gpu/shaders: update piet-gpu

Changes:
- faster implementation of RGBA output
- fix stroked clips and images

Signed-off-by: Elias Naur <mail@eliasnaur.com>

gpu/shaders/backdrop.comp

@@ -48,13 +48,13 @@ void main() {
     if (element_ix < conf.n_elements) {
         AnnotatedTag tag = Annotated_tag(conf.anno_alloc, ref);
         switch (tag.tag) {
+        case Annotated_Image:
+        case Annotated_BeginClip:
         case Annotated_Color:
             if (fill_mode_from_flags(tag.flags) != MODE_NONZERO) {
                 break;
             }
             // Fall through.
-        case Annotated_Image:
-        case Annotated_BeginClip:
             PathRef path_ref = PathRef(conf.tile_alloc.offset + element_ix * Path_size);
             Path path = Path_read(conf.tile_alloc, path_ref);
             sh_row_width[th_ix] = path.bbox.z - path.bbox.x;

gpu/shaders/coarse.comp

@@ -91,6 +91,23 @@ bool alloc_cmd(inout Alloc cmd_alloc, inout CmdRef cmd_ref, inout uint cmd_limit
     return true;
 }
 
+void write_fill(Alloc alloc, inout CmdRef cmd_ref, uint flags, Tile tile, float linewidth) {
+    if (fill_mode_from_flags(flags) == MODE_NONZERO) {
+        if (tile.tile.offset != 0) {
+            CmdFill cmd_fill = CmdFill(tile.tile.offset, tile.backdrop);
+            Cmd_Fill_write(alloc, cmd_ref, cmd_fill);
+            cmd_ref.offset += 4 + CmdFill_size;
+        } else {
+            Cmd_Solid_write(alloc, cmd_ref);
+            cmd_ref.offset += 4;
+        }
+    } else {
+        CmdStroke cmd_stroke = CmdStroke(tile.tile.offset, 0.5 * linewidth);
+        Cmd_Stroke_write(alloc, cmd_ref, cmd_stroke);
+        cmd_ref.offset += 4 + CmdStroke_size;
+    }
+}
+
 void main() {
     if (mem_error != NO_ERROR) {
         return;
@@ -135,6 +152,12 @@ void main() {
     uint part_start_ix = 0;
     uint ready_ix = 0;
 
