Betsy: Add caching and BC1 compression support

core/io/image.cpp

@@ -2751,6 +2751,19 @@ Error Image::compress_from_channels(CompressMode p_mode, UsedChannels p_channels
 
 			} break;
 
+			case COMPRESS_S3TC: {
+				// BC3 is unsupported currently.
+				if ((p_channels == USED_CHANNELS_RGB || p_channels == USED_CHANNELS_L) && _image_compress_bc_rd_func) {
+					Error result = _image_compress_bc_rd_func(this, p_channels);
+
+					// If the image was compressed successfully, we return here. If not, we fall back to the default compression scheme.
+					if (result == OK) {
+						return OK;
+					}
+				}
+
+			} break;
+
 			default: {
 			}
 		}
@@ -3138,6 +3151,7 @@ void (*Image::_image_compress_etc1_func)(Image *) = nullptr;
 void (*Image::_image_compress_etc2_func)(Image *, Image::UsedChannels) = nullptr;
 void (*Image::_image_compress_astc_func)(Image *, Image::ASTCFormat) = nullptr;
 Error (*Image::_image_compress_bptc_rd_func)(Image *, Image::UsedChannels) = nullptr;
+Error (*Image::_image_compress_bc_rd_func)(Image *, Image::UsedChannels) = nullptr;
 void (*Image::_image_decompress_bc)(Image *) = nullptr;
 void (*Image::_image_decompress_bptc)(Image *) = nullptr;
 void (*Image::_image_decompress_etc1)(Image *) = nullptr;

core/io/image.h

@@ -160,6 +160,7 @@ public:
 	static void (*_image_compress_astc_func)(Image *, ASTCFormat p_format);
 
 	static Error (*_image_compress_bptc_rd_func)(Image *, UsedChannels p_channels);
+	static Error (*_image_compress_bc_rd_func)(Image *, UsedChannels p_channels);
 
 	static void (*_image_decompress_bc)(Image *);
 	static void (*_image_decompress_bptc)(Image *);

doc/classes/ProjectSettings.xml

@@ -2890,10 +2890,13 @@
 		<member name="rendering/textures/lossless_compression/force_png" type="bool" setter="" getter="" default="false">
 			If [code]true[/code], the texture importer will import lossless textures using the PNG format. Otherwise, it will default to using WebP.
 		</member>
+		<member name="rendering/textures/vram_compression/cache_gpu_compressor" type="bool" setter="" getter="" default="true">
+			If [code]true[/code], the GPU texture compressor will cache the local RenderingDevice and its resources (shaders and pipelines), allowing for faster subsequent imports at a memory cost.
+		</member>
 		<member name="rendering/textures/vram_compression/compress_with_gpu" type="bool" setter="" getter="" default="true">
-			If [code]true[/code], the texture importer will utilize the GPU for compressing textures, which makes large textures import significantly faster.
+			If [code]true[/code], the texture importer will utilize the GPU for compressing textures, improving the import time of large images.
 			[b]Note:[/b] This setting requires either Vulkan or D3D12 available as a rendering backend.
-			[b]Note:[/b] Currently this only affects BC6H compression, which is used on Desktop and Console for HDR images.
+			[b]Note:[/b] Currently this only affects BC1 and BC6H compression, which are used on Desktop and Console for fully opaque and HDR images respectively.
 		</member>
 		<member name="rendering/textures/vram_compression/import_etc2_astc" type="bool" setter="" getter="" default="false">
 			If [code]true[/code], the texture importer will import VRAM-compressed textures using the Ericsson Texture Compression 2 algorithm for lower quality textures and normal maps and Adaptable Scalable Texture Compression algorithm for high quality textures (in 4×4 block size).

modules/betsy/SCsub

@@ -4,6 +4,7 @@ Import("env_modules")
 
 env_betsy = env_modules.Clone()
 env_betsy.GLSL_HEADER("bc6h.glsl")
+env_betsy.GLSL_HEADER("bc1.glsl")
 env_betsy.Depends(Glob("*.glsl.gen.h"), ["#glsl_builders.py"])
 
