mirrors/0ad
1
0
Fork 2
mirror of https://gitea.wildfiregames.com/0ad/0ad.git synced 2026-09-21 20:06:40 +00:00
Code Issues Packages Projects Releases Wiki Activity

color conversion: avoid type punning by returning SColor4ub directly. that required a non-extern-C declaration of the SSE function, so i replaced it with straightforward intrinsics.

Browse Source 
ia32: no longer needs #if since build system ensures it's not compiled
on amd64

This was SVN commit r6164.
This commit is contained in:
janwas
2008-06-30 19:08:29 +00:00
parent 10a6ef4e17
commit b79375a0c5
7 changed files with 58 additions and 72 deletions
Download Patch File Download Diff File Expand all files Collapse all files
source/graphics/Color.cpp
+38 -7
View File
@@ -12,25 +12,56 @@
#include "maths/MathUtil.h"
#include "graphics/SColor.h"
#include "lib/sysdep/arch/x86_x64/x86_x64.h"
static u32 fallback_ConvertRGBColorTo4ub(const RGBColor& src)
#if ARCH_IA32 || ARCH_AMD64
# include <emmintrin.h>
# include "lib/sysdep/arch/x86_x64/x86_x64.h"
#endif
static SColor4ub fallback_ConvertRGBColorTo4ub(const RGBColor& src)
{
SColor4ub result;
result.R=clamp(int(src.X*255),0,255);
result.G=clamp(int(src.Y*255),0,255);
result.B=clamp(int(src.Z*255),0,255);
result.A=0xff;
return *(u32*)&result;
return result;
}
// on IA32, this is replaced by an SSE assembly version in ia32.cpp
u32 (*ConvertRGBColorTo4ub)(const RGBColor& src) = fallback_ConvertRGBColorTo4ub;
SColor4ub (*ConvertRGBColorTo4ub)(const RGBColor& src) = fallback_ConvertRGBColorTo4ub;
// Assembler-optimized function for color conversion
#if ARCH_IA32
EXTERN_C u32 sse_ConvertRGBColorTo4ub(const RGBColor& src);
#if ARCH_IA32 || ARCH_AMD64
static SColor4ub sse_ConvertRGBColorTo4ub(const RGBColor& src)
{
const __m128 zero = _mm_setzero_ps();
const __m128 _255 = _mm_set_ss(255.0f);
__m128 r = _mm_load_ss(&src.X);
__m128 g = _mm_load_ss(&src.Y);
__m128 b = _mm_load_ss(&src.Z);
// C = min(255, 255*max(C, 0)) ( == clamp(255*C, 0, 255) )
r = _mm_max_ss(r, zero);
g = _mm_max_ss(g, zero);
b = _mm_max_ss(b, zero);
r = _mm_mul_ss(r, _255);
g = _mm_mul_ss(g, _255);
b = _mm_mul_ss(b, _255);
r = _mm_min_ss(r, _255);
g = _mm_min_ss(g, _255);
b = _mm_min_ss(b, _255);
// convert to integer and combine channels using bit logic
int ri = _mm_cvtss_si32(r);
int gi = _mm_cvtss_si32(g);
int bi = _mm_cvtss_si32(b);
return SColor4ub(ri, gi, bi, 0xFF);
}
#endif
void ColorActivateFastImpl()
@@ -38,7 +69,7 @@ void ColorActivateFastImpl()
if(0)
{
}
#if ARCH_IA32
#if ARCH_IA32 || ARCH_AMD64
else if (x86_x64_cap(X86_X64_CAP_SSE))
{
ConvertRGBColorTo4ub = sse_ConvertRGBColorTo4ub;