// basisu_enc.cpp
// Copyright (C) 2019-2026 Binomial LLC. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
//    http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "basisu_enc.h"
#include "basisu_resampler.h"
#include "basisu_resampler_filters.h"
#include "basisu_etc.h"
#include "../transcoder/basisu_transcoder.h"
#include "basisu_bc7enc.h"
#include "jpgd.h"
#include "pvpngreader.h"
#include "basisu_opencl.h"
#include "basisu_uastc_hdr_4x4_enc.h"
#include "basisu_astc_hdr_6x6_enc.h"
#include "basisu_astc_ldr_common.h"
#include "basisu_astc_ldr_encode.h"

#include <vector>

#define TINYEXR_USE_MINIZ (0)
#include "3rdparty/tinyexr.h"

#ifndef MINIZ_HEADER_FILE_ONLY
#define MINIZ_HEADER_FILE_ONLY
#endif
#ifndef MINIZ_NO_ZLIB_COMPATIBLE_NAMES
#define MINIZ_NO_ZLIB_COMPATIBLE_NAMES
#endif
#include "basisu_miniz.h"

#define QOI_IMPLEMENTATION
#include "3rdparty/qoi.h"

#if defined(_WIN32)
// For QueryPerformanceCounter/QueryPerformanceFrequency
#define WIN32_LEAN_AND_MEAN
#include <windows.h>
#endif

namespace basisu
{
	uint64_t interval_timer::g_init_ticks, interval_timer::g_freq;
	double interval_timer::g_timer_freq;

	bool g_cpu_supports_sse41 = false;

	fast_linear_to_srgb g_fast_linear_to_srgb;
		
	uint8_t g_hamming_dist[256] =
	{
		0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4,
		1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5,
		1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5,
		2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6,
		1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5,
		2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6,
		2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6,
		3, 4, 4, 5, 4, 5, 5, 6, 4, 5, 5, 6, 5, 6, 6, 7,
		1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5,
		2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6,
		2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6,
		3, 4, 4, 5, 4, 5, 5, 6, 4, 5, 5, 6, 5, 6, 6, 7,
		2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6,
		3, 4, 4, 5, 4, 5, 5, 6, 4, 5, 5, 6, 5, 6, 6, 7,
		3, 4, 4, 5, 4, 5, 5, 6, 4, 5, 5, 6, 5, 6, 6, 7,
		4, 5, 5, 6, 5, 6, 6, 7, 5, 6, 6, 7, 6, 7, 7, 8
	};

	// This is a Public Domain 8x8 font from here:
	// https://github.com/dhepper/font8x8/blob/master/font8x8_basic.h
	const uint8_t g_debug_font8x8_basic[127 - 32 + 1][8] = 
	{
	 { 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00},	// U+0020 ( )
	 { 0x18, 0x3C, 0x3C, 0x18, 0x18, 0x00, 0x18, 0x00},   // U+0021 (!)
	 { 0x36, 0x36, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00},   // U+0022 (")
	 { 0x36, 0x36, 0x7F, 0x36, 0x7F, 0x36, 0x36, 0x00},   // U+0023 (#)
	 { 0x0C, 0x3E, 0x03, 0x1E, 0x30, 0x1F, 0x0C, 0x00},   // U+0024 ($)
	 { 0x00, 0x63, 0x33, 0x18, 0x0C, 0x66, 0x63, 0x00},   // U+0025 (%)
	 { 0x1C, 0x36, 0x1C, 0x6E, 0x3B, 0x33, 0x6E, 0x00},   // U+0026 (&)
	 { 0x06, 0x06, 0x03, 0x00, 0x00, 0x00, 0x00, 0x00},   // U+0027 (')
	 { 0x18, 0x0C, 0x06, 0x06, 0x06, 0x0C, 0x18, 0x00},   // U+0028 (()
	 { 0x06, 0x0C, 0x18, 0x18, 0x18, 0x0C, 0x06, 0x00},   // U+0029 ())
	 { 0x00, 0x66, 0x3C, 0xFF, 0x3C, 0x66, 0x00, 0x00},   // U+002A (*)
	 { 0x00, 0x0C, 0x0C, 0x3F, 0x0C, 0x0C, 0x00, 0x00},   // U+002B (+)
	 { 0x00, 0x00, 0x00, 0x00, 0x00, 0x0C, 0x0C, 0x06},   // U+002C (,)
	 { 0x00, 0x00, 0x00, 0x3F, 0x00, 0x00, 0x00, 0x00},   // U+002D (-)
	 { 0x00, 0x00, 0x00, 0x00, 0x00, 0x0C, 0x0C, 0x00},   // U+002E (.)
	 { 0x60, 0x30, 0x18, 0x0C, 0x06, 0x03, 0x01, 0x00},   // U+002F (/)
	 { 0x3E, 0x63, 0x73, 0x7B, 0x6F, 0x67, 0x3E, 0x00},   // U+0030 (0)
	 { 0x0C, 0x0E, 0x0C, 0x0C, 0x0C, 0x0C, 0x3F, 0x00},   // U+0031 (1)
	 { 0x1E, 0x33, 0x30, 0x1C, 0x06, 0x33, 0x3F, 0x00},   // U+0032 (2)
	 { 0x1E, 0x33, 0x30, 0x1C, 0x30, 0x33, 0x1E, 0x00},   // U+0033 (3)
	 { 0x38, 0x3C, 0x36, 0x33, 0x7F, 0x30, 0x78, 0x00},   // U+0034 (4)
	 { 0x3F, 0x03, 0x1F, 0x30, 0x30, 0x33, 0x1E, 0x00},   // U+0035 (5)
	 { 0x1C, 0x06, 0x03, 0x1F, 0x33, 0x33, 0x1E, 0x00},   // U+0036 (6)
	 { 0x3F, 0x33, 0x30, 0x18, 0x0C, 0x0C, 0x0C, 0x00},   // U+0037 (7)
	 { 0x1E, 0x33, 0x33, 0x1E, 0x33, 0x33, 0x1E, 0x00},   // U+0038 (8)
	 { 0x1E, 0x33, 0x33, 0x3E, 0x30, 0x18, 0x0E, 0x00},   // U+0039 (9)
	 { 0x00, 0x0C, 0x0C, 0x00, 0x00, 0x0C, 0x0C, 0x00},   // U+003A (:)
	 { 0x00, 0x0C, 0x0C, 0x00, 0x00, 0x0C, 0x0C, 0x06},   // U+003B (;)
	 { 0x18, 0x0C, 0x06, 0x03, 0x06, 0x0C, 0x18, 0x00},   // U+003C (<)
	 { 0x00, 0x00, 0x3F, 0x00, 0x00, 0x3F, 0x00, 0x00},   // U+003D (=)
	 { 0x06, 0x0C, 0x18, 0x30, 0x18, 0x0C, 0x06, 0x00},   // U+003E (>)
	 { 0x1E, 0x33, 0x30, 0x18, 0x0C, 0x00, 0x0C, 0x00},   // U+003F (?)
	 { 0x3E, 0x63, 0x7B, 0x7B, 0x7B, 0x03, 0x1E, 0x00},   // U+0040 (@)
	 { 0x0C, 0x1E, 0x33, 0x33, 0x3F, 0x33, 0x33, 0x00},   // U+0041 (A)
	 { 0x3F, 0x66, 0x66, 0x3E, 0x66, 0x66, 0x3F, 0x00},   // U+0042 (B)
	 { 0x3C, 0x66, 0x03, 0x03, 0x03, 0x66, 0x3C, 0x00},   // U+0043 (C)
	 { 0x1F, 0x36, 0x66, 0x66, 0x66, 0x36, 0x1F, 0x00},   // U+0044 (D)
	 { 0x7F, 0x46, 0x16, 0x1E, 0x16, 0x46, 0x7F, 0x00},   // U+0045 (E)
	 { 0x7F, 0x46, 0x16, 0x1E, 0x16, 0x06, 0x0F, 0x00},   // U+0046 (F)
	 { 0x3C, 0x66, 0x03, 0x03, 0x73, 0x66, 0x7C, 0x00},   // U+0047 (G)
	 { 0x33, 0x33, 0x33, 0x3F, 0x33, 0x33, 0x33, 0x00},   // U+0048 (H)
	 { 0x1E, 0x0C, 0x0C, 0x0C, 0x0C, 0x0C, 0x1E, 0x00},   // U+0049 (I)
	 { 0x78, 0x30, 0x30, 0x30, 0x33, 0x33, 0x1E, 0x00},   // U+004A (J)
	 { 0x67, 0x66, 0x36, 0x1E, 0x36, 0x66, 0x67, 0x00},   // U+004B (K)
	 { 0x0F, 0x06, 0x06, 0x06, 0x46, 0x66, 0x7F, 0x00},   // U+004C (L)
	 { 0x63, 0x77, 0x7F, 0x7F, 0x6B, 0x63, 0x63, 0x00},   // U+004D (M)
	 { 0x63, 0x67, 0x6F, 0x7B, 0x73, 0x63, 0x63, 0x00},   // U+004E (N)
	 { 0x1C, 0x36, 0x63, 0x63, 0x63, 0x36, 0x1C, 0x00},   // U+004F (O)
	 { 0x3F, 0x66, 0x66, 0x3E, 0x06, 0x06, 0x0F, 0x00},   // U+0050 (P)
	 { 0x1E, 0x33, 0x33, 0x33, 0x3B, 0x1E, 0x38, 0x00},   // U+0051 (Q)
	 { 0x3F, 0x66, 0x66, 0x3E, 0x36, 0x66, 0x67, 0x00},   // U+0052 (R)
	 { 0x1E, 0x33, 0x07, 0x0E, 0x38, 0x33, 0x1E, 0x00},   // U+0053 (S)
	 { 0x3F, 0x2D, 0x0C, 0x0C, 0x0C, 0x0C, 0x1E, 0x00},   // U+0054 (T)
	 { 0x33, 0x33, 0x33, 0x33, 0x33, 0x33, 0x3F, 0x00},   // U+0055 (U)
	 { 0x33, 0x33, 0x33, 0x33, 0x33, 0x1E, 0x0C, 0x00},   // U+0056 (V)
	 { 0x63, 0x63, 0x63, 0x6B, 0x7F, 0x77, 0x63, 0x00},   // U+0057 (W)
	 { 0x63, 0x63, 0x36, 0x1C, 0x1C, 0x36, 0x63, 0x00},   // U+0058 (X)
	 { 0x33, 0x33, 0x33, 0x1E, 0x0C, 0x0C, 0x1E, 0x00},   // U+0059 (Y)
	 { 0x7F, 0x63, 0x31, 0x18, 0x4C, 0x66, 0x7F, 0x00},   // U+005A (Z)
	 { 0x1E, 0x06, 0x06, 0x06, 0x06, 0x06, 0x1E, 0x00},   // U+005B ([)
	 { 0x03, 0x06, 0x0C, 0x18, 0x30, 0x60, 0x40, 0x00},   // U+005C (\)
	 { 0x1E, 0x18, 0x18, 0x18, 0x18, 0x18, 0x1E, 0x00},   // U+005D (])
	 { 0x08, 0x1C, 0x36, 0x63, 0x00, 0x00, 0x00, 0x00},   // U+005E (^)
	 { 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xFF},   // U+005F (_)
	 { 0x0C, 0x0C, 0x18, 0x00, 0x00, 0x00, 0x00, 0x00},   // U+0060 (`)
	 { 0x00, 0x00, 0x1E, 0x30, 0x3E, 0x33, 0x6E, 0x00},   // U+0061 (a)
	 { 0x07, 0x06, 0x06, 0x3E, 0x66, 0x66, 0x3B, 0x00},   // U+0062 (b)
	 { 0x00, 0x00, 0x1E, 0x33, 0x03, 0x33, 0x1E, 0x00},   // U+0063 (c)
	 { 0x38, 0x30, 0x30, 0x3e, 0x33, 0x33, 0x6E, 0x00},   // U+0064 (d)
	 { 0x00, 0x00, 0x1E, 0x33, 0x3f, 0x03, 0x1E, 0x00},   // U+0065 (e)
	 { 0x1C, 0x36, 0x06, 0x0f, 0x06, 0x06, 0x0F, 0x00},   // U+0066 (f)
	 { 0x00, 0x00, 0x6E, 0x33, 0x33, 0x3E, 0x30, 0x1F},   // U+0067 (g)
	 { 0x07, 0x06, 0x36, 0x6E, 0x66, 0x66, 0x67, 0x00},   // U+0068 (h)
	 { 0x0C, 0x00, 0x0E, 0x0C, 0x0C, 0x0C, 0x1E, 0x00},   // U+0069 (i)
	 { 0x30, 0x00, 0x30, 0x30, 0x30, 0x33, 0x33, 0x1E},   // U+006A (j)
	 { 0x07, 0x06, 0x66, 0x36, 0x1E, 0x36, 0x67, 0x00},   // U+006B (k)
	 { 0x0E, 0x0C, 0x0C, 0x0C, 0x0C, 0x0C, 0x1E, 0x00},   // U+006C (l)
	 { 0x00, 0x00, 0x33, 0x7F, 0x7F, 0x6B, 0x63, 0x00},   // U+006D (m)
	 { 0x00, 0x00, 0x1F, 0x33, 0x33, 0x33, 0x33, 0x00},   // U+006E (n)
	 { 0x00, 0x00, 0x1E, 0x33, 0x33, 0x33, 0x1E, 0x00},   // U+006F (o)
	 { 0x00, 0x00, 0x3B, 0x66, 0x66, 0x3E, 0x06, 0x0F},   // U+0070 (p)
	 { 0x00, 0x00, 0x6E, 0x33, 0x33, 0x3E, 0x30, 0x78},   // U+0071 (q)
	 { 0x00, 0x00, 0x3B, 0x6E, 0x66, 0x06, 0x0F, 0x00},   // U+0072 (r)
	 { 0x00, 0x00, 0x3E, 0x03, 0x1E, 0x30, 0x1F, 0x00},   // U+0073 (s)
	 { 0x08, 0x0C, 0x3E, 0x0C, 0x0C, 0x2C, 0x18, 0x00},   // U+0074 (t)
	 { 0x00, 0x00, 0x33, 0x33, 0x33, 0x33, 0x6E, 0x00},   // U+0075 (u)
	 { 0x00, 0x00, 0x33, 0x33, 0x33, 0x1E, 0x0C, 0x00},   // U+0076 (v)
	 { 0x00, 0x00, 0x63, 0x6B, 0x7F, 0x7F, 0x36, 0x00},   // U+0077 (w)
	 { 0x00, 0x00, 0x63, 0x36, 0x1C, 0x36, 0x63, 0x00},   // U+0078 (x)
	 { 0x00, 0x00, 0x33, 0x33, 0x33, 0x3E, 0x30, 0x1F},   // U+0079 (y)
	 { 0x00, 0x00, 0x3F, 0x19, 0x0C, 0x26, 0x3F, 0x00},   // U+007A (z)
	 { 0x38, 0x0C, 0x0C, 0x07, 0x0C, 0x0C, 0x38, 0x00},   // U+007B ({)
	 { 0x18, 0x18, 0x18, 0x00, 0x18, 0x18, 0x18, 0x00},   // U+007C (|)
	 { 0x07, 0x0C, 0x0C, 0x38, 0x0C, 0x0C, 0x07, 0x00},   // U+007D (})
	 { 0x6E, 0x3B, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00},   // U+007E (~)
	 { 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00}    // U+007F
	};

	float g_srgb_to_linear_table[256];
		
	void init_srgb_to_linear_table()
	{
		for (int i = 0; i < 256; ++i)
			g_srgb_to_linear_table[i] = srgb_to_linear((float)i * (1.0f / 255.0f));
	}

	bool g_library_initialized;
	std::mutex g_encoder_init_mutex;
					
	// Encoder library initialization (just call once at startup)
	bool basisu_encoder_init(bool use_opencl, bool opencl_force_serialization)
	{
		std::lock_guard<std::mutex> lock(g_encoder_init_mutex);

		if (g_library_initialized)
			return true;

		detect_sse41();
				
		basist::basisu_transcoder_init();
		pack_etc1_solid_color_init();
		//uastc_init();
		bc7enc_compress_block_init(); // must be after uastc_init()

		// Don't bother initializing the OpenCL module at all if it's been completely disabled.
		if (use_opencl)
		{
			opencl_init(opencl_force_serialization);
		}

		interval_timer::init(); // make sure interval_timer globals are initialized from main thread to avoid TSAN reports

		astc_hdr_enc_init();
		basist::bc6h_enc_init();
		astc_6x6_hdr::global_init();
		astc_ldr::global_init();
		astc_ldr::encoder_init();

		init_srgb_to_linear_table();
				
		g_library_initialized = true;
		return true;
	}

	void basisu_encoder_deinit()
	{
		opencl_deinit();

		g_library_initialized = false;
	}
		
	void error_vprintf(const char* pFmt, va_list args)
	{
		const uint32_t BUF_SIZE = 256;
		char buf[BUF_SIZE];

		va_list args_copy;
		va_copy(args_copy, args);
		int total_chars = vsnprintf(buf, sizeof(buf), pFmt, args_copy);
		va_end(args_copy);

		if (total_chars < 0)
		{
			assert(0);
			return;
		}

		fflush(stdout);

		if (total_chars >= (int)BUF_SIZE)
		{
			basisu::vector<char> var_buf(total_chars + 1);
			
			va_copy(args_copy, args);
			int total_chars_retry = vsnprintf(var_buf.data(), var_buf.size(), pFmt, args_copy);
			va_end(args_copy);

			if (total_chars_retry < 0)
			{
				assert(0);
				return;
			}

			fprintf(stderr, "ERROR: %s", var_buf.data());
		}
		else
		{
			fprintf(stderr, "ERROR: %s", buf);
		}
	}

	void error_printf(const char *pFmt, ...)
	{
		va_list args;
		va_start(args, pFmt);
		error_vprintf(pFmt, args);
		va_end(args);
	}

#if defined(_WIN32)
	void platform_sleep(uint32_t ms)
	{
		Sleep(ms);
	}
#else
	void platform_sleep(uint32_t ms)
	{
		// TODO
		BASISU_NOTE_UNUSED(ms);
	}
#endif

#if defined(_WIN32)
	inline void query_counter(timer_ticks* pTicks)
	{
		QueryPerformanceCounter(reinterpret_cast<LARGE_INTEGER*>(pTicks));
	}
	inline void query_counter_frequency(timer_ticks* pTicks)
	{
		QueryPerformanceFrequency(reinterpret_cast<LARGE_INTEGER*>(pTicks));
	}
#elif defined(__APPLE__) || defined(__FreeBSD__) || defined(__OpenBSD__) || defined(__EMSCRIPTEN__)
#include <sys/time.h>
	inline void query_counter(timer_ticks* pTicks)
	{
		struct timeval cur_time;
		gettimeofday(&cur_time, NULL);
		*pTicks = static_cast<unsigned long long>(cur_time.tv_sec) * 1000000ULL + static_cast<unsigned long long>(cur_time.tv_usec);
	}
	inline void query_counter_frequency(timer_ticks* pTicks)
	{
		*pTicks = 1000000;
	}
#elif defined(__GNUC__)
#include <sys/timex.h>
	inline void query_counter(timer_ticks* pTicks)
	{
		struct timeval cur_time;
		gettimeofday(&cur_time, NULL);
		*pTicks = static_cast<unsigned long long>(cur_time.tv_sec) * 1000000ULL + static_cast<unsigned long long>(cur_time.tv_usec);
	}
	inline void query_counter_frequency(timer_ticks* pTicks)
	{
		*pTicks = 1000000;
	}
#else
#error TODO
#endif
				
	interval_timer::interval_timer() : m_start_time(0), m_stop_time(0), m_started(false), m_stopped(false)
	{
		if (!g_timer_freq)
			init();
	}

	void interval_timer::start()
	{
		query_counter(&m_start_time);
		m_started = true;
		m_stopped = false;
	}

	void interval_timer::stop()
	{
		assert(m_started);
		query_counter(&m_stop_time);
		m_stopped = true;
	}

	double interval_timer::get_elapsed_secs() const
	{
		assert(m_started);
		if (!m_started)
			return 0;

		timer_ticks stop_time = m_stop_time;
		if (!m_stopped)
			query_counter(&stop_time);

		timer_ticks delta = stop_time - m_start_time;
		return delta * g_timer_freq;
	}
		
	void interval_timer::init()
	{
		if (!g_timer_freq)
		{
			query_counter_frequency(&g_freq);
			g_timer_freq = 1.0f / g_freq;
			query_counter(&g_init_ticks);
		}
	}

	timer_ticks interval_timer::get_ticks()
	{
		if (!g_timer_freq)
			init();
		timer_ticks ticks;
		query_counter(&ticks);
		return ticks - g_init_ticks;
	}

	double interval_timer::ticks_to_secs(timer_ticks ticks)
	{
		if (!g_timer_freq)
			init();
		return ticks * g_timer_freq;
	}

	// Note this is linear<->sRGB, NOT REC709 which uses slightly different equations/transfer functions. 
	// However the gamuts/white points of REC709 and sRGB are the same.
	float linear_to_srgb(float l)
	{
		assert(l >= 0.0f && l <= 1.0f);
		if (l < .0031308f)
			return saturate(l * 12.92f);
		else
			return saturate(1.055f * powf(l, 1.0f / 2.4f) - .055f);
	}
		
	float srgb_to_linear(float s)
	{
		assert(s >= 0.0f && s <= 1.0f);
		if (s < .04045f)
			return saturate(s * (1.0f / 12.92f));
		else
			return saturate(powf((s + .055f) * (1.0f / 1.055f), 2.4f));
	}
		
	const uint32_t MAX_32BIT_ALLOC_SIZE = 250000000;
		
	bool load_tga(const char* pFilename, image& img)
	{
		int w = 0, h = 0, n_chans = 0;
		uint8_t* pImage_data = read_tga(pFilename, w, h, n_chans);
				
		if ((!pImage_data) || (!w) || (!h) || ((n_chans != 3) && (n_chans != 4)))
		{
			error_printf("Failed loading .TGA image \"%s\"!\n", pFilename);

			if (pImage_data)
				free(pImage_data);
						
			return false;
		}

		if (sizeof(void *) == sizeof(uint32_t))
		{
			if (((uint64_t)w * h * n_chans) > MAX_32BIT_ALLOC_SIZE)
			{
				error_printf("Image \"%s\" is too large (%ux%u) to process in a 32-bit build!\n", pFilename, w, h);

				if (pImage_data)
					free(pImage_data);

				return false;
			}
		}
		
		img.resize(w, h);

		const uint8_t *pSrc = pImage_data;
		for (int y = 0; y < h; y++)
		{
			color_rgba *pDst = &img(0, y);

			for (int x = 0; x < w; x++)
			{
				pDst->r = pSrc[0];
				pDst->g = pSrc[1];
				pDst->b = pSrc[2];
				pDst->a = (n_chans == 3) ? 255 : pSrc[3];

				pSrc += n_chans;
				++pDst;
			}
		}

		free(pImage_data);

		return true;
	}
		
	bool load_qoi(const uint8_t* pBuf, size_t buf_size, image& img)
	{
		qoi_desc desc;
		clear_obj(desc);

		void* p = qoi_decode(pBuf, (size_t)buf_size, &desc, 4);
		if (!p)
			return false;

		img.grant_ownership(static_cast<color_rgba*>(p), desc.width, desc.height);

		return true;
	}

	bool load_qoi(const char* pFilename, image& img)
	{
		qoi_desc desc;
		clear_obj(desc);

		void* p = qoi_read(pFilename, &desc, 4);
		if (!p)
			return false;

		img.grant_ownership(static_cast<color_rgba *>(p), desc.width, desc.height);

		return true;
	}

	bool load_png(const uint8_t *pBuf, size_t buf_size, image &img, const char *pFilename)
	{
		interval_timer tm;
		tm.start();
		
		if (!buf_size)
			return false;

		uint32_t width = 0, height = 0, num_chans = 0;
		void* pImage = pv_png::load_png(pBuf, buf_size, 4, width, height, num_chans);

		if (!pImage)
		{
			error_printf("pv_png::load_png failed while loading image \"%s\"\n", pFilename);
			return false;
		}

		img.grant_ownership(reinterpret_cast<color_rgba*>(pImage), width, height);

