a170120 · nvidia-rtx-godot

nvidia-rtx-godot commits / a170120

Provide dFdx/dFdy/fwidth stubs for RT hit shaders

Screen-space derivatives do not exist in ray tracing: a hit shader has no 2x2
pixel quad, so dFdx/dFdy/fwidth have no closest-hit overload and glslang rejects
any shader that calls them:

    RT shader compile failed (hg2_v0_closest_hit):
    ERROR: 0:3411: 'dFdx' : no matching overloaded function found

That fails the whole hit group, build_tlas skips every surface whose hit group is
not ready, and the geometry ends up invisible while its collision -- being
CPU-side -- still works. Both consumers hit it: dc_level.gdshader:117 and the
dc-terrain shader each take dFdx/dFdy of the world position and project it
through the face's projection axes to feed textureGrad.

Macros rather than functions, because redefining a built-in with an existing
signature is not legal GLSL and macros stay type-generic.

LIMITATION, deliberately not hidden: the derivative is zero, so textureGrad
selects LOD 0 and screen-footprint mip selection is lost under RT. Expect
aliasing on minified detail -- the seam-glint behaviour dc_level's textureGrad
was written for does not survive this. The correct replacement is ray
differentials / ray cones, which need the hit triangle's attribute gradients and
therefore cannot be expressed as a macro over an arbitrary expression; the
position case (dFdx(world_pos)) is the tractable one to do first.

Verified by compiling the real closest-hit stage offline with dc_level's exact
textureGrad path (dump lines 3407-3415) substituted into the template.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

a1701201030971c7489514a1017bf20b19872a79

by plankatron <fleeky@gmail.com> · 2026-07-29T14:55:43.000Z

Changed files (1) +21 -0

servers/rendering/renderer_rd/shaders/raytracing/raytracing_custom_globals_inc.glsl +21 -0

…	…	@@ -39,6 +39,27 @@ vec4 rt_frag_coord = vec4(0.0);
39	39	float alpha_antialiasing_edge = 0.0;
40	40	vec2 alpha_texture_coordinate = vec2(0.0);
41	41
	42	// Screen-space derivatives do not exist in ray tracing: a hit shader has no 2x2
	43	// pixel quad, so dFdx/dFdy/fwidth have no closest-hit overload and glslang
	44	// rejects any shader calling them -- failing the whole hit group, which
	45	// build_tlas then skips, leaving the geometry invisible but still collidable.
	46	//
	47	// These are macros rather than functions because redefining a built-in with an
	48	// existing signature is not legal GLSL, and macros stay type-generic.
	49	//
	50	// LIMITATION: the derivative is zero, so textureGrad/textureQueryLod select
	51	// LOD 0 and screen-footprint mip selection is lost under RT (expect aliasing on
	52	// minified detail). Doing this properly needs ray differentials or ray cones,
	53	// which require the hit triangle's attribute gradients and so cannot be
	54	// expressed as a macro over an arbitrary expression.
	55	#define dFdx(m_v) ((m_v) * 0.0)
	56	#define dFdy(m_v) ((m_v) * 0.0)
	57	#define dFdxCoarse(m_v) ((m_v) * 0.0)
	58	#define dFdyCoarse(m_v) ((m_v) * 0.0)
	59	#define dFdxFine(m_v) ((m_v) * 0.0)
	60	#define dFdyFine(m_v) ((m_v) * 0.0)
	61	#define fwidth(m_v) ((m_v) * 0.0)
	62
42	63	// Screen/depth textures are unavailable in RT -- alias to bindless slot 0
43	64	// so shaders that reference them still compile (reads return dummy values).
44	65	#define depth_buffer bindless_textures[0]