Author: post
Date: 2009-06-27 16:44:19 +0200 (Sat, 27 Jun 2009)
New Revision: 2514
Modified:
trunk/plugins/denoise/floatplanarimage-x86.cpp
trunk/plugins/denoise/floatplanarimage.cpp
Log:
Denoise: Enabled RGB to YUV assembler in x64 mode. Slightly faster on Conroe,
should favour Nehalem CPUs.
Modified: trunk/plugins/denoise/floatplanarimage-x86.cpp
===================================================================
--- trunk/plugins/denoise/floatplanarimage-x86.cpp 2009-06-27 12:00:24 UTC
(rev 2513)
+++ trunk/plugins/denoise/floatplanarimage-x86.cpp 2009-06-27 14:44:19 UTC
(rev 2514)
@@ -17,18 +17,17 @@
*/
#include "floatplanarimage.h"
-#if defined (__i386__) || defined (__x86_64__)
+#if defined (__x86_64__)
+// Only 64 bits, and only if pixelsize is 4
void FloatPlanarImage::unpackInterleavedYUV_SSE( const ImgConvertJob* j )
-{
+{
RS_IMAGE16* image = j->rs;
float* temp = p[0]->data;
- temp[0] = (0.299f* WB_R_CORR); temp[4] = 0.587f; temp[8] = (0.114f *
WB_B_CORR); temp[3] = 0.0f;
- temp[1] = (-0.169f* WB_R_CORR); temp[5] = -0.331f; temp[9] = (0.499 *
WB_B_CORR); temp[7] = 0.0f;
- temp[2] = (0.499f* WB_R_CORR); temp[6] = -0.418f; temp[10] =(-0.0813f *
WB_B_CORR); temp[11] = 0.0f;
+ temp[0] = (0.299* WB_R_CORR); temp[4] = 0.587; temp[8] = (0.114 *
WB_B_CORR); temp[3] = 0.0f;
+ temp[1] = (-0.169* WB_R_CORR); temp[5] = -0.331; temp[9] = (0.499 *
WB_B_CORR); temp[7] = 0.0f;
+ temp[2] = (0.499* WB_R_CORR); temp[6] = -0.418; temp[10] =(-0.0813 *
WB_B_CORR); temp[11] = 0.0f;
- float* xfer = (float*)fftwf_malloc(3*sizeof(float));
-
asm volatile
(
"movaps (%0), %%xmm5\n" // R values
@@ -43,39 +42,75 @@
gfloat *Y = p[0]->getAt(ox, y+oy);
gfloat *Cb = p[1]->getAt(ox, y+oy);
gfloat *Cr = p[2]->getAt(ox, y+oy);
+ gint w = (1+image->w) >>1;
+ asm volatile
+ (
+ "unpack_next_pixel:\n"
+ "movq (%0), %%rax\n"
+ "movq %%rax, %%rbx\n"
+ "movq %%rax, %%rcx\n"
+ "and $65535, %%rax\n"
+ "shr $16, %%rbx\n"
+ "shr $32, %%rcx\n"
+ "and $65535, %%rbx\n"
+ "and $65535, %%rcx\n"
+ "movss (%4,%%rax,4), %%xmm0\n"
+ "movss (%4,%%rbx,4), %%xmm1\n"
+ "movss (%4,%%rcx,4), %%xmm2\n"
- for (int x=0; x<image->w; x++) {
- xfer[0] = shortToFloat[(*pix)]; // r
- xfer[1] = shortToFloat[(*(pix+1))]; // g
- xfer[2] = shortToFloat[(*(pix+2))]; // b
- asm volatile
- (
- "movss (%0), %%xmm0\n" // Move r into xmm0 (load 1 to avoid
StoreLoadForward pentalty)
- "movss 4(%0), %%xmm1\n" // Move g into xmm1
- "movss 8(%0), %%xmm2\n" // Move r into xmm2
- "shufps $0, %%xmm0, %%xmm0\n" // Splat r
- "shufps $0, %%xmm1, %%xmm1\n" // Splat g
- "shufps $0, %%xmm2, %%xmm2\n" // Splat b
- "mulps %%xmm5, %%xmm0\n" // Multiply R
- "mulps %%xmm6, %%xmm1\n" // Multiply G
- "mulps %%xmm7, %%xmm2\n" // Multiply B
- "addps %%xmm0, %%xmm1\n" // Add first
- "addps %%xmm1, %%xmm2\n" // Add second
- "shufps $85, %%xmm2, %%xmm0\n" // Move Cb into xmm0 lower (85 =
