alex/vello
1
0
Fork 0
mirror of https://github.com/italicsjenga/vello.git synced 2025-01-09 20:31:29 +11:00
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #16 from linebender/new_merge

Browse source 
More parallel merging in coarse raster
This commit is contained in:
Raph Levien 2020-05-31 09:53:20 -07:00 committed by GitHub
parent 9c0bdd664d 2c185c3718
commit f6ef1c16ab
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
6 changed files with 206 additions and 255 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

4
piet-gpu/bin/cli.rs
View file

@ -184,8 +184,8 @@ fn main() -> Result<(), Error> {
/* /*
let mut data: Vec<u32> = Default::default(); let mut data: Vec<u32> = Default::default();
device.read_buffer(&renderer.ptcl_buf, &mut data).unwrap(); device.read_buffer(&renderer.ptcl_buf, &mut data).unwrap();
//piet_gpu::dump_k1_data(&data); piet_gpu::dump_k1_data(&data);
trace_ptcl(&data); //trace_ptcl(&data);
*/ */
let mut img_data: Vec<u8> = Default::default(); let mut img_data: Vec<u8> = Default::default();

300
piet-gpu/shader/binning.comp
View file

@ -43,10 +43,7 @@ layout(set = 0, binding = 3) buffer BinsBuf {
// Note: cudaraster has N_TILE + 1 to cut down on bank conflicts. // Note: cudaraster has N_TILE + 1 to cut down on bank conflicts.
shared uint bitmaps[N_SLICE][N_TILE]; shared uint bitmaps[N_SLICE][N_TILE];
shared uint count[N_SLICE][N_TILE]; shared uint count[N_SLICE][N_TILE];
shared uint sh_my_tile;
shared uint sh_chunk_start[N_TILE]; shared uint sh_chunk_start[N_TILE];
shared uint sh_chunk_end[N_TILE];
shared uint sh_chunk_jump[N_TILE];
shared float sh_right_edge[N_TILE]; shared float sh_right_edge[N_TILE];
@ -57,179 +54,140 @@ uint state_right_edge_index(uint partition_ix) {
} }
void main() { void main() {
BinChunkRef chunk_ref = BinChunkRef((gl_LocalInvocationID.x * N_WG + gl_WorkGroupID.x) * BIN_INITIAL_ALLOC);
uint wr_limit = chunk_ref.offset + BIN_INITIAL_ALLOC;
uint chunk_n = 0; uint chunk_n = 0;
uint my_n_elements = n_elements; uint my_n_elements = n_elements;
while (true) { uint my_partition = gl_WorkGroupID.x;
if (gl_LocalInvocationID.x == 0) {
sh_my_tile = atomicAdd(tile_ix, 1); for (uint i = 0; i < N_SLICE; i++) {
bitmaps[i][gl_LocalInvocationID.x] = 0;
}
barrier();
// Read inputs and determine coverage of bins
uint element_ix = my_partition * N_TILE + gl_LocalInvocationID.x;
AnnotatedRef ref = AnnotatedRef(element_ix * Annotated_size);
uint tag = Annotated_Nop;
if (element_ix < my_n_elements) {
tag = Annotated_tag(ref);
}
int x0 = 0, y0 = 0, x1 = 0, y1 = 0;
float my_right_edge = INFINITY;
bool crosses_edge = false;
switch (tag) {
case Annotated_FillLine:
case Annotated_StrokeLine:
AnnoStrokeLineSeg line = Annotated_StrokeLine_read(ref);
x0 = int(floor((min(line.p0.x, line.p1.x) - line.stroke.x) * SX));
y0 = int(floor((min(line.p0.y, line.p1.y) - line.stroke.y) * SY));
x1 = int(ceil((max(line.p0.x, line.p1.x) + line.stroke.x) * SX));
y1 = int(ceil((max(line.p0.y, line.p1.y) + line.stroke.y) * SY));
crosses_edge = tag == Annotated_FillLine && ceil(line.p0.y * TSY) != ceil(line.p1.y * TSY);
break;
case Annotated_Fill:
case Annotated_Stroke:
// Note: we take advantage of the fact that fills and strokes
// have compatible layout.
AnnoFill fill = Annotated_Fill_read(ref);
x0 = int(floor(fill.bbox.x * SX));
y0 = int(floor(fill.bbox.y * SY));
x1 = int(ceil(fill.bbox.z * SX));
y1 = int(ceil(fill.bbox.w * SY));
// It probably makes more sense to track x1, to avoid having to redo
// the rounding to tile coords.
my_right_edge = fill.bbox.z;
break;
}
// If the last element in this partition is a fill edge, then we need to do a
// look-forward to find the right edge of its corresponding fill. That data is
// recorded in aggregates computed in the element processing pass.
if (gl_LocalInvocationID.x == N_TILE - 1 && tag == Annotated_FillLine) {
uint aggregate_ix = (my_partition + 1) * ELEMENT_BINNING_RATIO;
// This is sequential but the expectation is that the amount of
// look-forward is small (performance may degrade in the case
// of massively complex paths).
do {
