Fixing frontend's selector codebook generator so it guarantees that all-flag selector clusters will be present in the output codebook.

Adding debug visualization flags to transoder
Hooking up debug visualization to WebGL demo

basisu_comp.cpp

@@ -18,7 +18,7 @@
 
 #define BASISU_USE_STB_IMAGE_RESIZE_FOR_MIPMAP_GEN 0
 #define DEBUG_CROP_TEXTURE_TO_64x64 (0)
-#define DEBUG_RESIZE_TEXTURE (0)
+#define DEBUG_RESIZE_TEXTURE (1)
 #define DEBUG_EXTRACT_SINGLE_BLOCK (0)
 
 namespace basisu
@@ -319,7 +319,7 @@ namespace basisu
 #endif
 
 #if DEBUG_RESIZE_TEXTURE
-			image temp_img((file_image.get_width() + 3) / 4, (file_image.get_height() + 3) / 4);
+			image temp_img((file_image.get_width() + 1) / 2, (file_image.get_height() + 1) / 2);
 			image_resample(file_image, temp_img, m_params.m_perceptual, "kaiser");
 			temp_img.swap(file_image);
 #endif

basisu_frontend.cpp

@@ -156,6 +156,8 @@ namespace basisu
 
 			find_optimal_selector_clusters_for_each_block();
 
+			introduce_special_selector_clusters();
+
 			if (!m_params.m_faster)
 			{
 				if (!refine_block_endpoints_given_selectors())
@@ -181,6 +183,80 @@ namespace basisu
 		return true;
 	}
 
+	void basisu_frontend::introduce_special_selector_clusters()
+	{
+		debug_printf("optimize_selector_codebook\n");
+
+		if (m_params.m_pGlobal_sel_codebook)
+			return;
+
+		uint32_t total_blocks_relocated = 0;
+
+		// Make sure the selector codebook always has pure flat blocks for each possible selector, to avoid obvious artifacts.
+		// optimize_selector_codebook() will clean up any redundant clusters we create here.
+		for (uint32_t sel = 0; sel < 4; sel++)
+		{
+			etc_block blk;
+			clear_obj(blk);
+			for (uint32_t j = 0; j < 16; j++)
+				blk.set_selector(j & 3, j >> 2, sel);
+
+			int k;
+			for (k = 0; k < m_optimized_cluster_selectors.size(); k++)
+				if (m_optimized_cluster_selectors[k].get_raw_selector_bits() == blk.get_raw_selector_bits())
+					break;
+			if (k < m_optimized_cluster_selectors.size())
+				continue;
+
+			const uint32_t new_selector_cluster_index = (uint32_t)m_optimized_cluster_selectors.size();
+
+			m_optimized_cluster_selectors.push_back(blk);
+			
+			vector_ensure_element_is_valid(m_selector_cluster_indices, new_selector_cluster_index);
+			
+			for (uint32_t block_index = 0; block_index < m_total_blocks; block_index++)
+			{
+				if (m_orig_encoded_blocks[block_index].get_raw_selector_bits() != blk.get_raw_selector_bits())
+					continue;
+
+				// See if using flat selectors actually decreases the block's error.
+				const uint32_t old_selector_cluster_index = m_block_selector_cluster_index[block_index];
+				
+				etc_block cur_blk;
+				const uint32_t endpoint_cluster_index = get_subblock_endpoint_cluster_index(block_index, 0);
+				cur_blk.set_block_color5_etc1s(get_endpoint_cluster_unscaled_color(endpoint_cluster_index, false));
+				cur_blk.set_inten_tables_etc1s(get_endpoint_cluster_inten_table(endpoint_cluster_index, false));
+				cur_blk.set_raw_selector_bits(get_selector_cluster_selector_bits(old_selector_cluster_index).get_raw_selector_bits());
+
+				const uint64_t cur_err = cur_blk.evaluate_etc1_error(get_source_pixel_block(block_index).get_ptr(), m_params.m_perceptual);
+
+				cur_blk.set_raw_selector_bits(blk.get_raw_selector_bits());
+
+				const uint64_t new_err = cur_blk.evaluate_etc1_error(get_source_pixel_block(block_index).get_ptr(), m_params.m_perceptual);
+
+				if (new_err >= cur_err)
+					continue;
+				
+				// Change the block to use the new cluster
+				m_block_selector_cluster_index[block_index] = new_selector_cluster_index;
+				
+				m_selector_cluster_indices[new_selector_cluster_index].push_back(block_index);
+
+				int j = vector_find(m_selector_cluster_indices[old_selector_cluster_index], block_index);
+				if (j >= 0)
+					m_selector_cluster_indices[old_selector_cluster_index].erase(m_selector_cluster_indices[old_selector_cluster_index].begin() + j);
+
+				total_blocks_relocated++;
+
+				m_encoded_blocks[block_index].set_raw_selector_bits(blk.get_raw_selector_bits());
+
+			} // block_index
+
+		} // sel
+
+		debug_printf("Total blocks relocated to new flat selector clusters: %u\n", total_blocks_relocated);
+	}
+
 	void basisu_frontend::optimize_selector_codebook()
 	{
 		debug_printf("optimize_selector_codebook\n");
@@ -233,21 +309,23 @@ namespace basisu
 			if (m_optimized_cluster_selector_global_cb_ids.size())
 				new_optimized_cluster_selector_global_cb_ids[i] = m_optimized_cluster_selector_global_cb_ids[new_to_old[i]];
 
