Run the view transform after the detail stage, in a pass of its own

The fused pass stops at "linear working values" when a sharpener, a
blur or a repair follows, and the detail passes convolve what it hands
on. Until now it handed on the rendering: the base curve, and since the
last commit the view transform, ran before the store. So every kernel
worked on display-referred values while its comments promised the
opposite — D19's second finding.

A fused pass composed for a detail stage now stops before the view
transform, and carries a second shader, `ComposedShader::view`, composed
from the same inputs. It runs the same prologue, for the positions a
fragment reads (a film's grain seeds from `source_px`) and the corners
it blacks out, takes its colour from the detail stage's result bound
where the sample cache would be, and runs the view transform, the
output transform and the mask reveal. `render_detailed` dispatches it
after the last detail pass, in the same encoder.

So no detail pass encodes any more. Every pass writes an intermediate,
the last one included, which retires three things that existed only to
make the last pass encode: `writes_output` and the runner's second
layout, the body-less resolve pass for an active kernel with nothing to
draw at this scale, and capture sharpening's pass-through, which now
emits no pass at all. An empty chain is a whole render: the view pass
reads the fused result directly. The detail stage no longer takes an
output space either, so `compose_detail_for` folds into
`compose_detail` and the space is named once, on the fused half.

The cost is one full-render read and write per frame when a detail
stage exists, and a third intermediate for a one-pass chain.

core/dr-gpu/tests/capture_sharpen.rs

@@ -87,8 +87,7 @@ fn sharpened(amount: f32, radius: f32, threshold: f32) -> EditGraph {
 fn render(pass: &mut AdjustPass, graph: &EditGraph, source: &DemosaicedImage, out: u32) -> Vec<u8> {
     let shader = graph.compose_for(ColourSpace::Srgb);
     let scale = graph.render_scale(source.size(), (out, out));
-    let detail =
-        graph.compose_detail_for(scale.full_size(), scale.render_size(), ColourSpace::Srgb);
+    let detail = graph.compose_detail(scale.full_size(), scale.render_size());
     let key = graph.invalidation().through(Affects::Colour);
     pass.render_detailed(source, &shader, out, out, None, &detail, key)
         .expect("render");
@@ -413,7 +412,7 @@ fn dragging_the_amount_recompiles_nothing_and_reallocates_nothing() {
     let pipelines = pass.cached_detail_pipelines();
     let allocations = pass.detail_allocations();
     assert_eq!(pipelines, 2, "one per axis of the separable mask");
-    assert_eq!(allocations, 2, "the colour result, and one hand-off");
+    assert_eq!(allocations, 3, "the colour result, and the ping-pong pair");
     assert_eq!(pass.detail_dispatches(), 2);
     assert_eq!(pass.colour_dispatches(), 1);
 
@@ -453,13 +452,12 @@ fn dragging_the_amount_recompiles_nothing_and_reallocates_nothing() {
 
 #[test]
 fn a_render_too_coarse_for_the_radius_still_reaches_the_screen() {
-    // The failure mode that the pass-through exists to prevent, proved on a
-    // device rather than argued about. With the radius finer than a render
-    // pixel the operation declines to sharpen — but it is still active, so the
-    // fused pass has already been composed to hand on unclipped linear values,
-    // and something must still perform the output transform. An empty chain
-    // here would not be a soft preview: it would be a hard error out of
-    // `render_detailed`, on the most ordinary develop view there is.
+    // Proved on a device rather than argued about. With the radius finer than
+    // a render pixel the operation declines to sharpen — but it is still
+    // active, so the fused pass has already been composed to hand on
+    // unclipped linear values, and something must still perform the output
+    // transform. Since D19 that is the view pass, whatever the chain holds:
+    // the chain is empty and the frame is still whole.
     let Some(ctx) = ctx() else { return };
     const SOURCE: u32 = 128;
     const RENDER: u32 = 32; // a quarter scale, as a fit view of a large frame
@@ -471,7 +469,16 @@ fn a_render_too_coarse_for_the_radius_still_reaches_the_screen() {
 
     let mut pass = AdjustPass::new(&ctx);
     let sharp = render(&mut pass, &graph, &source, RENDER);
-    assert_eq!(pass.detail_dispatches(), 1, "one pass, and it only encodes");
+    assert_eq!(
+        pass.detail_dispatches(),
+        0,
+        "nothing to sharpen at this scale"
+    );
+    assert_eq!(
+        pass.view_dispatches(),
+        1,
+        "and the view pass finishes the frame"
+    );
 