+    // Leave room for the fine rasterizer scratch allocation.
+    Alloc scratch_alloc = slice_mem(cmd_alloc, 0, Alloc_size);
+    cmd_ref.offset += Alloc_size;
+
+    uint num_begin_slots = 0;
+    uint begin_slot = 0;
     while (true) {
         for (uint i = 0; i < N_SLICE; i++) {
             sh_bitmaps[i][th_ix] = 0;
@@ -320,20 +343,7 @@ void main() {
                     if (!alloc_cmd(cmd_alloc, cmd_ref, cmd_limit)) {
                         break;
                     }
-                    if (fill_mode_from_flags(tag.flags) == MODE_NONZERO) {
+                    write_fill(cmd_alloc, cmd_ref, tag.flags, tile, fill.linewidth);
-                        if (tile.tile.offset != 0) {
-                            CmdFill cmd_fill = CmdFill(tile.tile.offset, tile.backdrop);
-                            Cmd_Fill_write(cmd_alloc, cmd_ref, cmd_fill);
-                            cmd_ref.offset += 4 + CmdFill_size;
-                        } else {
-                            Cmd_Solid_write(cmd_alloc, cmd_ref);
-                            cmd_ref.offset += 4;
-                        }
-                    } else {
-                        CmdStroke cmd_stroke = CmdStroke(tile.tile.offset, 0.5 * fill.linewidth);
-                        Cmd_Stroke_write(cmd_alloc, cmd_ref, cmd_stroke);
-                        cmd_ref.offset += 4 + CmdStroke_size;
-                    }
                     Cmd_Color_write(cmd_alloc, cmd_ref, CmdColor(fill.rgba_color));
                     cmd_ref.offset += 4 + CmdColor_size;
                     break;
@@ -344,20 +354,7 @@ void main() {
                     if (!alloc_cmd(cmd_alloc, cmd_ref, cmd_limit)) {
                         break;
                     }
-                    if (fill_mode_from_flags(tag.flags) == MODE_NONZERO) {
+                    write_fill(cmd_alloc, cmd_ref, tag.flags, tile, fill_img.linewidth);
-                        if (tile.tile.offset != 0) {
-                            CmdFill cmd_fill = CmdFill(tile.tile.offset, tile.backdrop);
-                            Cmd_Fill_write(cmd_alloc, cmd_ref, cmd_fill);
-                            cmd_ref.offset += 4 + CmdFill_size;
-                        } else {
-                            Cmd_Solid_write(cmd_alloc, cmd_ref);
-                            cmd_ref.offset += 4;
-                        }
-                    } else {
-                        CmdStroke cmd_stroke = CmdStroke(tile.tile.offset, 0.5 * fill_img.linewidth);
-                        Cmd_Stroke_write(cmd_alloc, cmd_ref, cmd_stroke);
-                        cmd_ref.offset += 4 + CmdStroke_size;
-                    }
                     Cmd_Image_write(cmd_alloc, cmd_ref, CmdImage(fill_img.index, fill_img.offset));
                     cmd_ref.offset += 4 + CmdImage_size;
                     break;
@@ -373,27 +370,14 @@ void main() {
                         if (!alloc_cmd(cmd_alloc, cmd_ref, cmd_limit)) {
                             break;
                         }
-                        if (fill_mode_from_flags(tag.flags) == MODE_NONZERO) {
+                        write_fill(cmd_alloc, cmd_ref, tag.flags, tile, begin_clip.linewidth);
-                            if (tile.tile.offset != 0) {
-                                CmdFill cmd_fill = CmdFill(tile.tile.offset, tile.backdrop);
-                                Cmd_Fill_write(cmd_alloc, cmd_ref, cmd_fill);
-                                cmd_ref.offset += 4 + CmdFill_size;
-                            } else {
-                                // TODO: here is where a bunch of optimization magic should happen
-                                float alpha = tile.backdrop == 0 ? 0.0 : 1.0;
-                                Cmd_Alpha_write(cmd_alloc, cmd_ref, CmdAlpha(alpha));
-                                cmd_ref.offset += 4 + CmdAlpha_size;
-                            }
-                        } else {
-                            CmdStroke cmd_stroke = CmdStroke(tile.tile.offset, 0.5 * begin_clip.linewidth);
-                            Cmd_Stroke_write(cmd_alloc, cmd_ref, cmd_stroke);
-                            cmd_ref.offset += 4 + CmdStroke_size;
-                        }
                         Cmd_BeginClip_write(cmd_alloc, cmd_ref);
                         cmd_ref.offset += 4;
                         if (clip_depth < 32) {
                             clip_one_mask &= ~(1 << clip_depth);
                         }
+                        begin_slot++;
+                        num_begin_slots = max(num_begin_slots, begin_slot);
                     }
                     clip_depth++;
                     break;
@@ -405,6 +389,7 @@ void main() {
                         }
                         Cmd_Solid_write(cmd_alloc, cmd_ref);
                         cmd_ref.offset += 4;
+                        begin_slot--;
                         Cmd_EndClip_write(cmd_alloc, cmd_ref);
                         cmd_ref.offset += 4;
                     }
@@ -432,5 +417,13 @@ void main() {
     }
     if (bin_tile_x + tile_x < conf.width_in_tiles && bin_tile_y + tile_y < conf.height_in_tiles) {
         Cmd_End_write(cmd_alloc, cmd_ref);
+        if (num_begin_slots > 0) {
+            // Write scratch allocation: one state per BeginClip per rasterizer chunk.
+            uint scratch_size = num_begin_slots * TILE_WIDTH_PX * TILE_HEIGHT_PX * CLIP_STATE_SIZE * 4;
+            MallocResult scratch = malloc(scratch_size);
+            // Ignore scratch.failed; we don't use the allocation and kernel4
+            // checks for memory overflow before using it.
+            alloc_write(scratch_alloc, scratch_alloc.offset, scratch.alloc);
+        }
     }
 }

gpu/shaders/kernel4.comp

@@ -13,9 +13,12 @@
 #include "mem.h"
 #include "setup.h"
 
-#define CHUNK 8
+#define CHUNK_X 2
-#define CHUNK_DY (TILE_HEIGHT_PX / CHUNK)
+#define CHUNK_Y 4
-layout(local_size_x = TILE_WIDTH_PX, local_size_y = CHUNK_DY) in;
+#define CHUNK CHUNK_X * CHUNK_Y
+#define CHUNK_DX (TILE_WIDTH_PX / CHUNK_X)
+#define CHUNK_DY (TILE_HEIGHT_PX / CHUNK_Y)
+layout(local_size_x = CHUNK_DX, local_size_y = CHUNK_DY) in;
 
 layout(set = 0, binding = 1) readonly buffer ConfigBuf {
     Config conf;
@@ -32,36 +35,6 @@ layout(rgba8, set = 0, binding = 3) uniform readonly image2D images[1];
 #include "ptcl.h"
 #include "tile.h"
 