-			if (m_selector_cluster_indices.size())
-				new_selector_cluster_indices[i] = m_selector_cluster_indices[new_to_old[i]];
-
 			if (m_selector_cluster_uses_global_cb.size())
 				new_selector_cluster_uses_global_cb[i] = m_selector_cluster_uses_global_cb[new_to_old[i]];
 		}
 
+		for (uint32_t i = 0; i < m_block_selector_cluster_index.size(); i++)
+		{
+			m_block_selector_cluster_index[i] = old_to_new[m_block_selector_cluster_index[i]];
+		
+			if (m_selector_cluster_indices.size())
+				new_selector_cluster_indices[m_block_selector_cluster_index[i]].push_back(i);
+		}
+
 		m_optimized_cluster_selectors.swap(new_optimized_cluster_selectors);
 		m_optimized_cluster_selector_global_cb_ids.swap(new_optimized_cluster_selector_global_cb_ids);
 		m_selector_cluster_indices.swap(new_selector_cluster_indices);
 		m_selector_cluster_uses_global_cb.swap(new_selector_cluster_uses_global_cb);
 
-		for (uint32_t i = 0; i < m_block_selector_cluster_index.size(); i++)
-			m_block_selector_cluster_index[i] = old_to_new[m_block_selector_cluster_index[i]];
-
 		debug_printf("optimize_selector_codebook: Before: %u After: %u\n", orig_total_selector_clusters, total_new_entries);
 	}
 
@@ -1722,7 +1800,8 @@ namespace basisu
 
 		for (uint32_t block_index = 0; block_index < m_total_blocks; block_index++)
 		{
-#define CHECK(x) do { if (!(x)) return false; } while(0)
+//#define CHECK(x) do { if (!(x)) { DebugBreak(); return false; } } while(0)
+#define CHECK(x) BASISU_FRONTEND_VERIFY(x);
 
 			CHECK(get_output_block(block_index).get_flip_bit() == true);
 			

basisu_frontend.h

@@ -290,6 +290,7 @@ namespace basisu
 		uint32_t refine_block_endpoints_given_selectors();
 		void finalize();
 		bool validate_output() const;
+		void introduce_special_selector_clusters();
 		void optimize_selector_codebook();
 		bool check_etc1s_constraints() const;
 	};

transcoder/basisu_transcoder.cpp

@@ -33,10 +33,12 @@
 #define BASISD_WRITE_NEW_DXT1_TABLES			0
 #define BASISD_WRITE_NEW_ETC2_EAC_A8_TABLES	0
 
+#define BASISD_ENABLE_DEBUG_FLAGS	1
+
 namespace basisu
 {
 	static bool g_debug_printf;
-
+	
 	void enable_debug_printf(bool enabled)
 	{
 		g_debug_printf = enabled;
@@ -60,7 +62,28 @@ namespace basisu
 namespace basist
 {
 	#include "basisu_transcoder_tables_bc7_m6.inc"
-				
+			
+#if BASISD_ENABLE_DEBUG_FLAGS
+	static uint32_t g_debug_flags = 0;
+#endif
+
+	uint32_t get_debug_flags()
+	{
+#if BASISD_ENABLE_DEBUG_FLAGS
+		return g_debug_flags;
+#else
+		return 0;
+#endif
+	}
+
+	void set_debug_flags(uint32_t f)
+	{
+		(void)f;
+#if BASISD_ENABLE_DEBUG_FLAGS
+		g_debug_flags = f;
+#endif
+	}
+
 	uint16_t crc16(const void *r, size_t size, uint16_t crc)
 	{
 		crc = ~crc;
@@ -1718,7 +1741,7 @@ namespace basist
 