 # Thirdparty source files

modules/betsy/bc1.glsl

@@ -0,0 +1,483 @@
+#[versions]
+
+standard = "";
+dithered = "#define BC1_DITHER";
+
+#[compute]
+#version 450
+
+#include "CrossPlatformSettings_piece_all.glsl"
+#include "UavCrossPlatform_piece_all.glsl"
+
+#define FLT_MAX 340282346638528859811704183484516925440.0f
+
+layout(binding = 0) uniform sampler2D srcTex;
+layout(binding = 1, rg32ui) uniform restrict writeonly uimage2D dstTexture;
+
+layout(std430, binding = 2) readonly restrict buffer globalBuffer {
+	float2 c_oMatch5[256];
+	float2 c_oMatch6[256];
+};
+
+layout(push_constant, std430) uniform Params {
+	uint p_numRefinements;
+	uint p_padding[3];
+}
+params;
+
+layout(local_size_x = 8, //
+		local_size_y = 8, //
+		local_size_z = 1) in;
+
+float3 rgb565to888(float rgb565) {
+	float3 retVal;
+	retVal.x = floor(rgb565 / 2048.0f);
+	retVal.y = floor(mod(rgb565, 2048.0f) / 32.0f);
+	retVal.z = floor(mod(rgb565, 32.0f));
+
+	// This is the correct 565 to 888 conversion:
+	//		rgb = floor( rgb * ( 255.0f / float3( 31.0f, 63.0f, 31.0f ) ) + 0.5f )
+	//
+	// However stb_dxt follows a different one:
+	//		rb = floor( rb * ( 256 / 32 + 8 / 32 ) );
+	//		g  = floor( g  * ( 256 / 64 + 4 / 64 ) );
+	//
+	// I'm not sure exactly why but it's possible this is how the S3TC specifies it should be decoded
+	// It's quite possible this is the reason:
+	//		http://www.ludicon.com/castano/blog/2009/03/gpu-dxt-decompression/
+	//
+	// Or maybe it's just because it's cheap to do with integer shifts.
+	// Anyway, we follow stb_dxt's conversion just in case
+	// (gives almost the same result, with 1 or -1 of difference for a very few values)
+	//
+	// Perhaps when we make 888 -> 565 -> 888 it doesn't matter
+	// because they end up mapping to the original number
+
+	return floor(retVal * float3(8.25f, 4.0625f, 8.25f));
+}
+
+float rgb888to565(float3 rgbValue) {
+	rgbValue.rb = floor(rgbValue.rb * 31.0f / 255.0f + 0.5f);
+	rgbValue.g = floor(rgbValue.g * 63.0f / 255.0f + 0.5f);
+
+	return rgbValue.r * 2048.0f + rgbValue.g * 32.0f + rgbValue.b;
+}
+
+// linear interpolation at 1/3 point between a and b, using desired rounding type
+float3 lerp13(float3 a, float3 b) {
+#ifdef STB_DXT_USE_ROUNDING_BIAS
+	// with rounding bias
+	return a + floor((b - a) * (1.0f / 3.0f) + 0.5f);
+#else
+	// without rounding bias
+	return floor((2.0f * a + b) / 3.0f);
+#endif
+}
+
+/// Unpacks a block of 4 colors from two 16-bit endpoints
+void EvalColors(out float3 colors[4], float c0, float c1) {
+	colors[0] = rgb565to888(c0);
+	colors[1] = rgb565to888(c1);
+	colors[2] = lerp13(colors[0], colors[1]);
+	colors[3] = lerp13(colors[1], colors[0]);
+}
+
+/** The color optimization function. (Clever code, part 1)
+@param outMinEndp16 [out]
+	Minimum endpoint, in RGB565
+@param outMaxEndp16 [out]
+	Maximum endpoint, in RGB565
+*/
+void OptimizeColorsBlock(const uint srcPixelsBlock[16], out float outMinEndp16, out float outMaxEndp16) {
+	// determine color distribution
+	float3 avgColor;
+	float3 minColor;
+	float3 maxColor;
+
+	avgColor = minColor = maxColor = unpackUnorm4x8(srcPixelsBlock[0]).xyz;
+	for (int i = 1; i < 16; ++i) {
+		const float3 currColorUnorm = unpackUnorm4x8(srcPixelsBlock[i]).xyz;
+		avgColor += currColorUnorm;
+		minColor = min(minColor, currColorUnorm);
+		maxColor = max(maxColor, currColorUnorm);
+	}
+
+	avgColor = round(avgColor * 255.0f / 16.0f);
+	maxColor *= 255.0f;
+	minColor *= 255.0f;
+
+	// determine covariance matrix
+	float cov[6];
+	for (int i = 0; i < 6; ++i)
+		cov[i] = 0;
+
+	for (int i = 0; i < 16; ++i) {
+		const float3 currColor = unpackUnorm4x8(srcPixelsBlock[i]).xyz * 255.0f;
+		float3 rgbDiff = currColor - avgColor;
+
+		cov[0] += rgbDiff.r * rgbDiff.r;
+		cov[1] += rgbDiff.r * rgbDiff.g;
+		cov[2] += rgbDiff.r * rgbDiff.b;
+		cov[3] += rgbDiff.g * rgbDiff.g;
+		cov[4] += rgbDiff.g * rgbDiff.b;
+		cov[5] += rgbDiff.b * rgbDiff.b;
+	}
+
+	// convert covariance matrix to float, find principal axis via power iter
+	for (int i = 0; i < 6; ++i)
+		cov[i] /= 255.0f;
+
+	float3 vF = maxColor - minColor;
+
+	const int nIterPower = 4;
+	for (int iter = 0; iter < nIterPower; ++iter) {
+		const float r = vF.r * cov[0] + vF.g * cov[1] + vF.b * cov[2];
+		const float g = vF.r * cov[1] + vF.g * cov[3] + vF.b * cov[4];
+		const float b = vF.r * cov[2] + vF.g * cov[4] + vF.b * cov[5];
+
+		vF.r = r;
+		vF.g = g;
+		vF.b = b;
+	}
+
+	float magn = max3(abs(vF.r), abs(vF.g), abs(vF.b));
+	float3 v;
+
+	if (magn < 4.0f) { // too small, default to luminance
+		v.r = 299.0f; // JPEG YCbCr luma coefs, scaled by 1000.
+		v.g = 587.0f;
+		v.b = 114.0f;
+	} else {
+		v = trunc(vF * (512.0f / magn));
+	}
+
+	// Pick colors at extreme points
+	float3 minEndpoint, maxEndpoint;
+	float minDot = FLT_MAX;
+	float maxDot = -FLT_MAX;
+	for (int i = 0; i < 16; ++i) {
+		const float3 currColor = unpackUnorm4x8(srcPixelsBlock[i]).xyz * 255.0f;
+		const float dotValue = dot(currColor, v);
+
+		if (dotValue < minDot) {
+			minDot = dotValue;
+			minEndpoint = currColor;
+		}
+
+		if (dotValue > maxDot) {
+			maxDot = dotValue;
+			maxEndpoint = currColor;
+		}
+	}
+
+	outMinEndp16 = rgb888to565(minEndpoint);