		//debug_printf("Total load_png() time: %3.3f secs\n", tm.get_elapsed_secs());

		return true;
	}
		
	bool load_png(const char* pFilename, image& img)
	{
		uint8_vec buffer;
		if (!read_file_to_vec(pFilename, buffer))
		{
			error_printf("load_png: Failed reading file \"%s\"!\n", pFilename);
			return false;
		}

		return load_png(buffer.data(), buffer.size(), img, pFilename);
	}

	bool load_jpg(const char *pFilename, image& img)
	{
		int width = 0, height = 0, actual_comps = 0;
		uint8_t *pImage_data = jpgd::decompress_jpeg_image_from_file(pFilename, &width, &height, &actual_comps, 4, jpgd::jpeg_decoder::cFlagLinearChromaFiltering);
		if (!pImage_data)
			return false;
		
		img.init(pImage_data, width, height, 4);
		
		free(pImage_data);

		return true;
	}

	bool load_jpg(const uint8_t* pBuf, size_t buf_size, image& img)
	{
		if (buf_size > INT_MAX)
		{
			assert(0);
			return false;
		}

		int width = 0, height = 0, actual_comps = 0;
		uint8_t* pImage_data = jpgd::decompress_jpeg_image_from_memory(pBuf, (int)buf_size, &width, &height, &actual_comps, 4, jpgd::jpeg_decoder::cFlagLinearChromaFiltering);
		if (!pImage_data)
			return false;

		img.init(pImage_data, width, height, 4);

		free(pImage_data);

		return true;
	}

	bool load_image(const char* pFilename, image& img)
	{
		std::string ext(string_get_extension(std::string(pFilename)));

		if (ext.length() == 0)
			return false;

		const char *pExt = ext.c_str();

		if (strcasecmp(pExt, "png") == 0)
			return load_png(pFilename, img);
		if (strcasecmp(pExt, "tga") == 0)
			return load_tga(pFilename, img);
		if (strcasecmp(pExt, "qoi") == 0)
			return load_qoi(pFilename, img);
		if ( (strcasecmp(pExt, "jpg") == 0) || (strcasecmp(pExt, "jfif") == 0) || (strcasecmp(pExt, "jpeg") == 0) )
			return load_jpg(pFilename, img);

		return false;
	}

	void convert_ldr_to_hdr_image(imagef &img, const image &ldr_img, bool ldr_srgb_to_linear, float linear_nit_multiplier, float ldr_black_bias)
	{
		img.resize(ldr_img.get_width(), ldr_img.get_height());

		for (uint32_t y = 0; y < ldr_img.get_height(); y++)
		{
			for (uint32_t x = 0; x < ldr_img.get_width(); x++)
			{
				const color_rgba& c = ldr_img(x, y);

				vec4F& d = img(x, y);
				if (ldr_srgb_to_linear)
				{
					float r = (float)c[0];
					float g = (float)c[1];
					float b = (float)c[2];

					if (ldr_black_bias > 0.0f)
					{
						// ASTC HDR is noticeably weaker dealing with blocks containing some pixels with components set to 0.
						// Add a very slight bias less than .5 to avoid this difficulity. When the HDR image is mapped to SDR sRGB and rounded back to 8-bits, this bias will still result in zero.
						// (FWIW, in reality, a physical monitor would be unlikely to have a perfectly zero black level.)
						// This is purely optional and on most images it doesn't matter visually.
						if (r == 0.0f)
							r = ldr_black_bias;
						if (g == 0.0f)
							g = ldr_black_bias;
						if (b == 0.0f)
							b = ldr_black_bias;
					}

					// Compute how much linear light would be emitted by a SDR 80-100 nit monitor.
					d[0] = srgb_to_linear(r * (1.0f / 255.0f)) * linear_nit_multiplier;
					d[1] = srgb_to_linear(g * (1.0f / 255.0f)) * linear_nit_multiplier;
					d[2] = srgb_to_linear(b * (1.0f / 255.0f)) * linear_nit_multiplier;
				}
				else
				{
					d[0] = c[0] * (1.0f / 255.0f) * linear_nit_multiplier;
					d[1] = c[1] * (1.0f / 255.0f) * linear_nit_multiplier;
					d[2] = c[2] * (1.0f / 255.0f) * linear_nit_multiplier;
				}
				d[3] = c[3] * (1.0f / 255.0f);
			}
		}
	}

	bool load_image_hdr(const void* pMem, size_t mem_size, imagef& img, uint32_t width, uint32_t height, hdr_image_type img_type, bool ldr_srgb_to_linear, float linear_nit_multiplier, float ldr_black_bias)
	{
		if ((!pMem) || (!mem_size))
		{
			assert(0);
			return false;
		}

		switch (img_type)
		{
		case hdr_image_type::cHITRGBAHalfFloat:
		{
			if (mem_size != (uint64_t)width * height * sizeof(basist::half_float) * 4)
			{
				assert(0);
				return false;
			}

			if ((!width) || (!height))
			{
				assert(0);
				return false;
			}

			const basist::half_float* pSrc_image_h = static_cast<const basist::half_float *>(pMem);

			img.resize(width, height);
			for (uint32_t y = 0; y < height; y++)
			{
				for (uint32_t x = 0; x < width; x++)
				{
					const basist::half_float* pSrc_pixel = &pSrc_image_h[x * 4];

					vec4F& dst = img(x, y);
					dst[0] = basist::half_to_float(pSrc_pixel[0]);
					dst[1] = basist::half_to_float(pSrc_pixel[1]);
					dst[2] = basist::half_to_float(pSrc_pixel[2]);
					dst[3] = basist::half_to_float(pSrc_pixel[3]);
				}
			
				pSrc_image_h += (width * 4);
			}

			break;
		}
		case hdr_image_type::cHITRGBAFloat:
		{
			if (mem_size != (uint64_t)width * height * sizeof(float) * 4)
			{
				assert(0);
				return false;
			}

			if ((!width) || (!height))
			{
				assert(0);
				return false;
			}

			img.resize(width, height);
			memcpy((void *)img.get_ptr(), pMem, width * height * sizeof(float) * 4);

			break;
		}
		case hdr_image_type::cHITJPGImage:
		{
			image ldr_img;
			if (!load_jpg(static_cast<const uint8_t*>(pMem), mem_size, ldr_img))
				return false;

			convert_ldr_to_hdr_image(img, ldr_img, ldr_srgb_to_linear, linear_nit_multiplier, ldr_black_bias);
			break;
		}
		case hdr_image_type::cHITPNGImage:
		{
			image ldr_img;
			if (!load_png(static_cast<const uint8_t *>(pMem), mem_size, ldr_img))
				return false;

			convert_ldr_to_hdr_image(img, ldr_img, ldr_srgb_to_linear, linear_nit_multiplier, ldr_black_bias);
			break;
		}
		case hdr_image_type::cHITQOIImage:
		{
			image ldr_img;
			if (!load_qoi(static_cast<const uint8_t*>(pMem), mem_size, ldr_img))
				return false;

			convert_ldr_to_hdr_image(img, ldr_img, ldr_srgb_to_linear, linear_nit_multiplier, ldr_black_bias);
			break;
		}
		case hdr_image_type::cHITRGBA8Image:
		{
			if (!width || !height)
				return false;

			const uint64_t expected_size = (uint64_t)width * height * sizeof(uint32_t);
			if (mem_size != expected_size)
				return false;

			image ldr_img(static_cast<const uint8_t*>(pMem), width, height, 4);
						
			convert_ldr_to_hdr_image(img, ldr_img, ldr_srgb_to_linear, linear_nit_multiplier, ldr_black_bias);
						
			break;
		}
		case hdr_image_type::cHITEXRImage:
		{
			if (!read_exr(pMem, mem_size, img))
				return false;

			break;
		}
		case hdr_image_type::cHITHDRImage:
		{
			uint8_vec buf(mem_size);
			memcpy(buf.get_ptr(), pMem, mem_size);

			rgbe_header_info hdr;
			if (!read_rgbe(buf, img, hdr))
				return false;

			break;
		}
		default:
			assert(0);
			return false;
		}

		return true;
	}

	bool is_image_filename_hdr(const char *pFilename)
	{
		std::string ext(string_get_extension(std::string(pFilename)));

		if (ext.length() == 0)
			return false;

		const char* pExt = ext.c_str();

		return ((strcasecmp(pExt, "hdr") == 0) || (strcasecmp(pExt, "exr") == 0));
	}
	
	// TODO: move parameters to struct, add a HDR clean flag to eliminate NaN's/Inf's
	bool load_image_hdr(const char* pFilename, imagef& img, bool ldr_srgb_to_linear, float linear_nit_multiplier, float ldr_black_bias)
	{
		std::string ext(string_get_extension(std::string(pFilename)));

		if (ext.length() == 0)
			return false;

		const char* pExt = ext.c_str();

		if (strcasecmp(pExt, "hdr") == 0)
		{
			rgbe_header_info rgbe_info;
			if (!read_rgbe(pFilename, img, rgbe_info))
				return false;
			return true;
		}
					
		if (strcasecmp(pExt, "exr") == 0)
		{
			int n_chans = 0;
			if (!read_exr(pFilename, img, n_chans))
				return false;
			return true;
		}

		// Try loading image as LDR, then optionally convert to linear light.
		{
			image ldr_img;
			if (!load_image(pFilename, ldr_img))
				return false;

			convert_ldr_to_hdr_image(img, ldr_img, ldr_srgb_to_linear, linear_nit_multiplier, ldr_black_bias);
		}

		return true;
	}
			
	bool save_png(const char* pFilename, const image &img, uint32_t image_save_flags, uint32_t grayscale_comp)
	{
		if (!img.get_total_pixels())
			return false;
				
		void* pPNG_data = nullptr;
		size_t PNG_data_size = 0;

		if (image_save_flags & cImageSaveGrayscale)
		{
			uint8_vec g_pixels(img.get_total_pixels());
			uint8_t* pDst = &g_pixels[0];

			for (uint32_t y = 0; y < img.get_height(); y++)
				for (uint32_t x = 0; x < img.get_width(); x++)
					*pDst++ = img(x, y)[grayscale_comp];

			pPNG_data = buminiz::tdefl_write_image_to_png_file_in_memory_ex(g_pixels.data(), img.get_width(), img.get_height(), 1, &PNG_data_size, 1, false);
		}
		else
		{
			bool has_alpha = false;
			
			if ((image_save_flags & cImageSaveIgnoreAlpha) == 0)
				has_alpha = img.has_alpha();

			if (!has_alpha)
			{
				uint8_vec rgb_pixels(img.get_total_pixels() * 3);
				uint8_t* pDst = &rgb_pixels[0];

				for (uint32_t y = 0; y < img.get_height(); y++)
				{
					const color_rgba* pSrc = &img(0, y);
					for (uint32_t x = 0; x < img.get_width(); x++)
					{
						pDst[0] = pSrc->r;
						pDst[1] = pSrc->g;
						pDst[2] = pSrc->b;
						
						pSrc++;
						pDst += 3;
					}
				}

				pPNG_data = buminiz::tdefl_write_image_to_png_file_in_memory_ex(rgb_pixels.data(), img.get_width(), img.get_height(), 3, &PNG_data_size, 1, false);
			}
			else
			{
				pPNG_data = buminiz::tdefl_write_image_to_png_file_in_memory_ex(img.get_ptr(), img.get_width(), img.get_height(), 4, &PNG_data_size, 1, false);
			}
		}

		if (!pPNG_data)
			return false;

		bool status = write_data_to_file(pFilename, pPNG_data, PNG_data_size);
		if (!status)
		{
			error_printf("save_png: Failed writing to filename \"%s\"!\n", pFilename);
		}

		free(pPNG_data);
						
		return status;
	}

	bool save_qoi(const char* pFilename, const image& img, uint32_t qoi_colorspace)
	{
		assert(img.get_width() && img.get_height());

		qoi_desc desc;
		clear_obj(desc);

		desc.width = img.get_width();
		desc.height = img.get_height();
		desc.channels = 4;
		desc.colorspace = (uint8_t)qoi_colorspace;
		
		int out_len = 0;
		void* pData = qoi_encode(img.get_ptr(), &desc, &out_len);
		if ((!pData) || (!out_len))
			return false;

		const bool status = write_data_to_file(pFilename, pData, out_len);

		QOI_FREE(pData);
		pData = nullptr;

		return status;
	}
		
	bool read_file_to_vec(const char* pFilename, uint8_vec& data)
	{
		FILE* pFile = nullptr;
#ifdef _WIN32
		fopen_s(&pFile, pFilename, "rb");
#else
		pFile = fopen(pFilename, "rb");
#endif
		if (!pFile)
			return false;
				
		fseek(pFile, 0, SEEK_END);
#ifdef _WIN32
		int64_t filesize = _ftelli64(pFile);
#else
		int64_t filesize = ftello(pFile);
#endif
		if (filesize < 0)
		{
			fclose(pFile);
			return false;
		}
		fseek(pFile, 0, SEEK_SET);

		if (sizeof(size_t) == sizeof(uint32_t))
		{
			if (filesize > 0x70000000)
			{
				// File might be too big to load safely in one alloc
				fclose(pFile);
				return false;
			}
		}

		if (!data.try_resize((size_t)filesize))
		{
			fclose(pFile);
			return false;
		}

		if (filesize)
		{
			if (fread(&data[0], 1, (size_t)filesize, pFile) != (size_t)filesize)
			{
				fclose(pFile);
				return false;
			}
		}

		fclose(pFile);
		return true;
	}

	bool read_file_to_data(const char* pFilename, void *pData, size_t len)
	{
		assert(pData && len);
		if ((!pData) || (!len))
			return false;

		FILE* pFile = nullptr;
#ifdef _WIN32
		fopen_s(&pFile, pFilename, "rb");
#else
		pFile = fopen(pFilename, "rb");
#endif
		if (!pFile)
			return false;

		fseek(pFile, 0, SEEK_END);
#ifdef _WIN32
		int64_t filesize = _ftelli64(pFile);
#else
		int64_t filesize = ftello(pFile);
#endif

		if ((filesize < 0) || ((size_t)filesize < len))
		{
			fclose(pFile);
			return false;
		}
		fseek(pFile, 0, SEEK_SET);
				
		if (fread(pData, 1, (size_t)len, pFile) != (size_t)len)
		{
			fclose(pFile);
			return false;
		}

		fclose(pFile);
		return true;
	}

	bool write_data_to_file(const char* pFilename, const void* pData, size_t len)
	{
		FILE* pFile = nullptr;
#ifdef _WIN32
		fopen_s(&pFile, pFilename, "wb");
#else
		pFile = fopen(pFilename, "wb");
#endif
		if (!pFile)
			return false;

		if (len)
		{
			if (fwrite(pData, 1, len, pFile) != len)
			{
				fclose(pFile);
				return false;
			}
		}

		return fclose(pFile) != EOF;
	}
		
	bool image_resample(const image &src, image &dst, bool srgb,
		const char *pFilter, float filter_scale, 
		bool wrapping,
		uint32_t first_comp, uint32_t num_comps, 
		float filter_scale_y)
	{
		assert((first_comp + num_comps) <= 4);

		const int cMaxComps = 4;
				
		const uint32_t src_w = src.get_width(), src_h = src.get_height();
		const uint32_t dst_w = dst.get_width(), dst_h = dst.get_height();
				
		if (maximum(src_w, src_h) > BASISU_RESAMPLER_MAX_DIMENSION)
		{
			printf("Image is too large!\n");
			return false;
		}

		if (!src_w || !src_h || !dst_w || !dst_h)
			return false;
				
		if ((num_comps < 1) || (num_comps > cMaxComps))
			return false;
				
		if ((minimum(dst_w, dst_h) < 1) || (maximum(dst_w, dst_h) > BASISU_RESAMPLER_MAX_DIMENSION))
		{
			printf("Image is too large!\n");
			return false;
		}

		if ( (src_w == dst_w) && (src_h == dst_h) && 
			(filter_scale == 1.0f) &&
			((filter_scale_y < 0.0f) || (filter_scale_y == 1.0f)) )
		{
			dst = src;
			return true;
		}

		float srgb_to_linear_table[256];
		if (srgb)
		{
			for (int i = 0; i < 256; ++i)
				srgb_to_linear_table[i] = srgb_to_linear((float)i * (1.0f/255.0f));
		}

		const int LINEAR_TO_SRGB_TABLE_SIZE = 8192;
		uint8_t linear_to_srgb_table[LINEAR_TO_SRGB_TABLE_SIZE];

		if (srgb)
		{
			for (int i = 0; i < LINEAR_TO_SRGB_TABLE_SIZE; ++i)
				linear_to_srgb_table[i] = (uint8_t)clamp<int>((int)(255.0f * linear_to_srgb((float)i * (1.0f / (LINEAR_TO_SRGB_TABLE_SIZE - 1))) + .5f), 0, 255);
		}

		std::vector<float> samples[cMaxComps];
		Resampler *resamplers[cMaxComps];
		
		resamplers[0] = new Resampler(src_w, src_h, dst_w, dst_h,
			wrapping ? Resampler::BOUNDARY_WRAP : Resampler::BOUNDARY_CLAMP, 0.0f, 1.0f,
			pFilter, nullptr, nullptr, 
			filter_scale, (filter_scale_y >= 0.0f) ? filter_scale_y : filter_scale, 0, 0);
		samples[0].resize(src_w);

		for (uint32_t i = 1; i < num_comps; ++i)
		{
			resamplers[i] = new Resampler(src_w, src_h, dst_w, dst_h,
				wrapping ? Resampler::BOUNDARY_WRAP : Resampler::BOUNDARY_CLAMP, 0.0f, 1.0f,
				pFilter, resamplers[0]->get_clist_x(), resamplers[0]->get_clist_y(), 
				filter_scale, (filter_scale_y >= 0.0f) ? filter_scale_y : filter_scale, 0, 0);
			samples[i].resize(src_w);
		}

		uint32_t dst_y = 0;

		for (uint32_t src_y = 0; src_y < src_h; ++src_y)
		{
			const color_rgba *pSrc = &src(0, src_y);

			// Put source lines into resampler(s)
			for (uint32_t x = 0; x < src_w; ++x)
			{
				for (uint32_t c = 0; c < num_comps; ++c)
				{
					const uint32_t comp_index = first_comp + c;
					const uint32_t v = (*pSrc)[comp_index];

					if (!srgb || (comp_index == 3))
						samples[c][x] = v * (1.0f / 255.0f);
					else
						samples[c][x] = srgb_to_linear_table[v];
				}

				pSrc++;
			}

			for (uint32_t c = 0; c < num_comps; ++c)
			{
				if (!resamplers[c]->put_line(&samples[c][0]))
				{
					for (uint32_t i = 0; i < num_comps; i++)
						delete resamplers[i];
					return false;
				}
			}

			// Now retrieve any output lines
			for (;;)
			{
				uint32_t c;
				for (c = 0; c < num_comps; ++c)
				{
					const uint32_t comp_index = first_comp + c;

					const float *pOutput_samples = resamplers[c]->get_line();
					if (!pOutput_samples)
						break;

					const bool linear_flag = !srgb || (comp_index == 3);
					
					color_rgba *pDst = &dst(0, dst_y);

					for (uint32_t x = 0; x < dst_w; x++)
					{
						// TODO: Add dithering
						if (linear_flag)
						{
							int j = (int)(255.0f * pOutput_samples[x] + .5f);
							(*pDst)[comp_index] = (uint8_t)clamp<int>(j, 0, 255);
						}
						else
						{
							int j = (int)((LINEAR_TO_SRGB_TABLE_SIZE - 1) * pOutput_samples[x] + .5f);
							(*pDst)[comp_index] = linear_to_srgb_table[clamp<int>(j, 0, LINEAR_TO_SRGB_TABLE_SIZE - 1)];
						}

						pDst++;
					}
				}
				if (c < num_comps)
					break;

				++dst_y;
			}
		}

		for (uint32_t i = 0; i < num_comps; ++i)
			delete resamplers[i];

		return true;
	}

	bool image_resample(const imagef& src, imagef& dst, 
		const char* pFilter, float filter_scale,
		bool wrapping,
		uint32_t first_comp, uint32_t num_comps)
	{
		assert((first_comp + num_comps) <= 4);

		const int cMaxComps = 4;

		const uint32_t src_w = src.get_width(), src_h = src.get_height();
		const uint32_t dst_w = dst.get_width(), dst_h = dst.get_height();

		if (maximum(src_w, src_h) > BASISU_RESAMPLER_MAX_DIMENSION)
		{
			printf("Image is too large!\n");
			return false;
		}

		if (!src_w || !src_h || !dst_w || !dst_h)
			return false;

		if ((num_comps < 1) || (num_comps > cMaxComps))
			return false;

		if ((minimum(dst_w, dst_h) < 1) || (maximum(dst_w, dst_h) > BASISU_RESAMPLER_MAX_DIMENSION))
		{
			printf("Image is too large!\n");
			return false;
		}

		if ((src_w == dst_w) && (src_h == dst_h) && (filter_scale == 1.0f))
		{
			dst = src;
			return true;
		}

		std::vector<float> samples[cMaxComps];
		Resampler* resamplers[cMaxComps];

		resamplers[0] = new Resampler(src_w, src_h, dst_w, dst_h,
			wrapping ? Resampler::BOUNDARY_WRAP : Resampler::BOUNDARY_CLAMP, 1.0f, 0.0f, // no clamping
			pFilter, nullptr, nullptr, filter_scale, filter_scale, 0, 0);
		samples[0].resize(src_w);

		for (uint32_t i = 1; i < num_comps; ++i)
		{
			resamplers[i] = new Resampler(src_w, src_h, dst_w, dst_h,
				wrapping ? Resampler::BOUNDARY_WRAP : Resampler::BOUNDARY_CLAMP, 1.0f, 0.0f, // no clamping
				pFilter, resamplers[0]->get_clist_x(), resamplers[0]->get_clist_y(), filter_scale, filter_scale, 0, 0);
			samples[i].resize(src_w);
		}

		uint32_t dst_y = 0;

		for (uint32_t src_y = 0; src_y < src_h; ++src_y)
		{
			const vec4F* pSrc = &src(0, src_y);

			// Put source lines into resampler(s)
			for (uint32_t x = 0; x < src_w; ++x)
			{
				for (uint32_t c = 0; c < num_comps; ++c)
				{
					const uint32_t comp_index = first_comp + c;
					const float v = (*pSrc)[comp_index];

					samples[c][x] = v;
				}

				pSrc++;
			}

			for (uint32_t c = 0; c < num_comps; ++c)
			{
				if (!resamplers[c]->put_line(&samples[c][0]))
				{
					for (uint32_t i = 0; i < num_comps; i++)
						delete resamplers[i];
					return false;
				}
			}

			// Now retrieve any output lines
			for (;;)
			{
				uint32_t c;
				for (c = 0; c < num_comps; ++c)
				{
					const uint32_t comp_index = first_comp + c;

					const float* pOutput_samples = resamplers[c]->get_line();
					if (!pOutput_samples)
						break;
										
					vec4F* pDst = &dst(0, dst_y);

					for (uint32_t x = 0; x < dst_w; x++)
					{
						(*pDst)[comp_index] = pOutput_samples[x];
						pDst++;
					}
				}
				if (c < num_comps)
					break;

				++dst_y;
			}
		}

		for (uint32_t i = 0; i < num_comps; ++i)
			delete resamplers[i];

		return true;
	}

	void canonical_huffman_calculate_minimum_redundancy(sym_freq *A, int num_syms)
	{
		// See the paper "In-Place Calculation of Minimum Redundancy Codes" by Moffat and Katajainen
		if (!num_syms)
			return;

		if (1 == num_syms)
		{
			A[0].m_key = 1;
			return;
		}
		
		A[0].m_key += A[1].m_key;
		
		int s = 2, r = 0, next;
		for (next = 1; next < (num_syms - 1); ++next)
		{
			if ((s >= num_syms) || (A[r].m_key < A[s].m_key))
			{
				A[next].m_key = A[r].m_key;
				A[r].m_key = next;
				++r;
			}
			else
			{
				A[next].m_key = A[s].m_key;
				++s;
			}

			if ((s >= num_syms) || ((r < next) && A[r].m_key < A[s].m_key))
			{
				A[next].m_key = A[next].m_key + A[r].m_key;
				A[r].m_key = next;
				++r;
			}
			else
			{
				A[next].m_key = A[next].m_key + A[s].m_key;
				++s;
			}
		}
		A[num_syms - 2].m_key = 0;

		for (next = num_syms - 3; next >= 0; --next)
		{
			A[next].m_key = 1 + A[A[next].m_key].m_key;
		}

		int num_avail = 1, num_used = 0, depth = 0;
		r = num_syms - 2;
		next = num_syms - 1;
		while (num_avail > 0)
		{
			for ( ; (r >= 0) && ((int)A[r].m_key == depth); ++num_used, --r )
				;

			for ( ; num_avail > num_used; --next, --num_avail)
				A[next].m_key = depth;

			num_avail = 2 * num_used;
			num_used = 0;
			++depth;
		}
	}

	void canonical_huffman_enforce_max_code_size(int *pNum_codes, int code_list_len, int max_code_size)
	{
		int i;
		uint32_t total = 0;
		if (code_list_len <= 1)
			return;

		for (i = max_code_size + 1; i <= cHuffmanMaxSupportedInternalCodeSize; i++)
			pNum_codes[max_code_size] += pNum_codes[i];

		for (i = max_code_size; i > 0; i--)
			total += (((uint32_t)pNum_codes[i]) << (max_code_size - i));

		while (total != (1UL << max_code_size))
		{
			pNum_codes[max_code_size]--;
			for (i = max_code_size - 1; i > 0; i--)
			{
				if (pNum_codes[i])
				{
					pNum_codes[i]--;
					pNum_codes[i + 1] += 2;
					break;
				}
			}

			total--;
		}
	}

	sym_freq *canonical_huffman_radix_sort_syms(uint32_t num_syms, sym_freq *pSyms0, sym_freq *pSyms1)
	{
		uint32_t total_passes = 2, pass_shift, pass, i, hist[256 * 2];
		sym_freq *pCur_syms = pSyms0, *pNew_syms = pSyms1;

		clear_obj(hist);

		for (i = 0; i < num_syms; i++)
		{
			uint32_t freq = pSyms0[i].m_key;
			
