Vulkan Forward+: tile cull uses viewport/tile pixel size (match PBR)

- forward_plus_tile_cull.comp: sphere_tile_overlap uses dynamic tilePx
  from lights SSBO viewport and pc.tileGrid (no hard-coded 16px)
- renderer_regression_check: forbid 16px literals in compute; note alignment
- RENDERER_2026_ARCHITECTURE_PASS: document compute+fragment tile sizing
- Regenerate SPIR-V for compute shader

Co-authored-by: Tim Fox <timfox@outlook.com>

Author: cursoragent

docs/RENDERER_2026_ARCHITECTURE_PASS.md

@@ -154,7 +154,7 @@ Use **Vulkan as the primary renderer architecture**, freeze OpenGL as compatibil
 
 ## Phase 2: Lighting Scale
 
-- Add Vulkan light records plus cluster/tile culling. **Incremental (engine):** `r_forwardPlus 1` (default 0) allocates light + tile SSBOs, packs **at most `MAX_DLIGHTS` (32)** lights from `backEnd.refdef` (indices align with `tess.dlightBits`; excess lights are omitted and a **developer** log notes when the source count exceeds the cap), and runs a **compute tile cull** (`forward_plus_tile_cull.comp`, 16×16 tiles, up to **8** index slots per tile; active count **4–8** via latched **`r_forwardPlusMaxPerTile`**, default **8**) after `RB_BeginDrawingView` inside the main render pass. Tile grid and NDC→pixel use **`vk_get_render_target_width/height`** (FBO / `r_renderScale`); the tile SSBO is **reallocated when that resolution changes** (no `vid_restart` for resize alone). **PBR fragment:** descriptor set 18 binds light + tile + **param** SSBOs (clip matrix for shading). `r_forwardPlusDebug` (0–1) = debug overlay; **`r_forwardPlusShade`** (0–4, default 0) adds **experimental diffuse + microfacet spec** (per-light `CalcSpecular`, scaled) from tile-culled **point and linear/spot** lights. When `r_forwardPlus` is on, **`pbrForwardPlus.y`** carries **`floatBitsToUint(tess.dlightBits)`** so the shader **skips** any packed light index whose bit is set in `tess.dlightBits` (first **32** indices only; matches typical `MAX_DLIGHTS` range). Primary direct is still **softly renormalized** vs Forward+ energy. Works with deluxe/lightmap; toggling shade **invalidates cached pipelines** next frame.
+- Add Vulkan light records plus cluster/tile culling. **Incremental (engine):** `r_forwardPlus 1` (default 0) allocates light + tile SSBOs, packs **at most `MAX_DLIGHTS` (32)** lights from `backEnd.refdef` (indices align with `tess.dlightBits`; excess lights are omitted and a **developer** log notes when the source count exceeds the cap), and runs a **compute tile cull** (`forward_plus_tile_cull.comp`, **`VK_FP_TILE_DIM` (16)** px tiles via `ceil(viewport / dim)`, up to **8** index slots per tile; active count **4–8** via latched **`r_forwardPlusMaxPerTile`**, default **8**) after `RB_BeginDrawingView` inside the main render pass. **Compute + PBR fragment** derive **tile pixel size** from the packed **viewport ÷ tile grid** so cull, debug overlay, and experimental shade stay aligned if tile policy changes. Tile grid and NDC→pixel use **`vk_get_render_target_width/height`** (FBO / `r_renderScale`); the tile SSBO is **reallocated when that resolution changes** (no `vid_restart` for resize alone). **PBR fragment:** descriptor set 18 binds light + tile + **param** SSBOs (clip matrix for shading). `r_forwardPlusDebug` (0–1) = debug overlay; **`r_forwardPlusShade`** (0–4, default 0) adds **experimental diffuse + microfacet spec** (per-light `CalcSpecular`, scaled) from tile-culled **point and linear/spot** lights. When `r_forwardPlus` is on, **`pbrForwardPlus.y`** carries **`floatBitsToUint(tess.dlightBits)`** so the shader **skips** any packed light index whose bit is set in `tess.dlightBits` (first **32** indices only; matches typical `MAX_DLIGHTS` range). Primary direct is still **softly renormalized** vs Forward+ energy. Works with deluxe/lightmap; toggling shade **invalidates cached pipelines** next frame.
 - Introduce Forward+ shading for local lights.
 - Keep shadow budgets conservative and explicit.
 

