugins/resample/resample-sse4.c
trunk/plugins/resample/resample.c
Log:
Resampler: Fix and enable SSE4+AVX code.
Modified: trunk/plugins/resample/resample-avx.c
===================================================================
--- trunk/plugins/resample/resample-avx.c 2012-01-29 15:29:03 UTC (rev
4112)
+++ trunk/plugins/resample/resample-avx.c 2012-01-31 16:22:10 UTC (rev
4113)
@@ -141,7 +141,7 @@
for (k=0; k<fir_filter_size; ++k)
{
gfloat total3 = total2 + lanczos_weight((start_pos+k -
ok_pos) * filter_step) / total;
- weights[i*fir_filter_size+k] = ((gint)
(total3*FPScale+0.5) - (gint) (total2*FPScale+0.5)) & 0xffff;
+ weights[i*fir_filter_size+k] = ((gint)
(total3*FPScale+0.5) - (gint) (total2*FPScale+0.5));
total2 = total3;
}
@@ -210,7 +210,7 @@
acc3 = _mm_add_epi32(acc3, src3i);
}
- /* Add rounder and subtract 32768 */
+ /* Add rounder*/
acc1_h = _mm_add_epi32(acc1_h, add_32);
acc2_h = _mm_add_epi32(acc2_h, add_32);
acc3_h = _mm_add_epi32(acc3_h, add_32);
@@ -317,7 +317,7 @@
for (k=0; k<fir_filter_size; ++k)
{
gfloat total3 = total2 + lanczos_weight((start_pos+k -
ok_pos) * filter_step) / total;
- weights[i*fir_filter_size+k] = ((gint)
(total3*FPScale+0.5) - (gint) (total2*FPScale+0.5)) & 0xffff;
+ weights[i*fir_filter_size+k] = ((gint)
(total3*FPScale+0.5) - (gint) (total2*FPScale+0.5));
total2 = total3;
}
Modified: trunk/plugins/resample/resample-sse4.c
===================================================================
--- trunk/plugins/resample/resample-sse4.c 2012-01-29 15:29:03 UTC (rev
4112)
+++ trunk/plugins/resample/resample-sse4.c 2012-01-31 16:22:10 UTC (rev
4113)
@@ -141,7 +141,7 @@
for (k=0; k<fir_filter_size; ++k)
{
gfloat total3 = total2 + lanczos_weight((start_pos+k -
ok_pos) * filter_step) / total;
- weights[i*fir_filter_size+k] = ((gint)
(total3*FPScale+0.5) - (gint) (total2*FPScale+0.5)) & 0xffff;
+ weights[i*fir_filter_size+k] = ((gint)
(total3*FPScale+0.5) - (gint) (total2*FPScale+0.5));
total2 = total3;
}
@@ -210,7 +210,7 @@
acc3 = _mm_add_epi32(acc3, src3i);
}
- /* Add rounder and subtract 32768 */
+ /* Add rounder */
acc1_h = _mm_add_epi32(acc1_h, add_32);
acc2_h = _mm_add_epi32(acc2_h, add_32);
acc3_h = _mm_add_epi32(acc3_h, add_32);
@@ -317,7 +317,7 @@
for (k=0; k<fir_filter_size; ++k)
{
gfloat total3 = total2 + lanczos_weight((start_pos+k -
ok_pos) * filter_step) / total;
- weights[i*fir_filter_size+k] = ((gint)
(total3*FPScale+0.5) - (gint) (total2*FPScale+0.5)) & 0xffff;
+ weights[i*fir_filter_size+k] = ((gint)
(total3*FPScale+0.5) - (gint) (total2*FPScale+0.5));
total2 = total3;
}
Modified: trunk/plugins/resample/resample.c
===================================================================
--- trunk/plugins/resample/resample.c 2012-01-29 15:29:03 UTC (rev 4112)
+++ trunk/plugins/resample/resample.c 2012-01-31 16:22:10 UTC (rev 4113)
@@ -328,20 +328,16 @@
if (t->input->h != t->output->h)
{
gboolean sse2_available = !!(rs_detect_cpu_features() &
RS_CPU_FLAG_SSE2);
-#if 0 // FIXME: Test and enable
gboolean sse4_available = !!(rs_detect_cpu_features() &
RS_CPU_FLAG_SSE4_1);
gboolean avx_available = !!(rs_detect_cpu_features() &
RS_CPU_FLAG_AVX);
-#endif
if (t->use_fast)
ResizeV_fast(t);
else if (t->use_compatible)
ResizeV_compatible(t);
-#if 0 // FIXME: Test and enable
else if (avx_available)
ResizeV_AVX(t);
else if (sse4_available)
ResizeV_SSE4(t);
-#endif
else if (sse2_available)
ResizeV_SSE2(t);
else
_______________________________________________
Rawstudio-commit mailing list
[email protected]
http://rawstudio.org/cgi-bin/mailman/listinfo/rawstudio-commit