     // And what reaches the screen is the unsharpened picture, not a black
     // frame, a linear one, or a guess.

core/dr-gpu/tests/detail_instances.rs

@@ -38,15 +38,17 @@ fn grey(ctx: &GpuContext) -> DemosaicedImage {
     DemosaicedImage::from_rgba8(ctx, &data, SIZE, SIZE).expect("upload")
 }
 
-/// A pass that sums the instance list into the red channel and writes the
-/// output. Deliberately trivial: the value on screen is then a direct readout
-/// of what arrived in the buffer.
+/// A pass that sums the instance list into the red channel and writes a
+/// linear intermediate, which the view pass then encodes (D19). Deliberately
+/// trivial: the value on screen is then a direct readout of what arrived in the
+/// buffer, through the sRGB encode — the source is an 8-bit upload, so the
+/// view transform is skipped for it and the encode is the only thing between.
 fn summing_pass(storage: Vec<[f32; 4]>, structure: u64) -> ComposedDetailPass {
     let source = "
 @group(0) @binding(0) var source: texture_2d<f32>;
 struct Params { detail_base: vec4<f32> }
 @group(0) @binding(1) var<uniform> u: Params;
-@group(0) @binding(2) var output: texture_storage_2d<rgba8unorm, write>;
+@group(0) @binding(2) var output: texture_storage_2d<rgba16float, write>;
 @group(0) @binding(3) var<storage, read> instances: array<vec4<f32>>;
 
 @compute @workgroup_size(8, 8, 1)
@@ -61,7 +63,7 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
     for (var i = 0u; i < n; i = i + 1u) {
         total = total + instances[i].x * f32(i + 1u);
     }
-    textureStore(output, vec2<i32>(gid.xy), vec4<f32>(total, f32(n) / 255.0, 0.0, 1.0));
+    textureStore(output, vec2<i32>(gid.xy), vec4<f32>(total, f32(n) * 0.1, 0.0, 1.0));
 }
 "
     .to_string();
@@ -73,13 +75,17 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
         uniforms: vec![SIZE as f32, SIZE as f32, 1.0, 0.0],
         storage,
         radius: 0,
-        writes_output: true,
         // Any distinct number: the hash is a cache key, and these tests are
         // what decide whether two chains share a pipeline.
         structure_hash: structure,
     }
 }
 
+/// A linear value as the view pass leaves it in the 8-bit output.
+fn encoded(linear: f32) -> u8 {
+    (dr_types::Transfer::Srgb.encode(linear) * 255.0).round() as u8
+}
+
 fn render(pass: &mut AdjustPass, source: &DemosaicedImage, chain: &ComposedDetail) -> Vec<u8> {
     // The fused half has to be composed knowing a detail stage follows it, or
     // it encodes its own output and the chain would quantise twice — a mismatch
@@ -118,12 +124,15 @@ fn a_pass_reads_the_list_it_was_given() {
     let pixels = render(&mut pass, &source, &chain);
     let (red, green) = (pixels[0], pixels[1]);
 
-    // 0.05·1 + 0.1·2 = 0.25, written straight to an rgba8 target.
+    // 0.05·1 + 0.1·2 = 0.25.
     assert!(
-        red.abs_diff((0.25 * 255.0) as u8) <= 1,
+        red.abs_diff(encoded(0.25)) <= 1,
         "the shader summed {red}, not the list it was handed"
     );
-    assert_eq!(green, 2, "arrayLength saw both entries");
+    assert!(
+        green.abs_diff(encoded(0.2)) <= 1,
+        "arrayLength saw both entries"
+    );
 }
 
 /// A convolution declares no list and must still run: it is bound to the
@@ -142,7 +151,10 @@ fn a_pass_with_no_list_still_runs() {
 
     let pixels = render(&mut pass, &source, &chain);
     assert_eq!(pixels[0], 0, "the placeholder is zeroed");
-    assert_eq!(pixels[1], 1, "and is exactly one element long");
+    assert!(
+        pixels[1].abs_diff(encoded(0.1)) <= 1,
+        "and is exactly one element long"
+    );
 }
 