-#define BLEND_STACK_SIZE 4
-
-// Layout of a clip scratch frame:
-// Each frame is WIDTH * HEIGHT ClipStates, then a link reference.
-
-struct ClipState {
-    uint srgb;
-    float area;
-};
-
-// Link offset and frame size in 32-bit words.
-#define CLIP_STATE_SIZE 2
-#define CLIP_LINK_OFFSET (TILE_WIDTH_PX * TILE_HEIGHT_PX * CLIP_STATE_SIZE)
-#define CLIP_BUF_SIZE (CLIP_LINK_OFFSET + 1)
-
-shared MallocResult sh_clip_alloc;
-
-// Allocate a scratch buffer for clipping.
-MallocResult alloc_clip_buf(uint link) {
-    if (gl_LocalInvocationID.x == 0 && gl_LocalInvocationID.y == 0) {
-        MallocResult m = malloc(CLIP_BUF_SIZE * 4);
-        if (!m.failed) {
-            write_mem(m.alloc, (m.alloc.offset >> 2) + CLIP_LINK_OFFSET, link);
-        }
-        sh_clip_alloc = m;
-    }
-    barrier();
-    return sh_clip_alloc;
-}
-
 vec3 tosRGB(vec3 rgb) {
     bvec3 cutoff = greaterThanEqual(rgb, vec3(0.0031308));
     vec3 below = vec3(12.92)*rgb;
@@ -90,10 +63,14 @@ uint packsRGB(vec4 rgba) {
     return packUnorm4x8(rgba.wzyx);
 }
 
+uvec2 chunk_offset(uint i) {
+    return uvec2(i % CHUNK_X * CHUNK_DX, i / CHUNK_X * CHUNK_DY);
+}
+
 vec4[CHUNK] fillImage(uvec2 xy, CmdImage cmd_img) {
     vec4 rgba[CHUNK];
     for (uint i = 0; i < CHUNK; i++) {
-        ivec2 uv = ivec2(xy.x, xy.y + i * CHUNK_DY) + cmd_img.offset;
+        ivec2 uv = ivec2(xy + chunk_offset(i)) + cmd_img.offset;
 #ifdef ENABLE_IMAGE_INDICES
         vec4 fg_rgba = imageLoad(images[cmd_img.index], uv);
 #else
@@ -114,23 +91,24 @@ void main() {
     Alloc cmd_alloc = slice_mem(conf.ptcl_alloc, tile_ix * PTCL_INITIAL_ALLOC, PTCL_INITIAL_ALLOC);
     CmdRef cmd_ref = CmdRef(cmd_alloc.offset);
 
-    uvec2 xy_uint = uvec2(gl_GlobalInvocationID.x, gl_LocalInvocationID.y + TILE_HEIGHT_PX * gl_WorkGroupID.y);
+    // Read scrach space allocation, written first in the command list.
+    Alloc scratch_alloc = alloc_read(cmd_alloc, cmd_ref.offset);
+    cmd_ref.offset += Alloc_size;
+
+    uvec2 xy_uint = uvec2(gl_LocalInvocationID.x + TILE_WIDTH_PX * gl_WorkGroupID.x, gl_LocalInvocationID.y + TILE_HEIGHT_PX * gl_WorkGroupID.y);
     vec2 xy = vec2(xy_uint);
     vec4 rgba[CHUNK];
-    ClipState blend_stack[BLEND_STACK_SIZE][CHUNK];
-    uint blend_spill = 0;
-    uint blend_sp = 0;
-    Alloc clip_tos = new_alloc(0, 0);
     for (uint i = 0; i < CHUNK; i++) {
         rgba[i] = vec4(0.0);
 #ifdef ENABLE_IMAGE_INDICES
         if (xy_uint.x < 1024 && xy_uint.y < 1024) {
-            rgba[i] = imageLoad(images[gl_WorkGroupID.x / 64], ivec2(xy_uint.x, xy_uint.y + CHUNK_DY * i)/4);
+            rgba[i] = imageLoad(images[gl_WorkGroupID.x / 64], ivec2(xy_uint + chunk_offset(i))/4);
         }
 #endif
     }
 