+	outMaxEndp16 = rgb888to565(maxEndpoint);
+}
+
+// The color matching function
+uint MatchColorsBlock(const uint srcPixelsBlock[16], float3 color[4]) {
+	uint mask = 0u;
+	float3 dir = color[0] - color[1];
+	float stops[4];
+
+	for (int i = 0; i < 4; ++i)
+		stops[i] = dot(color[i], dir);
+
+	// think of the colors as arranged on a line; project point onto that line, then choose
+	// next color out of available ones. we compute the crossover points for "best color in top
+	// half"/"best in bottom half" and then the same inside that subinterval.
+	//
+	// relying on this 1d approximation isn't always optimal in terms of euclidean distance,
+	// but it's very close and a lot faster.
+	// http://cbloomrants.blogspot.com/2008/12/12-08-08-dxtc-summary.html
+
+	float c0Point = trunc((stops[1] + stops[3]) * 0.5f);
+	float halfPoint = trunc((stops[3] + stops[2]) * 0.5f);
+	float c3Point = trunc((stops[2] + stops[0]) * 0.5f);
+
+#ifndef BC1_DITHER
+	// the version without dithering is straightforward
+	for (uint i = 16u; i-- > 0u;) {
+		const float3 currColor = unpackUnorm4x8(srcPixelsBlock[i]).xyz * 255.0f;
+
+		const float dotValue = dot(currColor, dir);
+		mask <<= 2u;
+
+		if (dotValue < halfPoint)
+			mask |= ((dotValue < c0Point) ? 1u : 3u);
+		else
+			mask |= ((dotValue < c3Point) ? 2u : 0u);
+	}
+#else
+	// with floyd-steinberg dithering
+	float4 ep1 = float4(0, 0, 0, 0);
+	float4 ep2 = float4(0, 0, 0, 0);
+
+	c0Point *= 16.0f;
+	halfPoint *= 16.0f;
+	c3Point *= 16.0f;
+
+	for (uint y = 0u; y < 4u; ++y) {
+		float ditherDot;
+		uint lmask, step;
+
+		float3 currColor;
+		float dotValue;
+
+		currColor = unpackUnorm4x8(srcPixelsBlock[y * 4 + 0]).xyz * 255.0f;
+		dotValue = dot(currColor, dir);
+
+		ditherDot = (dotValue * 16.0f) + (3 * ep2[1] + 5 * ep2[0]);
+		if (ditherDot < halfPoint)
+			step = (ditherDot < c0Point) ? 1u : 3u;
+		else
+			step = (ditherDot < c3Point) ? 2u : 0u;
+		ep1[0] = dotValue - stops[step];
+		lmask = step;
+
+		currColor = unpackUnorm4x8(srcPixelsBlock[y * 4 + 1]).xyz * 255.0f;
+		dotValue = dot(currColor, dir);
+
+		ditherDot = (dotValue * 16.0f) + (7 * ep1[0] + 3 * ep2[2] + 5 * ep2[1] + ep2[0]);
+		if (ditherDot < halfPoint)
+			step = (ditherDot < c0Point) ? 1u : 3u;
+		else
+			step = (ditherDot < c3Point) ? 2u : 0u;
+		ep1[1] = dotValue - stops[step];
+		lmask |= step << 2u;
+
+		currColor = unpackUnorm4x8(srcPixelsBlock[y * 4 + 2]).xyz * 255.0f;
+		dotValue = dot(currColor, dir);
+
+		ditherDot = (dotValue * 16.0f) + (7 * ep1[1] + 3 * ep2[3] + 5 * ep2[2] + ep2[1]);
+		if (ditherDot < halfPoint)
+			step = (ditherDot < c0Point) ? 1u : 3u;
+		else
+			step = (ditherDot < c3Point) ? 2u : 0u;
+		ep1[2] = dotValue - stops[step];
+		lmask |= step << 4u;
+
+		currColor = unpackUnorm4x8(srcPixelsBlock[y * 4 + 2]).xyz * 255.0f;
+		dotValue = dot(currColor, dir);
+
+		ditherDot = (dotValue * 16.0f) + (7 * ep1[2] + 5 * ep2[3] + ep2[2]);
+		if (ditherDot < halfPoint)
+			step = (ditherDot < c0Point) ? 1u : 3u;
+		else
+			step = (ditherDot < c3Point) ? 2u : 0u;
+		ep1[3] = dotValue - stops[step];
+		lmask |= step << 6u;
+
+		mask |= lmask << (y * 8u);
+		{
+			float4 tmp = ep1;
+			ep1 = ep2;
+			ep2 = tmp;
+		} // swap
+	}
+#endif
+
+	return mask;
+}
+
+// The refinement function. (Clever code, part 2)
+// Tries to optimize colors to suit block contents better.
+// (By solving a least squares system via normal equations+Cramer's rule)
+bool RefineBlock(const uint srcPixelsBlock[16], uint mask, inout float inOutMinEndp16,
+		inout float inOutMaxEndp16) {
+	float newMin16, newMax16;
+	const float oldMin = inOutMinEndp16;
+	const float oldMax = inOutMaxEndp16;
+
+	if ((mask ^ (mask << 2u)) < 4u) // all pixels have the same index?
+	{
+		// yes, linear system would be singular; solve using optimal
+		// single-color match on average color
+		float3 rgbVal = float3(8.0f / 255.0f, 8.0f / 255.0f, 8.0f / 255.0f);
+		for (int i = 0; i < 16; ++i)
+			rgbVal += unpackUnorm4x8(srcPixelsBlock[i]).xyz;
+
+		rgbVal = floor(rgbVal * (255.0f / 16.0f));
+
+		newMax16 = c_oMatch5[uint(rgbVal.r)][0] * 2048.0f + //
+				c_oMatch6[uint(rgbVal.g)][0] * 32.0f + //
+				c_oMatch5[uint(rgbVal.b)][0];
+		newMin16 = c_oMatch5[uint(rgbVal.r)][1] * 2048.0f + //
+				c_oMatch6[uint(rgbVal.g)][1] * 32.0f + //
+				c_oMatch5[uint(rgbVal.b)][1];
+	} else {
+		const float w1Tab[4] = { 3, 0, 2, 1 };
+		const float prods[4] = { 589824.0f, 2304.0f, 262402.0f, 66562.0f };
+		// ^some magic to save a lot of multiplies in the accumulating loop...
+		// (precomputed products of weights for least squares system, accumulated inside one 32-bit
+		// register)
+
+		float akku = 0.0f;
+		uint cm = mask;
+		float3 at1 = float3(0, 0, 0);
+		float3 at2 = float3(0, 0, 0);
+		for (int i = 0; i < 16; ++i, cm >>= 2u) {
+			const float3 currColor = unpackUnorm4x8(srcPixelsBlock[i]).xyz * 255.0f;
+
+			const uint step = cm & 3u;
+			const float w1 = w1Tab[step];
+			akku += prods[step];
+			at1 += currColor * w1;
+			at2 += currColor;
+		}
+
+		at2 = 3.0f * at2 - at1;
+
+		// extract solutions and decide solvability
+		const float xx = floor(akku / 65535.0f);
+		const float yy = floor(mod(akku, 65535.0f) / 256.0f);
+		const float xy = mod(akku, 256.0f);
