コミット: ec393489c021af1d5498d68597e98300079ebdcb - color_quantizer (git)

減色プログラム

コミットメタ情報

• リビジョン: ec393489c021af1d5498d68597e98300079ebdcb (tree)
• 日時: 2011-05-19 03:43:15
• 作者: berupon <berupon@gmai...>
• コミッター: berupon

ログメッセージ

optimized

変更サマリ

• modified: Array.h
• modified: color_quantizer.vcproj
• add: lib.cpp
• modified: main.cpp
• modified: quantize.cpp

差分

--- a/Array.h
+++ b/Array.h
@@ -51,19 +51,28 @@ struct Array2D
 		}
 	}
 	
+	__forceinline
 	T* operator[] (int row) {
 		return &pBuff_[row * width_];
 	}
+	
+	__forceinline
 	const T* operator[] (int row) const {
 		return &pBuff_[row * width_];
 	}
 	
 	Array2D<T>& operator *= (const T& scalar) {
+#if 1
+		for (size_t i=0; i<width_*height_; ++i) {
+			pBuff_[i] *= scalar;
+		}
+#else
 		for (int i=0; i<width_; i++) {
 			for (int j=0; j<height_; j++) {
 				(*this)[j][i] *= scalar;
 			}
 		}
+#endif
 		return *this;
 	}
 	
@@ -174,12 +183,30 @@ struct Array3D
 		delete pBuff_;
 	}
 
+/*
 	Array2D<T> operator[] (int depth) {
 		return Array2D<T>(width_, height_, &pBuff_[depth * width_ * height_]);
 	}
 	Array2D<T> operator[] (int depth) const {
 		return Array2D<T>(width_, height_, &pBuff_[depth * width_ * height_]);
 	}
+*/
+	__forceinline
+	T& operator() (size_t z, size_t y, size_t x) {
+		return pBuff_[
+			z * width_ * height_
+			+ y * width_
+			+ x
+		];
+	}
+	__forceinline
+	const T& operator() (size_t z, size_t y, size_t x) const {
+		return pBuff_[
+			z * width_ * height_
+			+ y * width_
+			+ x
+		];
+	}
 };
 
 

--- a/color_quantizer.vcproj
+++ b/color_quantizer.vcproj
@@ -119,7 +119,10 @@
 				AdditionalIncludeDirectories="./"
 				PreprocessorDefinitions="WIN32;NDEBUG;_CONSOLE;_CRT_SECURE_NO_WARNINGS"
 				RuntimeLibrary="2"
+				BufferSecurityCheck="false"
 				EnableFunctionLevelLinking="true"
+				EnableEnhancedInstructionSet="2"
+				FloatingPointModel="2"
 				UsePrecompiledHeader="2"
 				WarningLevel="3"
 				DebugInformationFormat="3"
@@ -175,6 +178,10 @@
 			UniqueIdentifier="{4FC737F1-C7A5-4376-A066-2A32D752A2FF}"
 			>
 			<File
+				RelativePath=".\lib.cpp"
+				>
+			</File>
+			<File
 				RelativePath=".\main.cpp"
 				>
 			</File>

--- /dev/null
+++ b/lib.cpp
@@ -0,0 +1,3 @@
+#include "stdafx.h"
+
+#pragma comment(lib, "winmm.lib")

--- a/main.cpp
+++ b/main.cpp
@@ -48,6 +48,8 @@ typedef Array2D<Color4f> Image4f;
 
 // http://proglab.aki.gs/mediaproc/colors.html
 
+size_t call_count = 0;
+
 int _tmain(int argc, _TCHAR* argv[])
 {
 	if (argc < 1+3 || argc > 1+5) {
@@ -183,8 +185,14 @@ int _tmain(int argc, _TCHAR* argv[])
 		NULL,
 		&filter5_weights
 	};
+
+	DWORD started = ::timeGetTime();
+	
 	spatial_color_quant(image, *filters[filter_size], quantized_image, palette, num_colors, coarse_variables, 1.0, 0.001, 3, 1);
 	
+	DWORD timeTaken = ::timeGetTime() - started;
+	printf("time taken : %d ms, call count : %d\n", timeTaken, call_count);
+
 	{
 		FILE* out = _tfopen(argv[2], _T("wb"));
 		if (out == NULL) {

--- a/quantize.cpp
+++ b/quantize.cpp
@@ -34,7 +34,7 @@ void fill_random(Array3D<float>& a)
 	for (int i=0; i<a.width_; i++) {
 		for (int j=0; j<a.height_; j++) {
 			for (int k=0; k<a.depth_; k++) {
-				a[k][j][i] = ((float)rand())/RAND_MAX;
+				a(k,j,i) = ((float)rand())/RAND_MAX;
 			}
 		}
 	}
@@ -118,6 +118,7 @@ void compute_b_array(
 	}
 }
 
