render: march sun shafts at quarter res with reach-proportional steps

src/native/shaders/hdr.hlsl

@@ -52,7 +52,8 @@ cbuffer C : register(b0) {
   // strength / phase g, p1.xy = proj m22 / m32, vp/vs0/vs1/vs2 = the
   // row-vector inverse view-projection (world = clip * M).
   float4 vp;    // tonemap/linearize: x = |m00|, y = |m11|, z = m22, w = m32
-  float4 vs0;   // tonemap: w = shaft intensity (xyz unused)
+  float4 vs0;   // tonemap: x = shaft plane resolution divisor (>= 2),
+                // w = shaft intensity (yz unused)
   float4 vs1;   // tonemap: xyz = haze tint (xe space), w = haze intensity
   float4 vs2;   // tonemap: xyz = sun direction in AO view space (toward the
                 // sun), w = haze density (1 / view unit)
@@ -248,7 +249,14 @@ float4 ps_vol_shafts(VSOut i) : SV_Target {
   float dist = length(ray);
   float3 rdir = ray / max(dist, 1e-4);
   float reach = min(dist, src.w);
-  float steps = clamp(p0.x, 8.0, 64.0);
+  // Constant per-meter sample density: a short ray to near geometry needs
+  // proportionally fewer steps for the same marched quality (the fixed
+  // count oversampled a 5 m ray 8x versus the full-reach case), and the
+  // path-fraction scale keeps near-surface terms small enough that the
+  // extra variance stays under the tent blur.
+  float steps_max = clamp(p0.x, 8.0, 64.0);
+  float steps = clamp(ceil(steps_max * reach / max(src.w, 1.0)), 8.0,
+                      steps_max);
   float jitter = Ign(i.pos.xy);
   float lit = 0.0;
   float tot = 0.0;
@@ -400,7 +408,8 @@ float4 ps_tonemap(VSOut i) : SV_Target {
   bool sc_vol = sc_mask < 0 || (sc_mask & 64) != 0;
   if (vs0.w > 0.0 && sc_vol) {
     float dpix = lin_depth.SampleLevel(smp_point, i.uv, 0);
-    float2 vt = size.zw * 2.0;  // one shaft-plane texel in uv
+    // One shaft-plane texel in uv (vs0.x = the march resolution divisor).
+    float2 vt = size.zw * max(vs0.x, 2.0);
     float2 base = (floor(i.uv / vt - 0.5) + 0.5) * vt;
     float2 f = saturate((i.uv - base) / vt);
     float acc = 0.0;

src/skate3_native_scene.cpp

@@ -249,10 +249,20 @@ REXCVAR_DEFINE_DOUBLE(skate3_native_render_scene_shafts_reach, 40.0, "Skate 3",
     .range(5.0, 300.0)
     .lifecycle(rex::cvar::Lifecycle::kHotReload);
 REXCVAR_DEFINE_INT32(skate3_native_render_scene_shafts_steps, 64, "Skate 3",
-                     "Sun-shaft march steps per pixel (half-res pass). More steps "
+                     "Sun-shaft march step ceiling per pixel; rays shorter than "
-                     "resolve thinner shadow volumes at proportional GPU cost.")
+                     "the full reach take proportionally fewer steps at the same "
+                     "per-meter density. More steps resolve thinner shadow "
+                     "volumes at proportional GPU cost.")
     .range(8, 64)
     .lifecycle(rex::cvar::Lifecycle::kHotReload);
+REXCVAR_DEFINE_INT32(skate3_native_render_scene_shafts_res, 4, "Skate 3",
+                     "Sun-shaft march resolution divisor (2 = half res, 4 = "
+                     "quarter res). The march output is low frequency (jittered "
+                     "average under a tent blur) and upsamples depth-aware, so "
+                     "higher divisors trade little quality for a proportional "
+                     "GPU saving.")
+    .range(2, 8)
+    .lifecycle(rex::cvar::Lifecycle::kHotReload);
 REXCVAR_DEFINE_BOOL(skate3_native_render_scene_sun_override, false, "Skate 3",
                     "Lighting lab: replace the captured sun direction with the "
                     "azimuth/elevation cvars. Every per-frame light transform is "

src/skate3_native_scene_post.cpp

@@ -71,6 +71,7 @@ REXCVAR_DECLARE(bool, skate3_native_render_scene_shafts);
 REXCVAR_DECLARE(double, skate3_native_render_scene_shafts_intensity);
 REXCVAR_DECLARE(double, skate3_native_render_scene_shafts_reach);
 REXCVAR_DECLARE(int32_t, skate3_native_render_scene_shafts_steps);
+REXCVAR_DECLARE(int32_t, skate3_native_render_scene_shafts_res);
 REXCVAR_DECLARE(bool, skate3_native_render_scene_ssao_debug);
 REXCVAR_DECLARE(bool, skate3_native_render_scene_ssao_full_res);
 REXCVAR_DECLARE(double, skate3_native_render_scene_ssao_intensity);
@@ -1087,9 +1088,13 @@ bool ApplyVolumetricPass(const NativeGuestOutputRenderContext& context,
     }
   }
 
-  // ---- Shaft planes (half res, march + blurred), rebuilt on resize.
+  // ---- Shaft planes (march + blurred), rebuilt on resize or divisor
-  const uint32_t vw = std::max(1u, (width + 1) / 2);
+  // change; the tonemap upsample reads the divisor from vs0.x.
-  const uint32_t vh = std::max(1u, (height + 1) / 2);
+  const uint32_t vol_div = uint32_t(
+      std::clamp<int32_t>(REXCVAR_GET(skate3_native_render_scene_shafts_res),
+                          2, 8));
+  const uint32_t vw = std::max(1u, (width + vol_div - 1) / vol_div);
+  const uint32_t vh = std::max(1u, (height + vol_div - 1) / vol_div);
   if (shafts &&
       (g_r.vol_width != vw || g_r.vol_height != vh || g_r.vol_tex == nullptr ||
        g_r.vol_tex_b == nullptr)) {
@@ -1272,7 +1277,7 @@ bool ApplyVolumetricPass(const NativeGuestOutputRenderContext& context,
   g_r.vol_rows[1] = std::fabs(pr[5]);
   g_r.vol_rows[2] = pr[10];
   g_r.vol_rows[3] = pr[14];
-  g_r.vol_rows[4] = 0.0f;
+  g_r.vol_rows[4] = float(vol_div);
   g_r.vol_rows[5] = 0.0f;
   g_r.vol_rows[6] = 0.0f;
   g_r.vol_rows[7] = shaft_plane ? shaft_int : 0.0f;