Vulkan: render-target post-FX texels; Forward+ tile size from SSBO

Merge PR #126.

Author: timfox

docs/GLTF.md

@@ -76,6 +76,12 @@ Larger assets are **silently clamped** during load.
 - **Skinned / morphed** primitives prefer the **GPU path** under Vulkan PBR when **`r_gltfGpu 1`** and constraints above are met; otherwise **`RB_GLTFSurface`** falls back to **CPU tess**.
 - **More than eight** non-zero morph weights per vertex: GPU path keeps only the **eight largest** weights per draw (same cap as IQM `IQM_MORPH_TOP_K`; tune `r_morphMaxActive` for IQM batching only).
 
+### 8. Roadmap: full GPU qtangent / MikkTSpace (not shipped yet)
+
+- **Today (Vulkan PBR GPU):** bind-pose tangents + **`r_gltfGpuTangentFix`** Gram–Schmidt **T** vs deformed **N** after skin+morph (`gen_vert.tmpl`). CPU tess still does **MikkTSpace-style qtangent** from deformed positions + **UV0** when PBR needs it.
+- **Next increment:** optional **compute** or **vertex-neighborhood** pass is required for true **MikkTSpace** on arbitrary meshes on the GPU (needs neighbor topology / shared-vertex groups, not just per-vertex attributes). Likely behind a **`r_gltfGpuQtangent`** (or similar) latched cvar with clear **startup** + **developer** logs when enabled or when falling back.
+- **Validation:** add Tier B scenes with known normal-map assets and compare against CPU tess / reference captures once a GPU path exists.
+
 ## Engine references
 
 - Loader / registration: `src/renderers/vulkan/tr_model_gltf.c`, `tr_model_gltf.h`

scripts/renderer_regression_check.sh

@@ -218,6 +218,17 @@ else
   pass "forward_plus_shade_strength uses constant_id=$fp_cid (shader + pipeline agree)"
 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 (${fp_dim}) in tile corner math"
+else
+  pass "VK_FP_TILE_DIM=$fp_dim matches forward_plus_tile_cull.comp tile grid math"
+fi
+
 echo ""
 echo "Vulkan temporal: reset bitmask vs reason_string / log table:"
 VK_TEMP_H="$PROJECT_ROOT/src/renderers/vulkan/vk_temporal.h"

src/renderers/vulkan/shaders/glsl/gen_frag.tmpl

@@ -1232,17 +1232,18 @@ void main() {
 		uint tilesY = uint( fp_lights.fp_light_data[1].y + 0.5 );
 		float vw = fp_lights.fp_light_data[1].z;
 		float vh = fp_lights.fp_light_data[1].w;
-		if ( tilesX > 0u && tilesY > 0u && vw > 1.0 && vh > 1.0 ) {
-			vec4 wc = fp_params.fp_clip_from_world * vec4( var_WorldPos, 1.0 );
-			if ( abs( wc.w ) > 1e-5 ) {
-				vec3 ndc = wc.xyz / wc.w;
-				if ( ndc.z >= -1.0 && ndc.z <= 1.0 && ndc.x >= -1.05 && ndc.x <= 1.05 && ndc.y >= -1.05 && ndc.y <= 1.05 ) {
-					vec2 px;
-					px.x = 0.5 * ( 1.0 + ndc.x ) * vw;
-					px.y = 0.5 * ( 1.0 + ndc.y ) * vh;
-					const float tilePx = 16.0;
-					uint tx = min( uint( px.x / tilePx ), tilesX - 1u );
-					uint ty = min( uint( px.y / tilePx ), tilesY - 1u );
+				if ( tilesX > 0u && tilesY > 0u && vw > 1.0 && vh > 1.0 ) {
+					float tilePxX = vw / max( float( tilesX ), 1.0 );
+					float tilePxY = vh / max( float( tilesY ), 1.0 );
+					vec4 wc = fp_params.fp_clip_from_world * vec4( var_WorldPos, 1.0 );
+					if ( abs( wc.w ) > 1e-5 ) {
+						vec3 ndc = wc.xyz / wc.w;
+						if ( ndc.z >= -1.0 && ndc.z <= 1.0 && ndc.x >= -1.05 && ndc.x <= 1.05 && ndc.y >= -1.05 && ndc.y <= 1.05 ) {
+							vec2 px;
+							px.x = 0.5 * ( 1.0 + ndc.x ) * vw;
+							px.y = 0.5 * ( 1.0 + ndc.y ) * vh;
+							uint tx = min( uint( px.x / tilePxX ), tilesX - 1u );
+							uint ty = min( uint( px.y / tilePxY ), tilesY - 1u );
 					uint tileId = ty * tilesX + tx;
 					uint tbase = tileId * 8u;
 					float nLights = fp_lights.fp_light_data[0].x;
@@ -1430,11 +1431,14 @@ void main() {
 		if ( fp_dbg > 1e-6 ) {
 			uint tilesX = uint( fp_lights.fp_light_data[1].x + 0.5 );
 			uint tilesY = uint( fp_lights.fp_light_data[1].y + 0.5 );
-			if ( tilesX > 0u && tilesY > 0u ) {
-				const float tilePx = 16.0;
+			float vwDbg = fp_lights.fp_light_data[1].z;
+			float vhDbg = fp_lights.fp_light_data[1].w;
+			if ( tilesX > 0u && tilesY > 0u && vwDbg > 1.0 && vhDbg > 1.0 ) {
+				float tilePxX = vwDbg / max( float( tilesX ), 1.0 );
+				float tilePxY = vhDbg / max( float( tilesY ), 1.0 );
 				vec2 px = gl_FragCoord.xy;
-				uint tx = min( uint( px.x / tilePx ), tilesX - 1u );
-				uint ty = min( uint( px.y / tilePx ), tilesY - 1u );
+				uint tx = min( uint( px.x / tilePxX ), tilesX - 1u );
+				uint ty = min( uint( px.y / tilePxY ), tilesY - 1u );
 				uint tileId = ty * tilesX + tx;
 				uint base = tileId * 8u;
 				uint maxPerTile = uint( max( fp_lights.fp_light_data[0].z + 0.5, 1.0 ) );
@@ -1456,7 +1460,7 @@ void main() {
 				} else {
 					col = mix( vec3( 1.0, 0.88, 0.05 ), vec3( 0.92, 0.08, 0.02 ), ( occ - 0.75 ) * 4.0 );
 				}
-				vec2 f = fract( px / tilePx );
+				vec2 f = fract( vec2( px.x / tilePxX, px.y / tilePxY ) );
 				float edge = min( min( f.x, 1.0 - f.x ), min( f.y, 1.0 - f.y ) );
 				float border = smoothstep( 0.0, 0.35, edge );
 				col = mix( vec3( 0.08 ), col, border );

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=2026-04-19T12:38:57Z
-shader_data_sha256=49a4145ab67781b295fff758553e7f345ebcaa15277a33baa7216e1ddf8fee9f
+generated_at=2026-04-20T19:11:32Z
+shader_data_sha256=d85565a5bab935619eea87c0372b8ad703ad0e7a8ba7ca3f8710f8c8e9d366e3
 shader_binding_sha256=960921d52493ab4fcf2f8b60edb34fa0b64a852c47b5ad2ae3abdf06c35ebe4a