+
+		float2 f_rb_g;
+		f_rb_g.x = 3.0f * 31.0f / 255.0f / (xx * yy - xy * xy);
+		f_rb_g.y = f_rb_g.x * 63.0f / 31.0f;
+
+		// solve.
+		const float3 newMaxVal = clamp(floor((at1 * yy - at2 * xy) * f_rb_g.xyx + 0.5f),
+				float3(0.0f, 0.0f, 0.0f), float3(31, 63, 31));
+		newMax16 = newMaxVal.x * 2048.0f + newMaxVal.y * 32.0f + newMaxVal.z;
+
+		const float3 newMinVal = clamp(floor((at2 * xx - at1 * xy) * f_rb_g.xyx + 0.5f),
+				float3(0.0f, 0.0f, 0.0f), float3(31, 63, 31));
+		newMin16 = newMinVal.x * 2048.0f + newMinVal.y * 32.0f + newMinVal.z;
+	}
+
+	inOutMinEndp16 = newMin16;
+	inOutMaxEndp16 = newMax16;
+
+	return oldMin != newMin16 || oldMax != newMax16;
+}
+
+#ifdef BC1_DITHER
+/// Quantizes 'srcValue' which is originally in 888 (full range),
+/// converting it to 565 and then back to 888 (quantized)
+float3 quant(float3 srcValue) {
+	srcValue = clamp(srcValue, 0.0f, 255.0f);
+	// Convert 888 -> 565
+	srcValue = floor(srcValue * float3(31.0f / 255.0f, 63.0f / 255.0f, 31.0f / 255.0f) + 0.5f);
+	// Convert 565 -> 888 back
+	srcValue = floor(srcValue * float3(8.25f, 4.0625f, 8.25f));
+
+	return srcValue;
+}
+
+void DitherBlock(const uint srcPixBlck[16], out uint dthPixBlck[16]) {
+	float3 ep1[4] = { float3(0, 0, 0), float3(0, 0, 0), float3(0, 0, 0), float3(0, 0, 0) };
+	float3 ep2[4] = { float3(0, 0, 0), float3(0, 0, 0), float3(0, 0, 0), float3(0, 0, 0) };
+
+	for (uint y = 0u; y < 16u; y += 4u) {
+		float3 srcPixel, dithPixel;
+
+		srcPixel = unpackUnorm4x8(srcPixBlck[y + 0u]).xyz * 255.0f;
+		dithPixel = quant(srcPixel + trunc((3 * ep2[1] + 5 * ep2[0]) * (1.0f / 16.0f)));
+		ep1[0] = srcPixel - dithPixel;
+		dthPixBlck[y + 0u] = packUnorm4x8(float4(dithPixel * (1.0f / 255.0f), 1.0f));
+
+		srcPixel = unpackUnorm4x8(srcPixBlck[y + 1u]).xyz * 255.0f;
+		dithPixel = quant(
+				srcPixel + trunc((7 * ep1[0] + 3 * ep2[2] + 5 * ep2[1] + ep2[0]) * (1.0f / 16.0f)));
+		ep1[1] = srcPixel - dithPixel;
+		dthPixBlck[y + 1u] = packUnorm4x8(float4(dithPixel * (1.0f / 255.0f), 1.0f));
+
+		srcPixel = unpackUnorm4x8(srcPixBlck[y + 2u]).xyz * 255.0f;
+		dithPixel = quant(
+				srcPixel + trunc((7 * ep1[1] + 3 * ep2[3] + 5 * ep2[2] + ep2[1]) * (1.0f / 16.0f)));
+		ep1[2] = srcPixel - dithPixel;
+		dthPixBlck[y + 2u] = packUnorm4x8(float4(dithPixel * (1.0f / 255.0f), 1.0f));
+
+		srcPixel = unpackUnorm4x8(srcPixBlck[y + 3u]).xyz * 255.0f;
+		dithPixel = quant(srcPixel + trunc((7 * ep1[2] + 5 * ep2[3] + ep2[2]) * (1.0f / 16.0f)));
+		ep1[3] = srcPixel - dithPixel;
+		dthPixBlck[y + 3u] = packUnorm4x8(float4(dithPixel * (1.0f / 255.0f), 1.0f));
+
+		// swap( ep1, ep2 )
+		for (uint i = 0u; i < 4u; ++i) {
+			float3 tmp = ep1[i];
+			ep1[i] = ep2[i];
+			ep2[i] = tmp;
+		}
+	}
+}
+#endif
+
+void main() {
+	uint srcPixelsBlock[16];
+
+	bool bAllColorsEqual = true;
+
+	// Load the whole 4x4 block
+	const uint2 pixelsToLoadBase = gl_GlobalInvocationID.xy << 2u;
+	for (uint i = 0u; i < 16u; ++i) {
+		const uint2 pixelsToLoad = pixelsToLoadBase + uint2(i & 0x03u, i >> 2u);
+		const float3 srcPixels0 = OGRE_Load2D(srcTex, int2(pixelsToLoad), 0).xyz;
+		srcPixelsBlock[i] = packUnorm4x8(float4(srcPixels0, 1.0f));
+		bAllColorsEqual = bAllColorsEqual && srcPixelsBlock[0] == srcPixelsBlock[i];
+	}
+
+	float maxEndp16, minEndp16;
+	uint mask = 0u;
+
+	if (bAllColorsEqual) {
+		const uint3 rgbVal = uint3(unpackUnorm4x8(srcPixelsBlock[0]).xyz * 255.0f);
+		mask = 0xAAAAAAAAu;
+		maxEndp16 =
+				c_oMatch5[rgbVal.r][0] * 2048.0f + c_oMatch6[rgbVal.g][0] * 32.0f + c_oMatch5[rgbVal.b][0];
+		minEndp16 =
+				c_oMatch5[rgbVal.r][1] * 2048.0f + c_oMatch6[rgbVal.g][1] * 32.0f + c_oMatch5[rgbVal.b][1];
+	} else {
+#ifdef BC1_DITHER
+		uint ditherPixelsBlock[16];
+		// first step: compute dithered version for PCA if desired
+		DitherBlock(srcPixelsBlock, ditherPixelsBlock);
+#else
+#define ditherPixelsBlock srcPixelsBlock
+#endif
+
+		// second step: pca+map along principal axis
+		OptimizeColorsBlock(ditherPixelsBlock, minEndp16, maxEndp16);
+		if (minEndp16 != maxEndp16) {
+			float3 colors[4];
+			EvalColors(colors, maxEndp16, minEndp16); // Note min/max are inverted
+			mask = MatchColorsBlock(srcPixelsBlock, colors);
+		}
+
+		// third step: refine (multiple times if requested)
+		bool bStopRefinement = false;
+		for (uint i = 0u; i < params.p_numRefinements && !bStopRefinement; ++i) {
+			const uint lastMask = mask;
+
+			if (RefineBlock(ditherPixelsBlock, mask, minEndp16, maxEndp16)) {
+				if (minEndp16 != maxEndp16) {
+					float3 colors[4];
+					EvalColors(colors, maxEndp16, minEndp16); // Note min/max are inverted
+					mask = MatchColorsBlock(srcPixelsBlock, colors);
+				} else {
+					mask = 0u;
+					bStopRefinement = true;
+				}
+			}
+
+			bStopRefinement = mask == lastMask || bStopRefinement;
+		}
+	}
+
+	// write the color block
+	if (maxEndp16 < minEndp16) {
+		const float tmpValue = minEndp16;
+		minEndp16 = maxEndp16;
+		maxEndp16 = tmpValue;
+		mask ^= 0x55555555u;
+	}
+
+	uint2 outputBytes;
+	outputBytes.x = uint(maxEndp16) | (uint(minEndp16) << 16u);
+	outputBytes.y = mask;
+
+	uint2 dstUV = gl_GlobalInvocationID.xy;
+	imageStore(dstTexture, int2(dstUV), uint4(outputBytes.xy, 0u, 0u));
+}