			// We scale all input frequencies to 16-bits.
			assert(freq <= UINT16_MAX);

			hist[freq & 0xFF]++;
			hist[256 + ((freq >> 8) & 0xFF)]++;
		}

		while ((total_passes > 1) && (num_syms == hist[(total_passes - 1) * 256]))
			total_passes--;

		for (pass_shift = 0, pass = 0; pass < total_passes; pass++, pass_shift += 8)
		{
			const uint32_t *pHist = &hist[pass << 8];
			uint32_t offsets[256], cur_ofs = 0;
			for (i = 0; i < 256; i++)
			{
				offsets[i] = cur_ofs;
				cur_ofs += pHist[i];
			}

			for (i = 0; i < num_syms; i++)
				pNew_syms[offsets[(pCur_syms[i].m_key >> pass_shift) & 0xFF]++] = pCur_syms[i];

			sym_freq *t = pCur_syms;
			pCur_syms = pNew_syms;
			pNew_syms = t;
		}

		return pCur_syms;
	}

	bool huffman_encoding_table::init(uint32_t num_syms, const uint16_t *pFreq, uint32_t max_code_size)
	{
		if (max_code_size > cHuffmanMaxSupportedCodeSize)
			return false;
		if ((!num_syms) || (num_syms > cHuffmanMaxSyms))
			return false;

		uint32_t total_used_syms = 0;
		for (uint32_t i = 0; i < num_syms; i++)
			if (pFreq[i])
				total_used_syms++;

		if (!total_used_syms)
			return false;

		std::vector<sym_freq> sym_freq0(total_used_syms), sym_freq1(total_used_syms);
		for (uint32_t i = 0, j = 0; i < num_syms; i++)
		{
			if (pFreq[i])
			{
				sym_freq0[j].m_key = pFreq[i];
				sym_freq0[j++].m_sym_index = static_cast<uint16_t>(i);
			}
		}

		sym_freq *pSym_freq = canonical_huffman_radix_sort_syms(total_used_syms, &sym_freq0[0], &sym_freq1[0]);

		canonical_huffman_calculate_minimum_redundancy(pSym_freq, total_used_syms);

		int num_codes[cHuffmanMaxSupportedInternalCodeSize + 1];
		clear_obj(num_codes);

		for (uint32_t i = 0; i < total_used_syms; i++)
		{
			if (pSym_freq[i].m_key > cHuffmanMaxSupportedInternalCodeSize)
				return false;

			num_codes[pSym_freq[i].m_key]++;
		}

		canonical_huffman_enforce_max_code_size(num_codes, total_used_syms, max_code_size);

		m_code_sizes.resize(0);
		m_code_sizes.resize(num_syms);

		m_codes.resize(0);
		m_codes.resize(num_syms);

		for (uint32_t i = 1, j = total_used_syms; i <= max_code_size; i++)
			for (uint32_t l = num_codes[i]; l > 0; l--)
				m_code_sizes[pSym_freq[--j].m_sym_index] = static_cast<uint8_t>(i);

		uint32_t next_code[cHuffmanMaxSupportedInternalCodeSize + 1];

		next_code[1] = 0;
		for (uint32_t j = 0, i = 2; i <= max_code_size; i++)
			next_code[i] = j = ((j + num_codes[i - 1]) << 1);

		for (uint32_t i = 0; i < num_syms; i++)
		{
			uint32_t rev_code = 0, code, code_size;
			if ((code_size = m_code_sizes[i]) == 0)
				continue;
			if (code_size > cHuffmanMaxSupportedInternalCodeSize)
				return false;
			code = next_code[code_size]++;
			for (uint32_t l = code_size; l > 0; l--, code >>= 1)
				rev_code = (rev_code << 1) | (code & 1);
			m_codes[i] = static_cast<uint16_t>(rev_code);
		}

		return true;
	}

	bool huffman_encoding_table::init(uint32_t num_syms, const uint32_t *pSym_freq, uint32_t max_code_size)
	{
		if ((!num_syms) || (num_syms > cHuffmanMaxSyms))
			return false;

		uint16_vec sym_freq(num_syms);

		uint32_t max_freq = 0;
		for (uint32_t i = 0; i < num_syms; i++)
			max_freq = maximum(max_freq, pSym_freq[i]);

		if (max_freq < UINT16_MAX)
		{
			for (uint32_t i = 0; i < num_syms; i++)
				sym_freq[i] = static_cast<uint16_t>(pSym_freq[i]);
		}
		else
		{
			for (uint32_t i = 0; i < num_syms; i++)
			{
				if (pSym_freq[i])
				{
					uint32_t f = static_cast<uint32_t>((static_cast<uint64_t>(pSym_freq[i]) * 65534U + (max_freq >> 1)) / max_freq);
					sym_freq[i] = static_cast<uint16_t>(clamp<uint32_t>(f, 1, 65534));
				}
			}
		}

		return init(num_syms, &sym_freq[0], max_code_size);
	}

	void bitwise_coder::end_nonzero_run(uint16_vec &syms, uint32_t &run_size, uint32_t len)
	{
		if (run_size)
		{
			if (run_size < cHuffmanSmallRepeatSizeMin)
			{
				while (run_size--)
					syms.push_back(static_cast<uint16_t>(len));
			}
			else if (run_size <= cHuffmanSmallRepeatSizeMax)
			{
				syms.push_back(static_cast<uint16_t>(cHuffmanSmallRepeatCode | ((run_size - cHuffmanSmallRepeatSizeMin) << 6)));
			}
			else
			{
				assert((run_size >= cHuffmanBigRepeatSizeMin) && (run_size <= cHuffmanBigRepeatSizeMax));
				syms.push_back(static_cast<uint16_t>(cHuffmanBigRepeatCode | ((run_size - cHuffmanBigRepeatSizeMin) << 6)));
			}
		}

		run_size = 0;
	}

	void bitwise_coder::end_zero_run(uint16_vec &syms, uint32_t &run_size)
	{
		if (run_size)
		{
			if (run_size < cHuffmanSmallZeroRunSizeMin)
			{
				while (run_size--)
					syms.push_back(0);
			}
			else if (run_size <= cHuffmanSmallZeroRunSizeMax)
			{
				syms.push_back(static_cast<uint16_t>(cHuffmanSmallZeroRunCode | ((run_size - cHuffmanSmallZeroRunSizeMin) << 6)));
			}
			else
			{
				assert((run_size >= cHuffmanBigZeroRunSizeMin) && (run_size <= cHuffmanBigZeroRunSizeMax));
				syms.push_back(static_cast<uint16_t>(cHuffmanBigZeroRunCode | ((run_size - cHuffmanBigZeroRunSizeMin) << 6)));
			}
		}

		run_size = 0;
	}

	uint32_t bitwise_coder::emit_huffman_table(const huffman_encoding_table &tab)
	{
		const uint64_t start_bits = m_total_bits;

		const uint8_vec &code_sizes = tab.get_code_sizes();

		uint32_t total_used = tab.get_total_used_codes();
		put_bits(total_used, cHuffmanMaxSymsLog2);
			
		if (!total_used)
			return 0;

		uint16_vec syms;
		syms.reserve(total_used + 16);

		uint32_t prev_code_len = UINT_MAX, zero_run_size = 0, nonzero_run_size = 0;

		for (uint32_t i = 0; i <= total_used; ++i)
		{
			const uint32_t code_len = (i == total_used) ? 0xFF : code_sizes[i];
			assert((code_len == 0xFF) || (code_len <= 16));

			if (code_len)
			{
				end_zero_run(syms, zero_run_size);

				if (code_len != prev_code_len)
				{
					end_nonzero_run(syms, nonzero_run_size, prev_code_len);
					if (code_len != 0xFF)
						syms.push_back(static_cast<uint16_t>(code_len));
				}
				else if (++nonzero_run_size == cHuffmanBigRepeatSizeMax)
					end_nonzero_run(syms, nonzero_run_size, prev_code_len);
			}
			else
			{
				end_nonzero_run(syms, nonzero_run_size, prev_code_len);

				if (++zero_run_size == cHuffmanBigZeroRunSizeMax)
					end_zero_run(syms, zero_run_size);
			}

			prev_code_len = code_len;
		}

		histogram h(cHuffmanTotalCodelengthCodes);
		for (uint32_t i = 0; i < syms.size(); i++)
			h.inc(syms[i] & 63);

		huffman_encoding_table ct;
		if (!ct.init(h, 7))
			return 0;

		assert(cHuffmanTotalSortedCodelengthCodes == cHuffmanTotalCodelengthCodes);

		uint32_t total_codelength_codes;
		for (total_codelength_codes = cHuffmanTotalSortedCodelengthCodes; total_codelength_codes > 0; total_codelength_codes--)
			if (ct.get_code_sizes()[g_huffman_sorted_codelength_codes[total_codelength_codes - 1]])
				break;

		assert(total_codelength_codes);

		put_bits(total_codelength_codes, 5);
		for (uint32_t i = 0; i < total_codelength_codes; i++)
			put_bits(ct.get_code_sizes()[g_huffman_sorted_codelength_codes[i]], 3);

		for (uint32_t i = 0; i < syms.size(); ++i)
		{
			const uint32_t l = syms[i] & 63, e = syms[i] >> 6;

			put_code(l, ct);
				
			if (l == cHuffmanSmallZeroRunCode)
				put_bits(e, cHuffmanSmallZeroRunExtraBits);
			else if (l == cHuffmanBigZeroRunCode)
				put_bits(e, cHuffmanBigZeroRunExtraBits);
			else if (l == cHuffmanSmallRepeatCode)
				put_bits(e, cHuffmanSmallRepeatExtraBits);
			else if (l == cHuffmanBigRepeatCode)
				put_bits(e, cHuffmanBigRepeatExtraBits);
		}

		return (uint32_t)(m_total_bits - start_bits);
	}

	bool huffman_test(int rand_seed)
	{
		histogram h(19);

		// Feed in a fibonacci sequence to force large codesizes
		h[0] += 1; h[1] += 1; h[2] += 2; h[3] += 3;
		h[4] += 5; h[5] += 8; h[6] += 13; h[7] += 21;
		h[8] += 34; h[9] += 55; h[10] += 89; h[11] += 144;
		h[12] += 233; h[13] += 377; h[14] += 610; h[15] += 987;
		h[16] += 1597; h[17] += 2584; h[18] += 4181;

		huffman_encoding_table etab;
		etab.init(h, 16);
		
		{
			bitwise_coder c;
			c.init(1024);

			c.emit_huffman_table(etab);
			for (int i = 0; i < 19; i++)
				c.put_code(i, etab);

			c.flush();

			basist::bitwise_decoder d;
			d.init(&c.get_bytes()[0], static_cast<uint32_t>(c.get_bytes().size()));

			basist::huffman_decoding_table dtab;
			bool success = d.read_huffman_table(dtab);
			if (!success)
			{
				assert(0);
				printf("Failure 5\n");
				return false;
			}

			for (uint32_t i = 0; i < 19; i++)
			{
				uint32_t s = d.decode_huffman(dtab);
				if (s != i)
				{
					assert(0);
					printf("Failure 5\n");
					return false;
				}
			}
		}

		basisu::rand r;
		r.seed(rand_seed);

		for (int iter = 0; iter < 500000; iter++)
		{
			printf("%u\n", iter);

			uint32_t max_sym = r.irand(0, 8193);
			uint32_t num_codes = r.irand(1, 10000);
			uint_vec syms(num_codes);

			for (uint32_t i = 0; i < num_codes; i++)
			{
				if (r.bit())
					syms[i] = r.irand(0, max_sym);
				else
				{
					int s = (int)(r.gaussian((float)max_sym / 2, (float)maximum<int>(1, max_sym / 2)) + .5f);
					s = basisu::clamp<int>(s, 0, max_sym);

					syms[i] = s;
				}

			}

			histogram h1(max_sym + 1);
			for (uint32_t i = 0; i < num_codes; i++)
				h1[syms[i]]++;

			huffman_encoding_table etab2;
			if (!etab2.init(h1, 16))
			{
				assert(0);
				printf("Failed 0\n");
				return false;
			}

			bitwise_coder c;
			c.init(1024);

			c.emit_huffman_table(etab2);

			for (uint32_t i = 0; i < num_codes; i++)
				c.put_code(syms[i], etab2);

			c.flush();

			basist::bitwise_decoder d;
			d.init(&c.get_bytes()[0], (uint32_t)c.get_bytes().size());

			basist::huffman_decoding_table dtab;
			bool success = d.read_huffman_table(dtab);
			if (!success)
			{
				assert(0);
				printf("Failed 2\n");
				return false;
			}

			for (uint32_t i = 0; i < num_codes; i++)
			{
				uint32_t s = d.decode_huffman(dtab);
				if (s != syms[i])
				{
					assert(0);
					printf("Failed 4\n");
					return false;
				}
			}

		}
		return true;
	}

	void palette_index_reorderer::init(uint32_t num_indices, const uint32_t *pIndices, uint32_t num_syms, pEntry_dist_func pDist_func, void *pCtx, float dist_func_weight)
	{
		assert((num_syms > 0) && (num_indices > 0));
		assert((dist_func_weight >= 0.0f) && (dist_func_weight <= 1.0f));

		clear();

		m_remap_table.resize(num_syms);
		m_entries_picked.reserve(num_syms);
		m_total_count_to_picked.resize(num_syms);

		if (num_indices <= 1)
			return;

		prepare_hist(num_syms, num_indices, pIndices);
		find_initial(num_syms);

		while (m_entries_to_do.size())
		{
			// Find the best entry to move into the picked list.
			uint32_t best_entry;
			double best_count;
			find_next_entry(best_entry, best_count, pDist_func, pCtx, dist_func_weight);

			// We now have chosen an entry to place in the picked list, now determine which side it goes on.
			const uint32_t entry_to_move = m_entries_to_do[best_entry];
								
			float side = pick_side(num_syms, entry_to_move, pDist_func, pCtx, dist_func_weight);
								
			// Put entry_to_move either on the "left" or "right" side of the picked entries
			if (side <= 0)
				m_entries_picked.push_back(entry_to_move);
			else
				m_entries_picked.insert(m_entries_picked.begin(), entry_to_move);

			// Erase best_entry from the todo list
			m_entries_to_do.erase(m_entries_to_do.begin() + best_entry);

			// We've just moved best_entry to the picked list, so now we need to update m_total_count_to_picked[] to factor the additional count to best_entry
			for (uint32_t i = 0; i < m_entries_to_do.size(); i++)
				m_total_count_to_picked[m_entries_to_do[i]] += get_hist(m_entries_to_do[i], entry_to_move, num_syms);
		}

		for (uint32_t i = 0; i < num_syms; i++)
			m_remap_table[m_entries_picked[i]] = i;
	}

	void palette_index_reorderer::prepare_hist(uint32_t num_syms, uint32_t num_indices, const uint32_t *pIndices)
	{
		m_hist.resize(0);
		m_hist.resize(num_syms * num_syms);

		for (uint32_t i = 0; i < num_indices; i++)
		{
			const uint32_t idx = pIndices[i];
			inc_hist(idx, (i < (num_indices - 1)) ? pIndices[i + 1] : -1, num_syms);
			inc_hist(idx, (i > 0) ? pIndices[i - 1] : -1, num_syms);
		}
	}

	void palette_index_reorderer::find_initial(uint32_t num_syms)
	{
		uint32_t max_count = 0, max_index = 0;
		for (uint32_t i = 0; i < num_syms * num_syms; i++)
			if (m_hist[i] > max_count)
				max_count = m_hist[i], max_index = i;

		uint32_t a = max_index / num_syms, b = max_index % num_syms;

		const size_t ofs = m_entries_picked.size();

		m_entries_picked.push_back(a);
		m_entries_picked.push_back(b);

		for (uint32_t i = 0; i < num_syms; i++)
			if ((i != m_entries_picked[ofs + 1]) && (i != m_entries_picked[ofs]))
				m_entries_to_do.push_back(i);

		for (uint32_t i = 0; i < m_entries_to_do.size(); i++)
			for (uint32_t j = 0; j < m_entries_picked.size(); j++)
				m_total_count_to_picked[m_entries_to_do[i]] += get_hist(m_entries_to_do[i], m_entries_picked[j], num_syms);
	}

	void palette_index_reorderer::find_next_entry(uint32_t &best_entry, double &best_count, pEntry_dist_func pDist_func, void *pCtx, float dist_func_weight)
	{
		best_entry = 0;
		best_count = 0;

		for (uint32_t i = 0; i < m_entries_to_do.size(); i++)
		{
			const uint32_t u = m_entries_to_do[i];
			double total_count = m_total_count_to_picked[u];

			if (pDist_func)
			{
				float w = maximum<float>((*pDist_func)(u, m_entries_picked.front(), pCtx), (*pDist_func)(u, m_entries_picked.back(), pCtx));
				assert((w >= 0.0f) && (w <= 1.0f));
				total_count = (total_count + 1.0f) * lerp(1.0f - dist_func_weight, 1.0f + dist_func_weight, w);
			}

			if (total_count <= best_count)
				continue;

			best_entry = i;
			best_count = total_count;
		}
	}

	float palette_index_reorderer::pick_side(uint32_t num_syms, uint32_t entry_to_move, pEntry_dist_func pDist_func, void *pCtx, float dist_func_weight)
	{
		float which_side = 0;

		int l_count = 0, r_count = 0;
		for (uint32_t j = 0; j < m_entries_picked.size(); j++)
		{
			const int count = get_hist(entry_to_move, m_entries_picked[j], num_syms), r = ((int)m_entries_picked.size() + 1 - 2 * (j + 1));
			which_side += static_cast<float>(r * count);
			if (r >= 0)
				l_count += r * count;
			else
				r_count += -r * count;
		}

		if (pDist_func)
		{
			float w_left = lerp(1.0f - dist_func_weight, 1.0f + dist_func_weight, (*pDist_func)(entry_to_move, m_entries_picked.front(), pCtx));
			float w_right = lerp(1.0f - dist_func_weight, 1.0f + dist_func_weight, (*pDist_func)(entry_to_move, m_entries_picked.back(), pCtx));
			which_side = w_left * l_count - w_right * r_count;
		}
		return which_side;
	}
	
	void image_metrics::calc(const imagef& a, const imagef& b, uint32_t first_chan, uint32_t total_chans, bool avg_comp_error, bool log)
	{
		assert((first_chan < 4U) && (first_chan + total_chans <= 4U));

		const uint32_t width = basisu::minimum(a.get_width(), b.get_width());
		const uint32_t height = basisu::minimum(a.get_height(), b.get_height());

		double max_e = -1e+30f;
		double sum = 0.0f, sum_sqr = 0.0f;

		m_width = width;
		m_height = height;
		
		m_has_neg = false;
		m_any_abnormal = false;
		m_hf_mag_overflow = false;
				
		for (uint32_t y = 0; y < height; y++)
		{
			for (uint32_t x = 0; x < width; x++)
			{
				const vec4F& ca = a(x, y), &cb = b(x, y);
								
				if (total_chans)
				{
					for (uint32_t c = 0; c < total_chans; c++)
					{
						float fa = ca[first_chan + c], fb = cb[first_chan + c];

						if ((fabs(fa) > basist::MAX_HALF_FLOAT) || (fabs(fb) > basist::MAX_HALF_FLOAT))
							m_hf_mag_overflow = true;

						if ((fa < 0.0f) || (fb < 0.0f))
							m_has_neg = true;

						if (std::isinf(fa) || std::isinf(fb) || std::isnan(fa) || std::isnan(fb))
							m_any_abnormal = true;
												
						const double delta = fabs(fa - fb);
						max_e = basisu::maximum<double>(max_e, delta);

						if (log)
						{
							double log2_delta = log2f(basisu::maximum(0.0f, fa) + 1.0f) - log2f(basisu::maximum(0.0f, fb) + 1.0f);

							sum += fabs(log2_delta);
							sum_sqr += log2_delta * log2_delta;
						}
						else
						{
							sum += fabs(delta);
							sum_sqr += delta * delta;
						}
					}
				}
				else
				{
					for (uint32_t c = 0; c < 3; c++)
					{
						float fa = ca[c], fb = cb[c];

						if ((fabs(fa) > basist::MAX_HALF_FLOAT) || (fabs(fb) > basist::MAX_HALF_FLOAT))
							m_hf_mag_overflow = true;

						if ((fa < 0.0f) || (fb < 0.0f))
							m_has_neg = true;

						if (std::isinf(fa) || std::isinf(fb) || std::isnan(fa) || std::isnan(fb))
							m_any_abnormal = true;
					}

					double ca_l = get_luminance(ca), cb_l = get_luminance(cb);
					
					double delta = fabs(ca_l - cb_l);
					max_e = basisu::maximum(max_e, delta);
					
					if (log)
					{
						double log2_delta = log2(basisu::maximum<double>(0.0f, ca_l) + 1.0f) - log2(basisu::maximum<double>(0.0f, cb_l) + 1.0f);

						sum += fabs(log2_delta);
						sum_sqr += log2_delta * log2_delta;
					}
					else
					{
						sum += delta;
						sum_sqr += delta * delta;
					}
				}
			}
		}

		m_max = (double)(max_e);

		double total_values = (double)width * (double)height;
		if (avg_comp_error)
			total_values *= (double)clamp<uint32_t>(total_chans, 1, 4);

		m_mean = (float)(sum / total_values);
		m_mean_squared = (float)(sum_sqr / total_values);
		m_rms = (float)sqrt(sum_sqr / total_values);
		
		const double max_val = 1.0f;
		m_psnr = m_rms ? (float)clamp<double>(log10(max_val / m_rms) * 20.0f, 0.0f, 1000.0f) : 1000.0f;
	}

	void image_metrics::calc_half(const imagef& a, const imagef& b, uint32_t first_chan, uint32_t total_chans, bool avg_comp_error)
	{
		assert(total_chans);
		assert((first_chan < 4U) && (first_chan + total_chans <= 4U));

		const uint32_t width = basisu::minimum(a.get_width(), b.get_width());
		const uint32_t height = basisu::minimum(a.get_height(), b.get_height());

		m_width = width;
		m_height = height;

		m_has_neg = false;
		m_hf_mag_overflow = false;
		m_any_abnormal = false;

		uint_vec hist(65536);
		
		for (uint32_t y = 0; y < height; y++)
		{
			for (uint32_t x = 0; x < width; x++)
			{
				const vec4F& ca = a(x, y), &cb = b(x, y);

				for (uint32_t i = 0; i < 4; i++)
				{
					if ((ca[i] < 0.0f) || (cb[i] < 0.0f))
						m_has_neg = true;
					
					if ((fabs(ca[i]) > basist::MAX_HALF_FLOAT) || (fabs(cb[i]) > basist::MAX_HALF_FLOAT))
						m_hf_mag_overflow = true;

					if (std::isnan(ca[i]) || std::isnan(cb[i]) || std::isinf(ca[i]) || std::isinf(cb[i]))
						m_any_abnormal = true;
				}

				int cah[4] = { basist::float_to_half(ca[0]), basist::float_to_half(ca[1]), basist::float_to_half(ca[2]), basist::float_to_half(ca[3]) };
				int cbh[4] = { basist::float_to_half(cb[0]), basist::float_to_half(cb[1]), basist::float_to_half(cb[2]), basist::float_to_half(cb[3]) };

				for (uint32_t c = 0; c < total_chans; c++)
					hist[iabs(cah[first_chan + c] - cbh[first_chan + c]) & 65535]++;

			} // x
		} // y

		m_max = 0;
		double sum = 0.0f, sum2 = 0.0f;
		for (uint32_t i = 0; i < 65536; i++)
		{
			if (hist[i])
			{
				m_max = basisu::maximum<double>(m_max, (double)i);
				double v = (double)i * (double)hist[i];
				sum += v;
				sum2 += (double)i * v;
			}
		}

		double total_values = (double)width * (double)height;
		if (avg_comp_error)
			total_values *= (double)clamp<uint32_t>(total_chans, 1, 4);

		const float max_val = 65535.0f;
		m_mean = (float)clamp<double>(sum / total_values, 0.0f, max_val);
		m_mean_squared = (float)clamp<double>(sum2 / total_values, 0.0f, max_val * max_val);
		m_rms = (float)sqrt(m_mean_squared);
		m_psnr = m_rms ? (float)clamp<double>(log10(max_val / m_rms) * 20.0f, 0.0f, 1000.0f) : 1000.0f;
	}

	// Alt. variant, same as calc_half(), for validation.
	void image_metrics::calc_half2(const imagef& a, const imagef& b, uint32_t first_chan, uint32_t total_chans, bool avg_comp_error)
	{
		assert(total_chans);
		assert((first_chan < 4U) && (first_chan + total_chans <= 4U));

		const uint32_t width = basisu::minimum(a.get_width(), b.get_width());
		const uint32_t height = basisu::minimum(a.get_height(), b.get_height());

		m_width = width;
		m_height = height;

		m_has_neg = false;
		m_hf_mag_overflow = false;
		m_any_abnormal = false;
				
		double sum = 0.0f, sum2 = 0.0f;
		m_max = 0;

		for (uint32_t y = 0; y < height; y++)
		{
			for (uint32_t x = 0; x < width; x++)
			{
				const vec4F& ca = a(x, y), & cb = b(x, y);

				for (uint32_t i = 0; i < 4; i++)
				{
					if ((ca[i] < 0.0f) || (cb[i] < 0.0f))
						m_has_neg = true;

					if ((fabs(ca[i]) > basist::MAX_HALF_FLOAT) || (fabs(cb[i]) > basist::MAX_HALF_FLOAT))
						m_hf_mag_overflow = true;

					if (std::isnan(ca[i]) || std::isnan(cb[i]) || std::isinf(ca[i]) || std::isinf(cb[i]))
						m_any_abnormal = true;
				}

				int cah[4] = { basist::float_to_half(ca[0]), basist::float_to_half(ca[1]), basist::float_to_half(ca[2]), basist::float_to_half(ca[3]) };
				int cbh[4] = { basist::float_to_half(cb[0]), basist::float_to_half(cb[1]), basist::float_to_half(cb[2]), basist::float_to_half(cb[3]) };

				for (uint32_t c = 0; c < total_chans; c++)
				{
					int diff = iabs(cah[first_chan + c] - cbh[first_chan + c]);
					if (diff)
						m_max = std::max<double>(m_max, (double)diff);

					sum += diff;
					sum2 += squarei(cah[first_chan + c] - cbh[first_chan + c]);
				}

