Betsy: Implement BC4 compression

Author: BlueCube3310

core/io/image.cpp

@@ -2751,7 +2751,7 @@ Error Image::compress_from_channels(CompressMode p_mode, UsedChannels p_channels
 
 			case COMPRESS_S3TC: {
 				// BC3 is unsupported currently.
-				if ((p_channels == USED_CHANNELS_RGB || p_channels == USED_CHANNELS_L) && _image_compress_bc_rd_func) {
+				if ((p_channels == USED_CHANNELS_R || 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.

doc/classes/ProjectSettings.xml

@@ -2903,8 +2903,8 @@
 		</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, 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 BC1 and BC6H compression, which are used on Desktop and Console for fully opaque and HDR images respectively.
+			[b]Note:[/b] This only functions on a device which supports either Vulkan, D3D12, or Metal available as a rendering backend.
+			[b]Note:[/b] Currently this only affects certain compressed formats (BC1, BC4, and BC6), all of which are exclusive to desktop platforms and consoles.
 		</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

@@ -5,6 +5,7 @@ Import("env_modules")
 env_betsy = env_modules.Clone()
 env_betsy.GLSL_HEADER("bc6h.glsl")
 env_betsy.GLSL_HEADER("bc1.glsl")
+env_betsy.GLSL_HEADER("bc4.glsl")
 env_betsy.Depends(Glob("*.glsl.gen.h"), ["#glsl_builders.py"])
 