my_right_edge = uintBitsToFloat(state[state_right_edge_index(aggregate_ix)]);
aggregate_ix++;
} while (isinf(my_right_edge));
}
// Now propagate right_edge backward, from fill to segment.
for (uint i = 0; i < LG_N_TILE; i++) {
// Note: we could try to cut down on write bandwidth here if the value hasn't
// changed, but not sure it's worth the complexity to track.
sh_right_edge[gl_LocalInvocationID.x] = my_right_edge;
barrier();
if (gl_LocalInvocationID.x + (1 << i) < N_TILE && isinf(my_right_edge)) {
my_right_edge = sh_right_edge[gl_LocalInvocationID.x + (1 << i)];
} }
barrier(); barrier();
uint my_tile = sh_my_tile; }
if (my_tile * N_TILE >= my_n_elements) { if (crosses_edge) {
break; x1 = int(ceil(my_right_edge * SX));
} }
for (uint i = 0; i < N_SLICE; i++) { // At this point, we run an iterator over the coverage area,
bitmaps[i][gl_LocalInvocationID.x] = 0; // trying to keep divergence low.
} // Right now, it's just a bbox, but we'll get finer with
barrier(); // segments.
x0 = clamp(x0, 0, N_TILE_X);
// Read inputs and determine coverage of bins x1 = clamp(x1, x0, N_TILE_X);
uint element_ix = my_tile * N_TILE + gl_LocalInvocationID.x; y0 = clamp(y0, 0, N_TILE_Y);
AnnotatedRef ref = AnnotatedRef(element_ix * Annotated_size); y1 = clamp(y1, y0, N_TILE_Y);
uint tag = Annotated_Nop; if (x0 == x1) y1 = y0;
if (element_ix < my_n_elements) { int x = x0, y = y0;
tag = Annotated_tag(ref); uint my_slice = gl_LocalInvocationID.x / 32;
} uint my_mask = 1 << (gl_LocalInvocationID.x & 31);
int x0 = 0, y0 = 0, x1 = 0, y1 = 0; while (y < y1) {
float my_right_edge = INFINITY; atomicOr(bitmaps[my_slice][y * N_TILE_X + x], my_mask);
bool crosses_edge = false; x++;
switch (tag) { if (x == x1) {
case Annotated_FillLine: x = x0;
case Annotated_StrokeLine: y++;
AnnoStrokeLineSeg line = Annotated_StrokeLine_read(ref); }
x0 = int(floor((min(line.p0.x, line.p1.x) - line.stroke.x) * SX)); }
y0 = int(floor((min(line.p0.y, line.p1.y) - line.stroke.y) * SY));
x1 = int(ceil((max(line.p0.x, line.p1.x) + line.stroke.x) * SX)); barrier();
y1 = int(ceil((max(line.p0.y, line.p1.y) + line.stroke.y) * SY)); // Allocate output segments.
crosses_edge = tag == Annotated_FillLine && ceil(line.p0.y * TSY) != ceil(line.p1.y * TSY); uint element_count = 0;
break; for (uint i = 0; i < N_SLICE; i++) {
case Annotated_Fill: element_count += bitCount(bitmaps[i][gl_LocalInvocationID.x]);
case Annotated_Stroke: count[i][gl_LocalInvocationID.x] = element_count;
// Note: we take advantage of the fact that fills and strokes }
// have compatible layout. // element_count is number of elements covering bin for this invocation.
AnnoFill fill = Annotated_Fill_read(ref); uint chunk_start = 0;
x0 = int(floor(fill.bbox.x * SX)); if (element_count != 0) {
y0 = int(floor(fill.bbox.y * SY)); // TODO: aggregate atomic adds (subgroup is probably fastest)
x1 = int(ceil(fill.bbox.z * SX)); chunk_start = atomicAdd(alloc, element_count * BinInstance_size);
y1 = int(ceil(fill.bbox.w * SY)); sh_chunk_start[gl_LocalInvocationID.x] = chunk_start;
// It probably makes more sense to track x1, to avoid having to redo }
// the rounding to tile coords. // Note: it might be more efficient for reading to do this in the
my_right_edge = fill.bbox.z; // other order (each bin is a contiguous sequence of partitions)
break; uint out_ix = (my_partition * N_TILE + gl_LocalInvocationID.x) * 2;
} bins[out_ix] = element_count;
bins[out_ix + 1] = chunk_start;
// If the last element in this partition is a fill edge, then we need to do a
// look-forward to find the right edge of its corresponding fill. That data is barrier();
// recorded in aggregates computed in the element processing pass. // Use similar strategy as Laine & Karras paper; loop over bbox of bins
if (gl_LocalInvocationID.x == N_TILE - 1 && tag == Annotated_FillLine) { // touched by this element
uint aggregate_ix = (my_tile + 1) * ELEMENT_BINNING_RATIO; x = x0;
// This is sequential but the expectation is that the amount of y = y0;
// look-forward is small (performance may degrade in the case while (y < y1) {
// of massively complex paths). uint bin_ix = y * N_TILE_X + x;