			} // x
		} // y
						
		double total_values = (double)width * (double)height;
		if (avg_comp_error)
			total_values *= (double)clamp<uint32_t>(total_chans, 1, 4);

		const float max_val = 65535.0f;
		m_mean = (float)clamp<double>(sum / total_values, 0.0f, max_val);
		m_mean_squared = (float)clamp<double>(sum2 / total_values, 0.0f, max_val * max_val);
		m_rms = (float)sqrt(m_mean_squared);
		m_psnr = m_rms ? (float)clamp<double>(log10(max_val / m_rms) * 20.0f, 0.0f, 1000.0f) : 1000.0f;
	}

	void image_metrics::calc(const image &a, const image &b, uint32_t first_chan, uint32_t total_chans, bool avg_comp_error, bool use_601_luma)
	{
		assert((first_chan < 4U) && (first_chan + total_chans <= 4U));

		const uint32_t width = basisu::minimum(a.get_width(), b.get_width());
		const uint32_t height = basisu::minimum(a.get_height(), b.get_height());

		m_width = width;
		m_height = height;

		double hist[256];
		clear_obj(hist);

		m_has_neg = false;
		m_any_abnormal = false;
		m_hf_mag_overflow = false;
		m_sum_a = 0;
		m_sum_b = 0;

		for (uint32_t y = 0; y < height; y++)
		{
			for (uint32_t x = 0; x < width; x++)
			{
				const color_rgba &ca = a(x, y), &cb = b(x, y);

				if (total_chans)
				{
					for (uint32_t c = 0; c < total_chans; c++)
					{
						hist[iabs(ca[first_chan + c] - cb[first_chan + c])]++;
						m_sum_a += ca[first_chan + c];
						m_sum_b += cb[first_chan + c];
					}
				}
				else
				{
					if (use_601_luma)
						hist[iabs(ca.get_601_luma() - cb.get_601_luma())]++;
					else
						hist[iabs(ca.get_709_luma() - cb.get_709_luma())]++;

					for (uint32_t c = 0; c < 3; c++)
					{
						m_sum_a += ca[c];
						m_sum_b += cb[c];
					}
				}
			}
		}

		m_max = 0;
		double sum = 0.0f, sum2 = 0.0f;
		for (uint32_t i = 0; i < 256; i++)
		{
			if (hist[i])
			{
				m_max = basisu::maximum<double>(m_max, (double)i);
				double v = i * hist[i];
				sum += v;
				sum2 += i * v;
			}
		}

		double total_values = (double)width * (double)height;
		if (avg_comp_error)
			total_values *= (double)clamp<uint32_t>(total_chans, 1, 4);

		m_mean = (float)clamp<double>(sum / total_values, 0.0f, 255.0);
		m_mean_squared = (float)clamp<double>(sum2 / total_values, 0.0f, 255.0f * 255.0f);
		m_rms = (float)sqrt(m_mean_squared);
		m_psnr = m_rms ? (float)clamp<double>(log10(255.0 / m_rms) * 20.0f, 0.0f, 100.0f) : 100.0f;
	}

	void print_image_metrics(const image& a, const image& b)
	{
		image_metrics im;
		im.calc(a, b, 0, 3);
		im.print("RGB    ");

		im.calc(a, b, 0, 4);
		im.print("RGBA   ");

		im.calc(a, b, 0, 1);
		im.print("R      ");

		im.calc(a, b, 1, 1);
		im.print("G      ");

		im.calc(a, b, 2, 1);
		im.print("B      ");

		im.calc(a, b, 3, 1);
		im.print("A      ");

		im.calc(a, b, 0, 0);
		im.print("Y 709  ");

		im.calc(a, b, 0, 0, true, true);
		im.print("Y 601  ");
	}

	// PSNR-HVS and PSNR-HVS-M references:
	// https://www.researchgate.net/profile/Vladimir_Lukin2/publication/251229783_A_NEW_FULL-REFERENCE_QUALITY_METRICS_BASED_ON_HVS/links/0046351f669a9c1869000000.pdf
    // https://www.ponomarenko.info/psnrhvsm.htm
    // https://github.com/lyckantropen/psnr_hvsm
	// Note: to Match the Python implementation, convert to 8-bit REC 601 like this and compute only on Y.
	// static inline uint8_t get_601_y(int r, int g, int b) { return (uint8_t)std::round(16.0f + 65.481f * (float)r / 255.0f + 128.553f * (float)g / 255.0f + 24.966f * (float)b / 255.0f); }
	// For testing (image must be divisible by 8 pixels on each dimension, can't be grayscale): python3.12 -m psnr_hvsm image_a.png image_b.png
	// Our 8-bit 601 Y metrics should very closely (within ~.001 dB) match the Python implementation (as of 4/1/2026) on RGB images divisible by 8 pixels on each dimension, otherwise there's something wrong.
		
	static const float g_csf[64] = 
	{
		1.608443f, 2.339554f, 2.573509f, 1.608443f, 1.072295f, 0.643377f, 0.504610f, 0.421887f,
		2.144591f, 2.144591f, 1.838221f, 1.354478f, 0.989811f, 0.443708f, 0.428918f, 0.467911f,
		1.838221f, 1.979622f, 1.608443f, 1.072295f, 0.643377f, 0.451493f, 0.372972f, 0.459555f,
		1.838221f, 1.513829f, 1.169777f, 0.887417f, 0.504610f, 0.295806f, 0.321689f, 0.415082f,
		1.429727f, 1.169777f, 0.695543f, 0.459555f, 0.378457f, 0.236102f, 0.249855f, 0.334222f,
		1.072295f, 0.735288f, 0.467911f, 0.402111f, 0.317717f, 0.247453f, 0.227744f, 0.279729f,
		0.525206f, 0.402111f, 0.329937f, 0.295806f, 0.249855f, 0.212687f, 0.214459f, 0.254803f,
		0.357432f, 0.279729f, 0.270896f, 0.262603f, 0.229778f, 0.257351f, 0.249855f, 0.259950f
	};

	static const float g_mask[64] =
	{
		0.390625f, 0.826446f, 1.000000f, 0.390625f, 0.173611f, 0.062500f, 0.038447f, 0.026874f,
		0.694444f, 0.694444f, 0.510204f, 0.277008f, 0.147929f, 0.029727f, 0.027778f, 0.033058f,
		0.510204f, 0.591716f, 0.390625f, 0.173611f, 0.062500f, 0.030779f, 0.021004f, 0.031888f,
		0.510204f, 0.346021f, 0.206612f, 0.118906f, 0.038447f, 0.013212f, 0.015625f, 0.026015f,
		0.308642f, 0.206612f, 0.073046f, 0.031888f, 0.021626f, 0.008417f, 0.009426f, 0.016866f,
		0.173611f, 0.081633f, 0.033058f, 0.024414f, 0.015242f, 0.009246f, 0.007831f, 0.011815f,
		0.041649f, 0.024414f, 0.016437f, 0.013212f, 0.009426f, 0.006830f, 0.006944f, 0.009803f,
		0.019290f, 0.011815f, 0.011080f, 0.010412f, 0.007972f, 0.010000f, 0.009426f, 0.010203f
	};

	static float vari_ddof1_times_n(const float* s, uint32_t n)
	{
		assert(n);

		if (n <= 1)
			return 0.0f;

		float mean = 0.0f;
		for (uint32_t i = 0; i < n; ++i)
			mean += s[i];
		mean /= static_cast<float>(n);

		float sum_sq = 0.0f;
		for (uint32_t i = 0; i < n; ++i)
		{
			const float d = s[i] - mean;
			sum_sq += d * d;
		}

		// sample variance * N = sum_sq/(N-1) * N (to match the Python implementation)
		return sum_sq * (static_cast<float>(n) / static_cast<float>(n - 1));
	}

	static float vari_8x8_ddof1_times_n(const float block[64])
	{
		return vari_ddof1_times_n(block, 64);
	}

	static float vari_4x4_ddof1_times_n(const float block[64], uint32_t x0, uint32_t y0)
	{
		float tmp[16];
		uint32_t k = 0;
		for (uint32_t y = 0; y < 4; ++y)
			for (uint32_t x = 0; x < 4; ++x)
				tmp[k++] = block[(y0 + y) * 8 + (x0 + x)];
		return vari_ddof1_times_n(tmp, 16);
	}

	static float compute_mask_strength(const float block[64], const float dct[64])
	{
		float mask = 0.0f;
		for (uint32_t i = 1; i < 64; ++i)
			mask += (dct[i] * dct[i]) * g_mask[i];

		float pop = vari_8x8_ddof1_times_n(block);
		if (pop != 0.0f)
		{
			const float qsum = vari_4x4_ddof1_times_n(block, 0, 0) + vari_4x4_ddof1_times_n(block, 4, 0) + vari_4x4_ddof1_times_n(block, 0, 4) + vari_4x4_ddof1_times_n(block, 4, 4);
			pop = qsum / pop;
		}

		return std::sqrt(mask * pop / 16.0f / 64.0f);
	}
		
	bool psnr_hvs_compute_chan(const image& a, const image& b, int chan, psnr_hvs_chan_metrics&res)
	{
		clear_obj(res);
				
		// we allow the inputs to differ due to block size padding (which we assume has been done with clamping beyond the valid edges)
		const uint32_t width = minimum(a.get_width(), b.get_width());
		const uint32_t height = minimum(a.get_height(), b.get_height());
		
		if (!width || !height)
		{
			assert(0);
			return false;
		}

		const uint32_t num_blocks_x = (width + 7) / 8;
		const uint32_t num_blocks_y = (height + 7) / 8;
				
		basist::astc_ldr_t::dct2f dct2d;
		
		bool status = dct2d.init(8, 8);
		assert(status);
			
		if (!status)
			return false;

		basist::astc_ldr_t::fvec dct_work;

		double sum_hvs = 0.0f, sum_hvsm = 0.0f;

		// Note: Python/Matlab variants only process full blocks, we process ALL blocks with clamping as needed.
		for (uint32_t by = 0; by < num_blocks_y; by++)
		{
			for (uint32_t bx = 0; bx < num_blocks_x; bx++)
			{
				color_rgba a_block_rgba[64];
				a.extract_block_clamped(a_block_rgba, bx * 8, by * 8, 8, 8);
																
				color_rgba b_block_rgba[64];
				b.extract_block_clamped(b_block_rgba, bx * 8, by * 8, 8, 8);

				float a_block[64], b_block[64];
				if (chan < 0)
				{
					if ((psnr_hvs_channel_use)chan == psnr_hvs_channel_use::cUse601Y8Bit)
					{
						// convert to BT.601 Y 8-bit to match the Python implementation for testing/verification
						for (uint32_t i = 0; i < 64; i++)
						{
							a_block[i] = (float)get_psnr_hvs_601_y(a_block_rgba[i]) * (1.0f / 255.0f);
							b_block[i] = (float)get_psnr_hvs_601_y(b_block_rgba[i]) * (1.0f / 255.0f);
						}
					}
					else
					{
						assert((psnr_hvs_channel_use)chan == psnr_hvs_channel_use::cUse601YFloat);

						// convert to BT.601 Y float for more precision (but doesn't match the Python implementation, resulting in significantly different output)
						for (uint32_t i = 0; i < 64; i++)
						{
							a_block[i] = get_psnr_hvs_601_yf(a_block_rgba[i]);
							b_block[i] = get_psnr_hvs_601_yf(b_block_rgba[i]);
						}
					}
				}
				else
				{
					for (uint32_t i = 0; i < 64; i++)
					{
						a_block[i] = (float)(a_block_rgba[i])[chan] * (1.0f / 255.0f);
						b_block[i] = (float)(b_block_rgba[i])[chan] * (1.0f / 255.0f);
					}
				}

				float a_dct[64], b_dct[64];
				dct2d.forward(a_block, a_dct, dct_work);
				dct2d.forward(b_block, b_dct, dct_work);

				float mask_a = compute_mask_strength(a_block, a_dct);
				float mask_b = compute_mask_strength(b_block, b_dct);
				if (mask_b > mask_a)
					mask_a = mask_b;
								
				for (uint32_t i = 0; i < 64; i++)
				{
					float u = std::fabs(a_dct[i] - b_dct[i]);

					// PSNR-HVS
					{
						const float weighted = u * g_csf[i];
						sum_hvs += static_cast<double>(weighted * weighted);
					}

					// PSNR-HVS-M
					if (i != 0)
					{
						const float threshold = mask_a / g_mask[i];
						if (u < threshold)
							u = 0.0f;
						else
							u = u - threshold;
					}

					{
						const float weighted = u * g_csf[i];
						sum_hvsm += static_cast<double>(weighted * weighted);
					}

				} // j

			} // bx
		} // by

		const uint32_t total_blocks = num_blocks_x * num_blocks_y;
		const uint32_t total_samples = total_blocks * 64;

		res.m_mseh_hvs = sum_hvs / double(total_samples);
		res.m_mseh_hvsm = sum_hvsm / double(total_samples);

		res.m_psnr_hvs = psnr_hvs_calc_psnr(res.m_mseh_hvs, 1.0f);
		res.m_psnr_hvsm = psnr_hvs_calc_psnr(res.m_mseh_hvsm, 1.0f);
				
		return true;
	}
				
	bool psnr_hvs_compute_metrics(const image& a, const image& b, psnr_hvs_metrics& metrics)
	{
		metrics.clear();

		// This should closely match the psnr_hvsm project's output, but see Issue #9: https://github.com/lyckantropen/psnr_hvsm/issues/9
		bool status = psnr_hvs_compute_chan(a, b, (int)psnr_hvs_channel_use::cUse601Y8Bit, metrics.m_y_601_8bit);
		if (!status)
			return false;
		
		// Now compute as 601 float - noticeably more precise, but doesn't match psnr_hvsm (python)
		status = psnr_hvs_compute_chan(a, b, (int)psnr_hvs_channel_use::cUse601YFloat, metrics.m_y_601_float);
		if (!status)
			return false;
								
		double sum_hvs_rgb = 0, sum_hvsm_rgb = 0;
		double sum_hvs_rgba = 0, sum_hvsm_rgba = 0;
		for (uint32_t c = 0; c < 4; c++)
		{
			auto& chan_metrics = metrics.m_chan[c];

			bool status2 = psnr_hvs_compute_chan(a, b, c, chan_metrics);
			if (!status2)
				return false;
									
			if (c < 3)
			{
				sum_hvs_rgb += chan_metrics.m_mseh_hvs;
				sum_hvsm_rgb += chan_metrics.m_mseh_hvsm;
			}
			sum_hvs_rgba += chan_metrics.m_mseh_hvs;
			sum_hvsm_rgba += chan_metrics.m_mseh_hvsm;
		}

		sum_hvs_rgb /= 3.0f;
		sum_hvsm_rgb /= 3.0f;

		sum_hvs_rgba /= 4.0f;
		sum_hvsm_rgba /= 4.0f;

		metrics.m_rgb.m_mseh_hvs = sum_hvs_rgb;
		metrics.m_rgb.m_mseh_hvsm = sum_hvsm_rgb;
		metrics.m_rgb.m_psnr_hvs = psnr_hvs_calc_psnr(sum_hvs_rgb, 1.0f);
		metrics.m_rgb.m_psnr_hvsm = psnr_hvs_calc_psnr(sum_hvsm_rgb, 1.0f);

		metrics.m_rgba.m_mseh_hvs = sum_hvs_rgba;
		metrics.m_rgba.m_mseh_hvsm = sum_hvsm_rgba;
		metrics.m_rgba.m_psnr_hvs = psnr_hvs_calc_psnr(sum_hvs_rgba, 1.0f);
		metrics.m_rgba.m_psnr_hvsm = psnr_hvs_calc_psnr(sum_hvsm_rgba, 1.0f);

		metrics.m_valid = true;
				
		return true;
	}

	void psnr_hvs_print_metrics(const psnr_hvs_metrics& metrics)
	{
		if (!metrics.m_valid)
		{
			fmt_printf("  PSNR-HVS metrics are not valid.\n");
			return;
		}

		fmt_printf("  Float Y 601 PSNR-HVS: {1.3} dB, PSNR-HVS-M: {1.3} dB\n", metrics.m_y_601_float.m_psnr_hvs, metrics.m_y_601_float.m_psnr_hvsm);
		fmt_printf("  8-Bit Y 601 PSNR-HVS: {1.3} dB, PSNR-HVS-M: {1.3} dB\n", metrics.m_y_601_8bit.m_psnr_hvs, metrics.m_y_601_8bit.m_psnr_hvsm);
		
		fmt_printf("    RGB  Avg. PSNR-HVS: {1.3} dB, PSNR-HVS-M: {1.3} dB\n", metrics.m_rgb.m_psnr_hvs, metrics.m_rgb.m_psnr_hvsm);
		fmt_printf("    RGBA Avg. PSNR-HVS: {1.3} dB, PSNR-HVS-M: {1.3} dB\n", metrics.m_rgba.m_psnr_hvs, metrics.m_rgba.m_psnr_hvsm);

		for (uint32_t c = 0; c < 4; c++)
		{
			fmt_printf("            {c} PSNR-HVS: {1.3} dB, PSNR-HVS-M: {1.3} dB\n", "RGBA"[c], metrics.m_chan[c].m_psnr_hvs, metrics.m_chan[c].m_psnr_hvsm);
		}
	}

	void print_psnr_hvs_image_metrics(const image& a, const image& b)
	{
		psnr_hvs_metrics metrics;
		if (!psnr_hvs_compute_metrics(a, b, metrics))
		{
			fmt_error_printf("print_psnr_hvs_image_metrics: psnr_hvs_compute_metrics() failed!\n");
			return;
		}

		psnr_hvs_print_metrics(metrics);
	}

	void fill_buffer_with_random_bytes(void *pBuf, size_t size, uint32_t seed)
	{
		rand r(seed);

		uint8_t *pDst = static_cast<uint8_t *>(pBuf);

		while (size >= sizeof(uint32_t))
		{
			*(uint32_t *)pDst = r.urand32();
			pDst += sizeof(uint32_t);
			size -= sizeof(uint32_t);
		}

		while (size)
		{
			*pDst++ = r.byte();
			size--;
		}
	}

	std::vector<float> bounded_samples(
		size_t num_samples,
		float half_span,
		float sigma,
		uint32_t seed,
		float min_tail_prob,
		float tail_amp_cap,
		bool fix_tail_count,
		bool recenter)
	{
		assert(num_samples > 0);
		assert(half_span > 0.0f);

		if (sigma < 0.0f)
			sigma = 0.0f;
		if (sigma > half_span)
			sigma = half_span;
		tail_amp_cap = clamp(tail_amp_cap, 0.0f, 1.0f);

		std::mt19937 rng(seed ? seed : std::random_device{}());
		std::uniform_real_distribution<float> uniform_dist(-half_span, half_span);
		std::bernoulli_distribution coin(0.5);
		auto unit_rand = [&]() { return std::generate_canonical<float, 24>(rng); };

		const float H = half_span;
		const float H2 = H * H;
		const float V = sigma * sigma;           // target variance
		const float varU = H2 / 3.0f;            // variance of Uniform[-H, H]
		const float Amax = tail_amp_cap;

		float p = 0.0f;     // tail probability
		float A = 0.0f;     // tail amplitude fraction actually used (<= Amax)
		float k2 = 0.0f;    // bulk scale^2 for the k*Uniform component

		if (V <= varU + 1e-12f)
		{
			// Small target sigma: a scaled uniform already covers the variance, unless occasional tails are requested.
			if (min_tail_prob <= 0.0f)
			{
				p = 0.0f;
				k2 = (varU > 0.0f) ? (V / varU) : 0.0f; // support +/-kH
			}
			else
			{
				// Enforce a small tail probability, shrinking the amplitude so p*A^2*H^2 <= V.
				p = clamp(min_tail_prob, 0.0f, 1.0f);

				// Max feasible amplitude for this p under the variance budget: A_needed <= sqrt(V/(p*H^2)).
				float A_needed = (p > 0.0f) ? std::sqrt(maximum(0.0f, V) / (p * H2)) : 0.0f;
				A = clamp(A_needed, 0.0f, Amax);
				const float varT = (A * A) * H2;

				// Remaining variance goes to the scaled uniform.
				const float denom = (1.0f - p) * varU;
				k2 = (denom > 0.0f) ? ((V - p * varT) / denom) : 0.0f;
				k2 = clamp(k2, 0.0f, 1.0f);
			}
		}
		else
		{
			// Large target sigma: need enough tail mass to exceed the uniform's variance.
			// Use the maximum tail amplitude, then solve for the required tail probability.
			A = Amax;
			const float varT = (A * A) * H2;
			const float denom = (varT - varU);
			float p_needed = (denom > 0.0f) ? ((V - varU) / denom) : 1.0f;
			p_needed = clamp(p_needed, 0.0f, 1.0f);

			// Take at least p_needed, but also honor any requested baseline tail probability.
			p = maximum(p_needed, clamp(min_tail_prob, 0.0f, 1.0f));
			if (p >= 1.0f - 1e-12f)
			{
				p = 1.0f;
				k2 = 0.0f;
			}
			else
			{
				const float denom2 = (1.0f - p) * varU;
				k2 = (denom2 > 0.0f) ? ((V - p * varT) / denom2) : 0.0f;
				k2 = clamp(k2, 0.0f, 1.0f);
			}
		}

		const float k = std::sqrt(k2);
		const float edge = A * H;

		std::vector<float> out;
		out.reserve(num_samples);

		if (p == 0.0f)
		{
			// Scaled uniform only (support +/-kH).
			for (size_t i = 0; i < num_samples; ++i)
				out.push_back(k * uniform_dist(rng));
		}
		else if (fix_tail_count)
		{
			size_t num_tail = static_cast<size_t>(std::round(p * num_samples));
			if (num_tail > num_samples)
				num_tail = num_samples;
			if (num_tail & 1)
			{
				// Keep the tail count even so the +edge / -edge halves stay symmetric.
				if (num_tail < num_samples)
					++num_tail;
				else
					--num_tail;
			}

			for (size_t i = 0; i < num_tail / 2; ++i)
				out.push_back(+edge);
			for (size_t i = 0; i < num_tail / 2; ++i)
				out.push_back(-edge);
			for (size_t i = out.size(); i < num_samples; ++i)
				out.push_back(k * uniform_dist(rng));
		}
		else
		{
			for (size_t i = 0; i < num_samples; ++i)
			{
				if (unit_rand() < p)
					out.push_back(coin(rng) ? +edge : -edge);
				else
					out.push_back(k * uniform_dist(rng));
			}
		}

		if (recenter)
		{
			float sum = 0.0f;
			for (float v : out)
				sum += v;

			const float mu = sum / static_cast<float>(num_samples);
			if (std::abs(mu) > 0.0f)
			{
				for (float& v : out)
				{
					v -= mu;
					if (v > H)
						v = H;
					else if (v < -H)
						v = -H;
				}
			}
		}

		return out;
	}

	job_pool::job_pool(uint32_t num_threads) : 
		m_num_active_jobs(0)
	{
		m_kill_flag.store(false);
		m_num_active_workers.store(0);

		assert(num_threads >= 1U);

		debug_printf("job_pool::job_pool: %u total threads\n", num_threads);

		if (num_threads > 1)
		{
			m_threads.resize(num_threads - 1);

			for (int i = 0; i < ((int)num_threads - 1); i++)
			   m_threads[i] = std::thread([this, i] { job_thread(i); });
		}
	}

	job_pool::~job_pool()
	{
		debug_printf("job_pool::~job_pool\n");
		
		// Notify all workers that they need to die right now.
		{
			std::lock_guard<std::mutex> lk(m_mutex);
			m_kill_flag.store(true);
		}
		
		m_has_work.notify_all();

#ifdef __EMSCRIPTEN__
		// Without this wait the join()'s aren't reliable, and I have no idea why. WASM threading in the browser is sometimes mysterious.
		const uint32_t max_iterations = 90;

		uint32_t iteration_index = 0;
		for ( ; ; )
		{
			if (m_num_active_workers.load() <= 0)
				break;
			std::this_thread::sleep_for(std::chrono::milliseconds(50));

			if (++iteration_index > max_iterations)
			{
				debug_printf("job_pool::~job_pool: wait timed out!\n");
				break;
			}
		}
		
		// At this point all worker threads should be exiting or exited.
		// We could call detach(), but this seems to just call join() anyway.
#endif