01010101)
- "movss %%xmm2, (%1)\n" // Store Y
- "movhlps %%xmm2, %%xmm1\n" // Move Cr into xmm1 low
- "movss %%xmm0, (%2)\n" // Store Cb
- "movss %%xmm1, (%3)\n" // Store Cr
+ "movq 8(%0), %%rax\n"
+ "movq %%rax, %%rbx\n"
+ "movq %%rax, %%rcx\n"
+ "and $65535, %%rax\n"
+ "shr $16, %%rbx\n"
+ "shr $32, %%rcx\n"
+ "and $65535, %%rbx\n"
+ "and $65535, %%rcx\n"
+ "shufps $0, %%xmm0, %%xmm0\n" // Splat r (1)
+ "movss (%4,%%rax,4), %%xmm10\n"
+ "shufps $0, %%xmm1, %%xmm1\n" // Splat g (1)
+ "movss (%4,%%rbx,4), %%xmm11\n"
+ "movss (%4,%%rcx,4), %%xmm12\n"
+
+ "shufps $0, %%xmm2, %%xmm2\n" // Splat b (1)
+ "mulps %%xmm5, %%xmm0\n" // Multiply R (1)
+ "shufps $0, %%xmm10, %%xmm10\n" // Splat r (2)
+ "mulps %%xmm6, %%xmm1\n" // Multiply G (1)
+ "shufps $0, %%xmm11, %%xmm11\n" // Splat g (2)
+ "mulps %%xmm7, %%xmm2\n" // Multiply B (1)
+ "shufps $0, %%xmm12, %%xmm12\n" // Splat b (2)
+ "mulps %%xmm5, %%xmm10\n" // Multiply R (2)
+ "addps %%xmm0, %%xmm1\n" // Add first (1)
+ "mulps %%xmm6, %%xmm11\n" // Multiply G (2)
+ "addps %%xmm1, %%xmm2\n" // Add second (1)
+ "mulps %%xmm7, %%xmm12\n" // Multiply B (2)
+
+ "addps %%xmm10, %%xmm11\n" // Add first (2)
+ "addps %%xmm11, %%xmm12\n" // Add second (2)
+
+ "movaps %%xmm2, %%xmm1\n" // Copy (1)
+ "movaps %%xmm12, %%xmm11\n" // Copy (2)
+
+ "unpcklps %%xmm12, %%xmm2\n" // Unpack (2) into (1) Y1,Y2, Cb1,
Cb2
+ "unpckhps %%xmm11, %%xmm1\n" // Unpack (2) into (1) Cr1,Cr2, xx,
xx
+
+ "movlps %%xmm2, (%1)\n" // Store Y
+ "movlps %%xmm1, (%3)\n" // Store Cr
+ "movhps %%xmm2, (%2)\n" // Store Cb
+
+ "add $16, %0\n"
+ "add $8, %1\n"
+ "add $8, %2\n"
+ "add $8, %3\n"
+ "dec %5\n"
+ "jnz unpack_next_pixel\n"
: // no output registers
- : "r" (&xfer[0]), "r" (Y), "r" (Cb), "r" (Cr)
- : // %0 %1 %2 %3
+ : "r" (pix), "r" (Y), "r" (Cb), "r" (Cr), "r" (&shortToFloat[0]), "r"
(w)
+ // %0 %1 %2 %3 %4
%5
+ : "%rax", "%rbx", "%rcx"
);
- Y++; Cb++; Cr++;
- pix += image->pixelsize;
- }
}
asm volatile ( "emms\n" );
- fftwf_free(xfer);
}
Modified: trunk/plugins/denoise/floatplanarimage.cpp
===================================================================
--- trunk/plugins/denoise/floatplanarimage.cpp 2009-06-27 12:00:24 UTC (rev
2513)
+++ trunk/plugins/denoise/floatplanarimage.cpp 2009-06-27 14:44:19 UTC (rev
2514)
@@ -153,11 +153,9 @@
{
RS_IMAGE16* image = j->rs;
-#if defined (__i386__) || defined (__x86_64__)
- // FIXME: Seems slower - and yellow cast - check??
-/* guint cpu = rs_detect_cpu_features();
- if (cpu & RS_CPU_FLAG_SSE)
- return unpackInterleavedYUV_SSE(j);*/
+#if defined (__x86_64__)
+ if (image->pixelsize == 4)
+ return unpackInterleavedYUV_SSE(j);
#endif
for (int y = j->start_y; y < j->end_y; y++ ) {
_______________________________________________
Rawstudio-commit mailing list
[email protected]
http://rawstudio.org/cgi-bin/mailman/listinfo/rawstudio-commit