do { uint out_mask = bitmaps[my_slice][bin_ix];
my_right_edge = uintBitsToFloat(state[state_right_edge_index(aggregate_ix)]); if ((out_mask & my_mask) != 0) {
aggregate_ix++; uint idx = bitCount(out_mask & (my_mask - 1));
} while (isinf(my_right_edge)); if (my_slice > 0) {
} idx += count[my_slice - 1][bin_ix];
}
// Now propagate right_edge backward, from fill to segment. uint out_offset = sh_chunk_start[bin_ix] + idx * BinInstance_size;
for (uint i = 0; i < LG_N_TILE; i++) { BinInstance_write(BinInstanceRef(out_offset), BinInstance(element_ix, my_right_edge));
// Note: we could try to cut down on write bandwidth here if the value hasn't }
// changed, but not sure it's worth the complexity to track. x++;
sh_right_edge[gl_LocalInvocationID.x] = my_right_edge; if (x == x1) {
barrier(); x = x0;
if (gl_LocalInvocationID.x + (1 << i) < N_TILE && isinf(my_right_edge)) { y++;
my_right_edge = sh_right_edge[gl_LocalInvocationID.x + (1 << i)];
}
barrier();
}
if (crosses_edge) {
x1 = int(ceil(my_right_edge * SX));
}
// At this point, we run an iterator over the coverage area,
// trying to keep divergence low.
// Right now, it's just a bbox, but we'll get finer with
// segments.
x0 = clamp(x0, 0, N_TILE_X);
x1 = clamp(x1, x0, N_TILE_X);
y0 = clamp(y0, 0, N_TILE_Y);
y1 = clamp(y1, y0, N_TILE_Y);
if (x0 == x1) y1 = y0;
int x = x0, y = y0;
uint my_slice = gl_LocalInvocationID.x / 32;
uint my_mask = 1 << (gl_LocalInvocationID.x & 31);
while (y < y1) {
atomicOr(bitmaps[my_slice][y * N_TILE_X + x], my_mask);
x++;
if (x == x1) {
x = x0;
y++;
}
}
barrier();
// Allocate output segments.
uint element_count = 0;
for (uint i = 0; i < N_SLICE; i++) {
element_count += bitCount(bitmaps[i][gl_LocalInvocationID.x]);
count[i][gl_LocalInvocationID.x] = element_count;
}
// element_count is number of elements covering bin for this invocation.
if (element_count != 0) {
uint chunk_end;
uint chunk_new_start;
// Refactor to reduce code duplication?
if (chunk_n > 0) {
uint next_chunk = chunk_ref.offset + BinChunk_size + chunk_n * BinInstance_size;
if (next_chunk + BinChunk_size + min(24, element_count * BinInstance_size) > wr_limit) {
uint alloc_amount = max(BIN_ALLOC, BinChunk_size + element_count * BinInstance_size);
// could try to reduce fragmentation if BIN_ALLOC is only a bit above needed
next_chunk = atomicAdd(alloc, alloc_amount);
wr_limit = next_chunk + alloc_amount;
}
BinChunk_write(chunk_ref, BinChunk(chunk_n, BinChunkRef(next_chunk)));
chunk_ref = BinChunkRef(next_chunk);
}
BinInstanceRef instance_ref = BinInstanceRef(chunk_ref.offset + BinChunk_size);
if (instance_ref.offset + element_count * BinInstance_size > wr_limit) {
chunk_end = wr_limit;
chunk_n = (wr_limit - instance_ref.offset) / BinInstance_size;
uint alloc_amount = max(BIN_ALLOC, BinChunk_size + (element_count - chunk_n) * BinInstance_size);
chunk_new_start = atomicAdd(alloc, alloc_amount);
wr_limit = chunk_new_start + alloc_amount;
BinChunk_write(chunk_ref, BinChunk(chunk_n, BinChunkRef(chunk_new_start)));
chunk_ref = BinChunkRef(chunk_new_start);
chunk_new_start += BinChunk_size;
chunk_n = element_count - chunk_n;
} else {
chunk_end = ~0;
chunk_new_start = ~0;
chunk_n = element_count;
}
sh_chunk_start[gl_LocalInvocationID.x] = instance_ref.offset;
sh_chunk_end[gl_LocalInvocationID.x] = chunk_end;
sh_chunk_jump[gl_LocalInvocationID.x] = chunk_new_start - chunk_end;
}
barrier();
// Use similar strategy as Laine & Karras paper; loop over bbox of bins
// touched by this element
x = x0;
y = y0;
while (y < y1) {
uint bin_ix = y * N_TILE_X + x;
uint out_mask = bitmaps[my_slice][bin_ix];
if ((out_mask & my_mask) != 0) {
uint idx = bitCount(out_mask & (my_mask - 1));
if (my_slice > 0) {
idx += count[my_slice - 1][bin_ix];
}
uint out_offset = sh_chunk_start[bin_ix] + idx * BinInstance_size;
if (out_offset >= sh_chunk_end[bin_ix]) {
out_offset += sh_chunk_jump[bin_ix];
}
BinInstance_write(BinInstanceRef(out_offset), BinInstance(element_ix, my_right_edge));
}
x++;
if (x == x1) {
x = x0;
y++;
}
} }
} }
BinChunk_write(chunk_ref, BinChunk(chunk_n, BinChunkRef(0)));
} }