     float area[CHUNK];
+    uint clip_depth = 0;
     while (true) {
         uint tag = Cmd_tag(cmd_alloc, cmd_ref).tag;
         if (tag == Cmd_End) {
@@ -148,7 +126,7 @@ void main() {
                 vec2 line_vec = seg.vector;
                 for (uint k = 0; k < CHUNK; k++) {
                     vec2 dpos = xy + vec2(0.5, 0.5) - seg.origin;
-                    dpos.y += float(k * CHUNK_DY);
+                    dpos += vec2(chunk_offset(k));
                     float t = clamp(dot(line_vec, dpos) / dot(line_vec, line_vec), 0.0, 1.0);
                     df[k] = min(df[k], length(line_vec * t - dpos));
                 }
@@ -167,7 +145,7 @@ void main() {
             do {
                 TileSeg seg = TileSeg_read(new_alloc(tile_seg_ref.offset, TileSeg_size), tile_seg_ref);
                 for (uint k = 0; k < CHUNK; k++) {
-                    vec2 my_xy = vec2(xy.x, xy.y + float(k * CHUNK_DY));
+                    vec2 my_xy = xy + vec2(chunk_offset(k));
                     vec2 start = seg.origin - my_xy;
                     vec2 end = start + seg.vector;
                     vec2 window = clamp(vec2(start.y, end.y), 0.0, 1.0);
@@ -223,47 +201,29 @@ void main() {
             cmd_ref.offset += 4 + CmdImage_size;
             break;
         case Cmd_BeginClip:
-            uint blend_slot = blend_sp % BLEND_STACK_SIZE;
+            uint base_ix = (scratch_alloc.offset >> 2) + CLIP_STATE_SIZE * (clip_depth * TILE_WIDTH_PX * TILE_HEIGHT_PX +
-            if (blend_sp == blend_spill + BLEND_STACK_SIZE) {
+                gl_LocalInvocationID.x + TILE_WIDTH_PX * gl_LocalInvocationID.y);
-                // spill to scratch buffer
-                MallocResult m = alloc_clip_buf(clip_tos.offset);
-                if (m.failed) {
-                    return;
-                }
-                clip_tos = m.alloc;
-                uint base_ix = (clip_tos.offset >> 2) + CLIP_STATE_SIZE * (gl_LocalInvocationID.x + TILE_WIDTH_PX * gl_LocalInvocationID.y);
-                for (uint k = 0; k < CHUNK; k++) {
-                    uint srgb = blend_stack[blend_slot][k].srgb;
-                    uint area = floatBitsToUint(blend_stack[blend_slot][k].area);
-                    write_mem(clip_tos, base_ix + 0 + k * CLIP_STATE_SIZE * TILE_WIDTH_PX * CHUNK_DY, srgb);
-                    write_mem(clip_tos, base_ix + 1 + k * CLIP_STATE_SIZE * TILE_WIDTH_PX * CHUNK_DY, area);
-                }
-                blend_spill++;
-            }
             for (uint k = 0; k < CHUNK; k++) {
-                blend_stack[blend_slot][k] = ClipState(packsRGB(rgba[k]), clamp(abs(area[k]), 0.0, 1.0));
+                uvec2 offset = chunk_offset(k);
+                uint srgb = packsRGB(vec4(rgba[k]));
+                float alpha = clamp(abs(area[k]), 0.0, 1.0);
+                write_mem(scratch_alloc, base_ix + 0 + CLIP_STATE_SIZE * (offset.x + offset.y * TILE_WIDTH_PX), srgb);
+                write_mem(scratch_alloc, base_ix + 1 + CLIP_STATE_SIZE * (offset.x + offset.y * TILE_WIDTH_PX), floatBitsToUint(alpha));
                 rgba[k] = vec4(0.0);
             }
-            blend_sp++;
+            clip_depth++;
             cmd_ref.offset += 4;
             break;
         case Cmd_EndClip:
-            blend_slot = (blend_sp - 1) % BLEND_STACK_SIZE;
+            clip_depth--;
-            if (blend_sp == blend_spill) {
+            base_ix = (scratch_alloc.offset >> 2) + CLIP_STATE_SIZE * (clip_depth * TILE_WIDTH_PX * TILE_HEIGHT_PX +
-                uint base_ix = (clip_tos.offset >> 2) + CLIP_STATE_SIZE * (gl_LocalInvocationID.x + TILE_WIDTH_PX * gl_LocalInvocationID.y);
+                gl_LocalInvocationID.x + TILE_WIDTH_PX * gl_LocalInvocationID.y);
-                for (uint k = 0; k < CHUNK; k++) {
-                    uint srgb = read_mem(clip_tos, base_ix + 0 + k * CLIP_STATE_SIZE * TILE_WIDTH_PX * CHUNK_DY);
-                    uint area = read_mem(clip_tos, base_ix + 1 + k * CLIP_STATE_SIZE * TILE_WIDTH_PX * CHUNK_DY);
-                    ClipState state = ClipState(srgb, uintBitsToFloat(area));
-                    blend_stack[blend_slot][k] = state;
-                }
-                clip_tos.offset = read_mem(clip_tos, (clip_tos.offset >> 2) + CLIP_LINK_OFFSET);
-                blend_spill--;
-            }
-            blend_sp--;
             for (uint k = 0; k < CHUNK; k++) {
-                vec4 bg = unpacksRGB(blend_stack[blend_slot][k].srgb);
+                uvec2 offset = chunk_offset(k);
-                vec4 fg = rgba[k] * area[k] * blend_stack[blend_slot][k].area;
+                uint srgb = read_mem(scratch_alloc, base_ix + 0 + CLIP_STATE_SIZE * (offset.x + offset.y * TILE_WIDTH_PX));
+                uint alpha = read_mem(scratch_alloc, base_ix + 1 + CLIP_STATE_SIZE * (offset.x + offset.y * TILE_WIDTH_PX));
+                vec4 bg = unpacksRGB(srgb);
+                vec4 fg = rgba[k] * area[k] * uintBitsToFloat(alpha);
                 rgba[k] = bg * (1.0 - fg.a) + fg;
             }
             cmd_ref.offset += 4;
@@ -276,6 +236,6 @@ void main() {
     }
 