 		const color32 base_color(pSrc_block->get_base5_color_unscaled());
 		const uint32_t inten_table = pSrc_block->get_inten_table(0);
-
+				
 		if (low_selector == high_selector)
 		{
 			color32 block_colors[4];
@@ -1742,7 +1765,7 @@ namespace basist
 				// Make l > h
 				if (min16 > 0)
 					min16--;
-				else 
+				else
 				{
 					// l = h = 0
 					assert(min16 == max16 && max16 == 0);
@@ -1751,7 +1774,7 @@ namespace basist
 					min16 = 0;
 					mask = 0x55;
 				}
-			
+
 				assert(max16 > min16);
 			}
 
@@ -1760,7 +1783,7 @@ namespace basist
 				std::swap(max16, min16);
 				mask ^= 0x55;
 			}
-						
+
 			pDst_block->set_low_color(static_cast<uint16_t>(max16));
 			pDst_block->set_high_color(static_cast<uint16_t>(min16));
 			pDst_block->m_selectors[0] = static_cast<uint8_t>(mask);
@@ -1796,28 +1819,28 @@ namespace basist
 				{
 					min16--;
 
-					l = 0; 
+					l = 0;
 					h = 0;
 				}
-				else 
+				else
 				{
 					// l = h = 0
 					assert(min16 == max16 && max16 == 0);
 
 					max16 = 1;
 					min16 = 0;
-					
+
 					l = 1;
 					h = 1;
 				}
-			
+
 				assert(max16 > min16);
 			}
 
 			if (max16 < min16)
 			{
 				std::swap(max16, min16);
-				l = 1; 
+				l = 1;
 				h = 0;
 			}
 
@@ -1966,6 +1989,28 @@ namespace basist
 		pDst_block->m_selectors[3] = (uint8_t)dxt1_sels3;
 #endif
 	}
+
+#if BASISD_ENABLE_DEBUG_FLAGS
+	static void convert_etc1s_to_dxt1_vis(dxt1_block* pDst_block, const decoder_etc_block* pSrc_block, const selector* pSelector, bool use_threecolor_blocks)
+	{
+		convert_etc1s_to_dxt1(pDst_block, pSrc_block, pSelector, use_threecolor_blocks);
+
+		if (g_debug_flags & cDebugFlagVisBC1Sels)
+		{
+			uint32_t l = dxt1_block::pack_unscaled_color(31, 63, 31);
+			uint32_t h = dxt1_block::pack_unscaled_color(0, 0, 0);
+
+			pDst_block->set_low_color(static_cast<uint16_t>(l));
+			pDst_block->set_high_color(static_cast<uint16_t>(h));
+		}
+		else if (g_debug_flags & cDebugFlagVisBC1Endpoints)
+		{
+			for (uint32_t y = 0; y < 4; y++)
+				for (uint32_t x = 0; x < 4; x++)
+					pDst_block->set_selector(x, y, (y < 2) ? 0 : 1);
+		}
+	}
+#endif
 #endif
 
 	static dxt_selector_range s_dxt5a_selector_ranges[] =
@@ -3927,15 +3972,18 @@ namespace basist
 				if (is_video)
 					(*pPrevFrameIndices)[block_x + block_y * num_blocks_x] = endpoint_index | (selector_index << 16);
 
-#if 0
-				// Visualize CR's
-				if ((is_video) && (pred == 2))
+#if BASISD_ENABLE_DEBUG_FLAGS
+				if ((g_debug_flags & cDebugFlagVisCRs) && ((fmt == cETC1) || (fmt == cBC1)))
 				{
-					decoder_etc_block* pDst_block = reinterpret_cast<decoder_etc_block*>(static_cast<uint8_t*>(pDst_blocks) + (block_x + block_y * num_blocks_x) * output_stride);
-					memset(pDst_block, 0xFF, 8);
-					continue;
+					// Visualize CR's
+					if ((is_video) && (pred == 2))
+					{
+						decoder_etc_block* pDst_block = reinterpret_cast<decoder_etc_block*>(static_cast<uint8_t*>(pDst_blocks) + (block_x + block_y * num_blocks_x) * output_stride);
+						memset(pDst_block, 0xFF, 8);
+						continue;
+					}
 				}
-#endif					
+#endif
 
 				const endpoint *pEndpoint0 = &m_endpoints[endpoint_index];
 												
