liz/rawTherapee
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Add Ingo's optimizations

Browse Source 
Add Ingo's SSE optimizations and clean up the non-SSE part of `getRGB()`
with `intp()`.
This commit is contained in:
Flössie
2016-04-26 20:48:11 +02:00
parent 56f8ea086c
commit 78c08e9e5c
1 changed files with 48 additions and 57 deletions
Download Patch File Download Diff File Expand all files Collapse all files

105
rtengine/clutstore.cc
View File

@@ -1,11 +1,9 @@
#include <algorithm> #include <algorithm>
#ifdef __SSE2__
#include <xmmintrin.h>
#endif
#include "clutstore.h" #include "clutstore.h"
#include "opthelper.h"
#include "rt_math.h"
#include "imagefloat.h" #include "imagefloat.h"
#include "stdimagesource.h" #include "stdimagesource.h"
#include "../rtgui/options.h" #include "../rtgui/options.h"
@@ -76,16 +74,14 @@ bool loadFile(
return res; return res;
} }
inline void posToIndex(unsigned int pos, size_t (&index)[2])
{
index[0] = static_cast<size_t>(pos) * 4;
index[1] = static_cast<size_t>(pos + 1) * 4;
}
#ifdef __SSE2__ #ifdef __SSE2__
inline __m128 getClutValue(const AlignedBuffer<std::uint16_t>& clut_image, size_t index) inline vfloat getClutValue(const AlignedBuffer<std::uint16_t>& clut_image, size_t index)
{ {
#ifdef __SSE4_1__
return _mm_cvtepi32_ps(_mm_cvtepu16_epi32(*reinterpret_cast<const __m128i*>(clut_image.data + index)));
#else
return _mm_cvtpu16_ps(*reinterpret_cast<const __m64*>(clut_image.data + index)); return _mm_cvtpu16_ps(*reinterpret_cast<const __m64*>(clut_image.data + index));
#endif
} }
#endif #endif
@@ -185,79 +181,74 @@ void rtengine::HaldCLUT::getRGB(float r, float g, float b, float out_rgbx[4]) co
g = g * flevel_minus_one - green; g = g * flevel_minus_one - green;
b = b * flevel_minus_one - blue; b = b * flevel_minus_one - blue;
size_t index[2]; size_t index = color * 4;
posToIndex(color, index);
float tmp1[4] ALIGNED16; float tmp1[4] ALIGNED16;
tmp1[0] = clut_image.data[index[0]] * (1 - r) + clut_image.data[index[1]] * r; tmp1[0] = intp<float>(r, clut_image.data[index + 4], clut_image.data[index]);
tmp1[1] = clut_image.data[index[0] + 1] * (1 - r) + clut_image.data[index[1] + 1] * r; tmp1[1] = intp<float>(r, clut_image.data[index + 5], clut_image.data[index + 1]);
tmp1[2] = clut_image.data[index[0] + 2] * (1 - r) + clut_image.data[index[1] + 2] * r; tmp1[2] = intp<float>(r, clut_image.data[index + 6], clut_image.data[index + 2]);
posToIndex(color + level, index); index = (color + level) * 4;
float tmp2[4] ALIGNED16; float tmp2[4] ALIGNED16;
tmp2[0] = clut_image.data[index[0]] * (1 - r) + clut_image.data[index[1]] * r; tmp2[0] = intp<float>(r, clut_image.data[index + 4], clut_image.data[index]);
tmp2[1] = clut_image.data[index[0] + 1] * (1 - r) + clut_image.data[index[1] + 1] * r; tmp2[1] = intp<float>(r, clut_image.data[index + 5], clut_image.data[index + 1]);
tmp2[2] = clut_image.data[index[0] + 2] * (1 - r) + clut_image.data[index[1] + 2] * r; tmp2[2] = intp<float>(r, clut_image.data[index + 6], clut_image.data[index + 2]);
out_rgbx[0] = tmp1[0] * (1 - g) + tmp2[0] * g; out_rgbx[0] = intp<float>(g, tmp2[0], tmp1[0]);
out_rgbx[1] = tmp1[1] * (1 - g) + tmp2[1] * g; out_rgbx[1] = intp<float>(g, tmp2[1], tmp1[1]);
out_rgbx[2] = tmp1[2] * (1 - g) + tmp2[2] * g; out_rgbx[2] = intp<float>(g, tmp2[2], tmp1[2]);
posToIndex(color + level_square, index); index = (color + level_square) * 4;
tmp1[0] = clut_image.data[index[0]] * (1 - r) + clut_image.data[index[1]] * r; tmp1[0] = intp<float>(r, clut_image.data[index + 4], clut_image.data[index]);
