color quantizer
Fork

(Original repository, No fork origin)

Commit

optimized

--- a/color_quantizer.sln

+++ b/color_quantizer.sln

		@@ -17,4 +17,9 @@ Global
17	17	GlobalSection(SolutionProperties) = preSolution
18	18	HideSolutionNode = FALSE
19	19	EndGlobalSection
	20	+ GlobalSection(ExtensibilityGlobals) = postSolution
	21	+ AMDCaPersistentConfig = Release\|Win32
	22	+ AMDCaPersistentStartup = color_quantizer
	23	+ AMDCaProjectFile = C:\projects\color_quantizer\CodeAnalyst\color_quantizer.caw
	24	+ EndGlobalSection
20	25	EndGlobal

--- a/quantize.cpp

+++ b/quantize.cpp

		@@ -282,6 +282,12 @@ void compute_initial_s(
282	282	int coarse_width = coarse_variables.width_;
283	283	int coarse_height = coarse_variables.height_;
284	284	int center_x = (b.width_-1)/2, center_y = (b.height_-1)/2;
	285	+
	286	+ printf(
	287	+ "%s started %d %d %d\n",
	288	+ __FUNCTION__, palette_size, coarse_width, coarse_height
	289	+ );
	290	+
285	291	Color4f center_b = b_value(b,0,0,0,0);
286	292	Color4f zero_vector;
287	293	zero_vector.zero();

		@@ -290,6 +296,36 @@ void compute_initial_s(
290	296	s[alpha][v] = zero_vector;
291	297	}
292	298	}
	299	+#if 1
	300	+ const size_t cv_plane_size = coarse_width * coarse_height;
	301	+ for (size_t i_y=0; i_y<coarse_height; i_y++) {
	302	+ for (size_t i_x=0; i_x<coarse_width; i_x++) {
	303	+ int max_j_x = min<int>(coarse_width, i_x - center_x + b.width_);
	304	+ int max_j_y = min<int>(coarse_height, i_y - center_y + b.height_);
	305	+ for (size_t j_y=max<int>(0, i_y - center_y); j_y<max_j_y; j_y++) {
	306	+ for (size_t j_x=max<int>(0, i_x - center_x); j_x<max_j_x; j_x++) {
	307	+ if (i_x == j_x && i_y == j_y) continue;
	308	+ Color4f b_ij = b_value(b,i_x,i_y,j_x,j_y);
	309	+ const float* p_icv = coarse_variables.pBuff_ + i_y * coarse_width + i_x;
	310	+ for (size_t v=0; v<palette_size; ++v) {
	311	+ Color4f b_ij2 = b_ij * (*p_icv);
	312	+ p_icv += cv_plane_size;
	313	+ const float* p_jcv = coarse_variables.pBuff_ + cv_plane_size * v + j_y * coarse_width + j_x;
	314	+ Color4f* ps = s.pBuff_ + v * palette_size + v;
	315	+ for (size_t alpha=v; alpha<palette_size; ++alpha) {
	316	+ ps += (p_jcv) * b_ij2;
	317	+ p_jcv += cv_plane_size;
	318	+ ps += palette_size;
	319	+ }
	320	+ }
	321	+ }
	322	+ }
	323	+ for (int v=0; v<palette_size; v++) {
	324	+ s[v][v] += coarse_variables(v,i_y,i_x)*center_b;
	325	+ }
	326	+ }
	327	+ }
	328	+#else
293	329	for (int i_y=0; i_y<coarse_height; i_y++) {
294	330	for (int i_x=0; i_x<coarse_width; i_x++) {
295	331	int max_j_x = min<int>(coarse_width, i_x - center_x + b.width_);