     for (uint i = 0; i < CHUNK; i++) {
-        imageStore(image, ivec2(xy_uint.x, xy_uint.y + CHUNK_DY * i), vec4(tosRGB(rgba[i].rgb), rgba[i].a));
+        imageStore(image, ivec2(xy_uint + chunk_offset(i)), vec4(tosRGB(rgba[i].rgb), rgba[i].a));
     }
 }

gpu/shaders/mem.h

@@ -21,7 +21,11 @@ layout(set = 0, binding = 0) buffer Memory {
 #define ERR_OUT_OF_BOUNDS 2
 #define ERR_UNALIGNED_ACCESS 3
 
+#ifdef MEM_DEBUG
+#define Alloc_size 16
+#else
 #define Alloc_size 8
+#endif
 
 // Alloc represents a memory allocation.
 struct Alloc {
@@ -39,7 +43,7 @@ struct MallocResult {
     bool failed;
 };
 
-// new_alloc synthesizes an Alloc when its offset and size are derived.
+// new_alloc synthesizes an Alloc from an offset and size.
 Alloc new_alloc(uint offset, uint size) {
     Alloc a;
     a.offset = offset;
@@ -118,3 +122,21 @@ Alloc slice_mem(Alloc a, uint offset, uint size) {
 #endif
     return new_alloc(a.offset + offset, size);
 }
+
+// alloc_write writes alloc to memory at offset bytes.
+void alloc_write(Alloc a, uint offset, Alloc alloc) {
+    write_mem(a, offset >> 2, alloc.offset);
+#ifdef MEM_DEBUG
+    write_mem(a, (offset >> 2) + 1, alloc.size);
+#endif
+}
+
+// alloc_read reads an Alloc from memory at offset bytes.
+Alloc alloc_read(Alloc a, uint offset) {
+    Alloc alloc;
+    alloc.offset = read_mem(a, offset >> 2);
+#ifdef MEM_DEBUG
+    alloc.size = read_mem(a, (offset >> 2) + 1);
+#endif
+    return alloc;
+}

gpu/shaders/setup.h

@@ -42,7 +42,10 @@ struct Config {
 #define MODE_NONZERO 0
 #define MODE_STROKE 1
 
+// Size of kernel4 clip state, in words.
+#define CLIP_STATE_SIZE 2
+
 // fill_mode_from_flags extracts the fill mode from tag flags.
 uint fill_mode_from_flags(uint flags) {
-	return flags & 0x1;
+    return flags & 0x1;
 }