BIN
piet-gpu/shader/binning.spv
View file

Binary file not shown.

146
piet-gpu/shader/coarse.comp
View file

@ -16,6 +16,7 @@ layout(set = 0, binding = 1) buffer BinsBuf {
}; };
layout(set = 0, binding = 2) buffer AllocBuf { layout(set = 0, binding = 2) buffer AllocBuf {
uint n_elements;
uint alloc; uint alloc;
}; };
@ -27,19 +28,15 @@ layout(set = 0, binding = 3) buffer PtclBuf {
#include "bins.h" #include "bins.h"
#include "ptcl.h" #include "ptcl.h"
#define N_RINGBUF 512 #define LG_N_PART_READ 8
#define N_PART_READ (1 << LG_N_PART_READ)
shared uint sh_elements[N_RINGBUF]; shared uint sh_elements[N_TILE];
shared float sh_right_edge[N_RINGBUF]; shared float sh_right_edge[N_TILE];
shared uint sh_chunk[N_WG];
shared uint sh_chunk_next[N_WG]; // Number of elements in the partition; prefix sum.
shared uint sh_chunk_n[N_WG]; shared uint sh_part_count[N_PART_READ];
shared uint sh_min_buf; shared uint sh_part_elements[N_PART_READ];
// Some of these are kept in shared memory to ease register
// pressure, but it could go either way.
shared uint sh_first_el[N_WG];
shared uint sh_selected_n;
shared uint sh_elements_ref;
shared uint sh_bitmaps[N_SLICE][N_TILE]; shared uint sh_bitmaps[N_SLICE][N_TILE];
shared uint sh_backdrop[N_SLICE][N_TILE]; shared uint sh_backdrop[N_SLICE][N_TILE];
@ -96,14 +93,16 @@ void main() {
// Could use either linear or 2d layouts for both dispatch and // Could use either linear or 2d layouts for both dispatch and
// invocations within the workgroup. We'll use variables to abstract. // invocations within the workgroup. We'll use variables to abstract.
uint bin_ix = N_TILE_X * gl_WorkGroupID.y + gl_WorkGroupID.x; uint bin_ix = N_TILE_X * gl_WorkGroupID.y + gl_WorkGroupID.x;
uint partition_ix = 0;
uint n_partitions = (n_elements + N_TILE - 1) / N_TILE;
// Top left coordinates of this bin. // Top left coordinates of this bin.
vec2 xy0 = vec2(N_TILE_X * TILE_WIDTH_PX * gl_WorkGroupID.x, N_TILE_Y * TILE_HEIGHT_PX * gl_WorkGroupID.y); vec2 xy0 = vec2(N_TILE_X * TILE_WIDTH_PX * gl_WorkGroupID.x, N_TILE_Y * TILE_HEIGHT_PX * gl_WorkGroupID.y);
uint th_ix = gl_LocalInvocationID.x; uint th_ix = gl_LocalInvocationID.x;
uint tile_x = N_TILE_X * gl_WorkGroupID.x + gl_LocalInvocationID.x % N_TILE_X; uint tile_x = N_TILE_X * gl_WorkGroupID.x + gl_LocalInvocationID.x % N_TILE_X;
uint tile_y = N_TILE_Y * gl_WorkGroupID.y + gl_LocalInvocationID.x / N_TILE_X; uint tile_y = N_TILE_Y * gl_WorkGroupID.y + gl_LocalInvocationID.x / N_TILE_X;
uint tile_ix = tile_y * WIDTH_IN_TILES + tile_x; uint this_tile_ix = tile_y * WIDTH_IN_TILES + tile_x;
CmdRef cmd_ref = CmdRef(tile_ix * PTCL_INITIAL_ALLOC); CmdRef cmd_ref = CmdRef(this_tile_ix * PTCL_INITIAL_ALLOC);
uint cmd_limit = cmd_ref.offset + PTCL_INITIAL_ALLOC - 2 * Cmd_size; uint cmd_limit = cmd_ref.offset + PTCL_INITIAL_ALLOC - 2 * Cmd_size;
// Allocation and management of segment output // Allocation and management of segment output