 # Thirdparty source files

modules/betsy/bc4.glsl

@@ -0,0 +1,151 @@
+#[versions]
+
+unsigned = "";
+signed = "#define SNORM";
+
+#[compute]
+#version 450
+
+#include "CrossPlatformSettings_piece_all.glsl"
+#include "UavCrossPlatform_piece_all.glsl"
+
+#VERSION_DEFINES
+
+shared float2 g_minMaxValues[4u * 4u * 4u];
+shared uint2 g_mask[4u * 4u];
+
+layout(binding = 0) uniform sampler2D srcTex;
+layout(binding = 1, rg32ui) uniform restrict writeonly uimage2D dstTexture;
+
+layout(push_constant, std430) uniform Params {
+	uint p_channelIdx;
+	uint p_padding[3];
+}
+params;
+
+layout(local_size_x = 4, //
+		local_size_y = 4, //
+		local_size_z = 4) in;
+
+/// Each block is 16 pixels
+/// Each thread works on 4 pixels
+/// Therefore each block needs 4 threads, generating 8 masks
+/// At the end these 8 masks get merged into 2 and results written to output
+///
+/// **Q: Why 4 pixels per thread? Why not 1 pixel per thread? Why not 2? Why not 16?**
+///
+/// A: It's a sweetspot.
+///  - Very short threads cannot fill expensive GPUs with enough work (dispatch bound)
+///  - Lots of threads means lots of synchronization (e.g. evaluating min/max, merging masks)
+///    overhead, and also more LDS usage which reduces occupancy.
+///  - Long threads (e.g. 1 thread per block) misses parallelism opportunities
+void main() {
+	float minVal, maxVal;
+	float4 srcPixel;
+
+	const uint blockThreadId = gl_LocalInvocationID.x;
+
+	const uint2 pixelsToLoadBase = gl_GlobalInvocationID.yz << 2u;
+
+	for (uint i = 0u; i < 4u; ++i) {
+		const uint2 pixelsToLoad = pixelsToLoadBase + uint2(i, blockThreadId);
+
+		const float4 value = OGRE_Load2D(srcTex, int2(pixelsToLoad), 0).xyzw;
+		srcPixel[i] = params.p_channelIdx == 0 ? value.x : (params.p_channelIdx == 1 ? value.y : value.w);
+		srcPixel[i] *= 255.0f;
+	}
+
+	minVal = min3(srcPixel.x, srcPixel.y, srcPixel.z);
+	maxVal = max3(srcPixel.x, srcPixel.y, srcPixel.z);
+	minVal = min(minVal, srcPixel.w);
+	maxVal = max(maxVal, srcPixel.w);
+
+	const uint minMaxIdxBase = (gl_LocalInvocationID.z << 4u) + (gl_LocalInvocationID.y << 2u);
+	const uint maskIdxBase = (gl_LocalInvocationID.z << 2u) + gl_LocalInvocationID.y;
+
+	g_minMaxValues[minMaxIdxBase + blockThreadId] = float2(minVal, maxVal);
+	g_mask[maskIdxBase] = uint2(0u, 0u);
+
+	memoryBarrierShared();
+	barrier();
+
+	// Have all 4 threads in the block grab the min/max value by comparing what all 4 threads uploaded
+	for (uint i = 0u; i < 4u; ++i) {
+		minVal = min(g_minMaxValues[minMaxIdxBase + i].x, minVal);
+		maxVal = max(g_minMaxValues[minMaxIdxBase + i].y, maxVal);
+	}
+
+	// determine bias and emit color indices
+	// given the choice of maxVal/minVal, these indices are optimal:
+	// http://fgiesen.wordpress.com/2009/12/15/dxt5-alpha-block-index-determination/
+	float dist = maxVal - minVal;
+	float dist4 = dist * 4.0f;
+	float dist2 = dist * 2.0f;
+	float bias = (dist < 8) ? (dist - 1) : (trunc(dist * 0.5f) + 2);
+	bias -= minVal * 7;
+
+	uint mask0 = 0u, mask1 = 0u;
+
+	for (uint i = 0u; i < 4u; ++i) {
+		float a = srcPixel[i] * 7.0f + bias;
+
+		int ind = 0;
+
+		// select index. this is a "linear scale" lerp factor between 0 (val=min) and 7 (val=max).
+		if (a >= dist4) {
+			ind = 4;
+			a -= dist4;
+		}
+
+		if (a >= dist2) {
+			ind += 2;
+			a -= dist2;
+		}
+
+		if (a >= dist)
+			ind += 1;
+
+		// turn linear scale into DXT index (0/1 are extremal pts)
+		ind = -ind & 7;
+		ind ^= (2 > ind) ? 1 : 0;
+
+		// write index
+		const uint bits = 16u + ((blockThreadId << 2u) + i) * 3u;
+		if (bits < 32u) {
+			mask0 |= uint(ind) << bits;
+			if (bits + 3u > 32u) {
+				mask1 |= uint(ind) >> (32u - bits);
+			}
+		} else {
+			mask1 |= uint(ind) << (bits - 32u);
+		}
+	}
+
+	if (mask0 != 0u)
+		atomicOr(g_mask[maskIdxBase].x, mask0);
+	if (mask1 != 0u)
+		atomicOr(g_mask[maskIdxBase].y, mask1);
+
+	memoryBarrierShared();
+	barrier();
+
+	if (blockThreadId == 0u) {
+		// Save data
+		uint2 outputBytes;
+
+#ifdef SNORM
+		outputBytes.x =
+				packSnorm4x8(float4(maxVal * (1.0f / 255.0f) * 2.0f - 1.0f,
+						minVal * (1.0f / 255.0f) * 2.0f - 1.0f, 0.0f, 0.0f));
+#else
+		outputBytes.x = packUnorm4x8(
+				float4(maxVal * (1.0f / 255.0f), minVal * (1.0f / 255.0f), 0.0f, 0.0f));
+#endif
+
+		outputBytes.x |= g_mask[maskIdxBase].x;
+		outputBytes.y = g_mask[maskIdxBase].y;
+
+		uint2 dstUV = gl_GlobalInvocationID.yz;
+		imageStore(dstTexture, int2(dstUV), uint4(outputBytes.xy, 0u, 0u));
+	}
+}

modules/betsy/image_compress_betsy.cpp

@@ -35,6 +35,7 @@
 #include "betsy_bc1.h"
 
 #include "bc1.glsl.gen.h"
+#include "bc4.glsl.gen.h"
 #include "bc6h.glsl.gen.h"
 
 static Mutex betsy_mutex;
@@ -165,6 +166,10 @@ static String get_shader_name(BetsyFormat p_format) {
 		case BETSY_FORMAT_BC3:
 			return "BC3";
 
+		case BETSY_FORMAT_BC4_SIGNED:
+		case BETSY_FORMAT_BC4_UNSIGNED:
+			return "BC4";
+
 		case BETSY_FORMAT_BC6_SIGNED:
 		case BETSY_FORMAT_BC6_UNSIGNED:
 			return "BC6";
@@ -202,6 +207,12 @@ Error BetsyCompressor::_compress(BetsyFormat p_format, Image *r_img) {
 			dest_format = Image::FORMAT_DXT1;
 			break;
 