modules/betsy/image_compress_betsy.cpp

@@ -30,6 +30,7 @@
 
 #include "image_compress_betsy.h"
 
+#include "core/config/project_settings.h"
 #include "servers/rendering/rendering_device_binds.h"
 #include "servers/rendering/rendering_server_default.h"
 
@@ -40,29 +41,68 @@
 #include "drivers/metal/rendering_context_driver_metal.h"
 #endif
 
+#include "betsy_bc1.h"
+
+#include "bc1.glsl.gen.h"
 #include "bc6h.glsl.gen.h"
 
-struct BC6PushConstant {
-	float sizeX;
-	float sizeY;
-	uint32_t padding[2];
-};
+// Static variables (for caching).
+
+static RenderingDevice *compress_rd = nullptr;
+static RenderingContextDriver *compress_rcd = nullptr;
+
+static Mutex rd_mutex;
+static Mutex shader_mutex;
+
+static HashMap<String, Ref<BetsyShader>> cached_shaders;
+
+// Betsy shader (for caching).
+
+BetsyShader::BetsyShader() {
+}
+
+BetsyShader::~BetsyShader() {
+	// Free just the shader, the pipelines will be cleared automatically.
+	if (compress_rd && compiled.is_valid()) {
+		compress_rd->free(compiled);
+	}
+}
+
+// Helper functions.
 
 static int get_next_multiple(int n, int m) {
 	return n + (m - (n % m));
 }
 
-Error _compress_betsy(BetsyFormat p_format, Image *r_img) {
+static String get_shader_name(BetsyFormat p_format) {
+	switch (p_format) {
+		case BETSY_FORMAT_BC1:
+		case BETSY_FORMAT_BC1_DITHER:
+			return "BC1";
+
+		case BETSY_FORMAT_BC3:
+			return "BC3";
+
+		case BETSY_FORMAT_BC6_SIGNED:
+		case BETSY_FORMAT_BC6_UNSIGNED:
+			return "BC6";
+
+		default:
+			return "";
+	}
+}
+
+Error compress_betsy(BetsyFormat p_format, Image *r_img) {
 	uint64_t start_time = OS::get_singleton()->get_ticks_msec();
 
 	if (r_img->is_compressed()) {
 		return ERR_INVALID_DATA;
 	}
 
-	ERR_FAIL_COND_V_MSG(r_img->get_format() < Image::FORMAT_RF || r_img->get_format() > Image::FORMAT_RGBE9995, ERR_INVALID_DATA, "Image is not an HDR image.");
-
 	Error err = OK;
 
+	rd_mutex.lock();
+	if (!compress_rd) {
 		// Create local RD.
 		RenderingContextDriver *rcd = nullptr;
 		RenderingDevice *rd = RenderingServer::get_singleton()->create_local_rendering_device();
@@ -97,27 +137,71 @@ Error _compress_betsy(BetsyFormat p_format, Image *r_img) {
 
 		ERR_FAIL_NULL_V_MSG(rd, err, "Unable to create a local RenderingDevice.");
 
-	Ref<RDShaderFile> compute_shader;
-	compute_shader.instantiate();
+		compress_rd = rd;
+		compress_rcd = rcd;
+	}
+	rd_mutex.unlock();
 
 	// Destination format.
 	Image::Format dest_format = Image::FORMAT_MAX;
+	RD::DataFormat dst_rd_format = RD::DATA_FORMAT_MAX;
 