@ -113,18 +112,14 @@ void main() {
SegmentRef last_chunk_segs = SegmentRef(0); SegmentRef last_chunk_segs = SegmentRef(0);
alloc_chunk_remaining = 0; alloc_chunk_remaining = 0;
uint wr_ix = 0; // I'm sure we can figure out how to do this with at least one fewer register...
// Items up to rd_ix have been read from sh_elements
uint rd_ix = 0; uint rd_ix = 0;
uint first_el; // Items up to wr_ix have been written into sh_elements
if (th_ix < N_WG) { uint wr_ix = 0;
uint start_chunk = (bin_ix * N_WG + th_ix) * BIN_INITIAL_ALLOC; // Items between part_start_ix and ready_ix are ready to be transferred from sh_part_elements
sh_chunk[th_ix] = start_chunk; uint part_start_ix = 0;
BinChunk chunk = BinChunk_read(BinChunkRef(start_chunk)); uint ready_ix = 0;
sh_chunk_n[th_ix] = chunk.n;
sh_chunk_next[th_ix] = chunk.next.offset;
sh_first_el[th_ix] = chunk.n > 0 ?
BinInstance_read(BinInstanceRef(start_chunk + BinChunk_size)).element_ix : ~0;
}
if (th_ix < N_SLICE) { if (th_ix < N_SLICE) {
sh_bd_sign[th_ix] = 0; sh_bd_sign[th_ix] = 0;
} }
@ -138,56 +133,56 @@ void main() {
sh_is_segment[th_ix] = 0; sh_is_segment[th_ix] = 0;
} }
while (wr_ix - rd_ix <= N_TILE) { // parallel read of input partitions
// Choose segment with least element. do {
uint my_min; if (ready_ix == wr_ix && partition_ix < n_partitions) {
if (th_ix < N_WG) { part_start_ix = ready_ix;
if (th_ix == 0) { uint count = 0;
sh_selected_n = 0; if (th_ix < N_PART_READ && partition_ix + th_ix < n_partitions) {
sh_min_buf = ~0; uint in_ix = ((partition_ix + th_ix) * N_TILE + bin_ix) * 2;
count = bins[in_ix];
sh_part_elements[th_ix] = bins[in_ix + 1];
} }
// prefix sum of counts
for (uint i = 0; i < LG_N_PART_READ; i++) {
if (th_ix < N_PART_READ) {
sh_part_count[th_ix] = count;
}
barrier();
if (th_ix < N_PART_READ) {
if (th_ix >= (1 << i)) {
count += sh_part_count[th_ix - (1 << i)];
}
}
barrier();
}
if (th_ix < N_PART_READ) {
sh_part_count[th_ix] = part_start_ix + count;
}
barrier();
ready_ix = sh_part_count[N_PART_READ - 1];
partition_ix += N_PART_READ;
} }
barrier(); // use binary search to find element to read
// Tempting to do this with subgroups, but atomic should be good enough. uint ix = rd_ix + th_ix;
if (th_ix < N_WG) { if (ix >= wr_ix && ix < ready_ix) {
my_min = sh_first_el[th_ix]; uint part_ix = 0;
atomicMin(sh_min_buf, my_min); for (uint i = 0; i < LG_N_PART_READ; i++) {
} uint probe = part_ix + ((N_PART_READ / 2) >> i);
barrier(); if (ix >= sh_part_count[probe - 1]) {
if (th_ix < N_WG) { part_ix = probe;
if (my_min == sh_min_buf && my_min != ~0) {
sh_elements_ref = sh_chunk[th_ix] + BinChunk_size;
uint selected_n = sh_chunk_n[th_ix];
sh_selected_n = selected_n;
uint next_chunk = sh_chunk_next[th_ix];
if (next_chunk == 0) {
sh_first_el[th_ix] = ~0;
} else {
sh_chunk[th_ix] = next_chunk;
BinChunk chunk = BinChunk_read(BinChunkRef(next_chunk));
sh_chunk_n[th_ix] = chunk.n;
sh_chunk_next[th_ix] = chunk.next.offset;
sh_first_el[th_ix] = BinInstance_read(