+		case BETSY_FORMAT_BC4_UNSIGNED:
+			version = "unsigned";
+			dst_rd_format = RD::DATA_FORMAT_R32G32_UINT;
+			dest_format = Image::FORMAT_RGTC_R;
+			break;
+
 		case BETSY_FORMAT_BC6_SIGNED:
 			version = "signed";
 			dst_rd_format = RD::DATA_FORMAT_R32G32B32A32_UINT;
@@ -235,8 +246,13 @@ Error BetsyCompressor::_compress(BetsyFormat p_format, Image *r_img) {
 				err = source->parse_versions_from_text(bc1_shader_glsl);
 				break;
 
-			case BETSY_FORMAT_BC6_UNSIGNED:
+			case BETSY_FORMAT_BC4_SIGNED:
+			case BETSY_FORMAT_BC4_UNSIGNED:
+				err = source->parse_versions_from_text(bc4_shader_glsl);
+				break;
+
 			case BETSY_FORMAT_BC6_SIGNED:
+			case BETSY_FORMAT_BC6_UNSIGNED:
 				err = source->parse_versions_from_text(bc6h_shader_glsl);
 				break;
 
@@ -430,26 +446,45 @@ Error BetsyCompressor::_compress(BetsyFormat p_format, Image *r_img) {
 		compress_rd->compute_list_bind_compute_pipeline(compute_list, shader.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));
+		switch (dest_format) {
+			case Image::FORMAT_BPTC_RGBFU:
+			case 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));
+				compress_rd->compute_list_dispatch(compute_list, get_next_multiple(width, 32) / 32, get_next_multiple(height, 32) / 32, 1);
+			} break;
+
+			case Image::FORMAT_DXT1: {
+				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));
+				compress_rd->compute_list_dispatch(compute_list, get_next_multiple(width, 32) / 32, get_next_multiple(height, 32) / 32, 1);
+			} break;
+
+			case Image::FORMAT_RGTC_R: {
+				BC4PushConstant push_constant;
+				push_constant.channel_idx = 0;
+				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(BC4PushConstant));
+				compress_rd->compute_list_dispatch(compute_list, 1, get_next_multiple(width, 16) / 16, get_next_multiple(height, 16) / 16);
+			} break;
+
+			default: {
+			} break;
 		}
 
-		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();
@@ -511,13 +546,14 @@ Error _betsy_compress_s3tc(Image *r_img, Image::UsedChannels p_channels) {
 
 	switch (p_channels) {
 		case Image::USED_CHANNELS_RGB:
-			result = betsy->compress(BETSY_FORMAT_BC1_DITHER, r_img);
-			break;
-
 		case Image::USED_CHANNELS_L:
 			result = betsy->compress(BETSY_FORMAT_BC1, r_img);
 			break;
 
+		case Image::USED_CHANNELS_R:
+			result = betsy->compress(BETSY_FORMAT_BC4_UNSIGNED, r_img);
+			break;
+
 		default:
 			break;
 	}

modules/betsy/image_compress_betsy.h

@@ -50,6 +50,8 @@ enum BetsyFormat {
 	BETSY_FORMAT_BC1,
 	BETSY_FORMAT_BC1_DITHER,
 	BETSY_FORMAT_BC3,
+	BETSY_FORMAT_BC4_SIGNED,
+	BETSY_FORMAT_BC4_UNSIGNED,
 	BETSY_FORMAT_BC6_SIGNED,
 	BETSY_FORMAT_BC6_UNSIGNED,
 };
@@ -65,6 +67,11 @@ struct BC1PushConstant {
 	uint32_t padding[3];
 };
 
+struct BC4PushConstant {
+	uint32_t channel_idx;
+	uint32_t padding[3];
+};
+
 void free_device();
 
 Error _betsy_compress_bptc(Image *r_img, Image::UsedChannels p_channels);

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`, `bc1.glsl`, `CrossPlatformSettings_piece_all.glsl` and `UavCrossPlatform_piece_all.glsl`.
+- `bc6h.glsl`, `bc1.glsl`, `bc4.glsl`, `CrossPlatformSettings_piece_all.glsl` and `UavCrossPlatform_piece_all.glsl`.
 - `LICENSE.md`