Add draft kernel 4

Render from ptcl rather than original scene.

piet-gpu/shader/build.ninja

@@ -12,3 +12,5 @@ build image.spv: glsl image.comp | scene.h
 build kernel1.spv: glsl kernel1.comp | scene.h tilegroup.h
 
 build kernel3.spv: glsl kernel3.comp | scene.h tilegroup.h ptcl.h
+
+build kernel4.spv: glsl kernel4.comp | ptcl.h

piet-gpu/shader/kernel3.comp

@@ -1,3 +1,6 @@
+// This is "kernel 3" in a 4-kernel pipeline. It walks the active items
+// for the tilegroup and produces a per-tile command list for each tile.
+
 #version 450
 #extension GL_GOOGLE_include_directive : enable
 

piet-gpu/shader/kernel4.comp

@@ -0,0 +1,61 @@
+// This is "kernel 4" in a 4-kernel pipeline. It renders the commands
+// in the per-tile command list to an image.
+
+// Right now, this kernel stores the image in a buffer, but a better
+// plan is to use a texture. This is because of limited support.
+
+#version 450
+#extension GL_GOOGLE_include_directive : enable
+
+layout(local_size_x = 16, local_size_y = 16) in;
+
+// Same concern that this should be readonly as in kernel 3.
+layout(set = 0, binding = 0) buffer PtclBuf {
+    uint[] ptcl;
+};
+
+layout(set = 0, binding = 1) buffer ImageBuf {
+    uint[] image;
+};
+
+#include "ptcl.h"
+
+// TODO: make the image size dynamic.
+#define IMAGE_WIDTH 2048
+#define IMAGE_HEIGHT 1535
+
+#define WIDTH_IN_TILES 128
+
+#define PTCL_INITIAL_ALLOC 4096
+
+void main() {
+    uint tile_ix = gl_WorkGroupID.y * WIDTH_IN_TILES + gl_WorkGroupID.x;
+    CmdRef cmd_ref = CmdRef(tile_ix * PTCL_INITIAL_ALLOC);
+
+    uvec2 xy_uint = gl_GlobalInvocationID.xy;
+    vec2 xy = vec2(xy_uint);
+    vec2 uv = xy * vec2(1.0 / IMAGE_WIDTH, 1.0 / IMAGE_HEIGHT);
+    vec3 rgb = uv.xyy;
+
+    while (true) {
+        uint tag = Cmd_tag(cmd_ref);
+        if (tag == Cmd_End) {
+            break;
+        }
+        switch (tag) {
+        case Cmd_Circle:
+            CmdCircle circle = Cmd_Circle_read(cmd_ref);
+            float r = length(xy + vec2(0.5, 0.5) - circle.center.xy);
+            float alpha = clamp(0.5 + circle.radius - r, 0.0, 1.0);
+            vec4 fg_rgba = unpackUnorm4x8(circle.rgba_color);
+            // TODO: sRGB
+            rgb = mix(rgb, fg_rgba.rgb, alpha * fg_rgba.a);
+        }
+        cmd_ref.offset += Cmd_size;
+    }
+
+    // TODO: sRGB
+    uvec4 s = uvec4(round(vec4(rgb, 1.0) * 255.0));
+    uint rgba_packed = s.r | (s.g << 8) | (s.b << 16) | (s.a << 24);
+    image[xy_uint.y * IMAGE_WIDTH + xy_uint.x] = rgba_packed;
+}

piet-gpu/src/main.rs

@@ -16,7 +16,7 @@ const HEIGHT: usize = 1536;
 const TILE_W: usize = 16;
 const TILE_H: usize = 16;
 
-const N_CIRCLES: usize = 100;
+const N_CIRCLES: usize = 3000;
 
 fn make_scene() -> Vec<u8> {
     let mut rng = rand::thread_rng();
@@ -119,10 +119,10 @@ fn main() {
             .create_descriptor_set(&k3_pipeline, &[&scene_dev, &tilegroup_buf, &ptcl_buf])
             .unwrap();
 
-        let code = include_bytes!("../shader/image.spv");
-        let pipeline = device.create_simple_compute_pipeline(code, 2).unwrap();
+        let k4_code = include_bytes!("../shader/kernel4.spv");
+        let pipeline = device.create_simple_compute_pipeline(k4_code, 2).unwrap();
         let descriptor_set = device
-            .create_descriptor_set(&pipeline, &[&scene_dev, &image_dev])
+            .create_descriptor_set(&pipeline, &[&ptcl_buf, &image_dev])
             .unwrap();
         let query_pool = device.create_query_pool(4).unwrap();
         let mut cmd_buf = device.create_cmd_buf().unwrap();
@@ -167,9 +167,11 @@ fn main() {
             (timestamps[2] - timestamps[1]) * 1e3
         );
 
+        /*
         let mut k1_data: Vec<u32> = Default::default();
         device.read_buffer(&ptcl_buf, &mut k1_data).unwrap();
         dump_k1_data(&k1_data);
+        */
 
         let mut img_data: Vec<u8> = Default::default();
         // Note: because png can use a `&[u8]` slice, we could avoid an extra copy