BinInstanceRef(next_chunk + BinChunk_size)).element_ix;
} }
} }
ix -= part_ix > 0 ? sh_part_count[part_ix - 1] : part_start_ix;
BinInstanceRef inst_ref = BinInstanceRef(sh_part_elements[part_ix]);
BinInstance inst = BinInstance_read(BinInstance_index(inst_ref, ix));
sh_elements[th_ix] = inst.element_ix;
sh_right_edge[th_ix] = inst.right_edge;
} }
barrier(); barrier();
uint chunk_n = sh_selected_n;
if (chunk_n == 0) { wr_ix = min(rd_ix + N_TILE, ready_ix);
// All chunks consumed } while (wr_ix - rd_ix < N_TILE && (wr_ix < ready_ix || partition_ix < n_partitions));
break;
}
BinInstanceRef inst_ref = BinInstanceRef(sh_elements_ref);
if (th_ix < chunk_n) {
BinInstance inst = BinInstance_read(BinInstance_index(inst_ref, th_ix));
uint wr_el_ix = (wr_ix + th_ix) % N_RINGBUF;
sh_elements[wr_el_ix] = inst.element_ix;
sh_right_edge[wr_el_ix] = inst.right_edge;
}
wr_ix += chunk_n;
}
barrier();
// We've done the merge and filled the buffer. // We've done the merge and filled the buffer.
@ -196,9 +191,8 @@ void main() {
AnnotatedRef ref; AnnotatedRef ref;
float right_edge = 0.0; float right_edge = 0.0;
if (th_ix + rd_ix < wr_ix) { if (th_ix + rd_ix < wr_ix) {
uint rd_el_ix = (rd_ix + th_ix) % N_RINGBUF; uint element_ix = sh_elements[th_ix];
uint element_ix = sh_elements[rd_el_ix]; right_edge = sh_right_edge[th_ix];
right_edge = sh_right_edge[rd_el_ix];
ref = AnnotatedRef(element_ix * Annotated_size); ref = AnnotatedRef(element_ix * Annotated_size);
tag = Annotated_tag(ref); tag = Annotated_tag(ref);
} }
@ -356,7 +350,7 @@ void main() {
} }
} }
uint out_offset = seg_alloc + Segment_size * ix + SegChunk_size; uint out_offset = seg_alloc + Segment_size * ix + SegChunk_size;
uint rd_el_ix = (rd_ix + slice_ix * 32 + bit_ix) % N_RINGBUF; uint rd_el_ix = slice_ix * 32 + bit_ix;
uint element_ix = sh_elements[rd_el_ix]; uint element_ix = sh_elements[rd_el_ix];
ref = AnnotatedRef(element_ix * Annotated_size); ref = AnnotatedRef(element_ix * Annotated_size);
AnnoFillLineSeg line = Annotated_FillLine_read(ref); AnnoFillLineSeg line = Annotated_FillLine_read(ref);
@ -408,7 +402,7 @@ void main() {
} }
} }
uint element_ref_ix = slice_ix * 32 + findLSB(nonseg_bitmap); uint element_ref_ix = slice_ix * 32 + findLSB(nonseg_bitmap);
uint element_ix = sh_elements[(rd_ix + element_ref_ix) % N_RINGBUF]; uint element_ix = sh_elements[element_ref_ix];
// Bits up to and including the lsb // Bits up to and including the lsb
uint bd_mask = (nonseg_bitmap - 1) ^ nonseg_bitmap; uint bd_mask = (nonseg_bitmap - 1) ^ nonseg_bitmap;
@ -526,9 +520,7 @@ void main() {
barrier(); barrier();
rd_ix += N_TILE; rd_ix += N_TILE;
// The second disjunct is there as a strange workaround on Nvidia. If it is if (rd_ix >= ready_ix && partition_ix >= n_partitions) break;
// removed, then the kernel fails with ERROR_DEVICE_LOST.
if (rd_ix >= wr_ix || bin_ix == ~0) break;
} }
Cmd_End_write(cmd_ref); Cmd_End_write(cmd_ref);
} }

