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

dump SSE code in boxblur2 because new code is faster with auto-vectorization, also pad bufferwidth for boxblur2 to a multiple of 16

Browse Source
This commit is contained in:
Ingo Weyrich
2019-07-08 23:57:24 +02:00
parent 4d6c3f2ce2
commit 54fdbe41ea
1 changed files with 52 additions and 137 deletions
Download Patch File Download Diff File Expand all files Collapse all files

189
rtengine/hilite_recon.cc
View File

@@ -22,6 +22,10 @@
// //
//////////////////////////////////////////////////////////////// ////////////////////////////////////////////////////////////////
#ifndef NDEBUG
#include <cassert>
#endif
#include <cstddef> #include <cstddef>
#include <cmath> #include <cmath>
#include "array2D.h" #include "array2D.h"
@@ -32,8 +36,11 @@
#include "StopWatch.h" #include "StopWatch.h"
namespace { namespace {
void boxblur2(float** src, float** dst, float** temp, int startY, int startX, int H, int W, int box ) void boxblur2(const float* const* src, float** dst, float** temp, int startY, int startX, int H, int W, int box)
{ {
constexpr int numCols = 16;
assert((W % numCols) == 0);
//box blur image channel; box size = 2*box+1 //box blur image channel; box size = 2*box+1
//horizontal blur //horizontal blur
#ifdef _OPENMP #ifdef _OPENMP
@@ -63,148 +70,52 @@ void boxblur2(float** src, float** dst, float** temp, int startY, int startX, in
} }
} }
#ifdef __SSE2__
//vertical blur //vertical blur
#ifdef _OPENMP #ifdef _OPENMP
#pragma omp parallel #pragma omp parallel
#endif #endif
{ {
float len = box + 1; float tempvalN[numCols] ALIGNED64;
vfloat lenv = F2V( len );
vfloat lenp1v = F2V( len + 1.0f );
vfloat onev = F2V( 1.0f );
vfloat tempv, temp2v;
#ifdef _OPENMP #ifdef _OPENMP
#pragma omp for nowait #pragma omp for
#endif #endif
for (int col = 0; col < W - numCols + 1; col += numCols) {
for (int col = 0; col < W - 7; col += 8) { float len = box + 1;
tempv = LVFU(temp[0][col]) / lenv; for(int n = 0; n < numCols; n++) {
temp2v = LVFU(temp[0][col + 4]) / lenv; tempvalN[n] = temp[0][col + n] / len;
}
for (int i = 1; i <= box; i++) { for (int i = 1; i <= box; i++) {
tempv = tempv + LVFU(temp[i][col]) / lenv; for(int n = 0; n < numCols; n++) {
temp2v = temp2v + LVFU(temp[i][col + 4]) / lenv; tempvalN[n] += temp[i][col + n] / len;
}
}
for(int n = 0; n < numCols; n++) {
dst[0][col + n] = tempvalN[n];
} }
_mm_storeu_ps( &dst[0][col], tempv);
_mm_storeu_ps( &dst[0][col + 4], temp2v);
for (int row = 1; row <= box; row++) { for (int row = 1; row <= box; row++) {
tempv = (tempv * lenv + LVFU(temp[(row + box)][col])) / lenp1v; for(int n = 0; n < numCols; n++) {
temp2v = (temp2v * lenv + LVFU(temp[(row + box)][col + 4])) / lenp1v; tempvalN[n] = (tempvalN[n] * len + temp[(row + box)][col + n]) / (len + 1);
_mm_storeu_ps( &dst[row][col], tempv); dst[row][col + n] = tempvalN[n];
_mm_storeu_ps( &dst[row][col + 4], temp2v); }
lenv = lenp1v; len ++;
lenp1v = lenp1v + onev;
} }
const float rlen = 1.f / len;
for (int row = box + 1; row < H - box; row++) { for (int row = box + 1; row < H - box; row++) {
tempv = tempv + (LVFU(temp[(row + box)][col]) - LVFU(temp[(row - box - 1)][col])) / lenv; for(int n = 0; n < numCols; n++) {
temp2v = temp2v + (LVFU(temp[(row + box)][col + 4]) - LVFU(temp[(row - box - 1)][col + 4])) / lenv; tempvalN[n] = tempvalN[n] + (temp[(row + box)][col + n] - temp[(row - box - 1)][col + n]) * rlen;
_mm_storeu_ps( &dst[row][col], tempv); dst[row][col + n] = tempvalN[n];
_mm_storeu_ps( &dst[row][col + 4], temp2v); }
} }
for (int row = H - box; row < H; row++) { for (int row = H - box; row < H; row++) {