scripts/renderer_regression_check.sh

@@ -219,14 +219,14 @@ else
 fi
 
 echo ""
-echo "Forward+ tile pixel size: VK_FP_TILE_DIM vs compute shader literals:"
-fp_dim="$(sed -n 's/^#define VK_FP_TILE_DIM[[:space:]]*\([0-9][0-9]*\)u*$/\1/p' "$FP_C" | head -1)"
-if [[ -z "$fp_dim" ]]; then
-  fail "could not parse VK_FP_TILE_DIM from vk_forward_plus.c"
-elif ! grep -q "tileX \* ${fp_dim}u" "$FP_COMP" 2>/dev/null; then
-  fail "forward_plus_tile_cull.comp sphere_tile_overlap must use VK_FP_TILE_DIM (${fp_dim}) in tile corner math"
+echo "Forward+ tile cull: sphere overlap uses viewport-derived tile pixels (not hard-coded 16px):"
+if grep -q 'tileX \* 16u\|tileY \* 16u' "$FP_COMP" 2>/dev/null; then
+  fail "forward_plus_tile_cull.comp must not hard-code 16px tiles; derive tilePx from viewport / tileGrid (see vk_forward_plus VK_FP_TILE_DIM)"
+elif ! grep -q 'tilePxX' "$FP_COMP" 2>/dev/null || ! grep -q 'sphere_tile_overlap' "$FP_COMP" 2>/dev/null; then
+  fail "forward_plus_tile_cull.comp expected tilePxX/tilePxY sphere_tile_overlap path"
 else
-  pass "VK_FP_TILE_DIM=$fp_dim matches forward_plus_tile_cull.comp tile grid math"
+  fp_dim="$(sed -n 's/^#define VK_FP_TILE_DIM[[:space:]]*\([0-9][0-9]*\)u*$/\1/p' "$FP_C" | head -1)"
+  pass "forward_plus_tile_cull uses dynamic tile pixels (VK_FP_TILE_DIM=$fp_dim on host grid)"
 fi
 
 echo ""

src/renderers/vulkan/shaders/glsl/forward_plus_tile_cull.comp

@@ -1,7 +1,7 @@
 #version 450
 
 /*
- * Forward+ scaffolding: per-tile light index list (max 8 per 16x16 tile).
+ * Forward+ scaffolding: per-tile light index list (max 8 per tile; tile pixel size = viewport / tileGrid).
  * CPU-packed lights: binding 0 — vec4 (x=count,...), vec4 (tiles_x, tiles_y, viewport_w, viewport_h), then count * 4 vec4 records.
  */
 
@@ -31,11 +31,11 @@ layout(push_constant) uniform Push {
 	uint maxPerTile;
 } pc;
 
-bool sphere_tile_overlap(vec2 centerPx, float radiusPx, uint tileX, uint tileY) {
-	float minX = float(tileX * 16u);
-	float minY = float(tileY * 16u);
-	float maxX = minX + 16.0;
-	float maxY = minY + 16.0;
+bool sphere_tile_overlap(vec2 centerPx, float radiusPx, uint tileX, uint tileY, float tilePxX, float tilePxY) {
+	float minX = float(tileX) * tilePxX;
+	float minY = float(tileY) * tilePxY;
+	float maxX = minX + tilePxX;
+	float maxY = minY + tilePxY;
 	vec2 closest = clamp(centerPx, vec2(minX, minY), vec2(maxX, maxY));
 	return distance(centerPx, closest) <= radiusPx + 1e-3;
 }
@@ -47,6 +47,7 @@ void main() {
 	}
 
 	uint tilesX = pc.tileGrid.x;
+	uint tilesY = pc.tileGrid.y;
 	uint tileX = tileId % tilesX;
 	uint tileY = tileId / tilesX;
 
@@ -66,6 +67,8 @@ void main() {
 
 	float vw = lights.data[1].z;
 	float vh = lights.data[1].w;
+	float tilePxX = vw / max(float(tilesX), 1.0);
+	float tilePxY = vh / max(float(tilesY), 1.0);
 
 	uint outCount = 0u;
 	for (uint li = 0u; li < n && outCount < cap; li++) {
@@ -90,7 +93,7 @@ void main() {
 		float distW = max(abs(clip.w), 1e-3);
 		float radiusPx = clamp(radius * (0.5 * vw) / distW, 2.0, vw + vh);
 
-		if (!sphere_tile_overlap(centerPx, radiusPx, tileX, tileY)) {
+		if (!sphere_tile_overlap(centerPx, radiusPx, tileX, tileY, tilePxX, tilePxY)) {
 			continue;
 		}
 

src/renderers/vulkan/shaders/spirv/generated/shader_build_meta.txt

@@ -1,5 +1,5 @@
 glslang_validator_path=/usr/bin/glslangValidator
 glslang_validator_version=Glslang Version: 11:15.1.0 
-generated_at=2026-04-20T19:11:32Z
-shader_data_sha256=d85565a5bab935619eea87c0372b8ad703ad0e7a8ba7ca3f8710f8c8e9d366e3
+generated_at=2026-04-20T19:18:59Z
+shader_data_sha256=b6c7f805a5810cddfd7365a786824d42ea896f25f1d35c3ec412e157eaeee493
 shader_binding_sha256=960921d52493ab4fcf2f8b60edb34fa0b64a852c47b5ad2ae3abdf06c35ebe4a