BIN
piet-gpu/shader/coarse.spv
View file

Binary file not shown.

11
piet-gpu/src/lib.rs
View file

@ -176,12 +176,12 @@ impl<D: Device> Renderer<D> {
let bin_alloc_buf_dev = device.create_buffer(12, dev)?; let bin_alloc_buf_dev = device.create_buffer(12, dev)?;
// TODO: constants // TODO: constants
let bin_alloc_start = 256 * 64 * N_WG; let bin_alloc_start = ((n_elements + 255) & !255) * 8;
device device
.write_buffer(&bin_alloc_buf_host, &[ .write_buffer(&bin_alloc_buf_host, &[
n_elements as u32, n_elements as u32,
0, 0,
bin_alloc_start, bin_alloc_start as u32,
]) ])
?; ?;
let bin_code = include_bytes!("../shader/binning.spv"); let bin_code = include_bytes!("../shader/binning.spv");
@ -192,12 +192,13 @@ impl<D: Device> Renderer<D> {
&[], &[],
)?; )?;
let coarse_alloc_buf_host = device.create_buffer(4, host)?; let coarse_alloc_buf_host = device.create_buffer(8, host)?;
let coarse_alloc_buf_dev = device.create_buffer(4, dev)?; let coarse_alloc_buf_dev = device.create_buffer(8, dev)?;
let coarse_alloc_start = WIDTH_IN_TILES * HEIGHT_IN_TILES * PTCL_INITIAL_ALLOC; let coarse_alloc_start = WIDTH_IN_TILES * HEIGHT_IN_TILES * PTCL_INITIAL_ALLOC;
device device
.write_buffer(&coarse_alloc_buf_host, &[ .write_buffer(&coarse_alloc_buf_host, &[
n_elements as u32,
coarse_alloc_start as u32, coarse_alloc_start as u32,
]) ])
?; ?;
@ -264,7 +265,7 @@ impl<D: Device> Renderer<D> {
cmd_buf.dispatch( cmd_buf.dispatch(
&self.bin_pipeline, &self.bin_pipeline,
&self.bin_ds, &self.bin_ds,
(N_WG, 1, 1), (((self.n_elements + 255) / 256) as u32, 1, 1),
); );
cmd_buf.write_timestamp(&query_pool, 2); cmd_buf.write_timestamp(&query_pool, 2);
cmd_buf.memory_barrier(); cmd_buf.memory_barrier();