tmp1[1] = clut_image.data[index[0] + 1] * (1 - r) + clut_image.data[index[1] + 1] * r; tmp1[1] = intp<float>(r, clut_image.data[index + 5], clut_image.data[index + 1]);
tmp1[2] = clut_image.data[index[0] + 2] * (1 - r) + clut_image.data[index[1] + 2] * r; tmp1[2] = intp<float>(r, clut_image.data[index + 6], clut_image.data[index + 2]);
posToIndex(color + level + level_square, index); index = (color + level + level_square) * 4;
tmp2[0] = clut_image.data[index[0]] * (1 - r) + clut_image.data[index[1]] * r; tmp2[0] = intp<float>(r, clut_image.data[index + 4], clut_image.data[index]);
tmp2[1] = clut_image.data[index[0] + 1] * (1 - r) + clut_image.data[index[1] + 1] * r; tmp2[1] = intp<float>(r, clut_image.data[index + 5], clut_image.data[index + 1]);
tmp2[2] = clut_image.data[index[0] + 2] * (1 - r) + clut_image.data[index[1] + 2] * r; tmp2[2] = intp<float>(r, clut_image.data[index + 6], clut_image.data[index + 2]);
tmp1[0] = tmp1[0] * (1 - g) + tmp2[0] * g; tmp1[0] = intp<float>(g, tmp2[0], tmp1[0]);
tmp1[1] = tmp1[1] * (1 - g) + tmp2[1] * g; tmp1[1] = intp<float>(g, tmp2[1], tmp1[1]);
tmp1[2] = tmp1[2] * (1 - g) + tmp2[2] * g; tmp1[2] = intp<float>(g, tmp2[2], tmp1[2]);
out_rgbx[0] = out_rgbx[0] * (1 - b) + tmp1[0] * b; out_rgbx[0] = intp<float>(b, tmp1[0], out_rgbx[0]);
out_rgbx[1] = out_rgbx[1] * (1 - b) + tmp1[1] * b; out_rgbx[1] = intp<float>(b, tmp1[1], out_rgbx[1]);
out_rgbx[2] = out_rgbx[2] * (1 - b) + tmp1[2] * b; out_rgbx[2] = intp<float>(b, tmp1[2], out_rgbx[2]);
#else #else
const __m128 v_rgb = _mm_set_ps(0.0f, b, g, r) *_mm_load_ps1(&flevel_minus_one) - _mm_set_ps(0.0f, blue, green, red); const vfloat v_tmp = _mm_set_ps(0.0f, b, g, r) * _mm_load_ps1(&flevel_minus_one);
const vfloat v_rgb = v_tmp - _mm_cvtepi32_ps(_mm_cvttps_epi32(_mm_min_ps(_mm_load_ps1(&flevel_minus_two), v_tmp)));
size_t index[2]; size_t index = color * 4;
posToIndex(color, index);
const __m128 v_r = _mm_shuffle_ps(v_rgb, v_rgb, 0x00); const vfloat v_r = PERMUTEPS(v_rgb, 0x00);
__m128 v_cv0 = getClutValue(clut_image, index[0]); vfloat v_tmp1 = vintpf(v_r, getClutValue(clut_image, index + 4), getClutValue(clut_image, index));
__m128 v_tmp1 = v_r * (getClutValue(clut_image, index[1]) - v_cv0) + v_cv0;
posToIndex(color + level, index); index = (color + level) * 4;
v_cv0 = getClutValue(clut_image, index[0]); vfloat v_tmp2 = vintpf(v_r, getClutValue(clut_image, index + 4), getClutValue(clut_image, index));
__m128 v_tmp2 = v_r * (getClutValue(clut_image, index[1]) - v_cv0) + v_cv0;
const __m128 v_g = _mm_shuffle_ps(v_rgb, v_rgb, 0x55); const vfloat v_g = PERMUTEPS(v_rgb, 0x55);
__m128 v_out = v_g * (v_tmp2 - v_tmp1) + v_tmp1; vfloat v_out = vintpf(v_g, v_tmp2, v_tmp1);
posToIndex(color + level_square, index); index = (color + level_square) * 4;
v_cv0 = getClutValue(clut_image, index[0]); v_tmp1 = vintpf(v_r, getClutValue(clut_image, index + 4), getClutValue(clut_image, index));
v_tmp1 = v_r * (getClutValue(clut_image, index[1]) - v_cv0) + v_cv0;
posToIndex(color + level + level_square, index); index = (color + level + level_square) * 4;
v_cv0 = getClutValue(clut_image, index[0]); v_tmp2 = vintpf(v_r, getClutValue(clut_image, index + 4), getClutValue(clut_image, index));
v_tmp2 = v_r * (getClutValue(clut_image, index[1]) - v_cv0) + v_cv0;
v_tmp1 = v_g * (v_tmp2 - v_tmp1) + v_tmp1; v_tmp1 = vintpf(v_g, v_tmp2, v_tmp1);
const __m128 v_b = _mm_shuffle_ps(v_rgb, v_rgb, 0xAA); const vfloat v_b = PERMUTEPS(v_rgb, 0xAA);
_mm_store_ps(out_rgbx, v_b * (v_tmp1 - v_out) + v_out); _mm_store_ps(out_rgbx, vintpf(v_b, v_tmp1, v_out));
#endif #endif
} }