		// Wait for all worker threads to exit.
		for (uint32_t i = 0; i < m_threads.size(); i++)
			m_threads[i].join();

		debug_printf("job_pool::~job_pool: joined OK\n");
	}
				
	void job_pool::add_job(const std::function<void()>& job)
	{
		std::unique_lock<std::mutex> lock(m_mutex);

		m_queue.emplace_back(job);

		const size_t queue_size = m_queue.size();

		lock.unlock();

		if (queue_size > 1)
			m_has_work.notify_one();
	}

	void job_pool::add_job(std::function<void()>&& job)
	{
		std::unique_lock<std::mutex> lock(m_mutex);

		m_queue.emplace_back(std::move(job));
						
		const size_t queue_size = m_queue.size();

		lock.unlock();

		if (queue_size > 1)
		{
			m_has_work.notify_one();
		}
	}

	void job_pool::wait_for_all()
	{
		std::unique_lock<std::mutex> lock(m_mutex);

		// Drain the job queue on the calling thread.
		while (!m_queue.empty())
		{
			std::function<void()> job(m_queue.back());
			m_queue.pop_back();

			lock.unlock();

			job();

			lock.lock();
		}

		// The queue is empty, now wait for all active jobs to finish up.
#ifndef __EMSCRIPTEN__
		m_no_more_jobs.wait(lock, [this]{ return !m_num_active_jobs; } );
#else
		// Avoid infinite blocking
		for (; ; )
		{
			if (m_no_more_jobs.wait_for(lock, std::chrono::milliseconds(50), [this] { return !m_num_active_jobs; }))
			{
				break;
			}
		}
#endif
	}

	void job_pool::job_thread(uint32_t index)
	{
		BASISU_NOTE_UNUSED(index);
		//debug_printf("job_pool::job_thread: starting %u\n", index);

		m_num_active_workers.fetch_add(1);
		
		while (!m_kill_flag)
		{
			std::unique_lock<std::mutex> lock(m_mutex);

			// Wait for any jobs to be issued.
#if 0
			m_has_work.wait(lock, [this] { return m_kill_flag || m_queue.size(); } );
#else
			// For more safety vs. buggy RTL's. Worse case we stall for a second vs. locking up forever if something goes wrong.
			m_has_work.wait_for(lock, std::chrono::milliseconds(1000), [this] {
				return m_kill_flag || !m_queue.empty();
				});
#endif

			// Check to see if we're supposed to exit.
			if (m_kill_flag)
				break;

			if (m_queue.empty())
				continue;

			// Get the job and execute it.
			std::function<void()> job(m_queue.back());
			m_queue.pop_back();

			++m_num_active_jobs;

			lock.unlock();

			job();

			lock.lock();

			--m_num_active_jobs;

			// Now check if there are no more jobs remaining. 
			const bool all_done = m_queue.empty() && !m_num_active_jobs;
			
			lock.unlock();

			if (all_done)
				m_no_more_jobs.notify_all();
		}

		m_num_active_workers.fetch_add(-1);

		//debug_printf("job_pool::job_thread: exiting\n");
	}

	// .TGA image loading
	#pragma pack(push)
	#pragma pack(1)
	struct tga_header
	{
		uint8_t			m_id_len;
		uint8_t			m_cmap;
		uint8_t			m_type;
		packed_uint<2>	m_cmap_first;
		packed_uint<2> m_cmap_len;
		uint8_t			m_cmap_bpp;
		packed_uint<2> m_x_org;
		packed_uint<2> m_y_org;
		packed_uint<2> m_width;
		packed_uint<2> m_height;
		uint8_t			m_depth;
		uint8_t			m_desc;
	};
	#pragma pack(pop)

	const uint32_t MAX_TGA_IMAGE_SIZE = 16384;

	enum tga_image_type
	{
		cITPalettized = 1,
		cITRGB = 2,
		cITGrayscale = 3
	};

	uint8_t *read_tga(const uint8_t *pBuf, uint32_t buf_size, int &width, int &height, int &n_chans)
	{
		width = 0;
		height = 0;
		n_chans = 0;

		if (buf_size <= sizeof(tga_header))
			return nullptr;

		const tga_header &hdr = *reinterpret_cast<const tga_header *>(pBuf);

		if ((!hdr.m_width) || (!hdr.m_height) || (hdr.m_width > MAX_TGA_IMAGE_SIZE) || (hdr.m_height > MAX_TGA_IMAGE_SIZE))
			return nullptr;

		if (hdr.m_desc >> 6)
			return nullptr;

		// Simple validation
		if ((hdr.m_cmap != 0) && (hdr.m_cmap != 1))
			return nullptr;
		
		if (hdr.m_cmap)
		{
			if ((hdr.m_cmap_bpp == 0) || (hdr.m_cmap_bpp > 32))
				return nullptr;

			// Nobody implements CMapFirst correctly, so we're not supporting it. Never seen it used, either.
			if (hdr.m_cmap_first != 0)
				return nullptr;
		}

		const bool x_flipped = (hdr.m_desc & 0x10) != 0;
		const bool y_flipped = (hdr.m_desc & 0x20) == 0;

		bool rle_flag = false;
		int file_image_type = hdr.m_type;
		if (file_image_type > 8)
		{
			file_image_type -= 8;
			rle_flag = true;
		}

		const tga_image_type image_type = static_cast<tga_image_type>(file_image_type);

		switch (file_image_type)
		{
		case cITRGB:
			if (hdr.m_depth == 8)
				return nullptr;
			break;
		case cITPalettized:
			if ((hdr.m_depth != 8) || (hdr.m_cmap != 1) || (hdr.m_cmap_len == 0))
				return nullptr;
			break;
		case cITGrayscale:
			if ((hdr.m_cmap != 0) || (hdr.m_cmap_len != 0))
				return nullptr;
			if ((hdr.m_depth != 8) && (hdr.m_depth != 16))
				return nullptr;
			break;
		default:
			return nullptr;
		}

		uint32_t tga_bytes_per_pixel = 0;

		switch (hdr.m_depth)
		{
		case 32:
			tga_bytes_per_pixel = 4;
			n_chans = 4;
			break;
		case 24:
			tga_bytes_per_pixel = 3;
			n_chans = 3;
			break;
		case 16:
		case 15:
			tga_bytes_per_pixel = 2;
			// For compatibility with stb_image_write.h
			n_chans = ((file_image_type == cITGrayscale) && (hdr.m_depth == 16)) ? 4 : 3;
			break;
		case 8:
			tga_bytes_per_pixel = 1;
			// For palettized RGBA support, which both FreeImage and stb_image support.
			n_chans = ((file_image_type == cITPalettized) && (hdr.m_cmap_bpp == 32)) ? 4 : 3;
			break;
		default:
			return nullptr;
		}

		//const uint32_t bytes_per_line = hdr.m_width * tga_bytes_per_pixel;

		const uint8_t *pSrc = pBuf + sizeof(tga_header);
		uint32_t bytes_remaining = buf_size - sizeof(tga_header);

		if (hdr.m_id_len)
		{
			if (bytes_remaining < hdr.m_id_len)
				return nullptr;
			pSrc += hdr.m_id_len;
			bytes_remaining += hdr.m_id_len;
		}

		color_rgba pal[256];
		for (uint32_t i = 0; i < 256; i++)
			pal[i].set(0, 0, 0, 255);

		if ((hdr.m_cmap) && (hdr.m_cmap_len))
		{
			if (image_type == cITPalettized)
			{
				// Note I cannot find any files using 32bpp palettes in the wild (never seen any in ~30 years).
				if ( ((hdr.m_cmap_bpp != 32) && (hdr.m_cmap_bpp != 24) && (hdr.m_cmap_bpp != 15) && (hdr.m_cmap_bpp != 16)) || (hdr.m_cmap_len > 256) )
					return nullptr;

				if (hdr.m_cmap_bpp == 32)
				{
					const uint32_t pal_size = hdr.m_cmap_len * 4;
					if (bytes_remaining < pal_size)
						return nullptr;

					for (uint32_t i = 0; i < hdr.m_cmap_len; i++)
					{
						pal[i].r = pSrc[i * 4 + 2];
						pal[i].g = pSrc[i * 4 + 1];
						pal[i].b = pSrc[i * 4 + 0];
						pal[i].a = pSrc[i * 4 + 3];
					}

					bytes_remaining -= pal_size;
					pSrc += pal_size;
				}
				else if (hdr.m_cmap_bpp == 24)
				{
					const uint32_t pal_size = hdr.m_cmap_len * 3;
					if (bytes_remaining < pal_size)
						return nullptr;

					for (uint32_t i = 0; i < hdr.m_cmap_len; i++)
					{
						pal[i].r = pSrc[i * 3 + 2];
						pal[i].g = pSrc[i * 3 + 1];
						pal[i].b = pSrc[i * 3 + 0];
						pal[i].a = 255;
					}

					bytes_remaining -= pal_size;
					pSrc += pal_size;
				}
				else
				{
					const uint32_t pal_size = hdr.m_cmap_len * 2;
					if (bytes_remaining < pal_size)
						return nullptr;

					for (uint32_t i = 0; i < hdr.m_cmap_len; i++)
					{
						const uint32_t v = pSrc[i * 2 + 0] | (pSrc[i * 2 + 1] << 8);

						pal[i].r = (((v >> 10) & 31) * 255 + 15) / 31;
						pal[i].g = (((v >> 5) & 31) * 255 + 15) / 31;
						pal[i].b = ((v & 31) * 255 + 15) / 31;
						pal[i].a = 255;
					}

					bytes_remaining -= pal_size;
					pSrc += pal_size;
				}
			}
			else
			{
				const uint32_t bytes_to_skip = (hdr.m_cmap_bpp >> 3) * hdr.m_cmap_len;
				if (bytes_remaining < bytes_to_skip)
					return nullptr;
				pSrc += bytes_to_skip;
				bytes_remaining += bytes_to_skip;
			}
		}
		
		width = hdr.m_width;
		height = hdr.m_height;

		const uint32_t source_pitch = width * tga_bytes_per_pixel;
		const uint32_t dest_pitch = width * n_chans;
		
		uint8_t *pImage = (uint8_t *)malloc(dest_pitch * height);
		if (!pImage)
			return nullptr;

		std::vector<uint8_t> input_line_buf;
		if (rle_flag)
			input_line_buf.resize(source_pitch);

		int run_type = 0, run_remaining = 0;
		uint8_t run_pixel[4];
		memset(run_pixel, 0, sizeof(run_pixel));

		for (int y = 0; y < height; y++)
		{
			const uint8_t *pLine_data;

			if (rle_flag)
			{
				int pixels_remaining = width;
				uint8_t *pDst = &input_line_buf[0];

				do 
				{
					if (!run_remaining)
					{
						if (bytes_remaining < 1)
						{
							free(pImage);
							return nullptr;
						}

						int v = *pSrc++;
						bytes_remaining--;

						run_type = v & 0x80;
						run_remaining = (v & 0x7F) + 1;

						if (run_type)
						{
							if (bytes_remaining < tga_bytes_per_pixel)
							{
								free(pImage);
								return nullptr;
							}

							memcpy(run_pixel, pSrc, tga_bytes_per_pixel);
							pSrc += tga_bytes_per_pixel;
							bytes_remaining -= tga_bytes_per_pixel;
						}
					}

					const uint32_t n = basisu::minimum<uint32_t>(pixels_remaining, run_remaining);
					pixels_remaining -= n;
					run_remaining -= n;

					if (run_type)
					{
						for (uint32_t i = 0; i < n; i++)
							for (uint32_t j = 0; j < tga_bytes_per_pixel; j++)
								*pDst++ = run_pixel[j];
					}
					else
					{
						const uint32_t bytes_wanted = n * tga_bytes_per_pixel;

						if (bytes_remaining < bytes_wanted)
						{
							free(pImage);
							return nullptr;
						}

						memcpy(pDst, pSrc, bytes_wanted);
						pDst += bytes_wanted;

						pSrc += bytes_wanted;
						bytes_remaining -= bytes_wanted;
					}

				} while (pixels_remaining);

				assert((pDst - &input_line_buf[0]) == (int)(width * tga_bytes_per_pixel));

				pLine_data = &input_line_buf[0];
			}
			else
			{
				if (bytes_remaining < source_pitch)
				{
					free(pImage);
					return nullptr;
				}

				pLine_data = pSrc;
				bytes_remaining -= source_pitch;
				pSrc += source_pitch;
			}

			// Convert to 24bpp RGB or 32bpp RGBA.
			uint8_t *pDst = pImage + (y_flipped ? (height - 1 - y) : y) * dest_pitch + (x_flipped ? (width - 1) * n_chans : 0);
			const int dst_stride = x_flipped ? -((int)n_chans) : n_chans;

			switch (hdr.m_depth)
			{
			case 32:
				assert(tga_bytes_per_pixel == 4 && n_chans == 4);
				for (int i = 0; i < width; i++, pLine_data += 4, pDst += dst_stride)
				{
					pDst[0] = pLine_data[2];
					pDst[1] = pLine_data[1];
					pDst[2] = pLine_data[0];
					pDst[3] = pLine_data[3];
				}
				break;
			case 24:
				assert(tga_bytes_per_pixel == 3 && n_chans == 3);
				for (int i = 0; i < width; i++, pLine_data += 3, pDst += dst_stride)
				{
					pDst[0] = pLine_data[2];
					pDst[1] = pLine_data[1];
					pDst[2] = pLine_data[0];
				}
				break;
			case 16:
			case 15:
				if (image_type == cITRGB)
				{
					assert(tga_bytes_per_pixel == 2 && n_chans == 3);
					for (int i = 0; i < width; i++, pLine_data += 2, pDst += dst_stride)
					{
						const uint32_t v = pLine_data[0] | (pLine_data[1] << 8);
						pDst[0] = (((v >> 10) & 31) * 255 + 15) / 31;
						pDst[1] = (((v >> 5) & 31) * 255 + 15) / 31;
						pDst[2] = ((v & 31) * 255 + 15) / 31;
					}
				}
				else
				{
					assert(image_type == cITGrayscale && tga_bytes_per_pixel == 2 && n_chans == 4);
					for (int i = 0; i < width; i++, pLine_data += 2, pDst += dst_stride)
					{
						pDst[0] = pLine_data[0];
						pDst[1] = pLine_data[0];
						pDst[2] = pLine_data[0];
						pDst[3] = pLine_data[1];
					}
				}
				break;
			case 8:
				assert(tga_bytes_per_pixel == 1);
				if (image_type == cITPalettized)
				{
					if (hdr.m_cmap_bpp == 32)
					{
						assert(n_chans == 4);
						for (int i = 0; i < width; i++, pLine_data++, pDst += dst_stride)
						{
							const uint32_t c = *pLine_data;
							pDst[0] = pal[c].r;
							pDst[1] = pal[c].g;
							pDst[2] = pal[c].b;
							pDst[3] = pal[c].a;
						}
					}
					else
					{
						assert(n_chans == 3);
						for (int i = 0; i < width; i++, pLine_data++, pDst += dst_stride)
						{
							const uint32_t c = *pLine_data;
							pDst[0] = pal[c].r;
							pDst[1] = pal[c].g;
							pDst[2] = pal[c].b;
						}
					}
				}
				else
				{
					assert(n_chans == 3);
					for (int i = 0; i < width; i++, pLine_data++, pDst += dst_stride)
					{
						const uint8_t c = *pLine_data;
						pDst[0] = c;
						pDst[1] = c;
						pDst[2] = c;
					}
				}
				break;
			default:
				assert(0);
				break;
			}
		} // y

		return pImage;
	}

	uint8_t *read_tga(const char *pFilename, int &width, int &height, int &n_chans)
	{
		width = height = n_chans = 0;

		uint8_vec filedata;
		if (!read_file_to_vec(pFilename, filedata))
			return nullptr;

		if (!filedata.size() || (filedata.size() > UINT32_MAX))
			return nullptr;
		
		return read_tga(&filedata[0], (uint32_t)filedata.size(), width, height, n_chans);
	}

	static inline void hdr_convert(const color_rgba& rgbe, vec4F& c)
	{
		if (rgbe[3] != 0)
		{
			float scale = ldexp(1.0f, rgbe[3] - 128 - 8);
			c.set((float)rgbe[0] * scale, (float)rgbe[1] * scale, (float)rgbe[2] * scale, 1.0f);
		}
		else
		{
			c.set(0.0f, 0.0f, 0.0f, 1.0f);
		}
	}

	bool string_begins_with(const std::string& str, const char* pPhrase)
	{
		const size_t str_len = str.size();

		const size_t phrase_len = strlen(pPhrase);
		assert(phrase_len);

		if (str_len >= phrase_len)
		{
#ifdef _MSC_VER
			if (_strnicmp(pPhrase, str.c_str(), phrase_len) == 0)
#else
			if (strncasecmp(pPhrase, str.c_str(), phrase_len) == 0)
#endif
				return true;
		}

		return false;
	}

	// Radiance RGBE (.HDR) image reading.
	// This code tries to preserve the original logic in Radiance's ray/src/common/color.c code:
	// https://www.radiance-online.org/cgi-bin/viewcvs.cgi/ray/src/common/color.c?revision=2.26&view=markup&sortby=log
	// Also see: https://flipcode.com/archives/HDR_Image_Reader.shtml.
	// https://github.com/LuminanceHDR/LuminanceHDR/blob/master/src/Libpfs/io/rgbereader.cpp.
	// https://radsite.lbl.gov/radiance/refer/filefmts.pdf
	// Buggy readers:
	// stb_image.h: appears to be a clone of rgbe.c, but with goto's (doesn't support old format files, doesn't support mixture of RLE/non-RLE scanlines)
	// http://www.graphics.cornell.edu/~bjw/rgbe.html - rgbe.c/h
	// http://www.graphics.cornell.edu/online/formats/rgbe/ - rgbe.c/.h - buggy
	bool read_rgbe(const uint8_vec &filedata, imagef& img, rgbe_header_info& hdr_info)
	{
		hdr_info.clear();

		const uint32_t MAX_SUPPORTED_DIM = 65536;

		if (filedata.size() < 4)
			return false;

		// stb_image.h checks for the string "#?RADIANCE" or "#?RGBE" in the header.
		// The original Radiance header code doesn't care about the specific string.
		// opencv's reader only checks for "#?", so that's what we're going to do.
		if ((filedata[0] != '#') || (filedata[1] != '?'))
			return false;

		//uint32_t width = 0, height = 0;
		bool is_rgbe = false;
		size_t cur_ofs = 0;

		// Parse the lines until we encounter a blank line.
		std::string cur_line;
		for (; ; )
		{
			if (cur_ofs >= filedata.size())
				return false;

			const uint32_t HEADER_TOO_BIG_SIZE = 4096;
			if (cur_ofs >= HEADER_TOO_BIG_SIZE)
			{
				// Header seems too large - something is likely wrong. Return failure.
				return false;
			}

			uint8_t c = filedata[cur_ofs++];

			if (c == '\n')
			{
				if (!cur_line.size())
					break;

				if ((cur_line[0] == '#') && (!string_begins_with(cur_line, "#?")) && (!hdr_info.m_program.size()))
				{
					cur_line.erase(0, 1);
					while (cur_line.size() && (cur_line[0] == ' '))
						cur_line.erase(0, 1);

					hdr_info.m_program = cur_line;
				}
				else if (string_begins_with(cur_line, "EXPOSURE=") && (cur_line.size() > 9))
				{
					hdr_info.m_exposure = atof(cur_line.c_str() + 9);
					hdr_info.m_has_exposure = true;
				}
				else if (string_begins_with(cur_line, "GAMMA=") && (cur_line.size() > 6))
				{
					hdr_info.m_exposure = atof(cur_line.c_str() + 6);
					hdr_info.m_has_gamma = true;
				}
				else if (cur_line == "FORMAT=32-bit_rle_rgbe")
				{
					is_rgbe = true;
				}

				cur_line.resize(0);
			}
			else
				cur_line.push_back((char)c);
		}

		if (!is_rgbe)
			return false;

		// Assume and require the final line to have the image's dimensions. We're not supporting flipping.
		for (; ; )
		{
			if (cur_ofs >= filedata.size())
				return false;
			uint8_t c = filedata[cur_ofs++];
			if (c == '\n')
				break;
			cur_line.push_back((char)c);
		}

		int comp[2] = { 1, 0 }; // y, x (major, minor)
		int dir[2] = { -1, 1 }; // -1, 1, (major, minor), for y -1=up
		uint32_t major_dim = 0, minor_dim = 0;

		// Parse the dimension string, normally it'll be "-Y # +X #" (major, minor), rarely it differs
		for (uint32_t d = 0; d < 2; d++) // 0=major, 1=minor
		{
			const bool is_neg_x = (strncmp(&cur_line[0], "-X ", 3) == 0);
			const bool is_pos_x = (strncmp(&cur_line[0], "+X ", 3) == 0);
			const bool is_x = is_neg_x || is_pos_x;

			const bool is_neg_y = (strncmp(&cur_line[0], "-Y ", 3) == 0);
			const bool is_pos_y = (strncmp(&cur_line[0], "+Y ", 3) == 0);
			const bool is_y = is_neg_y || is_pos_y;

			if (cur_line.size() < 3)
				return false;
			
			if (!is_x && !is_y)
				return false;

			comp[d] = is_x ? 0 : 1;
			dir[d] = (is_neg_x || is_neg_y) ? -1 : 1;
			
			uint32_t& dim = d ? minor_dim : major_dim;

			cur_line.erase(0, 3);

			while (cur_line.size())
			{
				char c = cur_line[0];
				if (c != ' ')
					break;
				cur_line.erase(0, 1);
			}

			bool has_digits = false;
			while (cur_line.size())
			{
				char c = cur_line[0];
				cur_line.erase(0, 1);

				if (c == ' ')
					break;

				if ((c < '0') || (c > '9'))
					return false;

				const uint32_t prev_dim = dim;
				dim = dim * 10 + (c - '0');
				if (dim < prev_dim)
					return false;

				has_digits = true;
			}
			if (!has_digits)
				return false;

			if ((dim < 1) || (dim > MAX_SUPPORTED_DIM))
				return false;
		}
				
		// temp image: width=minor, height=major
		img.resize(minor_dim, major_dim);

		std::vector<color_rgba> temp_scanline(minor_dim);

		// Read the scanlines.
		for (uint32_t y = 0; y < major_dim; y++)
		{
			vec4F* pDst = &img(0, y);

			if ((filedata.size() - cur_ofs) < 4)
				return false;

			// Determine if the line uses the new or old format. See the logic in color.c.
			bool old_decrunch = false;
			if ((minor_dim < 8) || (minor_dim > 0x7FFF))
			{
				// Line is too short or long; must be old format.
				old_decrunch = true;
			}
			else if (filedata[cur_ofs] != 2)
			{
				// R is not 2, must be old format
				old_decrunch = true;
			}
			else
			{
				// c[0]/red is 2.Check GB and E for validity.				
				color_rgba c;
				memcpy(&c, &filedata[cur_ofs], 4);

				if ((c[1] != 2) || (c[2] & 0x80))
				{
					// G isn't 2, or the high bit of B is set which is impossible (image's > 0x7FFF pixels can't get here). Use old format.
					old_decrunch = true;
				}
				else
				{
					// Check B and E. If this isn't the minor_dim in network order, something is wrong. The pixel would also be denormalized, and invalid.
					uint32_t w = (c[2] << 8) | c[3];
					if (w != minor_dim)
						return false;

					cur_ofs += 4;
				}
			}

			if (old_decrunch)
			{
				uint32_t rshift = 0, x = 0;

				while (x < minor_dim)
				{
					if ((filedata.size() - cur_ofs) < 4)
						return false;

					color_rgba c;
					memcpy(&c, &filedata[cur_ofs], 4);
					cur_ofs += 4;

					if ((c[0] == 1) && (c[1] == 1) && (c[2] == 1))
					{
						// We'll allow RLE matches to cross scanlines, but not on the very first pixel.
						if ((!x) && (!y))
							return false;

						const uint32_t run_len = c[3] << rshift;
						const vec4F run_color(pDst[-1]);

						if ((x + run_len) > minor_dim)
							return false;

						for (uint32_t i = 0; i < run_len; i++)
							*pDst++ = run_color;

						rshift += 8;
						x += run_len;
					}
					else
					{
						rshift = 0;

						hdr_convert(c, *pDst);
						pDst++;
						x++;
					}
				}
				continue;
			}

			// New format
			for (uint32_t s = 0; s < 4; s++)
			{
				uint32_t x_ofs = 0;
				while (x_ofs < minor_dim)
				{
					uint32_t num_remaining = minor_dim - x_ofs;

					if (cur_ofs >= filedata.size())
						return false;

					uint8_t count = filedata[cur_ofs++];
					if (count > 128)
					{
						count -= 128;
						if (count > num_remaining)
							return false;

						if (cur_ofs >= filedata.size())
							return false;
						const uint8_t val = filedata[cur_ofs++];

						for (uint32_t i = 0; i < count; i++)
							temp_scanline[x_ofs + i][s] = val;

						x_ofs += count;
					}
					else
					{
						if ((!count) || (count > num_remaining))
							return false;

						for (uint32_t i = 0; i < count; i++)
						{
							if (cur_ofs >= filedata.size())
								return false;
							const uint8_t val = filedata[cur_ofs++];

							temp_scanline[x_ofs + i][s] = val;
						}

						x_ofs += count;
					}
				} // while (x_ofs < minor_dim)
			} // c

			// Convert all the RGBE pixels to float now
			for (uint32_t x = 0; x < minor_dim; x++, pDst++)
				hdr_convert(temp_scanline[x], *pDst);

			assert((pDst - &img(0, y)) == (int)minor_dim);