 	String version = "";
 
 	switch (p_format) {
-		case BETSY_FORMAT_BC6: {
-			err = compute_shader->parse_versions_from_text(bc6h_shader_glsl);
+		case BETSY_FORMAT_BC1:
+			version = "standard";
+			dst_rd_format = RD::DATA_FORMAT_R32G32_UINT;
+			dest_format = Image::FORMAT_DXT1;
+			break;
 
-			if (r_img->detect_signed(true)) {
-				dest_format = Image::FORMAT_BPTC_RGBF;
+		case BETSY_FORMAT_BC1_DITHER:
+			version = "dithered";
+			dst_rd_format = RD::DATA_FORMAT_R32G32_UINT;
+			dest_format = Image::FORMAT_DXT1;
+			break;
+
+		case BETSY_FORMAT_BC6_SIGNED:
 			version = "signed";
-			} else {
-				dest_format = Image::FORMAT_BPTC_RGBFU;
+			dst_rd_format = RD::DATA_FORMAT_R32G32B32A32_UINT;
+			dest_format = Image::FORMAT_BPTC_RGBF;
+			break;
+
+		case BETSY_FORMAT_BC6_UNSIGNED:
 			version = "unsigned";
+			dst_rd_format = RD::DATA_FORMAT_R32G32B32A32_UINT;
+			dest_format = Image::FORMAT_BPTC_RGBFU;
+			break;
+
+		default:
+			err = ERR_INVALID_PARAMETER;
+			break;
 	}
 
-		} break;
+	const String shader_name = get_shader_name(p_format) + "-" + version;
+	const BetsyShader *shader_ptr;
+
+	shader_mutex.lock();
+	if (cached_shaders.has(shader_name)) {
+		shader_ptr = cached_shaders[shader_name].ptr();
+
+	} else {
+		Ref<BetsyShader> shader;
+		shader.instantiate();
+
+		Ref<RDShaderFile> source;
+		source.instantiate();
+
+		switch (p_format) {
+			case BETSY_FORMAT_BC1:
+			case BETSY_FORMAT_BC1_DITHER:
+				err = source->parse_versions_from_text(bc1_shader_glsl);
+				break;
+
+			case BETSY_FORMAT_BC6_UNSIGNED:
+			case BETSY_FORMAT_BC6_SIGNED:
+				err = source->parse_versions_from_text(bc6h_shader_glsl);
+				break;
 
 			default:
 				err = ERR_INVALID_PARAMETER;
@@ -125,28 +209,30 @@ Error _compress_betsy(BetsyFormat p_format, Image *r_img) {
 		}
 
 		if (err != OK) {
-		compute_shader->print_errors("Betsy compress shader");
-		memdelete(rd);
-		if (rcd != nullptr) {
-			memdelete(rcd);
-		}
-
+			source->print_errors("Betsy compress shader");
 			return err;
 		}
 
 		// Compile the shader, return early if invalid.
-	RID shader = rd->shader_create_from_spirv(compute_shader->get_spirv_stages(version));
-
-	if (shader.is_null()) {
-		memdelete(rd);
-		if (rcd != nullptr) {
-			memdelete(rcd);
+		shader->compiled = compress_rd->shader_create_from_spirv(source->get_spirv_stages(version));
+		if (shader->compiled.is_null()) {
+			return ERR_CANT_CREATE;
 		}
 
-		return err;
+		// Compile the pipeline, return early if invalid.
+		shader->pipeline = compress_rd->compute_pipeline_create(shader->compiled);
+		if (shader->pipeline.is_null()) {
+			return ERR_CANT_CREATE;
 		}
 
-	RID pipeline = rd->compute_pipeline_create(shader);
+		cached_shaders[shader_name] = shader;
+		shader_ptr = cached_shaders[shader_name].ptr();
+	}
+	shader_mutex.unlock();
+
+	if (shader_ptr->compiled.is_null() || shader_ptr->pipeline.is_null()) {
+		return ERR_INVALID_DATA;
+	}
 
 	// src_texture format information.
 	RD::TextureFormat src_texture_format;
@@ -159,6 +245,33 @@ Error _compress_betsy(BetsyFormat p_format, Image *r_img) {
 	}
 
 	switch (r_img->get_format()) {
+		case Image::FORMAT_L8:
+			r_img->convert(Image::FORMAT_RGBA8);
+			src_texture_format.format = RD::DATA_FORMAT_R8G8B8A8_UNORM;
+			break;
+
+		case Image::FORMAT_LA8:
+			r_img->convert(Image::FORMAT_RGBA8);
+			src_texture_format.format = RD::DATA_FORMAT_R8G8B8A8_UNORM;
+			break;
+
+		case Image::FORMAT_R8:
+			src_texture_format.format = RD::DATA_FORMAT_R8_UNORM;
+			break;
+
+		case Image::FORMAT_RG8:
+			src_texture_format.format = RD::DATA_FORMAT_R8G8_UNORM;
+			break;
+
+		case Image::FORMAT_RGB8:
+			r_img->convert(Image::FORMAT_RGBA8);
+			src_texture_format.format = RD::DATA_FORMAT_R8G8B8A8_UNORM;
+			break;
+
+		case Image::FORMAT_RGBA8:
+			src_texture_format.format = RD::DATA_FORMAT_R8G8B8A8_UNORM;
+			break;
+
 		case Image::FORMAT_RH:
 			src_texture_format.format = RD::DATA_FORMAT_R16_SFLOAT;
 			break;
@@ -198,13 +311,6 @@ Error _compress_betsy(BetsyFormat p_format, Image *r_img) {
 			break;
 
 		default: {
-			rd->free(shader);
-
-			memdelete(rd);
-			if (rcd != nullptr) {
-				memdelete(rcd);
-			}
-
 			return err;
 		}
 	}
@@ -219,12 +325,25 @@ Error _compress_betsy(BetsyFormat p_format, Image *r_img) {
 		src_sampler_state.mip_filter = RD::SAMPLER_FILTER_NEAREST;
 	}
 
