// MinecraftPT — Bloom_CS
// Bloom compute passes: two downsample levels + axial blur X/Y.
// Uses two alternating images (bloomA/bloomB) to avoid read-write conflicts.

#include "/Lib/Settings.glsl"
#include "/Lib/Utilities.glsl"

const ivec3 workGroups = ivec3(int(ceil(viewWidth * 0.03125)), int(ceil(viewHeight * 0.03125)), 1);
layout (local_size_x = 8, local_size_y = 8) in;

layout (rgba16f) uniform writeonly image2D img_bloomA;
layout (rgba16f) uniform writeonly image2D img_bloomB;

uniform sampler2D colortex12;
uniform sampler2D bloomA;
uniform sampler2D bloomB;

// Downsample pass
#ifdef PROGRAM_BLOOM_DOWNSAMPLE
	void main(){
		ivec2 texel = ivec2(gl_GlobalInvocationID.xy);

		vec3 color = vec3(0.0);

		#if PROGRAM_BLOOM_DOWNSAMPLE_LEVEL == 1
			// Level 1: 2x2 box from the HDR scene -> bloomA
			for (int i = 0; i < 2; i++){
			for (int j = 0; j < 2; j++){
				color += texelFetch(colortex12, texel * 2 + ivec2(i, j), 0).rgb;
			}}
			color *= 0.25;

			float brightness = luminance(color);
			color *= saturate(brightness * BLOOM_CLAMP_STRENGTH);

			ivec2 imgSize = imageSize(img_bloomA);
			if (all(lessThan(texel, imgSize))){
				imageStore(img_bloomA, texel, vec4(color, 1.0));
			}

		#else
			// Level 2: 2x2 box from bloomA -> bloomB
			for (int i = 0; i < 2; i++){
			for (int j = 0; j < 2; j++){
				color += texelFetch(bloomA, texel * 2 + ivec2(i, j), 0).rgb;
			}}
			color *= 0.25;

			float brightness = luminance(color);
			color *= saturate(brightness * BLOOM_CLAMP_STRENGTH);

			ivec2 imgSize = imageSize(img_bloomB);
			if (all(lessThan(texel, imgSize))){
				imageStore(img_bloomB, texel, vec4(color, 1.0));
			}

		#endif
	}
#endif

// Axial blur pass (X or Y depending on defines)
#ifdef PROGRAM_BLOOM_AXIALBLUR
	void main(){
		ivec2 texel = ivec2(gl_GlobalInvocationID.xy);
		ivec2 imgSize = imageSize(img_bloomA);
		if (any(greaterThanEqual(texel, imgSize))) return;

		vec2 uv = (vec2(texel) + 0.5) / vec2(imgSize);

		// Which buffer to read (BLOOM_AXIAL_READ_A defined for the pass reading bloomA)
		#ifdef BLOOM_AXIAL_READ_A
			sampler2D src = bloomA;
		#else
			sampler2D src = bloomB;
		#endif

		#ifdef PROGRAM_BLOOM_AXIALBLUR_X
			vec2 axis = vec2(1.0, 0.0);
		#else
			vec2 axis = vec2(0.0, 1.0);
		#endif

		vec3 color = vec3(0.0);
		float weightSum = 0.0;

		for (int i = -8; i <= 8; i++){
			vec2 coord = uv + axis * float(i) / vec2(imgSize);
			float w = exp(-float(i * i) * 0.05);
			color += textureLod(src, coord, 0.0).rgb * w;
			weightSum += w;
		}

		// Write to the opposite buffer of the read source
		#ifdef BLOOM_AXIAL_READ_A
			imageStore(img_bloomB, texel, vec4(color / weightSum, 1.0));
		#else
			imageStore(img_bloomA, texel, vec4(color / weightSum, 1.0));
		#endif
	}
#endif