lenp1v = lenv; for(int n = 0; n < numCols; n++) {
lenv = lenv - onev; tempvalN[n] = (dst[(row - 1)][col + n] * len - temp[(row - box - 1)][col + n]) / (len - 1);
tempv = (tempv * lenp1v - LVFU(temp[(row - box - 1)][col])) / lenv; dst[row][col + n] = tempvalN[n];
temp2v = (temp2v * lenp1v - LVFU(temp[(row - box - 1)][col + 4])) / lenv;
_mm_storeu_ps( &dst[row][col], tempv );
_mm_storeu_ps( &dst[row][col + 4], temp2v );
}
}
#ifdef _OPENMP
#pragma omp single
#endif
{
for (int col = W - (W % 8); col < W - 3; col += 4) {
tempv = LVFU(temp[0][col]) / lenv;
for (int i = 1; i <= box; i++) {
tempv = tempv + LVFU(temp[i][col]) / lenv;
}
_mm_storeu_ps( &dst[0][col], tempv);
for (int row = 1; row <= box; row++) {
tempv = (tempv * lenv + LVFU(temp[(row + box)][col])) / lenp1v;
_mm_storeu_ps( &dst[row][col], tempv);
lenv = lenp1v;
lenp1v = lenp1v + onev;
}
for (int row = box + 1; row < H - box; row++) {
tempv = tempv + (LVFU(temp[(row + box)][col]) - LVFU(temp[(row - box - 1)][col])) / lenv;
_mm_storeu_ps( &dst[row][col], tempv);
}
for (int row = H - box; row < H; row++) {
lenp1v = lenv;
lenv = lenv - onev;
tempv = (tempv * lenp1v - LVFU(temp[(row - box - 1)][col])) / lenv;
_mm_storeu_ps( &dst[row][col], tempv );
}
}
for (int col = W - (W % 4); col < W; col++) {
int len = box + 1;
dst[0][col] = temp[0][col] / len;
for (int i = 1; i <= box; i++) {
dst[0][col] += temp[i][col] / len;
}
for (int row = 1; row <= box; row++) {
dst[row][col] = (dst[(row - 1)][col] * len + temp[(row + box)][col]) / (len + 1);
len ++;
}
for (int row = box + 1; row < H - box; row++) {
dst[row][col] = dst[(row - 1)][col] + (temp[(row + box)][col] - temp[(row - box - 1)][col]) / len;
}
for (int row = H - box; row < H; row++) {
dst[row][col] = (dst[(row - 1)][col] * len - temp[(row - box - 1)][col]) / (len - 1);
len --;
} }
len --;
} }
} }
} }
#else
//vertical blur
#ifdef _OPENMP
#pragma omp parallel for
#endif
for (int col = 0; col < W; col++) {
int len = box + 1;
dst[0][col] = temp[0][col] / len;
for (int i = 1; i <= box; i++) {
dst[0][col] += temp[i][col] / len;
}
for (int row = 1; row <= box; row++) {
dst[row][col] = (dst[(row - 1)][col] * len + temp[(row + box)][col]) / (len + 1);
len ++;
}
for (int row = box + 1; row < H - box; row++) {
dst[row][col] = dst[(row - 1)][col] + (temp[(row + box)][col] - temp[(row - box - 1)][col]) / len;
}
for (int row = H - box; row < H; row++) {
dst[row][col] = (dst[(row - 1)][col] * len - temp[(row - box - 1)][col]) / (len - 1);
len --;
}
}
#endif
} }
void boxblur_resamp(float **src, float **dst, float ** temp, int H, int W, int box, int samp ) void boxblur_resamp(float **src, float **dst, float ** temp, int H, int W, int box, int samp )
@@ -263,19 +174,19 @@ void boxblur_resamp(float **src, float **dst, float ** temp, int H, int W, int b
} }
} }
static const int numCols = 8; // process numCols columns at once for better L1 CPU cache usage constexpr int numCols = 8; // process numCols columns at once for better L1 CPU cache usage
#ifdef _OPENMP #ifdef _OPENMP
#pragma omp parallel #pragma omp parallel
#endif #endif
{ {
float tempvalN[numCols] ALIGNED16; float tempvalN[numCols] ALIGNED64;
#ifdef _OPENMP #ifdef _OPENMP
#pragma omp for nowait #pragma omp for nowait
#endif #endif
//vertical blur //vertical blur
for (int col = 0; col < (W / samp) - (numCols - 1); col += numCols) { for (int col = 0; col < (W / samp) - (numCols - 1); col += numCols) {
int len = box + 1; float len = box + 1;
for(int n = 0; n < numCols; n++) { for(int n = 0; n < numCols; n++) {
tempvalN[n] = temp[0][col + n] / len; tempvalN[n] = temp[0][col + n] / len;