+__forceinline
 Color4f b_value(const Image4f& b, int i_x, int i_y, int j_x, int j_y)
 {
 	int radius_width = (b.width_ - 1)/2,
@@ -210,11 +211,11 @@ int best_match_color(
 	)
 {
 	int max_v = 0;
-	float max_weight = vars[0][i_y][i_x];
+	float max_weight = vars(0,i_y,i_x);
 	for (size_t v=1; v<num_colors; ++v) {
-		if (vars[v][i_y][i_x] > max_weight) {
+		if (vars(v,i_y,i_x) > max_weight) {
 			max_v = v;
-			max_weight = vars[v][i_y][i_x];
+			max_weight = vars(v,i_y,i_x);
 		}
 	}
 	return max_v;
@@ -228,34 +229,43 @@ void zoom(const Array3D<float>& small, Array3D<float>& big)
 	// is 1.2 fine pixels wide and high.
 	for (int y=0; y<big.height_/2*2; y++) {
 		for (int x=0; x<big.width_/2*2; x++) {
-			float left = max(0.0, (x-0.1)/2.0), right	= min(small.width_-0.001, (x+1.1)/2.0);
-			float top	= max(0.0, (y-0.1)/2.0), bottom = min(small.height_-0.001, (y+1.1)/2.0);
+			float left = max(0.0, (x-0.1)/2.0);
+			float right	= min(small.width_-0.001, (x+1.1)/2.0);
+			float top	= max(0.0, (y-0.1)/2.0);
+			float bottom = min(small.height_-0.001, (y+1.1)/2.0);
 			int x_left = (int)floor(left), x_right	= (int)floor(right);
 			int y_top  = (int)floor(top),  y_bottom = (int)floor(bottom);
 			float area = (right-left)*(bottom-top);
-			float top_left_weight	= (ceil(left) - left)*(ceil(top) - top)/area;
-			float top_right_weight = (right - floor(right))*(ceil(top) - top)/area;
-			float bottom_left_weight  = (ceil(left) - left)*(bottom - floor(bottom))/area;
-			float bottom_right_weight = (right - floor(right))*(bottom - floor(bottom))/area;
-			float top_weight	  = (right-left)*(ceil(top) - top)/area;
-			float bottom_weight  = (right-left)*(bottom - floor(bottom))/area;
-			float left_weight	  = (bottom-top)*(ceil(left) - left)/area;
-			float right_weight	  = (bottom-top)*(right - floor(right))/area;
+			float inv_area	= 1.0 / area;
+			float left2		= (ceil(left) - left);
+			float top2		= (ceil(top) - top);
+			float bottom2	= (bottom - floor(bottom));
+			float right2	= (right - floor(right));
+			float top_left_weight		= left2 * top2 * inv_area;
+			float top_right_weight		= right2 * top2 * inv_area;
+			float bottom_left_weight	= left2 * bottom2 * inv_area;
+			float bottom_right_weight	= right2 * bottom2 * inv_area;
+			float top_weight			= (right-left) * top2 * inv_area;
+			float bottom_weight			= (right-left) * bottom2 * inv_area;
+			float left_weight			= (bottom-top) * left2 * inv_area;
+			float right_weight			= (bottom-top) * right2 * inv_area;
 			for (int z=0; z<big.depth_; z++) {
+				float val;
 				if (x_left == x_right && y_top == y_bottom) {
-					big[z][y][x] = small[z][y_top][x_left];
+					val =	small(z,y_top,x_left);
 				} else if (x_left == x_right) {
-					big[z][y][x] = top_weight*small[z][y_top][x_left] +
-								   bottom_weight*small[z][y_bottom][x_left];
+					val =	top_weight * small(z,y_top,x_left) +
+								   bottom_weight * small(z,y_bottom,x_left);
 				} else if (y_top == y_bottom) {
-					big[z][y][x] = left_weight*small[z][y_top][x_left] +
-								   right_weight*small[z][y_top][x_right];
+					val =	left_weight * small(z,y_top,x_left) +
+								   right_weight * small(z,y_top,x_right);
 				} else {
-					big[z][y][x] = top_left_weight*small[z][y_top][x_left] +
-								   top_right_weight*small[z][y_top][x_right] +
-								   bottom_left_weight*small[z][y_bottom][x_left] +
-								   bottom_right_weight*small[z][y_bottom][x_right];