-	RID src_sampler = rd->sampler_create(src_sampler_state);
+	RID src_sampler = compress_rd->sampler_create(src_sampler_state);
 
 	// For the destination format just copy the source format and change the usage bits.
 	RD::TextureFormat dst_texture_format = src_texture_format;
 	dst_texture_format.usage_bits = RD::TEXTURE_USAGE_COLOR_ATTACHMENT_BIT | RD::TEXTURE_USAGE_STORAGE_BIT | RD::TEXTURE_USAGE_CAN_COPY_FROM_BIT | RD::TEXTURE_USAGE_CAN_COPY_TO_BIT | RD::TEXTURE_USAGE_CAN_UPDATE_BIT;
-	dst_texture_format.format = RD::DATA_FORMAT_R32G32B32A32_UINT;
+	dst_texture_format.format = dst_rd_format;
+
+	RID encoding_table_buffer;
+	bool uses_encoding_table = false;
+
+	// Encoding table setup.
+	if (dest_format == Image::FORMAT_DXT1) {
+		Vector<uint8_t> data;
+		data.resize(1024 * 4);
+		memcpy(data.ptrw(), dxt1_encoding_table, 1024 * 4);
+
+		encoding_table_buffer = compress_rd->storage_buffer_create(1024 * 4, data);
+		uses_encoding_table = true;
+	}
 
 	const int mip_count = r_img->get_mipmap_count() + 1;
 
@@ -256,15 +375,8 @@ Error _compress_betsy(BetsyFormat p_format, Image *r_img) {
 		memcpy(src_image_ptr[0].ptrw(), r_img->ptr() + ofs, size);
 
 		// Create the textures on the GPU.
-		RID src_texture = rd->texture_create(src_texture_format, RD::TextureView(), src_images);
-		RID dst_texture = rd->texture_create(dst_texture_format, RD::TextureView());
-
-		if (dest_format == Image::FORMAT_BPTC_RGBFU || dest_format == Image::FORMAT_BPTC_RGBF) {
-			BC6PushConstant push_constant;
-			push_constant.sizeX = 1.0f / width;
-			push_constant.sizeY = 1.0f / height;
-			push_constant.padding[0] = 0;
-			push_constant.padding[1] = 0;
+		RID src_texture = compress_rd->texture_create(src_texture_format, RD::TextureView(), src_images);
+		RID dst_texture = compress_rd->texture_create(dst_texture_format, RD::TextureView());
 
 		Vector<RD::Uniform> uniforms;
 		{
@@ -283,30 +395,56 @@ Error _compress_betsy(BetsyFormat p_format, Image *r_img) {
 				u.append_id(dst_texture);
 				uniforms.push_back(u);
 			}
+
+			if (uses_encoding_table) {
+				RD::Uniform u;
+				u.uniform_type = RD::UNIFORM_TYPE_STORAGE_BUFFER;
+				u.binding = 2;
+				u.append_id(encoding_table_buffer);
+				uniforms.push_back(u);
+			}
 		}
 
-			RID uniform_set = rd->uniform_set_create(uniforms, shader, 0);
-			RD::ComputeListID compute_list = rd->compute_list_begin();
+		RID uniform_set = compress_rd->uniform_set_create(uniforms, shader_ptr->compiled, 0);
+		RD::ComputeListID compute_list = compress_rd->compute_list_begin();
 
-			rd->compute_list_bind_compute_pipeline(compute_list, pipeline);
-			rd->compute_list_bind_uniform_set(compute_list, uniform_set, 0);
-			rd->compute_list_set_push_constant(compute_list, &push_constant, sizeof(BC6PushConstant));
-			rd->compute_list_dispatch(compute_list, get_next_multiple(width, 32) / 32, get_next_multiple(height, 32) / 32, 1);
-			rd->compute_list_end();
+		compress_rd->compute_list_bind_compute_pipeline(compute_list, shader_ptr->pipeline);
+		compress_rd->compute_list_bind_uniform_set(compute_list, uniform_set, 0);
+
+		if (dest_format == Image::FORMAT_BPTC_RGBFU || dest_format == Image::FORMAT_BPTC_RGBF) {
+			BC6PushConstant push_constant;
+			push_constant.sizeX = 1.0f / width;
+			push_constant.sizeY = 1.0f / height;
+			push_constant.padding[0] = 0;
+			push_constant.padding[1] = 0;
+
+			compress_rd->compute_list_set_push_constant(compute_list, &push_constant, sizeof(BC6PushConstant));
+
+		} else {
+			BC1PushConstant push_constant;
+			push_constant.num_refines = 2;
+			push_constant.padding[0] = 0;
+			push_constant.padding[1] = 0;
+			push_constant.padding[2] = 0;
+
+			compress_rd->compute_list_set_push_constant(compute_list, &push_constant, sizeof(BC1PushConstant));
 		}
 
-		rd->submit();
-		rd->sync();
+		compress_rd->compute_list_dispatch(compute_list, get_next_multiple(width, 32) / 32, get_next_multiple(height, 32) / 32, 1);
+		compress_rd->compute_list_end();
+
+		compress_rd->submit();
+		compress_rd->sync();
 
 		// Copy data from the GPU to the buffer.
-		const Vector<uint8_t> texture_data = rd->texture_get_data(dst_texture, 0);
+		const Vector<uint8_t> texture_data = compress_rd->texture_get_data(dst_texture, 0);
 		int64_t dst_ofs = Image::get_image_mipmap_offset(r_img->get_width(), r_img->get_height(), dest_format, i);
 
 		memcpy(dst_data_ptr + dst_ofs, texture_data.ptr(), texture_data.size());
 
 		// Free the source and dest texture.
-		rd->free(dst_texture);
-		rd->free(src_texture);
+		compress_rd->free(dst_texture);
+		compress_rd->free(src_texture);
 	}
 
 	src_images.clear();
@@ -315,14 +453,11 @@ Error _compress_betsy(BetsyFormat p_format, Image *r_img) {
 	r_img->set_data(r_img->get_width(), r_img->get_height(), r_img->has_mipmaps(), dest_format, dst_data);
 