@@ -304,10 +215,10 @@ void boxblur_resamp(float **src, float **dst, float ** temp, int H, int W, int b
len ++; len ++;
} }
const float rlen = 1.f / len;
for (int row = box + 1; row < H - box; row++) { for (int row = box + 1; row < H - box; row++) {
for(int n = 0; n < numCols; n++) { for(int n = 0; n < numCols; n++) {
tempvalN[n] = tempvalN[n] + (temp[(row + box)][col + n] - temp[(row - box - 1)][col + n]) / len; tempvalN[n] = tempvalN[n] + (temp[(row + box)][col + n] - temp[(row - box - 1)][col + n]) * rlen;
} }
if(row % samp == 0) { if(row % samp == 0) {
@@ -511,28 +422,32 @@ void RawImageSource :: HLRecovery_inpaint (float** red, float** green, float** b
maxy = std::min(height - 1, maxy + blurBorder); maxy = std::min(height - 1, maxy + blurBorder);
const int blurWidth = maxx - minx + 1; const int blurWidth = maxx - minx + 1;
const int blurHeight = maxy - miny + 1; const int blurHeight = maxy - miny + 1;
const int bufferWidth = blurWidth + ((16 - (blurWidth % 16)) & 15);
std::cout << "blurWidth : " << blurWidth << std::endl;
std::cout << "bufferWidth : " << bufferWidth << std::endl;
std::cout << "Corrected area reduced by factor: " << (((float)width * height) / (blurWidth * blurHeight)) << std::endl; std::cout << "Corrected area reduced by factor: " << (((float)width * height) / (blurWidth * blurHeight)) << std::endl;
multi_array2D<float, 3> channelblur(blurWidth, blurHeight, 0, 48); multi_array2D<float, 3> channelblur(bufferWidth, blurHeight, 0, 48);
array2D<float> temp(blurWidth, blurHeight); // allocate temporary buffer array2D<float> temp(bufferWidth, blurHeight); // allocate temporary buffer
// blur RGB channels // blur RGB channels
boxblur2(red, channelblur[0], temp, miny, minx, blurHeight, blurWidth, 4); boxblur2(red, channelblur[0], temp, miny, minx, blurHeight, bufferWidth, 4);
if(plistener) { if(plistener) {
progress += 0.05; progress += 0.05;
plistener->setProgress(progress); plistener->setProgress(progress);
} }
boxblur2(green, channelblur[1], temp, miny, minx, blurHeight, blurWidth, 4); boxblur2(green, channelblur[1], temp, miny, minx, blurHeight, bufferWidth, 4);
if(plistener) { if(plistener) {
progress += 0.05; progress += 0.05;
plistener->setProgress(progress); plistener->setProgress(progress);
} }
boxblur2(blue, channelblur[2], temp, miny, minx, blurHeight, blurWidth, 4); boxblur2(blue, channelblur[2], temp, miny, minx, blurHeight, bufferWidth, 4);
if(plistener) { if(plistener) {
progress += 0.05; progress += 0.05;
@@ -558,7 +473,7 @@ void RawImageSource :: HLRecovery_inpaint (float** red, float** green, float** b
plistener->setProgress(progress); plistener->setProgress(progress);
} }
multi_array2D<float, 4> hilite_full(blurWidth, blurHeight, ARRAY2D_CLEAR_DATA, 32); multi_array2D<float, 4> hilite_full(bufferWidth, blurHeight, ARRAY2D_CLEAR_DATA, 32);
if(plistener) { if(plistener) {
progress += 0.10; progress += 0.10;
@@ -598,10 +513,10 @@ void RawImageSource :: HLRecovery_inpaint (float** red, float** green, float** b
plistener->setProgress(progress); plistener->setProgress(progress);
} }
array2D<float> hilite_full4(blurWidth, blurHeight); array2D<float> hilite_full4(bufferWidth, blurHeight);
//%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% //%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
//blur highlight data //blur highlight data
boxblur2(hilite_full[3], hilite_full4, temp, 0, 0, blurHeight, blurWidth, 1); boxblur2(hilite_full[3], hilite_full4, temp, 0, 0, blurHeight, bufferWidth, 1);
temp.free(); // free temporary buffer temp.free(); // free temporary buffer