 /// The property that makes placing the tenth spot as cheap as moving a slider:

core/dr-gpu/tests/detail_stage.rs

@@ -101,8 +101,7 @@ fn render(
     let _ = ctx;
     let shader = graph.compose_for(ColourSpace::Srgb);
     let scale = graph.render_scale(source.size(), (out, out));
-    let detail =
-        graph.compose_detail_for(scale.full_size(), scale.render_size(), ColourSpace::Srgb);
+    let detail = graph.compose_detail(scale.full_size(), scale.render_size());
     let key = graph.invalidation().through(Affects::Colour);
     pass.render_detailed(source, &shader, out, out, None, &detail, key)
         .expect("render");
@@ -301,7 +300,7 @@ fn dragging_a_slider_recompiles_nothing_and_reallocates_nothing() {
     let pipelines = pass.cached_detail_pipelines();
     let allocations = pass.detail_allocations();
     assert_eq!(pipelines, 2, "one per pass of the separable blur");
-    assert_eq!(allocations, 2, "the colour result, and one hand-off");
+    assert_eq!(allocations, 3, "the colour result, and the ping-pong pair");
 
     for radius in [0.06, 0.07, 0.08, 0.09] {
         graph.set_param(PROBE, RADIUS, radius);
@@ -403,8 +402,7 @@ fn an_empty_chain_falls_through_to_the_ordinary_render() {
 
     let shader = graph.compose_for(ColourSpace::Srgb);
     let scale = graph.render_scale((SIZE, SIZE), (SIZE, SIZE));
-    let detail =
-        graph.compose_detail_for(scale.full_size(), scale.render_size(), ColourSpace::Srgb);
+    let detail = graph.compose_detail(scale.full_size(), scale.render_size());
     assert!(detail.is_empty());
 
     pass.render_detailed(&source, &shader, SIZE, SIZE, None, &detail, 0)

core/dr-gpu/tests/local_contrast.rs

@@ -109,8 +109,7 @@ fn row_rgb(pixels: &[u8], size: u32, y: u32) -> Vec<[u8; 3]> {
 fn render(pass: &mut AdjustPass, graph: &EditGraph, source: &DemosaicedImage, out: u32) -> Vec<u8> {
     let shader = graph.compose_for(ColourSpace::Srgb);
     let scale = graph.render_scale(source.size(), (out, out));
-    let detail =
-        graph.compose_detail_for(scale.full_size(), scale.render_size(), ColourSpace::Srgb);
+    let detail = graph.compose_detail(scale.full_size(), scale.render_size());
     let key = graph.invalidation().through(Affects::Colour);
     pass.render_detailed(source, &shader, out, out, None, &detail, key)
         .expect("render");
@@ -683,28 +682,21 @@ fn texture_contributes_nothing_where_its_scale_does_not_exist() {
     // `render_masked`, and was rejected for handing a linear-working shader
     // to the plain path — so texture alone on a thumbnail did not render.
     //
-    // The seam was closed where that note said it would have to be, at the
-    // composition boundary: `compose_detail` now emits a bodyless
-    // `detail/resolve` pass in exactly this case, which reads only the pixel
-    // it writes and performs the output transform the fused pass declined to
-    // do. So the chain is no longer empty — it carries precisely the one pass
-    // that finishes the render and no kernel at all, which is the honest
-    // description of "a two-pixel surface structure is not present in a
-    // 128-pixel rendering".
+    // The seam was closed at the composition boundary, and closed again,
+    // more simply, by D19: no detail pass encodes any more, the fused pass's
+    // view pass performs the output transform whatever the chain holds, and
+    // so the empty chain is a whole render. That is the honest description of
+    // "a two-pixel surface structure is not present in a 128-pixel
+    // rendering".
     let scale = graph.render_scale(source.size(), (128, 128));
-    let composed =
-        graph.compose_detail_for(scale.full_size(), scale.render_size(), ColourSpace::Srgb);
-    assert_eq!(
-        composed.len(),
-        1,
-        "the chain must carry the resolve pass and nothing else"
-    );
-    assert_eq!(composed.passes[0].label, "detail/resolve");
-    assert_eq!(
-        composed.radius(),
-        0,
+    let composed = graph.compose_detail(scale.full_size(), scale.render_size());
+    assert!(
+        composed.is_empty(),
         "texture claimed a kernel it cannot draw"
     );
+    let mut pass = AdjustPass::new(&ctx);
+    render(&mut pass, &graph, &source, 128);
+    assert_eq!(pass.view_dispatches(), 1, "the view pass still finishes it");
 