		} // y

		// at here:
		// img(width,height)=image pixels as read from file, x=minor axis, y=major axis
		// width=minor axis dimension
		// height=major axis dimension
		// in file, pixels are emitted in minor order, them major (so major=scanlines in the file)
		
		imagef final_img;
		if (comp[0] == 0) // if major axis is X
			final_img.resize(major_dim, minor_dim);
		else // major axis is Y, minor is X
			final_img.resize(minor_dim, major_dim);

		// TODO: optimize the identity case
		for (uint32_t major_iter = 0; major_iter < major_dim; major_iter++)
		{
			for (uint32_t minor_iter = 0; minor_iter < minor_dim; minor_iter++)
			{
				const vec4F& p = img(minor_iter, major_iter);

				uint32_t dst_x = 0, dst_y = 0;

				// is the minor dim output x?
				if (comp[1] == 0) 
				{
					// minor axis is x, major is y
					
					// is minor axis (which is output x) flipped?
					if (dir[1] < 0)
						dst_x = minor_dim - 1 - minor_iter;
					else
						dst_x = minor_iter;

					// is major axis (which is output y) flipped? -1=down in raster order, 1=up
					if (dir[0] < 0)
						dst_y = major_iter;
					else
						dst_y = major_dim - 1 - major_iter;
				}
				else
				{
					// minor axis is output y, major is output x

					// is minor axis (which is output y) flipped?
					if (dir[1] < 0)
						dst_y = minor_iter;
					else
						dst_y = minor_dim - 1 - minor_iter;

					// is major axis (which is output x) flipped?
					if (dir[0] < 0)
						dst_x = major_dim - 1 - major_iter;
					else
						dst_x = major_iter;
				}

				final_img(dst_x, dst_y) = p;
			}
		}

		final_img.swap(img);

		return true;
	}

	bool read_rgbe(const char* pFilename, imagef& img, rgbe_header_info& hdr_info)
	{
		uint8_vec filedata;
		if (!read_file_to_vec(pFilename, filedata))
			return false;
		return read_rgbe(filedata, img, hdr_info);
	}

	static uint8_vec& append_string(uint8_vec& buf, const char* pStr)
	{
		const size_t str_len = strlen(pStr);
		if (!str_len)
			return buf;

		const size_t ofs = buf.size();
		buf.resize(ofs + str_len);
		memcpy(&buf[ofs], pStr, str_len);

		return buf;
	}
	
	static uint8_vec& append_string(uint8_vec& buf, const std::string& str)
	{
		if (!str.size())
			return buf;
		return append_string(buf, str.c_str());
	}

	static inline void float2rgbe(color_rgba &rgbe, const vec4F &c)
	{
		const float red = c[0], green = c[1], blue = c[2];
		assert(red >= 0.0f && green >= 0.0f && blue >= 0.0f);

		const float max_v = basisu::maximumf(basisu::maximumf(red, green), blue);

		if (max_v < 1e-32f)
			rgbe.clear();
		else 
		{
			int e;
			const float scale = frexp(max_v, &e) * 256.0f / max_v;
			rgbe[0] = (uint8_t)(clamp<int>((int)(red * scale), 0, 255));
			rgbe[1] = (uint8_t)(clamp<int>((int)(green * scale), 0, 255));
			rgbe[2] = (uint8_t)(clamp<int>((int)(blue * scale), 0, 255));
			rgbe[3] = (uint8_t)(e + 128);
		}
	}

	const bool RGBE_FORCE_RAW = false;
	const bool RGBE_FORCE_OLD_CRUNCH = false; // note must readers (particularly stb_image.h's) don't properly support this, when they should
		
	bool write_rgbe(uint8_vec &file_data, imagef& img, rgbe_header_info& hdr_info)
	{
		if (!img.get_width() || !img.get_height())
			return false;

		const uint32_t width = img.get_width(), height = img.get_height();
		
		file_data.resize(0);
		file_data.reserve(1024 + img.get_width() * img.get_height() * 4);

		append_string(file_data, "#?RADIANCE\n");

		if (hdr_info.m_has_exposure)
			append_string(file_data, string_format("EXPOSURE=%g\n", hdr_info.m_exposure));

		if (hdr_info.m_has_gamma)
			append_string(file_data, string_format("GAMMA=%g\n", hdr_info.m_gamma));

		append_string(file_data, "FORMAT=32-bit_rle_rgbe\n\n");
		append_string(file_data, string_format("-Y %u +X %u\n", height, width));

		if (((width < 8) || (width > 0x7FFF)) || (RGBE_FORCE_RAW))
		{
			for (uint32_t y = 0; y < height; y++)
			{
				for (uint32_t x = 0; x < width; x++)
				{
					color_rgba rgbe;
					float2rgbe(rgbe, img(x, y));
					append_vector(file_data, (const uint8_t *)&rgbe, sizeof(rgbe));
				}
			}
		}
		else if (RGBE_FORCE_OLD_CRUNCH)
		{
			for (uint32_t y = 0; y < height; y++)
			{
				int prev_r = -1, prev_g = -1, prev_b = -1, prev_e = -1;
				uint32_t cur_run_len = 0;
				
				for (uint32_t x = 0; x < width; x++)
				{
					color_rgba rgbe;
					float2rgbe(rgbe, img(x, y));

					if ((rgbe[0] == prev_r) && (rgbe[1] == prev_g) && (rgbe[2] == prev_b) && (rgbe[3] == prev_e))
					{
						if (++cur_run_len == 255)
						{
							// this ensures rshift stays 0, it's lame but this path is only for testing readers
							color_rgba f(1, 1, 1, cur_run_len - 1);
							append_vector(file_data, (const uint8_t*)&f, sizeof(f));
							append_vector(file_data, (const uint8_t*)&rgbe, sizeof(rgbe)); 
							cur_run_len = 0;
						}
					}
					else
					{
						if (cur_run_len > 0)
						{
							color_rgba f(1, 1, 1, cur_run_len);
							append_vector(file_data, (const uint8_t*)&f, sizeof(f));
							
							cur_run_len = 0;
						}
						
						append_vector(file_data, (const uint8_t*)&rgbe, sizeof(rgbe));
																		
						prev_r = rgbe[0];
						prev_g = rgbe[1];
						prev_b = rgbe[2];
						prev_e = rgbe[3];
					}
				} // x

				if (cur_run_len > 0)
				{
					color_rgba f(1, 1, 1, cur_run_len);
					append_vector(file_data, (const uint8_t*)&f, sizeof(f));
				}
			} // y
		}
		else
		{
			uint8_vec temp[4];
			for (uint32_t c = 0; c < 4; c++)
				temp[c].resize(width);

			for (uint32_t y = 0; y < height; y++)
			{
				color_rgba rgbe(2, 2, width >> 8, width & 0xFF);
				append_vector(file_data, (const uint8_t*)&rgbe, sizeof(rgbe));
								
				for (uint32_t x = 0; x < width; x++)
				{
					float2rgbe(rgbe, img(x, y));

					for (uint32_t c = 0; c < 4; c++)
						temp[c][x] = rgbe[c];
				}

				for (uint32_t c = 0; c < 4; c++)
				{
					int raw_ofs = -1;
					
					uint32_t x = 0;
					while (x < width)
					{
						const uint32_t num_bytes_remaining = width - x;
						const uint32_t max_run_len = basisu::minimum<uint32_t>(num_bytes_remaining, 127);
						const uint8_t cur_byte = temp[c][x];

						uint32_t run_len = 1;
						while (run_len < max_run_len)
						{
							if (temp[c][x + run_len] != cur_byte)
								break;
							run_len++;
						}
												
						const uint32_t cost_to_keep_raw = ((raw_ofs != -1) ? 0 : 1) + run_len; // 0 or 1 bytes to start a raw run, then the repeated bytes issued as raw
						const uint32_t cost_to_take_run = 2 + 1; // 2 bytes to issue the RLE, then 1 bytes to start whatever follows it (raw or RLE)

						if ((run_len >= 3) && (cost_to_take_run < cost_to_keep_raw))
						{
							file_data.push_back((uint8_t)(128 + run_len));
							file_data.push_back(cur_byte);

							x += run_len;
							raw_ofs = -1;
						}
						else
						{
							if (raw_ofs < 0)
							{
								raw_ofs = (int)file_data.size();
								file_data.push_back(0);
							}

							if (++file_data[raw_ofs] == 128)
								raw_ofs = -1;

							file_data.push_back(cur_byte);
							
							x++;
						}
					} // x

				} // c
			} // y
		}

		return true;
	}

	bool write_rgbe(const char* pFilename, imagef& img, rgbe_header_info& hdr_info)
	{
		uint8_vec file_data;
		if (!write_rgbe(file_data, img, hdr_info))
			return false;
		return write_vec_to_file(pFilename, file_data);
	}

	bool read_exr(const char* pFilename, imagef& img, int& n_chans)
	{
		n_chans = 0;

		int width = 0, height = 0;
		float* out_rgba = nullptr;
		const char* err = nullptr;

		int status = LoadEXRWithLayer(&out_rgba, &width, &height, pFilename, nullptr, &err);
		if (status != 0)
		{
			error_printf("Failed loading .EXR image \"%s\"! (TinyEXR error: %s)\n", pFilename, err ? err : "?");
			FreeEXRErrorMessage(err);
			free(out_rgba);
			return false;
		}

		const uint32_t MAX_SUPPORTED_DIM = 32768;
		if ((width < 1) || (height < 1) || (width > (int)MAX_SUPPORTED_DIM) || (height > (int)MAX_SUPPORTED_DIM))
		{
			error_printf("Invalid dimensions of .EXR image \"%s\"!\n", pFilename);
			free(out_rgba);
			return false;
		}

		img.resize(width, height);

		memcpy((void*)img.get_ptr(), out_rgba, static_cast<size_t>(sizeof(float) * 4 * img.get_total_pixels()));
		
		free(out_rgba);
		out_rgba = nullptr;

		uint32_t total_all_same_rgba = 0, total_all_same_rgb = 0, total_has_alpha = 0;

		for (int y = 0; y < height; y++)
		{
			for (int x = 0; x < width; x++)
			{
				const vec4F& p = img(x, y);

				if ((p[0] == p[1]) && (p[0] == p[2]))
					total_all_same_rgb++;

				const float a = p[3];

				if ((a == p[0]) && (a == p[1]) && (a == p[2]))
					total_all_same_rgba++;
								
				if (a != 1.0f)
					total_has_alpha++;

			} // x
		} // y

		const uint32_t total_pixels = width * height;
		if (total_all_same_rgba == total_pixels)
		{
			// TinyEXR loads single channel EXR images into all output channels (including alpha) - assume they are luminance and fix our alpha.
			// Odds are this is an opaque luminance-only image, not a true alpha channel image. (As of early 2026 we don't support any HDR format with alpha, anyway.)
			for (int y = 0; y < height; y++)
				for (int x = 0; x < width; x++)
					img(x, y)[3] = 1.0f;
			
			n_chans = 1;
		}
		else if (total_has_alpha)
		{
			n_chans = 4;
		}
		else if (total_all_same_rgb == total_pixels)
		{
			n_chans = 1;
		}
		else
		{
			n_chans = 3;
		}

		//fmt_printf("Number of detected EXR channels: {}\n", n_chans);
				
		return true;
	}

	bool read_exr(const void* pMem, size_t mem_size, imagef& img)
	{
		float* out_rgba = nullptr;
		int width = 0, height = 0;
		const char* pErr = nullptr;
		int res = LoadEXRFromMemory(&out_rgba, &width, &height, (const uint8_t*)pMem, mem_size, &pErr);
		if (res < 0)
		{
			error_printf("Failed loading .EXR image from memory! (TinyEXR error: %s)\n", pErr ? pErr : "?");
			FreeEXRErrorMessage(pErr);
			free(out_rgba);
			return false;
		}

		img.resize(width, height);
		memcpy((void *)img.get_ptr(), out_rgba, width * height * sizeof(float) * 4);
		free(out_rgba);

		// TODO: detect luminance-only etc.

		return true;
	}

	bool write_exr(const char* pFilename, const imagef& img, uint32_t n_chans, uint32_t flags)
	{
		assert((n_chans == 1) || (n_chans == 3) || (n_chans == 4));

		const bool linear_hint = (flags & WRITE_EXR_LINEAR_HINT) != 0, 
			store_float = (flags & WRITE_EXR_STORE_FLOATS) != 0,
			no_compression = (flags & WRITE_EXR_NO_COMPRESSION) != 0;
								
		const uint32_t width = img.get_width(), height = img.get_height();
		assert(width && height);
		
		if (!width || !height)
			return false;
		
		float_vec layers[4];
		float* image_ptrs[4];
		for (uint32_t c = 0; c < n_chans; c++)
		{
			layers[c].resize(width * height);
			image_ptrs[c] = layers[c].get_ptr();
		}

		// ABGR
		int chan_order[4] = { 3, 2, 1, 0 };

		if (n_chans == 1)
		{
			// Y
			chan_order[0] = 0;
		}
		else if (n_chans == 3)
		{
			// BGR
			chan_order[0] = 2;
			chan_order[1] = 1;
			chan_order[2] = 0;
		}
		else if (n_chans != 4)
		{
			assert(0);
			return false;
		}
		
		for (uint32_t y = 0; y < height; y++)
		{
			for (uint32_t x = 0; x < width; x++)
			{
				const vec4F& p = img(x, y);

				for (uint32_t c = 0; c < n_chans; c++)
					layers[c][x + y * width] = p[chan_order[c]];
			} // x
		} // y

		EXRHeader header;
		InitEXRHeader(&header);

		EXRImage image;
		InitEXRImage(&image);

		image.num_channels = n_chans;
		image.images = (unsigned char**)image_ptrs;
		image.width = width;
		image.height = height;

		header.num_channels = n_chans;
		
		header.channels = (EXRChannelInfo*)calloc(header.num_channels, sizeof(EXRChannelInfo));

		// Must be (A)BGR order, since most of EXR viewers expect this channel order.
		for (uint32_t i = 0; i < n_chans; i++)
		{
			char c = 'Y';
			if (n_chans == 3)
				c = "BGR"[i];
			else if (n_chans == 4)
				c = "ABGR"[i];
						
			header.channels[i].name[0] = c;
			header.channels[i].name[1] = '\0';

			header.channels[i].p_linear = linear_hint;
		}
		
		header.pixel_types = (int*)calloc(header.num_channels, sizeof(int));
		header.requested_pixel_types = (int*)calloc(header.num_channels, sizeof(int));
		
		if (!no_compression)
			header.compression_type = TINYEXR_COMPRESSIONTYPE_ZIP;

		for (int i = 0; i < header.num_channels; i++) 
		{
			// pixel type of input image
			header.pixel_types[i] = TINYEXR_PIXELTYPE_FLOAT; 

			// pixel type of output image to be stored in .EXR
			header.requested_pixel_types[i] = store_float ? TINYEXR_PIXELTYPE_FLOAT : TINYEXR_PIXELTYPE_HALF; 
		}

		const char* pErr_msg = nullptr;

		int ret = SaveEXRImageToFile(&image, &header, pFilename, &pErr_msg);
		if (ret != TINYEXR_SUCCESS) 
		{
			error_printf("Save EXR err: %s\n", pErr_msg);
			FreeEXRErrorMessage(pErr_msg);
		}
				
		free(header.channels);
		free(header.pixel_types);
		free(header.requested_pixel_types);

		return (ret == TINYEXR_SUCCESS);
	}

	void image::debug_text(uint32_t x_ofs, uint32_t y_ofs, uint32_t scale_x, uint32_t scale_y, const color_rgba& fg, const color_rgba* pBG, bool alpha_only, const char* pFmt, ...)
	{
		char buf[2048];

		va_list args;
		va_start(args, pFmt);
#ifdef _WIN32		
		vsprintf_s(buf, sizeof(buf), pFmt, args);
#else
		vsnprintf(buf, sizeof(buf), pFmt, args);
#endif
		va_end(args);

		const char* p = buf;

		const uint32_t orig_x_ofs = x_ofs;

		while (*p)
		{
			uint8_t c = *p++;
			if ((c < 32) || (c > 127))
				c = '.';

			const uint8_t* pGlpyh = &g_debug_font8x8_basic[c - 32][0];

			for (uint32_t y = 0; y < 8; y++)
			{
				uint32_t row_bits = pGlpyh[y];
				for (uint32_t x = 0; x < 8; x++)
				{
					const uint32_t q = row_bits & (1 << x);
										
					const color_rgba* pColor = q ? &fg : pBG;
					if (!pColor)
						continue;

					if (alpha_only)
						fill_box_alpha(x_ofs + x * scale_x, y_ofs + y * scale_y, scale_x, scale_y, *pColor);
					else
						fill_box(x_ofs + x * scale_x, y_ofs + y * scale_y, scale_x, scale_y, *pColor);
				}
			}

			x_ofs += 8 * scale_x;
			if ((x_ofs + 8 * scale_x) > m_width)
			{
				x_ofs = orig_x_ofs;
				y_ofs += 8 * scale_y;
			}
		}
	}
	
	// Very basic global Reinhard tone mapping, output converted to sRGB with no dithering, alpha is carried through unchanged. 
	// Only used for debugging/development.
	void tonemap_image_reinhard(image &ldr_img, const imagef &hdr_img, float exposure, bool add_noise, bool per_component, bool luma_scaling)
	{
		uint32_t width = hdr_img.get_width(), height = hdr_img.get_height();

		ldr_img.resize(width, height);

		rand r;
		r.seed(128);
				
		for (uint32_t y = 0; y < height; y++)
		{
			for (uint32_t x = 0; x < width; x++)
			{
				vec4F c(hdr_img(x, y));

				if (per_component)
				{
					for (uint32_t t = 0; t < 3; t++)
					{
						if (c[t] <= 0.0f)
						{
							c[t] = 0.0f;
						}
						else
						{
							c[t] *= exposure;
							c[t] = c[t] / (1.0f + c[t]);
						}
					}
				}
				else
				{
					c[0] *= exposure;
					c[1] *= exposure;
					c[2] *= exposure;

					const float L = 0.2126f * c[0] + 0.7152f * c[1] + 0.0722f * c[2];

					float Lmapped = 0.0f;
					if (L > 0.0f)
					{
						//Lmapped = L / (1.0f + L);
						//Lmapped /= L;
						
						Lmapped = 1.0f / (1.0f + L);
					}

					c[0] = c[0] * Lmapped;
					c[1] = c[1] * Lmapped;
					c[2] = c[2] * Lmapped;

					if (luma_scaling)
					{
						// Keeps the ratio of r/g/b intact
						float m = maximum(c[0], c[1], c[2]);
						if (m > 1.0f)
						{
							c /= m;
						}
					}
				}

				c.clamp(0.0f, 1.0f);

				c[3] = c[3] * 255.0f;

				color_rgba& o = ldr_img(x, y);

				if (add_noise)
				{
					c[0] = linear_to_srgb(c[0]) * 255.0f;
					c[1] = linear_to_srgb(c[1]) * 255.0f;
					c[2] = linear_to_srgb(c[2]) * 255.0f;

					const float NOISE_AMP = .5f;
					c[0] += r.frand(-NOISE_AMP, NOISE_AMP);
					c[1] += r.frand(-NOISE_AMP, NOISE_AMP);
					c[2] += r.frand(-NOISE_AMP, NOISE_AMP);

					c.clamp(0.0f, 255.0f);

					o[0] = (uint8_t)fast_roundf_int(c[0]);
					o[1] = (uint8_t)fast_roundf_int(c[1]);
					o[2] = (uint8_t)fast_roundf_int(c[2]);
					o[3] = (uint8_t)fast_roundf_int(c[3]);
				}
				else
				{
					o[0] = g_fast_linear_to_srgb.convert(c[0]);
					o[1] = g_fast_linear_to_srgb.convert(c[1]);
					o[2] = g_fast_linear_to_srgb.convert(c[2]);
					o[3] = (uint8_t)fast_roundf_int(c[3]);
				}
			}
		}
	}

	bool tonemap_image_compressive(image& dst_img, const imagef& hdr_test_img)
	{
		const uint32_t width = hdr_test_img.get_width();
		const uint32_t height = hdr_test_img.get_height();

		uint16_vec orig_half_img(width * 3 * height);
		uint16_vec half_img(width * 3 * height);

		int max_shift = 32;

		for (uint32_t y = 0; y < height; y++)
		{
			for (uint32_t x = 0; x < width; x++)
			{
				const vec4F& p = hdr_test_img(x, y);

				for (uint32_t i = 0; i < 3; i++)
				{
					if (p[i] < 0.0f)
						return false;
					if (p[i] > basist::MAX_HALF_FLOAT)
						return false;

					uint32_t h = basist::float_to_half(p[i]);
					//uint32_t orig_h = h;

					orig_half_img[(x + y * width) * 3 + i] = (uint16_t)h;

					// Rotate sign bit into LSB
					//h = rot_left16((uint16_t)h, 1);
					//assert(rot_right16((uint16_t)h, 1) == orig_h);
					h <<= 1;

					half_img[(x + y * width) * 3 + i] = (uint16_t)h;

					// Determine # of leading zero bits, ignoring the sign bit
					if (h)
					{
						int lz = clz(h) - 16;
						assert(lz >= 0 && lz <= 16);

						assert((h << lz) <= 0xFFFF);

						max_shift = basisu::minimum<int>(max_shift, lz);
					}
				} // i
			} // x
		} // y

		//printf("tonemap_image_compressive: Max leading zeros: %i\n", max_shift);

		uint32_t high_hist[256];
		clear_obj(high_hist);

		for (uint32_t y = 0; y < height; y++)
		{
			for (uint32_t x = 0; x < width; x++)
			{
				for (uint32_t i = 0; i < 3; i++)
				{
					uint16_t& hf = half_img[(x + y * width) * 3 + i];

					assert(((uint32_t)hf << max_shift) <= 65535);

					hf <<= max_shift;

					uint32_t h = (uint8_t)(hf >> 8);
					high_hist[h]++;
				}
			} // x
		} // y

		uint32_t total_vals_used = 0;
		int remap_old_to_new[256];
		for (uint32_t i = 0; i < 256; i++)
			remap_old_to_new[i] = -1;

		for (uint32_t i = 0; i < 256; i++)
		{
			if (high_hist[i] != 0)
			{
				remap_old_to_new[i] = total_vals_used;
				total_vals_used++;
			}
		}

		assert(total_vals_used >= 1);

		//printf("tonemap_image_compressive: Total used high byte values: %u, unused: %u\n", total_vals_used, 256 - total_vals_used);

		bool val_used[256];
		clear_obj(val_used);

		int remap_new_to_old[256];
		for (uint32_t i = 0; i < 256; i++)
			remap_new_to_old[i] = -1;
		BASISU_NOTE_UNUSED(remap_new_to_old);

		int prev_c = -1;
		BASISU_NOTE_UNUSED(prev_c);
		for (uint32_t i = 0; i < 256; i++)
		{
			if (remap_old_to_new[i] >= 0)
			{
				int c;
				if (total_vals_used <= 1)
					c = remap_old_to_new[i];
				else
				{
					c = (remap_old_to_new[i] * 255 + ((total_vals_used - 1) / 2)) / (total_vals_used - 1);

					assert(c > prev_c);
				}

				assert(!val_used[c]);

				remap_new_to_old[c] = i;

				remap_old_to_new[i] = c;
				prev_c = c;

				//printf("%u ", c);

				val_used[c] = true;
			}
		} // i
		//printf("\n");

		dst_img.resize(width, height);

		for (uint32_t y = 0; y < height; y++)
		{
			for (uint32_t x = 0; x < width; x++)
			{
				for (uint32_t c = 0; c < 3; c++)
				{
					uint16_t& v16 = half_img[(x + y * width) * 3 + c];

					uint32_t hb = v16 >> 8;
					//uint32_t lb = v16 & 0xFF;

					assert(remap_old_to_new[hb] != -1);
					assert(remap_old_to_new[hb] <= 255);
					assert(remap_new_to_old[remap_old_to_new[hb]] == (int)hb);

					hb = remap_old_to_new[hb];

					//v16 = (uint16_t)((hb << 8) | lb);

					dst_img(x, y)[c] = (uint8_t)hb;
				}
			} // x
		} // y

		return true;
	}

	bool tonemap_image_compressive2(image& dst_img, const imagef& hdr_test_img)
	{
		const uint32_t width = hdr_test_img.get_width();
		const uint32_t height = hdr_test_img.get_height();

		dst_img.resize(width, height);
		dst_img.set_all(color_rgba(0, 0, 0, 255));

		basisu::vector<basist::half_float> half_img(width * 3 * height);
				
		uint32_t low_h = UINT32_MAX, high_h = 0;

		for (uint32_t y = 0; y < height; y++)
		{
			for (uint32_t x = 0; x < width; x++)
			{
				const vec4F& p = hdr_test_img(x, y);

				for (uint32_t i = 0; i < 3; i++)
				{
					float f = p[i];

					if (std::isnan(f) || std::isinf(f))
						f = 0.0f;
					else if (f < 0.0f)
						f = 0.0f;
					else if (f > basist::MAX_HALF_FLOAT)
						f = basist::MAX_HALF_FLOAT;

					uint32_t h = basist::float_to_half(f);

					low_h = minimum(low_h, h);
					high_h = maximum(high_h, h);
					
					half_img[(x + y * width) * 3 + i] = (basist::half_float)h;