@@ -3964,8 +4012,15 @@ namespace basist
 					block.set_raw_selector_bits(pSelector->m_bytes[0], pSelector->m_bytes[1], pSelector->m_bytes[2], pSelector->m_bytes[3]);
 
 					void* pDst_block = static_cast<uint8_t*>(pDst_blocks) + (block_x + block_y * num_blocks_x) * output_stride;
+
 #if BASISD_SUPPORT_DXT1
-					convert_etc1s_to_dxt1(static_cast<dxt1_block*>(pDst_block), &block, pSelector, bc1_allow_threecolor_blocks);
+
+#if BASISD_ENABLE_DEBUG_FLAGS
+					if (g_debug_flags & (cDebugFlagVisBC1Sels | cDebugFlagVisBC1Endpoints))
+						convert_etc1s_to_dxt1_vis(static_cast<dxt1_block*>(pDst_block), &block, pSelector, bc1_allow_threecolor_blocks);
+					else
+#endif
+						convert_etc1s_to_dxt1(static_cast<dxt1_block*>(pDst_block), &block, pSelector, bc1_allow_threecolor_blocks);
 #else
 					assert(0);
 #endif

transcoder/basisu_transcoder.h

@@ -304,4 +304,13 @@ namespace basist
 	// basisu_transcoder_init() must be called before a .basis file can be transcoded.
 	void basisu_transcoder_init();
 
+	enum debug_flags_t
+	{
+		cDebugFlagVisCRs = 1,
+		cDebugFlagVisBC1Sels = 2,
+		cDebugFlagVisBC1Endpoints = 4
+	};
+	uint32_t get_debug_flags();
+	void set_debug_flags(uint32_t f);
+
 } // namespace basisu

webgl_videotest/basis_wrappers.cpp

@@ -29,6 +29,9 @@ extern "C"
   uint basis_transcode_image(void *h, void *dst, uint dst_size_in_bytes, 
   uint image_index, uint level_index, uint format, 
   uint pvrtc_wrap_addressing, uint get_alpha_for_opaque_formats);
+  
+  void basis_set_debug_flags(uint f) { basist::set_debug_flags(f); }
+  uint basis_get_debug_flags() { return basist::get_debug_flags(); }
 }
 
 void basis_init()

webgl_videotest/build.bat

@@ -1,3 +1,3 @@
 @REM -O0 -s ASSERTIONS=1 -s DEMANGLE_SUPPORT=1
 @REM -O2 -s ASSERTIONS=0
-emcc -s EXPORTED_FUNCTIONS="['allocate', '_malloc', '_free', '_basis_init','_basis_open','_basis_close','_basis_get_has_alpha','_basis_get_num_images','_basis_get_num_levels','_basis_get_image_width','_basis_get_image_height','_basis_get_image_transcoded_size_in_bytes','_basis_transcode_image','_basis_start_transcoding']" -o basis.js ../transcoder/basisu_transcoder.cpp basis_wrappers.cpp -s TOTAL_MEMORY=60000000 -O2 -s ASSERTIONS=0 -I ../transcoder
+emcc -s EXPORTED_FUNCTIONS="['allocate', '_malloc', '_free', '_basis_init','_basis_open','_basis_close','_basis_get_has_alpha','_basis_get_num_images','_basis_get_num_levels','_basis_get_image_width','_basis_get_image_height','_basis_get_image_transcoded_size_in_bytes','_basis_transcode_image','_basis_start_transcoding','_basis_get_debug_flags','_basis_set_debug_flags']" -o basis.js ../transcoder/basisu_transcoder.cpp basis_wrappers.cpp -s TOTAL_MEMORY=80000000 -O2 -s ASSERTIONS=0 -I ../transcoder

webgl_videotest/build.sh

@@ -4,7 +4,7 @@
 
 ~/emscripten_tutorial/emscripten/emcc \
   -s EXPORTED_FUNCTIONS="['allocate', '_malloc', '_free', '_basis_init','_basis_open','_basis_close','_basis_get_has_alpha','_basis_get_num_images','_basis_get_num_levels','_basis_get_image_width','_basis_get_image_height','_basis_get_image_transcoded_size_in_bytes','_basis_transcode_image','_basis_start_transcoding']" \
-  -s TOTAL_MEMORY=60000000 -O2 -s ASSERTIONS=0 -I ../transcoder \
+  -s TOTAL_MEMORY=80000000 -O2 -s ASSERTIONS=0 -I ../transcoder \
   -o basis.js \
   ../transcoder/basisu_transcoder.cpp basis_wrappers.cpp && \
 chmod -R a+rX .