 	// Free the shader (dependencies will be cleared automatically).
-	rd->free(src_sampler);
-	rd->free(shader);
-
-	memdelete(rd);
-	if (rcd != nullptr) {
-		memdelete(rcd);
+	if (uses_encoding_table) {
+		compress_rd->free(encoding_table_buffer);
 	}
 
+	compress_rd->free(src_sampler);
 	print_verbose(vformat("Betsy: Encoding took %d ms.", OS::get_singleton()->get_ticks_msec() - start_time));
 
 	return OK;
@@ -330,10 +465,61 @@ Error _compress_betsy(BetsyFormat p_format, Image *r_img) {
 
 Error _betsy_compress_bptc(Image *r_img, Image::UsedChannels p_channels) {
 	Image::Format format = r_img->get_format();
+	Error result = ERR_UNAVAILABLE;
 
 	if (format >= Image::FORMAT_RF && format <= Image::FORMAT_RGBE9995) {
-		return _compress_betsy(BETSY_FORMAT_BC6, r_img);
+		if (r_img->detect_signed()) {
+			result = compress_betsy(BETSY_FORMAT_BC6_SIGNED, r_img);
+		} else {
+			result = compress_betsy(BETSY_FORMAT_BC6_UNSIGNED, r_img);
+		}
 	}
 
-	return ERR_UNAVAILABLE;
+	if (!GLOBAL_GET("rendering/textures/vram_compression/cache_gpu_compressor")) {
+		free_device();
+	}
+
+	return result;
+}
+
+Error _betsy_compress_s3tc(Image *r_img, Image::UsedChannels p_channels) {
+	Error result = ERR_UNAVAILABLE;
+
+	switch (p_channels) {
+		case Image::USED_CHANNELS_RGB:
+			result = compress_betsy(BETSY_FORMAT_BC1_DITHER, r_img);
+			break;
+
+		case Image::USED_CHANNELS_L:
+			result = compress_betsy(BETSY_FORMAT_BC1, r_img);
+			break;
+
+		default:
+			break;
+	}
+
+	if (!GLOBAL_GET("rendering/textures/vram_compression/cache_gpu_compressor")) {
+		free_device();
+	}
+
+	return result;
+}
+
+void free_device() {
+	if (compress_rd != nullptr) {
+		// Clear the shader cache, shaders will be unreferenced automatically.
+		shader_mutex.lock();
+		cached_shaders.clear();
+		shader_mutex.unlock();
+
+		// Free the RD (and RCD if necessary).
+		rd_mutex.lock();
+		memdelete(compress_rd);
+		compress_rd = nullptr;
+		if (compress_rcd != nullptr) {
+			memdelete(compress_rcd);
+			compress_rcd = nullptr;
+		}
+		rd_mutex.unlock();
+	}
 }

modules/betsy/image_compress_betsy.h

@@ -34,11 +34,38 @@
 #include "core/io/image.h"
 
 enum BetsyFormat {
-	BETSY_FORMAT_BC6,
+	BETSY_FORMAT_BC1,
+	BETSY_FORMAT_BC1_DITHER,
+	BETSY_FORMAT_BC3,
+	BETSY_FORMAT_BC6_SIGNED,
+	BETSY_FORMAT_BC6_UNSIGNED,
 };
 
-Error _compress_betsy(BetsyFormat p_format, Image *r_img);
+class BetsyShader : public RefCounted {
+public:
+	RID compiled;
+	RID pipeline;
+
+	BetsyShader();
+	~BetsyShader();
+};
+
+struct BC6PushConstant {
+	float sizeX;
+	float sizeY;
+	uint32_t padding[2];
+};
+
+struct BC1PushConstant {
+	uint32_t num_refines;
+	uint32_t padding[3];
+};
+
+void free_device();
+
+Error compress_betsy(BetsyFormat p_format, Image *r_img);
 
 Error _betsy_compress_bptc(Image *r_img, Image::UsedChannels p_channels);
+Error _betsy_compress_s3tc(Image *r_img, Image::UsedChannels p_channels);
 
 #endif // IMAGE_COMPRESS_BETSY_H

modules/betsy/register_types.cpp

@@ -38,10 +38,13 @@ void initialize_betsy_module(ModuleInitializationLevel p_level) {
 	}
 
 	Image::_image_compress_bptc_rd_func = _betsy_compress_bptc;
+	Image::_image_compress_bc_rd_func = _betsy_compress_s3tc;
 }
 
 void uninitialize_betsy_module(ModuleInitializationLevel p_level) {
 	if (p_level != MODULE_INITIALIZATION_LEVEL_SCENE) {
 		return;
 	}
+
+	free_device();
 }

servers/rendering_server.cpp

@@ -3528,6 +3528,7 @@ void RenderingServer::init() {
 	GLOBAL_DEF_RST("rendering/textures/vram_compression/import_s3tc_bptc", false);
 	GLOBAL_DEF_RST("rendering/textures/vram_compression/import_etc2_astc", false);
 	GLOBAL_DEF("rendering/textures/vram_compression/compress_with_gpu", true);
+	GLOBAL_DEF("rendering/textures/vram_compression/cache_gpu_compressor", true);
 
 	GLOBAL_DEF("rendering/textures/lossless_compression/force_png", false);
 

thirdparty/README.md

@@ -78,7 +78,7 @@ fix build with our own copy of zstd (patch in `patches`).
 
 Files extracted from upstream source:
 
-- `bc6h.glsl`, `CrossPlatformSettings_piece_all.glsl` and `UavCrossPlatform_piece_all.glsl`.
+- `bc6h.glsl`, `bc1.glsl`, `CrossPlatformSettings_piece_all.glsl` and `UavCrossPlatform_piece_all.glsl`.
 - `LICENSE.md`