				} // i
			} // x
		} // y

		if (low_h == high_h)
			return false;

		for (uint32_t y = 0; y < height; y++)
		{
			for (uint32_t x = 0; x < width; x++)
			{
				for (uint32_t i = 0; i < 3; i++)
				{
					basist::half_float h = half_img[(x + y * width) * 3 + i];
					
					float f = (float)(h - low_h) / (float)(high_h - low_h);

					int iv = basisu::clamp<int>((int)std::round(f * 255.0f), 0, 255);

					dst_img(x, y)[i] = (uint8_t)iv;

				} // i
			} // x
		} // y

		return true;
	}

	bool arith_test()
	{
		basist::arith_fastbits_f32::init();

		fmt_printf("random bit test\n");

		const uint32_t N = 1000;

		// random bit test
		for (uint32_t i = 0; i < N; i++)
		{
			basist::arith::arith_enc enc;
			enc.init(4096);

			{
				basisu::rand r;
				r.seed(i + 1);
				uint32_t num_vals = r.irand(1, 20000);

				for (uint32_t j = 0; j < num_vals; j++)
					enc.put_bit(r.bit());

				enc.flush();
			}

			{
				basisu::rand r;
				r.seed(i + 1);
				uint32_t num_vals = r.irand(1, 20000);

				basist::arith::arith_dec dec;
				dec.init(enc.get_data_buf().get_ptr(), enc.get_data_buf().size());

				for (uint32_t j = 0; j < num_vals; j++)
				{
					uint32_t t = r.bit();

					uint32_t a = dec.get_bit();
					if (t != a)
					{
						fmt_printf("error!");
						return false;
					}
				}
			}
		}

		fmt_printf("Random bit test OK\n");

		fmt_printf("random bits test\n");

		// random bits test
		for (uint32_t i = 0; i < N; i++)
		{
			basist::arith::arith_enc enc;
			enc.init(4096);

			{
				basisu::rand r;
				r.seed(i + 1);
				uint32_t num_vals = r.irand(1, 20000);
				uint32_t num_bits = r.irand(1, 20);

				for (uint32_t j = 0; j < num_vals; j++)
					enc.put_bits(r.urand32() & ((1 << num_bits) - 1), num_bits);

				enc.flush();
			}

			{
				basisu::rand r;
				r.seed(i + 1);
				uint32_t num_vals = r.irand(1, 20000);
				uint32_t num_bits = r.irand(1, 20);

				basist::arith::arith_dec dec;
				dec.init(enc.get_data_buf().get_ptr(), enc.get_data_buf().size());

				for (uint32_t j = 0; j < num_vals; j++)
				{
					uint32_t t = r.urand32() & ((1 << num_bits) - 1);

					uint32_t a = dec.get_bits(num_bits);
					if (t != a)
					{
						fmt_printf("error!");
						return false;
					}
				}
			}
		}

		fmt_printf("Random bits test OK\n");

		fmt_printf("random adaptive bit model test\n");

		// adaptive bit model random test
		for (uint32_t i = 0; i < N; i++)
		{
			basist::arith::arith_enc enc;
			enc.init(4096);

			{
				basisu::rand r;
				r.seed(i + 1);
				uint32_t num_vals = r.irand(1, 20000);

				basist::arith::arith_bit_model bm;
				bm.init();

				for (uint32_t j = 0; j < num_vals; j++)
					enc.encode(r.bit(), bm);

				enc.flush();
			}

			{
				basisu::rand r;
				r.seed(i + 1);
				uint32_t num_vals = r.irand(1, 20000);

				basist::arith::arith_dec dec;
				dec.init(enc.get_data_buf().get_ptr(), enc.get_data_buf().size());

				basist::arith::arith_bit_model bm;
				bm.init();

				for (uint32_t j = 0; j < num_vals; j++)
				{
					uint32_t t = r.bit();

					uint32_t a = dec.decode_bit(bm);
					if (t != a)
					{
						fmt_printf("error!");
						return false;
					}
				}
			}
		}
		fmt_printf("Random adaptive bits test OK\n");

		fmt_printf("random adaptive bit model 0 or 1 run test\n");

		// adaptive bit model 0 or 1 test
		for (uint32_t i = 0; i < N; i++)
		{
			basist::arith::arith_enc enc;
			enc.init(4096);

			{
				basisu::rand r;
				r.seed(i + 1);
				uint32_t num_vals = r.irand(1, 20000);

				basist::arith::arith_bit_model bm;
				bm.init();

				for (uint32_t j = 0; j < num_vals; j++)
					enc.encode(i & 1, bm);

				enc.flush();
			}

			{
				basisu::rand r;
				r.seed(i + 1);
				uint32_t num_vals = r.irand(1, 20000);

				basist::arith::arith_dec dec;
				dec.init(enc.get_data_buf().get_ptr(), enc.get_data_buf().size());

				basist::arith::arith_bit_model bm;
				bm.init();

				for (uint32_t j = 0; j < num_vals; j++)
				{
					uint32_t t = i & 1;

					uint32_t a = dec.decode_bit(bm);
					if (t != a)
					{
						fmt_printf("error!");
						return false;
					}
				}
			}
		}

		fmt_printf("Adaptive bit model 0 or 1 run test OK\n");

		fmt_printf("random adaptive bit model 0 or 1 run 2 test\n");

		// adaptive bit model 0 or 1 run test
		for (uint32_t i = 0; i < N; i++)
		{
			basist::arith::arith_enc enc;
			enc.init(4096);

			{
				basisu::rand r;
				r.seed(i + 1);
				uint32_t num_vals = r.irand(1, 2000);

				basist::arith::arith_bit_model bm;
				bm.init();

				for (uint32_t j = 0; j < num_vals; j++)
				{
					const uint32_t run_len = r.irand(1, 128);
					const uint32_t t = r.bit();
					for (uint32_t k = 0; k < run_len; k++)
						enc.encode(t, bm);
				}

				if (r.frand(0.0f, 1.0f) < .1f)
				{
					for (uint32_t q = 0; q < 1000; q++)
						enc.encode(0, bm);
				}

				enc.flush();
			}

			{
				basisu::rand r;
				r.seed(i + 1);
				uint32_t num_vals = r.irand(1, 2000);

				basist::arith::arith_dec dec;
				dec.init(enc.get_data_buf().get_ptr(), enc.get_data_buf().size());

				basist::arith::arith_bit_model bm;
				bm.init();

				for (uint32_t j = 0; j < num_vals; j++)
				{
					const uint32_t run_len = r.irand(1, 128);
					const uint32_t t = r.bit();

					for (uint32_t k = 0; k < run_len; k++)
					{
						uint32_t a = dec.decode_bit(bm);
						if (a != t)
						{
							fmt_printf("adaptive bit model random run test failed!\n");
							return false;
						}
					}
				}

				if (r.frand(0.0f, 1.0f) < .1f)
				{
					for (uint32_t q = 0; q < 1000; q++)
					{
						uint32_t d = dec.decode_bit(bm);
						if (d != 0)
						{
							fmt_printf("adaptive bit model random run test failed!\n");
							return false;
						}
					}
				}
			}
		}

		fmt_printf("Random data model test\n");

		// random data model test
		for (uint32_t i = 0; i < N; i++)
		{
			basist::arith::arith_enc enc;
			enc.init(4096);

			{
				basisu::rand r;
				r.seed(i + 1);
				const uint32_t num_vals = r.irand(1, 60000);

				uint32_t num_syms = r.irand(2, basist::arith::ArithMaxSyms);

				basist::arith::arith_data_model dm;
				dm.init(num_syms);

				for (uint32_t j = 0; j < num_vals; j++)
					enc.encode(r.irand(0, num_syms - 1), dm);

				enc.flush();
			}

			{
				basisu::rand r;
				r.seed(i + 1);
				uint32_t num_vals = r.irand(1, 60000);

				const uint32_t num_syms = r.irand(2, basist::arith::ArithMaxSyms);

				basist::arith::arith_dec dec;
				dec.init(enc.get_data_buf().get_ptr(), enc.get_data_buf().size());

				basist::arith::arith_data_model dm;
				dm.init(num_syms);

				for (uint32_t j = 0; j < num_vals; j++)
				{
					uint32_t expected = r.irand(0, num_syms - 1);
					uint32_t actual = dec.decode_sym(dm);
					if (actual != expected)
					{
						fmt_printf("adaptive data model random test failed!\n");
						return false;
					}
				}
			}
		}

		fmt_printf("Adaptive data model random test OK\n");

		fmt_printf("Overall OK\n");
		return true;
	}

	static void rasterize_line(image& dst, int xs, int ys, int xe, int ye, int pred, int inc_dec, int e, int e_inc, int e_no_inc, const color_rgba& color)
	{
		int start, end, var;

		if (pred)
		{
			start = ys; end = ye; var = xs;
			for (int i = start; i <= end; i++)
			{
				dst.set_clipped(var, i, color);
				if (e < 0)
					e += e_no_inc;
				else
				{
					var += inc_dec;
					e += e_inc;
				}
			}
		}
		else
		{
			start = xs; end = xe; var = ys;
			for (int i = start; i <= end; i++)
			{
				dst.set_clipped(i, var, color);
				if (e < 0)
					e += e_no_inc;
				else
				{
					var += inc_dec;
					e += e_inc;
				}
			}
		}
	}

	void draw_line(image& dst, int xs, int ys, int xe, int ye, const color_rgba& color)
	{
		if (xs > xe)
		{
			std::swap(xs, xe);
			std::swap(ys, ye);
		}

		int dx = xe - xs, dy = ye - ys;
		if (!dx)
		{
			if (ys > ye)
				std::swap(ys, ye);
			for (int i = ys; i <= ye; i++)
				dst.set_clipped(xs, i, color);
		}
		else if (!dy)
		{
			for (int i = xs; i < xe; i++)
				dst.set_clipped(i, ys, color);
		}
		else if (dy > 0)
		{
			if (dy <= dx)
			{
				int e = 2 * dy - dx, e_no_inc = 2 * dy, e_inc = 2 * (dy - dx);
				rasterize_line(dst, xs, ys, xe, ye, 0, 1, e, e_inc, e_no_inc, color);
			}
			else
			{
				int e = 2 * dx - dy, e_no_inc = 2 * dx, e_inc = 2 * (dx - dy);
				rasterize_line(dst, xs, ys, xe, ye, 1, 1, e, e_inc, e_no_inc, color);
			}
		}
		else
		{
			dy = -dy;
			if (dy <= dx)
			{
				int e = 2 * dy - dx, e_no_inc = 2 * dy, e_inc = 2 * (dy - dx);
				rasterize_line(dst, xs, ys, xe, ye, 0, -1, e, e_inc, e_no_inc, color);
			}
			else
			{
				int e = 2 * dx - dy, e_no_inc = (2 * dx), e_inc = 2 * (dx - dy);
				rasterize_line(dst, xe, ye, xs, ys, 1, -1, e, e_inc, e_no_inc, color);
			}
		}
	}

	// Used for generating random test data
	void draw_circle(image& dst, int cx, int cy, int r, const color_rgba& color)
	{
		assert(r >= 0);
		if (r < 0)
			return;

		int x = r;
		int y = 0;
		int err = 1 - x;

		while (x >= y) 
		{
			dst.set_clipped(cx + x, cy + y, color);
			dst.set_clipped(cx + y, cy + x, color);
			dst.set_clipped(cx - y, cy + x, color);
			dst.set_clipped(cx - x, cy + y, color);
			dst.set_clipped(cx - x, cy - y, color);
			dst.set_clipped(cx - y, cy - x, color);
			dst.set_clipped(cx + y, cy - x, color);
			dst.set_clipped(cx + x, cy - y, color);

			++y;

			if (err < 0) 
			{
				err += 2 * y + 1;
			}
			else 
			{
				--x;
				err += 2 * (y - x) + 1;
			}
		}
	}

	void set_image_alpha(image& img, uint32_t a)
	{
		for (uint32_t y = 0; y < img.get_height(); y++)
			for (uint32_t x = 0; x < img.get_width(); x++)
				img(x, y).a = (uint8_t)a;
	}

	// red=3 subsets, blue=2 subsets, green=mode 6, white=mode 7, purple = 2 plane
	const color_rgba g_bc7_mode_vis_colors[8] =
	{
		color_rgba(190, 0,   0, 255), // 0
		color_rgba(0, 0, 255, 255), // 1
		color_rgba(255, 0, 0, 255), // 2
		color_rgba(0, 0,  130, 255), // 3
		color_rgba(255, 0, 255, 255), // 4 
		color_rgba(190,  0,  190, 255), // 5
		color_rgba(50, 167, 30, 255), // 6
		color_rgba(255,   255,   255, 255)  // 7
	};

	void create_bc7_debug_images(
		uint32_t width, uint32_t height, 
		const void *pBlocks, 
		const char *pFilename_prefix)
	{
		assert(width && height && pBlocks );

		const uint32_t num_bc7_blocks_x = (width + 3) >> 2;
		const uint32_t num_bc7_blocks_y = (height + 3) >> 2;
		const uint32_t total_bc7_blocks = num_bc7_blocks_x * num_bc7_blocks_y;

		image bc7_mode_vis(width, height);

		uint32_t bc7_mode_hist[9] = {};

		uint32_t mode4_index_hist[2] = {};
		uint32_t mode4_rot_hist[4] = {};
		uint32_t mode5_rot_hist[4] = {};

		uint32_t num_2subsets = 0, num_3subsets = 0, num_dp = 0;

		uint32_t total_solid_bc7_blocks = 0;
		uint32_t num_unpack_failures = 0;

		for (uint32_t by = 0; by < num_bc7_blocks_y; by++)
		{
			const uint32_t base_y = by * 4;

			for (uint32_t bx = 0; bx < num_bc7_blocks_x; bx++)
			{
				const uint32_t base_x = bx * 4;
								
				const basist::bc7_block& blk = ((const basist::bc7_block *)pBlocks)[bx + by * num_bc7_blocks_x];

				color_rgba unpacked_pixels[16];
				bool status = basist::bc7u::unpack_bc7(&blk, (basist::color_rgba*)unpacked_pixels);
				if (!status)
					num_unpack_failures++;

				int mode_index = basist::bc7u::determine_bc7_mode(&blk);

				bool is_solid = false;
				
				// assumes our transcoder's analytical BC7 encoder wrote the solid block
				if (mode_index == 5)
				{
					const uint8_t* pBlock_bytes = (const uint8_t *)&blk;
										
					if (pBlock_bytes[0] == 0b00100000)
					{
						static const uint8_t s_tail_bytes[8] = { 0xac, 0xaa, 0xaa, 0xaa, 0, 0, 0, 0 };
						if ((pBlock_bytes[8] & ~3) == (s_tail_bytes[0] & ~3))
						{
							if (memcmp(pBlock_bytes + 9, s_tail_bytes + 1, 7) == 0)
							{
								is_solid = true;
							}
						}
					}
				}

				total_solid_bc7_blocks += is_solid;

				if ((mode_index == 0) || (mode_index == 2))
					num_3subsets++;
				else if ((mode_index == 1) || (mode_index == 3))
					num_2subsets++;

				bc7_mode_hist[mode_index + 1]++;

				if (mode_index == 4)
				{
					num_dp++;
					mode4_index_hist[range_check(basist::bc7u::determine_bc7_mode_4_index_mode(&blk), 0, 1)]++;
					mode4_rot_hist[range_check(basist::bc7u::determine_bc7_mode_4_or_5_rotation(&blk), 0, 3)]++;
				}
				else if (mode_index == 5)
				{
					num_dp++;
					mode5_rot_hist[range_check(basist::bc7u::determine_bc7_mode_4_or_5_rotation(&blk), 0, 3)]++;
				}

				color_rgba c((mode_index < 0) ? g_black_color : g_bc7_mode_vis_colors[mode_index]);

				if (is_solid)
					c.set(64, 0, 64, 255);

				bc7_mode_vis.fill_box(base_x, base_y, 4, 4, c);

			} // bx

		} // by

		fmt_debug_printf("--------- BC7 statistics:\n");
		fmt_debug_printf("\nTotal BC7 unpack failures: {}\n", num_unpack_failures);
		fmt_debug_printf("Total solid blocks: {} {3.2}%\n", total_solid_bc7_blocks, (float)total_solid_bc7_blocks * (float)100.0f / (float)total_bc7_blocks);

		fmt_debug_printf("\nTotal 2-subsets: {} {3.2}%\n", num_2subsets, (float)num_2subsets * 100.0f / (float)total_bc7_blocks);
		fmt_debug_printf("Total 3-subsets: {} {3.2}%\n", num_3subsets, (float)num_3subsets * 100.0f / (float)total_bc7_blocks);
		fmt_debug_printf("Total Dual Plane: {} {3.2}%\n", num_dp, (float)num_dp * 100.0f / (float)total_bc7_blocks);

		fmt_debug_printf("\nBC7 mode histogram:\n");
		for (int i = -1; i <= 7; i++)
		{
			fmt_debug_printf(" {}: {} {3.3}%\n", i, bc7_mode_hist[1 + i], (float)bc7_mode_hist[1 + i] * 100.0f / (float)total_bc7_blocks);
		}

		fmt_debug_printf("\nMode 4 index bit histogram: {} {3.2}%, {} {3.2}%\n",
			mode4_index_hist[0], (float)mode4_index_hist[0] * 100.0f / (float)total_bc7_blocks,
			mode4_index_hist[1], (float)mode4_index_hist[1] * 100.0f / (float)total_bc7_blocks);

		fmt_debug_printf("\nMode 4 rotation histogram:\n");
		for (uint32_t i = 0; i < 4; i++)
		{
			fmt_debug_printf(" {}: {} {3.2}%\n", i, mode4_rot_hist[i], (float)mode4_rot_hist[i] * 100.0f / (float)total_bc7_blocks);
		}

		fmt_debug_printf("\nMode 5 rotation histogram:\n");
		for (uint32_t i = 0; i < 4; i++)
		{
			fmt_debug_printf(" {}: {} {3.2}%\n", i, mode5_rot_hist[i], (float)mode5_rot_hist[i] * 100.0f / (float)total_bc7_blocks);
		}
				
		if (pFilename_prefix)
		{
			std::string mode_vis_filename(std::string(pFilename_prefix) + "_bc7_mode_vis.png");
			save_png(mode_vis_filename, bc7_mode_vis);

			fmt_debug_printf("Wrote BC7 mode visualization to PNG file {}\n", mode_vis_filename);
		}
		
		fmt_debug_printf("--------- End BC7 statistics\n");
		fmt_debug_printf("\n");
	}

	static inline float edge(const vec2F& a, const vec2F& b, const vec2F& pos)
	{
		return (pos[0] - a[0]) * (b[1] - a[1]) - (pos[1] - a[1]) * (b[0] - a[0]);
	}

	void draw_tri2(image& dst, const image* pTex, const tri2& tri, bool alpha_blend)
	{
		assert(dst.get_total_pixels());

		float area = edge(tri.p0, tri.p1, tri.p2);
		if (std::fabs(area) < 1e-6f)
			return;

		const float oo_area = 1.0f / area;

		int minx = (int)std::floor(basisu::minimum(tri.p0[0], tri.p1[0], tri.p2[0] ));
		int miny = (int)std::floor(basisu::minimum(tri.p0[1], tri.p1[1], tri.p2[1] ));

		int maxx = (int)std::ceil(basisu::maximum(tri.p0[0], tri.p1[0], tri.p2[0]));
		int maxy = (int)std::ceil(basisu::maximum(tri.p0[1], tri.p1[1], tri.p2[1]));

		auto clamp8 = [&](float fv) { int v = (int)(fv + .5f); if (v < 0) v = 0; else if (v > 255) v = 255;  return (uint8_t)v; };

		if ((maxx < 0) || (maxy < 0))
			return;
		if ((minx >= (int)dst.get_width()) || (miny >= (int)dst.get_height()))
			return;

		if (minx < 0)
			minx = 0;
		if (maxx >= (int)dst.get_width())
			maxx = dst.get_width() - 1;
		if (miny < 0)
			miny = 0;
		if (maxy >= (int)dst.get_height())
			maxy = dst.get_height() - 1;

		vec4F tex(1.0f);

		for (int y = miny; y <= maxy; ++y)
		{
			assert((y >= 0) && (y < (int)dst.get_height()));

			for (int x = minx; x <= maxx; ++x)
			{
				assert((x >= 0) && (x < (int)dst.get_width()));

				vec2F p{ (float)x + 0.5f, (float)y + 0.5f };

				float w0 = edge(tri.p1, tri.p2, p) * oo_area;
				float w1 = edge(tri.p2, tri.p0, p) * oo_area;
				float w2 = edge(tri.p0, tri.p1, p) * oo_area;

				if ((w0 < 0) || (w1 < 0) || (w2 < 0))
					continue;

				float u = tri.t0[0] * w0 + tri.t1[0] * w1 + tri.t2[0] * w2;
				float v = tri.t0[1] * w0 + tri.t1[1] * w1 + tri.t2[1] * w2;

				if (pTex)
					tex = pTex->get_filtered_vec4F(u * float(pTex->get_width()), v * float(pTex->get_height())) * (1.0f / 255.0f);

				float r = (float)tri.c0.r * w0 + (float)tri.c1.r * w1 + (float)tri.c2.r * w2;
				float g = (float)tri.c0.g * w0 + (float)tri.c1.g * w1 + (float)tri.c2.g * w2;
				float b = (float)tri.c0.b * w0 + (float)tri.c1.b * w1 + (float)tri.c2.b * w2;
				float a = (float)tri.c0.a * w0 + (float)tri.c1.a * w1 + (float)tri.c2.a * w2;

				r *= tex[0];
				g *= tex[1];
				b *= tex[2];
				a *= tex[3];

				if (alpha_blend)
				{
					color_rgba dst_color(dst(x, y));

					const float fa = (float)a * (1.0f / 255.0f);

					r = lerp((float)dst_color[0], r, fa);
					g = lerp((float)dst_color[1], g, fa);
					b = lerp((float)dst_color[2], b, fa);
					a = lerp((float)dst_color[3], a, fa);

					dst(x, y) = color_rgba(clamp8(r), clamp8(g), clamp8(b), clamp8(a));
				}
				else
				{
					dst(x, y) = color_rgba(clamp8(r), clamp8(g), clamp8(b), clamp8(a));
				}

			} // x
		} // y
	}
		
	static inline float eval_rotated_gaussian(float x, float y, float cos_t, float sin_t, float sigma_par, float sigma_perp)
	{
		// Project (x,y) onto the rotated axes
		const float u = cos_t * x + sin_t * y; // parallel to blur direction
		const float v = -sin_t * x + cos_t * y; // perpendicular

		const float inv_par2 = 1.0f / (2.0f * sigma_par * sigma_par);
		const float inv_perp2 = 1.0f / (2.0f * sigma_perp * sigma_perp);

		return std::exp(-(u * u * inv_par2 + v * v * inv_perp2));
	}

	DirectionalKernel make_directional_kernel(float angle_deg, float sigma_par, float sigma_perp)
	{
		assert(sigma_par > 0.0f && "sigma_par must be positive");
		assert(sigma_perp > 0.0f && "sigma_perp must be positive");

		// Convert angle to radians
		const float angle_rad = angle_deg * (3.14159265358979323846f / 180.0f);
		const float cos_t = std::cos(angle_rad);
		const float sin_t = std::sin(angle_rad);

		// Determine kernel half-size and total size
		const float sigma_x = std::sqrt(sigma_par * sigma_par * cos_t * cos_t + sigma_perp * sigma_perp * sin_t * sin_t);
		const float sigma_y = std::sqrt(sigma_par * sigma_par * sin_t * sin_t + sigma_perp * sigma_perp * cos_t * cos_t);
		const float reach = 3.0f * basisu::maximum(sigma_x, sigma_y) + 0.5f;
		int half = maximum(1, static_cast<int>(std::ceil(reach)));

		const int size = 2 * half + 1; // always odd
		assert(size >= 3 && "kernel must be at least 3x3");

		DirectionalKernel kernel;
		kernel.m_size = size;
		kernel.m_data.resize(size * size, 0.0f);

		// Fill with un-normalised Gaussian values, accumulate sum for normalisation
		float sum = 0.0f;
		for (int row = 0; row < size; ++row)
		{
			// y offset from kernel centre (positive = downward in image space)
			const float y = static_cast<float>(row - half);

			for (int col = 0; col < size; ++col)
			{
				const float x = static_cast<float>(col - half);
				const float w = eval_rotated_gaussian(x, y, cos_t, sin_t, sigma_par, sigma_perp);
				kernel.m_data[row * size + col] = w;
				sum += w;
			}
		}
				
		assert(sum > 0.0f && "kernel sum is zero - sigma values too small?");
		const float inv_sum = 1.0f / sum;
		for (float& v : kernel.m_data)
			v *= inv_sum;

#if defined(DEBUG) || defined(_DEBUG)
		{
			float check = 0.0f;
			for (float v : kernel.m_data) check += v;
			assert(std::abs(check - 1.0f) < 1e-4f && "kernel normalisation failed");
		}
#endif

		return kernel;
	}

	void directional_gaussian_blur(const image& src_img, image& dst_img, float angle_deg, float sigma_par, float sigma_perp)
	{
		assert((sigma_par > 0.0f) && (sigma_par > 0.0f));

		dst_img.match_dimensions(src_img);

		const int width = src_img.get_width();
		const int height = src_img.get_height();