webgl_videotest/index.html

@@ -98,6 +98,12 @@ BASIS_FORMAT = {
 	cTFBC5: 7,		
 };
 
+BASIS_DEBUG_FLAGS = {
+	DEBUG_FLAG_VIS_CR: 1,
+	DEBUG_FLAG_VIS_SELS: 2,
+   	DEBUG_FLAG_VIS_ENDPOINTS: 4
+};
+
 BASIS_FORMAT_NAMES = {};
 for (var name in BASIS_FORMAT) {
   BASIS_FORMAT_NAMES[BASIS_FORMAT[name]] = name;
@@ -190,9 +196,9 @@ function frame(now)
   		redraw()
 		
 		timeSinceLastFrame += deltaTime;
-		if (timeSinceLastFrame > 1.0/25.0)
+		if (timeSinceLastFrame > 1.0/24.0)
 		{
-			timeSinceLastFrame -= 1.0/25.0;
+			timeSinceLastFrame -= 1.0/24.0;
 			
 			curImageIndex++;
 			if (curImageIndex == images)
@@ -265,6 +271,9 @@ function dataLoaded(data)
   	log('Decoding .basis data to BC1');
   }
 
+  log('.basis file size: ' + srcSize);
+  if (levels == 1)
+  	log('average bits/texel: ' + ((srcSize * 8.0) / (width * height * images)));
   log('width: ' + width);
   log('height: ' + height);
   log('images: ' + images);
@@ -347,6 +356,27 @@ function alphaBlend() { drawMode = 0; redraw(); }
 function viewRGB() { drawMode = 1; redraw(); }
 function viewAlpha() { drawMode = 2; redraw(); }
 
+function visCR() 
+{ 
+    var i = Module._basis_get_debug_flags();
+	i = i ^ BASIS_DEBUG_FLAGS.DEBUG_FLAG_VIS_CR;
+	Module._basis_set_debug_flags(i);
+}
+
+function visSels() 
+{ 
+    var i = Module._basis_get_debug_flags();
+	i = i ^ BASIS_DEBUG_FLAGS.DEBUG_FLAG_VIS_SELS;
+	Module._basis_set_debug_flags(i);
+}
+
+function visEndpoints() 
+{ 
+    var i = Module._basis_get_debug_flags();
+	i ^= BASIS_DEBUG_FLAGS.DEBUG_FLAG_VIS_ENDPOINTS;
+	Module._basis_set_debug_flags(i);
+}
+
 </script>
 </head>
 <body onload="initialize()">
@@ -360,14 +390,19 @@ function viewAlpha() { drawMode = 2; redraw(); }
   <br>
   <br>
       .basis file:
-      <input id="file" type="text" size=30 value="pic0001.basis"></input>
+      <input id="file" type="text" size=30 value="2010_2.basis"></input>
       <input type="button" value="Run!" onclick="run()"></input>
   <br>
   <br>
       <input type="button" value="Alpha blend" onclick="alphaBlend()"></input>
       <input type="button" value="View RGB" onclick="viewRGB()"></input>
 	  <input type="button" value="View Alpha" onclick="viewAlpha()"></input>
-	  
+  <br>
+  <br>	  
+	  <input type="button" value="Visualize CR's" onclick="visCR()"></input>
+	  <input type="button" value="Visualize Selectors" onclick="visSels()"></input>
+	  <input type="button" value="Visualize Endpoints" onclick="visEndpoints()"></input>
+  
   <div style="position:absolute; left: 525px; top:130px; font-size: 20pt; font-weight: bold; color: red">
     <div id="nodxt" style="display: none; width: 768px; font-size: 20pt; font-weight: bold; color: red">
       NOTE: Your browser does not support DXT, so using RGB565