+					val =	top_left_weight * small(z,y_top,x_left) +
+							top_right_weight * small(z,y_top,x_right) +
+							bottom_left_weight * small(z,y_bottom,x_left) +
+							bottom_right_weight * small(z,y_bottom,x_right);
 				}
+				big(z,y,x) = val;
 			}
 		}
 	}
@@ -289,16 +299,16 @@ void compute_initial_s(
 					if (i_x == j_x && i_y == j_y) continue;
 					Color4f b_ij = b_value(b,i_x,i_y,j_x,j_y);
 					for (int v=0; v<palette_size; v++) {
-						float vcv = coarse_variables[v][i_y][i_x];
+						float vcv = coarse_variables(v,i_y,i_x);
 						for (int alpha=v; alpha<palette_size; alpha++) {
-							float mult = vcv * coarse_variables[alpha][j_y][j_x];
+							float mult = vcv * coarse_variables(alpha,j_y,j_x);
 							s[alpha][v] += mult * b_ij;
 						}
 					}
 				}
 			}		
 			for (int v=0; v<palette_size; v++) {
-				s[v][v] += coarse_variables[v][i_y][i_x]*center_b;
+				s[v][v] += coarse_variables(v,i_y,i_x)*center_b;
 			}
 		}
 	}
@@ -326,11 +336,11 @@ void update_s(
 			if (i_x == j_x && i_y == j_y) continue;
 			Color4f* ps = s[alpha];
 			for (int v=0; v <= alpha; v++) {
-				float mult = coarse_variables[v][i_y][i_x];
+				float mult = coarse_variables(v,i_y,i_x);
 				ps[v] += mult * delta_b_ij;
 			}
 			for (int v=alpha; v<palette_size; v++) {
-				float mult = coarse_variables[v][i_y][i_x];
+				float mult = coarse_variables(v,i_y,i_x);
 				s[v][alpha] += mult * delta_b_ij;
 			}
 		}
@@ -359,7 +369,7 @@ void refine_palette(
 		sum.zero();
 		for (int i_y=0; i_y<coarse_variables.height_; i_y++) {
 			for (int i_x=0; i_x<coarse_variables.width_; i_x++) {
-				float cv = coarse_variables[v][i_y][i_x];
+				float cv = coarse_variables(v,i_y,i_x);
 				Color4f av = a[i_y][i_x];
 				Color4f result = cv * av;
 				sum += result;
@@ -371,7 +381,7 @@ void refine_palette(
 	for (unsigned int k=0; k<3; k++) {
 		Array2D<float> S_k = extract_vector_layer_2d(s, k);
 		vector<float> R_k = extract_vector_layer_1d(&r[0], num_colors, k);
-		vector<float> palette_channel = -1.0f*((2.0f*S_k).matrix_inverse())*R_k;
+		vector<float> palette_channel = -1.0f * ((2.0f*S_k).matrix_inverse()) * R_k;
 		for (unsigned int v=0; v<num_colors; v++) {
 			float val = palette_channel[v];
 			if (val < 0) val = 0;
@@ -399,7 +409,7 @@ void compute_initial_j_palette_sum(
 			 Color4f palette_sum;
 			 palette_sum.zero();
 			 for (size_t alpha=0; alpha<num_colors; ++alpha) {
-				 palette_sum += coarse_variables[alpha][j_y][j_x]*palette[alpha];
+				 palette_sum += coarse_variables(alpha,j_y,j_x)*palette[alpha];
 			 }
 			 j_palette_sum[j_y][j_x] = palette_sum;
 		 }
@@ -559,8 +569,8 @@ void spatial_color_quant(
 					// Prevent the matrix S from becoming singular
 					if (new_val <= 0) new_val = 1e-10;
 					if (new_val >= 1) new_val = 1 - 1e-10;
-					float delta_m_iv = new_val - coarse_variables[v][i_y][i_x];
-					coarse_variables[v][i_y][i_x] = new_val;
+					float delta_m_iv = new_val - coarse_variables(v,i_y,i_x);
+					coarse_variables(v,i_y,i_x) = new_val;
 					j_pal += delta_m_iv * palette[v];
 					if (abs(delta_m_iv) > 0.001 && !skip_palette_maintenance) {
 						update_s(s, coarse_variables, b, i_x, i_y, v, delta_m_iv);
@@ -636,7 +646,7 @@ void spatial_color_quant(
 			temperature *= temperature_multiplier;
 		}
 	}
-
+	
 	// This is normally not used, but is handy sometimes for debugging
 	while (coarse_level > 0) {
 		coarse_level--;