		DirectionalKernel kernel(make_directional_kernel(angle_deg, sigma_par, sigma_perp));
		const int kernel_size = kernel.m_size;
		const int half_kernel_size = kernel_size >> 1;

		for (int y = 0; y < height; y++)
		{
			for (int x = 0; x < width; x++)
			{
				float sum_r = 0, sum_g = 0, sum_b = 0, sum_a = 0;

				for (int ky = 0; ky < kernel_size; ky++)
				{
					const int sy = clamp(y + (ky - half_kernel_size), 0, height - 1);

					for (int kx = 0; kx < kernel_size; kx++)
					{
						const int sx = clamp(x + (kx - half_kernel_size), 0, width - 1);

						float weight = kernel.m_data[kx + ky * kernel_size];

						const color_rgba& pixel = src_img(sx, sy);

						sum_r += (float)pixel.r * weight;
						sum_g += (float)pixel.g * weight;
						sum_b += (float)pixel.b * weight;
						sum_a += (float)pixel.a * weight;

					} // kx

				} // ky

				color_rgba& dst_pixel = dst_img(x, y);
				dst_pixel.set(basisu::fast_roundf_pos_int(sum_r), basisu::fast_roundf_pos_int(sum_g), basisu::fast_roundf_pos_int(sum_b), basisu::fast_roundf_pos_int(sum_a));

			} // x
		} // y
	}

	// macro sent by CMakeLists.txt file when (TARGET_WASM AND WASM_THREADING)
#if BASISU_WASI_THREADS
	// Default to 8 - seems reasonable.
	static int g_num_wasi_threads = 8;
#else
	static int g_num_wasi_threads = 0;
#endif

	void set_num_wasi_threads(uint32_t num_threads)
	{
		g_num_wasi_threads = num_threads;
	}

	int get_num_hardware_threads()
	{
#ifdef __wasi__
		int num_threads = g_num_wasi_threads;
#else
		int num_threads = std::thread::hardware_concurrency();
#endif
				
		return num_threads;
	}
		
	bool display_astc_statistics(
		const vector2D<astc_helpers::astc_block>& blocks,
		uint32_t block_width, uint32_t block_height, uint32_t image_width, uint32_t image_height, bool verbose)
	{
		const uint32_t total_block_pixels = block_width * block_height;

		fmt_printf("------- display_astc_statistics:\n");
		fmt_printf("Image dimensions in pixels: {}x{}, blocks: {}x{}\n", image_width, image_height, blocks.get_width(), blocks.get_height());
		fmt_printf("Block dimensions in pixels: {}x{}, {} total pixels\n", block_width, block_height, total_block_pixels);

		fmt_printf("Extra cols/rows to pad image to ASTC block dimensions: {}x{}\n",
			blocks.get_width() * block_width - image_width,
			blocks.get_height() * block_height - image_height);

		image dec_image_srgb(image_width, image_height);
		image dec_image_linear(image_width, image_height);
		imagef dec_image_float(image_width, image_height);

		uint32_t cem_hist[16] = { };
		uint32_t cem_dp_hist[16] = { };
		uint32_t cem_used_bc_hist[16] = { };
		uint32_t total_dp = 0;
		uint32_t cem_ccs_hist[16][4] = { };
		uint32_t cem_part_hist[16][4] = { }; // 1-4 subsets

		uint32_t total_solid_blocks_ldr = 0;
		uint32_t total_solid_blocks_hdr = 0;
		uint32_t total_normal_blocks = 0;

		uint32_t part_hist[4] = { };
		uint32_t used_endpoint_levels_hist[astc_helpers::LAST_VALID_ENDPOINT_ISE_RANGE - astc_helpers::FIRST_VALID_ENDPOINT_ISE_RANGE + 1] = { };
		uint32_t used_weight_levels_hist[astc_helpers::LAST_VALID_WEIGHT_ISE_RANGE - astc_helpers::FIRST_VALID_WEIGHT_ISE_RANGE + 1] = { };

		uint32_t total_unequal_cem_blocks = 0;
		uint32_t total_unequal_cem_blocks_2subsets = 0;
		uint32_t total_unequal_cem_blocks_3subsets = 0;
		uint32_t total_unequal_cem_blocks_4subsets = 0;

		uint32_t highest_part_seed = 0;

		uint32_t total_suboptimal_cem_blocks = 0;
		uint32_t total_unnecessary_suboptimal_cem_blocks = 0;
		uint32_t total_useful_suboptimal_cem_blocks = 0;

		int min_weight_grid_width = INT_MAX, min_weight_grid_height = INT_MAX;
		int max_weight_grid_width = 0, max_weight_grid_height = 0;

		uint32_t total_ldr_blocks = 0, total_hdr_blocks = 0;

		basisu::hash_map<uint32_t, uint32_t> weight_grid_histogram;

		basisu::hash_map<uint32_t, uint32_t> part_seed_hash;

		struct log_astc_block_config_cmp_t
		{
			bool operator()(const astc_helpers::log_astc_block& a,
				const astc_helpers::log_astc_block& b) const
			{
				// This only compares the ASTC configuration for equality, NOT the contents.
				if (a.m_error_flag != b.m_error_flag)
					return false;
				if (a.m_error_flag)
					return true;

				if (a.m_grid_width != b.m_grid_width)
					return false;
				if (a.m_grid_height != b.m_grid_height)
					return false;

				if (a.m_solid_color_flag_ldr != b.m_solid_color_flag_ldr)
					return false;
				if (a.m_solid_color_flag_hdr != b.m_solid_color_flag_hdr)
					return false;

				if (a.m_solid_color_flag_ldr || a.m_solid_color_flag_hdr)
					return true;

				if (a.m_dual_plane != b.m_dual_plane)
					return false;
				if (a.m_color_component_selector != b.m_color_component_selector)
					return false;

				if (a.m_num_partitions != b.m_num_partitions)
					return false;
				if (a.m_uses_suboptimal_cem_encoding != b.m_uses_suboptimal_cem_encoding)
					return false;

				if (a.m_endpoint_ise_range != b.m_endpoint_ise_range)
					return false;
				if (a.m_weight_ise_range != b.m_weight_ise_range)
					return false;

				for (uint32_t i = 0; i < a.m_num_partitions; i++)
					if (a.m_color_endpoint_modes[i] != b.m_color_endpoint_modes[i])
						return false;

				return true;
			}
		};

		basisu::hash_map<astc_helpers::log_astc_block, uint32_t, basist::bit_hasher<astc_helpers::log_astc_block>, log_astc_block_config_cmp_t > unique_config_histogram;

		uint32_t total_subsets = 0;

		for (uint32_t by = 0; by < blocks.get_height(); by++)
		{
			for (uint32_t bx = 0; bx < blocks.get_width(); bx++)
			{
				astc_helpers::log_astc_block log_blk;

				if (!astc_helpers::unpack_block(&blocks(bx, by), log_blk, block_width, block_height))
				{
					fmt_error_printf("astc_helpers::unpack_block() failed on block {}x{}\n", bx, by);
					return false;
				}

				if (log_blk.m_error_flag)
				{
					fmt_error_printf("astc_helpers::unpack_block() returned an error flag on block {}x{}\n", bx, by);
					return false;
				}

				if (log_blk.m_uses_suboptimal_cem_encoding)
				{
					total_suboptimal_cem_blocks++;

					astc_helpers::log_astc_block temp_log_blk(log_blk);
					temp_log_blk.m_uses_suboptimal_cem_encoding = false;

					astc_helpers::astc_block temp_phys_block;

					int expected_endpoint_range = -1;

					bool pack_status = astc_helpers::pack_astc_block(temp_phys_block, temp_log_blk, &expected_endpoint_range);

					// If the packing succeeded without the suboptimal CEM encoding, it means the BISE endpoint range didn't change, and it was unnecessary to use the suboptimal CEM encoding in the first place.
					if (pack_status)
					{
						total_unnecessary_suboptimal_cem_blocks++;
					}
					else
					{
						// the endpoint range should have changed, and be valid
						assert(expected_endpoint_range != -1);
						assert(expected_endpoint_range != log_blk.m_endpoint_ise_range);
						total_useful_suboptimal_cem_blocks++;
					}
				}

				{
					astc_helpers::log_astc_block scrubbed_log_blk;
					memset(&scrubbed_log_blk, 0, sizeof(scrubbed_log_blk));

					// just record the config, not the contents, so only the config hashes
					scrubbed_log_blk.m_solid_color_flag_ldr = log_blk.m_solid_color_flag_ldr;
					scrubbed_log_blk.m_solid_color_flag_hdr = log_blk.m_solid_color_flag_hdr;
					scrubbed_log_blk.m_dual_plane = log_blk.m_dual_plane;
					scrubbed_log_blk.m_color_component_selector = log_blk.m_color_component_selector;
					scrubbed_log_blk.m_grid_width = log_blk.m_grid_width;
					scrubbed_log_blk.m_grid_height = log_blk.m_grid_height;
					scrubbed_log_blk.m_num_partitions = log_blk.m_num_partitions;
					scrubbed_log_blk.m_uses_suboptimal_cem_encoding = log_blk.m_uses_suboptimal_cem_encoding;
					scrubbed_log_blk.m_color_endpoint_modes[0] = log_blk.m_color_endpoint_modes[0];
					scrubbed_log_blk.m_color_endpoint_modes[1] = log_blk.m_color_endpoint_modes[1];
					scrubbed_log_blk.m_color_endpoint_modes[2] = log_blk.m_color_endpoint_modes[2];
					scrubbed_log_blk.m_color_endpoint_modes[3] = log_blk.m_color_endpoint_modes[3];
					scrubbed_log_blk.m_weight_ise_range = log_blk.m_weight_ise_range;
					scrubbed_log_blk.m_endpoint_ise_range = log_blk.m_endpoint_ise_range;

					auto ins_res(unique_config_histogram.insert(scrubbed_log_blk, 0));
					(ins_res.first)->second = (ins_res.first)->second + 1;
				}

				bool is_hdr = log_blk.m_solid_color_flag_hdr;

				if (log_blk.m_solid_color_flag_ldr)
				{
					total_solid_blocks_ldr++;
					total_ldr_blocks++;
				}
				else if (log_blk.m_solid_color_flag_hdr)
				{
					total_solid_blocks_hdr++;
					total_hdr_blocks++;
				}
				else
				{
					total_normal_blocks++;

					min_weight_grid_width = minimum<int>(min_weight_grid_width, log_blk.m_grid_width);
					min_weight_grid_height = minimum<int>(min_weight_grid_height, log_blk.m_grid_height);

					max_weight_grid_width = maximum<int>(max_weight_grid_width, log_blk.m_grid_width);
					max_weight_grid_height = maximum<int>(max_weight_grid_height, log_blk.m_grid_height);

					{
						uint32_t weight_grid_hash_key = log_blk.m_grid_width | (log_blk.m_grid_height << 8);
						auto ins_res(weight_grid_histogram.insert(weight_grid_hash_key, 0));
						(ins_res.first)->second = (ins_res.first)->second + 1;
					}

					if (log_blk.m_dual_plane)
					{
						total_dp++;
						cem_ccs_hist[log_blk.m_color_endpoint_modes[0]][log_blk.m_color_component_selector]++;
					}

					cem_part_hist[log_blk.m_color_endpoint_modes[0]][log_blk.m_num_partitions - 1]++;

					part_hist[log_blk.m_num_partitions - 1]++;

					// For debugging seed packing bugs
					highest_part_seed = basisu::maximum<uint32_t>(highest_part_seed, log_blk.m_partition_id);

					if (log_blk.m_num_partitions > 1)
					{
						auto ins_it = part_seed_hash.insert(log_blk.m_partition_id, 0);
						(ins_it.first)->second = (ins_it.first)->second + 1;
					}

					uint32_t cur_endpoint_ofs = 0;
					bool has_unequal_cems = false;

					total_subsets += log_blk.m_num_partitions;

					for (uint32_t p = 0; p < log_blk.m_num_partitions; p++)
					{
						if (astc_helpers::is_cem_hdr(log_blk.m_color_endpoint_modes[p]))
							is_hdr = true;

						cem_hist[log_blk.m_color_endpoint_modes[p]]++;

						if (log_blk.m_dual_plane)
							cem_dp_hist[log_blk.m_color_endpoint_modes[p]]++;

						if ((p) && (log_blk.m_color_endpoint_modes[p] != log_blk.m_color_endpoint_modes[0]))
						{
							has_unequal_cems = true;
						}

						if (astc_helpers::is_cem_ldr(log_blk.m_color_endpoint_modes[p]))
						{
							bool uses_bc = astc_helpers::used_blue_contraction(log_blk.m_color_endpoint_modes[p], log_blk.m_endpoints + cur_endpoint_ofs, log_blk.m_endpoint_ise_range);

							cem_used_bc_hist[log_blk.m_color_endpoint_modes[p]] += uses_bc;
						}

						cur_endpoint_ofs += astc_helpers::get_num_cem_values(log_blk.m_color_endpoint_modes[p]);
					}

					if (log_blk.m_num_partitions >= 2)
					{
						total_unequal_cem_blocks += has_unequal_cems;

						if (log_blk.m_num_partitions == 2)
							total_unequal_cem_blocks_2subsets += has_unequal_cems;
						else if (log_blk.m_num_partitions == 3)
							total_unequal_cem_blocks_3subsets += has_unequal_cems;
						else if (log_blk.m_num_partitions == 4)
							total_unequal_cem_blocks_4subsets += has_unequal_cems;
					}

					used_weight_levels_hist[open_range_check<int>(log_blk.m_weight_ise_range - astc_helpers::FIRST_VALID_WEIGHT_ISE_RANGE, std::size(used_weight_levels_hist))]++;
					used_endpoint_levels_hist[open_range_check<int>(log_blk.m_endpoint_ise_range - astc_helpers::FIRST_VALID_ENDPOINT_ISE_RANGE, std::size(used_endpoint_levels_hist))]++;
				}

				if (is_hdr)
				{
					total_hdr_blocks++;
				}
				else
				{
					total_ldr_blocks++;

					color_rgba block_pixels[astc_helpers::MAX_BLOCK_PIXELS];

					// sRGB8 decode profile unpack
					bool status = astc_helpers::decode_block(log_blk, block_pixels, block_width, block_height, astc_helpers::cDecodeModeSRGB8);
					if (!status)
					{
						fmt_error_printf("astc_helpers::decode_block() failed on block {}x{}\n", bx, by);
						return false;
					}

					dec_image_srgb.set_block_clipped(block_pixels, bx * block_width, by * block_height, block_width, block_height);

					// linear8 decode profile unpack
					status = astc_helpers::decode_block(log_blk, block_pixels, block_width, block_height, astc_helpers::cDecodeModeLDR8);
					if (!status)
					{
						fmt_error_printf("astc_helpers::decode_block() failed on block {}x{}\n", bx, by);
						return false;
					}

					dec_image_linear.set_block_clipped(block_pixels, bx * block_width, by * block_height, block_width, block_height);
				}

				// half float unpack
				{
					basist::half_float block_pixels_half[astc_helpers::MAX_BLOCK_PIXELS][4];

					bool status = astc_helpers::decode_block(log_blk, block_pixels_half, block_width, block_height, astc_helpers::cDecodeModeHDR16);
					if (!status)
					{
						fmt_error_printf("astc_helpers::decode_block() failed on block {}x{}\n", bx, by);
						return false;
					}

					vec4F block_pixels_float[astc_helpers::MAX_BLOCK_PIXELS];
					for (uint32_t i = 0; i < total_block_pixels; i++)
						for (uint32_t j = 0; j < 4; j++)
							block_pixels_float[i][j] = basist::half_to_float(block_pixels_half[i][j]);

					dec_image_float.set_block_clipped(block_pixels_float, bx * block_width, by * block_height, block_width, block_height);
				}

			} // bx

		} //by

		fmt_printf("Total LDR blocks: {}, total HDR blocks: {}\n", total_ldr_blocks, total_hdr_blocks);

		if (verbose)
		{
			save_png("astc_decoded_srgb8_ldr.png", dec_image_srgb);
			fmt_printf("Wrote astc_decoded_srgb8_ldr.png\n");

			save_png("astc_decoded_linear8_ldr.png", dec_image_linear);
			fmt_printf("Wrote astc_decoded_linear8_ldr.png\n");

			write_exr("astc_decoded_half.exr", dec_image_float, 4, 0);
			fmt_printf("Wrote astc_decoded_half.exr\n");
		}

		fmt_printf("\nASTC file statistics:\n");

		const uint32_t total_blocks = (uint32_t)blocks.size();

		fmt_printf("Total blocks: {}, total void extent LDR: {}, total void extent HDR: {}, total normal: {}\n", total_blocks, total_solid_blocks_ldr, total_solid_blocks_hdr, total_normal_blocks);
		fmt_printf("Total dual plane: {} {3.2}%\n", total_dp, total_dp * 100.0f / (float)total_blocks);

		fmt_printf("Total blocks using suboptimal CEM encodings: {} {3.2}%\n", total_suboptimal_cem_blocks, total_suboptimal_cem_blocks * 100.0f / (float)total_blocks);
		fmt_printf("Total blocks using unnecessary suboptimal CEM encodings: {} {3.2}%\n", total_unnecessary_suboptimal_cem_blocks, total_unnecessary_suboptimal_cem_blocks * 100.0f / (float)total_blocks);
		fmt_printf("Total blocks using useful suboptimal CEM encodings: {} {3.2}%\n", total_useful_suboptimal_cem_blocks, total_useful_suboptimal_cem_blocks * 100.0f / (float)total_blocks);

		fmt_printf("Total subsets across all blocks: {}, Avg. subsets per block: {}\n", total_subsets, (float)total_subsets / (float)total_blocks);

		fmt_printf("Min weight grid usage bounds: {}x{}\n", min_weight_grid_width, min_weight_grid_height);
		fmt_printf("Max weight grid usage bounds: {}x{}\n", max_weight_grid_width, max_weight_grid_height);

		fmt_printf("\nPartition usage histogram:\n");
		for (uint32_t i = 0; i < 4; i++)
			fmt_printf("{}: {} {3.2}%\n", i + 1, part_hist[i], (float)part_hist[i] * 100.0f / (float)total_blocks);

		fmt_printf("\nCEM usage histogram (percentages relative to total overall subsets used in texture):\n");
		for (uint32_t i = 0; i < 15; i++)
		{
			fmt_printf("{}: {} {3.2}%, total BC: {} {3.2}%, total DP: {} {3.2}% (R:{} G:{} B:{} A:{}), parts: {} {} {} {})\n", i,
				cem_hist[i], (float)cem_hist[i] * 100.0f / (float)total_subsets,
				cem_used_bc_hist[i], (float)cem_used_bc_hist[i] * 100.0f / (float)total_subsets,
				cem_dp_hist[i], (float)cem_dp_hist[i] * 100.0f / (float)total_subsets,
				cem_ccs_hist[i][0], cem_ccs_hist[i][1], cem_ccs_hist[i][2], cem_ccs_hist[i][3],
				cem_part_hist[i][0], cem_part_hist[i][1], cem_part_hist[i][2], cem_part_hist[i][3]);
		}

		fmt_printf("\nUsed endpoint ISE levels:\n");
		for (uint32_t i = 0; i < std::size(used_endpoint_levels_hist); i++)
			fmt_printf("{} levels: {}\n", astc_helpers::get_ise_levels(astc_helpers::FIRST_VALID_ENDPOINT_ISE_RANGE + i), used_endpoint_levels_hist[i]);

		fmt_printf("\nUsed weight ISE levels:\n");
		for (uint32_t i = 0; i < std::size(used_weight_levels_hist); i++)
			fmt_printf("{} levels: {}\n", astc_helpers::get_ise_levels(astc_helpers::FIRST_VALID_WEIGHT_ISE_RANGE + i), used_weight_levels_hist[i]);

		fmt_printf("\nTotal 2+ subset blocks using unequal CEM's: {} {3.2}%\n", total_unequal_cem_blocks, (float)total_unequal_cem_blocks * 100.0f / (float)total_blocks);
		fmt_printf("Total 2 subset blocks using unequal CEM's: {} {3.2}%\n", total_unequal_cem_blocks_2subsets, (float)total_unequal_cem_blocks_2subsets * 100.0f / (float)total_blocks);
		fmt_printf("Total 3 subset blocks using unequal CEM's: {} {3.2}%\n", total_unequal_cem_blocks_3subsets, (float)total_unequal_cem_blocks_3subsets * 100.0f / (float)total_blocks);
		fmt_printf("Total 4 subset blocks using unequal CEM's: {} {3.2}%\n", total_unequal_cem_blocks_4subsets, (float)total_unequal_cem_blocks_4subsets * 100.0f / (float)total_blocks);

		fmt_printf("\nHighest part ID seed: {}, 0x{0x}\n", highest_part_seed, highest_part_seed);

		fmt_printf("Total used partition seed ID's: {}\n", part_seed_hash.size_u32());
		if (verbose)
		{
			for (auto it = part_seed_hash.begin(); it != part_seed_hash.end(); ++it)
				fmt_printf("  Seed ID {} used {} times\n", it->first, it->second);
		}

		fmt_printf("\nWeight grid usage histogram:\n");

		uint64_vec v;
		for (auto it = weight_grid_histogram.begin(); it != weight_grid_histogram.end(); ++it)
			v.push_back(((uint64_t)it->first << 32) | it->second);

		v.sort();

		for (uint32_t i = 0; i < v.size(); i++)
			fmt_printf("  {}x{}: total blocks {}\n", (v[i] >> 32) & 0xFF, (v[i] >> 40) & 0xFF, v[i] & UINT32_MAX);

		fmt_printf("\nTotal unique ASTC configurations: {}\n", unique_config_histogram.size_u32());

		if (verbose)
		{
			uint32_t config_idx = 0;
			for (auto it = unique_config_histogram.begin(); it != unique_config_histogram.end(); ++it)
			{
				const auto& l = it->first;
				const uint32_t total = it->second;

				fmt_printf("  {}. Used {} {3.2}% times: Solid LDR: {} HDR: {}, Grid: {}x{}, Dual Plane: {}, CCS: {}, NumParts: {}, SuboptimalCEM: {}, CEMS: {} {} {} {}, WeightISERange: {} ({} levels), EndpointISERange: {} ({} levels)\n",
					config_idx, total, float(total) * 100.0f / total_blocks,
					l.m_solid_color_flag_ldr, l.m_solid_color_flag_hdr,
					l.m_grid_width, l.m_grid_height,
					l.m_dual_plane, l.m_color_component_selector,
					l.m_num_partitions, l.m_uses_suboptimal_cem_encoding,
					l.m_color_endpoint_modes[0], l.m_color_endpoint_modes[1], l.m_color_endpoint_modes[2], l.m_color_endpoint_modes[3],
					l.m_weight_ise_range, astc_helpers::get_ise_levels(l.m_weight_ise_range),
					l.m_endpoint_ise_range, astc_helpers::get_ise_levels(l.m_endpoint_ise_range));

				config_idx++;
			}
		}

		fmt_printf("------- display_astc_statistics: OK\n");
		return true;
	}

	bool display_astc_statistics(
		const vector2D<astc_helpers::log_astc_block>& blocks,
		uint32_t block_width, uint32_t block_height, uint32_t image_width, uint32_t image_height, bool verbose)
	{
		vector2D<astc_helpers::astc_block> phys_blocks(blocks.get_width(), blocks.get_height());

		// ugh, but it's just for development/testing
		for (uint32_t y = 0; y < blocks.get_height(); y++)
			for (uint32_t x = 0; x < blocks.get_width(); x++)
				if (!astc_helpers::pack_astc_block(phys_blocks(x, y), blocks(x, y)))
					return false;

		return display_astc_statistics(
			phys_blocks,
			block_width, block_height, image_width, image_height, verbose);
	}

	basisu::vector<convar*>& get_convars()
	{
		static basisu::vector<convar*> s_convars;
		return s_convars;
	}

	void list_convars()
	{
		fmt_printf("{} convars:\n", get_convars().size_u32());

		for (size_t i = 0; i < get_convars().size(); i++)
		{
			const convar* p = get_convars()[i];

			fmt_printf("convar: {} type: {} value: {}\n", p->get_name(), get_convar_type_string(p->get_type()), p->get_val_as_string());
		}
	}

	static convar* find_convar(const std::string& name)
	{
		for (size_t i = 0; i < get_convars().size(); i++)
			if (name == get_convars()[i]->get_name())
				return get_convars()[i];

		return nullptr;
	}

	void print_convar(const std::string& name)
	{
		convar* p = find_convar(name);
		if (!p)
		{
			fmt_printf("error: convar \"{}\" not found\n", name);
			return;
		}
		
		fmt_printf("convar: {} type: {} value: {}\n", name, get_convar_type_string(p->get_type()), p->get_val_as_string());
	}

	void reset_convar(const std::string& name)
	{
		convar* p = find_convar(name);
		if (!p)
		{
			fmt_printf("error: convar \"{}\" not found\n", name);
			return;
		}

		p->reset();

		fmt_printf("OK\n");
	}

	void set_convar(const std::string& name, const std::string& val)
	{
		convar* p = find_convar(name);
		if (!p)
		{
			fmt_printf("error: convar \"{}\" not found\n", name);
			return;
		}

		if (!val.size())
		{
			fmt_printf("error: empty value for convar \"{}\"\n", name);
			return;
		}

		if (p->get_type() == cConvarFloat)
		{
			p->set((float)atof(val.c_str()));
		}
		else
		{
			p->set(atoi(val.c_str()));
		}

		fmt_printf("OK\n");
	}
									
} // namespace basisu