     // With clarity on as well the edit is renderable again, and the dispatch
     // count says what the assertion above says: two passes, not four. Texture

core/dr-gpu/tests/noise_reduction.rs

@@ -73,8 +73,7 @@ fn render_at(
     scale: RenderScale,
 ) -> Vec<u8> {
     let shader = graph.compose_for(ColourSpace::Srgb);
-    let detail =
-        graph.compose_detail_for(scale.full_size(), scale.render_size(), ColourSpace::Srgb);
+    let detail = graph.compose_detail(scale.full_size(), scale.render_size());
     let key = graph.invalidation().through(Affects::Colour);
     pass.render_detailed(source, &shader, out.0, out.1, None, &detail, key)
         .expect("render");

core/dr-gpu/tests/spot_removal.rs

@@ -72,7 +72,7 @@ fn render(pass: &mut AdjustPass, graph: &EditGraph, source: &DemosaicedImage, ou
     let shader = graph.compose_for(ColourSpace::Srgb);
     let (w, h) = graph.output_size(source.size().0, source.size().1);
     let (w, h) = (w.min(out), h.min(out));
-    let detail = graph.compose_detail_for(source.size(), (w, h), ColourSpace::Srgb);
+    let detail = graph.compose_detail(source.size(), (w, h));
     let key = graph.invalidation().through(Affects::Colour);
     pass.render_detailed(source, &shader, w, h, None, &detail, key)
         .expect("render");

core/dr-gpu/tests/view_transform.rs

@@ -127,3 +127,52 @@ fn the_rendering_is_monotone_through_the_whole_range() {
         last = got;
     }
 }
+
+/// Render `graph` over a flat frame through `render_detailed`, the path every
+/// frontend takes, and return the centre pixel's red.
+fn rendered_detailed(ctx: &GpuContext, level: u16, graph: &EditGraph) -> (u8, AdjustPass) {
+    let source = Demosaicer::new(ctx)
+        .expect("demosaicer")
+        .run(&flat_raw(level))
+        .expect("demosaic");
+    let shader = graph.compose_for(dr_types::ColourSpace::Srgb);
+    let detail = graph.compose_detail(source.size(), (SIZE, SIZE));
+    let key = graph.invalidation().through(dr_pipeline::Affects::Colour);
+    let mut adjust = AdjustPass::new(ctx);
+    adjust
+        .render_detailed(&source, &shader, SIZE, SIZE, None, &detail, key)
+        .expect("render");
+    let (pixels, _, _) = adjust.export_pixels().expect("readback");
+    let centre = ((SIZE / 2) * SIZE + SIZE / 2) * 4;
+    (pixels[centre as usize], adjust)
+}
+
+#[test]
+fn a_detail_stage_renders_through_the_view_pass_unchanged() {
+    // TRACES: FR-DEV-3j | FR-DEV-2
+    // With a detail stage the view transform is a dispatch of its own after
+    // it (D19). Sharpening a flat field changes nothing, so the same frame
+    // with and without it must render the same: the view pass read the detail
+    // stage's result, applied the view transform once, and encoded once.
+    let Some(ctx) = ctx() else {
+        eprintln!("skipping: no GPU adapter");
+        return;
+    };
+    let plain = rendered(&ctx, 8_520, &EditGraph::default_chain());
+
+    let mut sharpened = EditGraph::default_chain();
+    let id = dr_pipeline::ops::capture_sharpen::ID;
+    sharpened.set_param(id, dr_pipeline::ops::capture_sharpen::AMOUNT, 100.0);
+    sharpened.set_param(id, dr_pipeline::ops::capture_sharpen::RADIUS, 1.0);
+    let (detailed, pass) = rendered_detailed(&ctx, 8_520, &sharpened);
+
+    assert!(
+        pass.detail_dispatches() > 0,
+        "the premise: a detail stage ran"
+    );
+    assert_eq!(pass.view_dispatches(), 1);
+    assert!(
+        detailed.abs_diff(plain) <= 1,
+        "with a detail stage {detailed}, without {plain}"
+    );
+}