AlexIndustrial/mesa
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

d3d12: Move "fake" SO buffer handling to compute transforms instead of CPU readback

Browse Source 
Reviewed-by: Sil Vilerino <sivileri@microsoft.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/14787>
This commit is contained in:
Jesse Natalie
2022-01-26 14:34:47 -08:00
parent c1b52d8c3a
commit 7009d23865
2 changed files with 77 additions and 34 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/gallium/drivers/d3d12/d3d12_context.cpp
+77 -33
View File
@@ -1544,7 +1544,7 @@ d3d12_set_stream_output_targets(struct pipe_context *pctx,
if (target) {
/* Sub-allocate a new fill buffer each time to avoid GPU/CPU synchronization */
if (offsets[i] != ~0u) {
u_suballocator_alloc(&ctx->so_allocator, sizeof(uint32_t), 4,
u_suballocator_alloc(&ctx->so_allocator, sizeof(uint32_t), 16,
&target->fill_buffer_offset, &target->fill_buffer);
update_so_fill_buffer_count(ctx, target->fill_buffer, target->fill_buffer_offset, offsets[i]);
}
@@ -1769,7 +1769,6 @@ d3d12_enable_fake_so_buffers(struct d3d12_context *ctx, unsigned factor)
pipe_resource_reference(&fake_target->base.buffer, prev_target->base.buffer);
pipe_resource_reference(&fake_target->fill_buffer, prev_target->fill_buffer);
fake_target->fill_buffer_offset = prev_target->fill_buffer_offset;
fake_target->cached_filled_size = prev_target->cached_filled_size;
break;
}
}
@@ -1780,16 +1779,14 @@ d3d12_enable_fake_so_buffers(struct d3d12_context *ctx, unsigned factor)
PIPE_BIND_STREAM_OUTPUT,
PIPE_USAGE_STAGING,
target->base.buffer->width0 * factor);
u_suballocator_alloc(&ctx->so_allocator, sizeof(uint32_t), 4,
u_suballocator_alloc(&ctx->so_allocator, sizeof(uint32_t) * 5, 256,
&fake_target->fill_buffer_offset, &fake_target->fill_buffer);
update_so_fill_buffer_count(ctx, fake_target->fill_buffer, fake_target->fill_buffer_offset, 0);
pipe_buffer_read(&ctx->base, target->fill_buffer,
target->fill_buffer_offset, sizeof(uint32_t),
&fake_target->cached_filled_size);
}
fake_target->base.buffer_offset = target->base.buffer_offset * factor;
fake_target->base.buffer_size = (target->base.buffer_size - fake_target->cached_filled_size) * factor;
/* TODO: This will mess with SO statistics/overflow queries, but we're already missing things there */
fake_target->base.buffer_size = target->base.buffer_size * factor;
ctx->fake_so_targets[i] = &fake_target->base;
fill_stream_output_buffer_view(&ctx->fake_so_buffer_views[i], fake_target);
}
@@ -1808,40 +1805,84 @@ d3d12_disable_fake_so_buffers(struct d3d12_context *ctx)
d3d12_flush_cmdlist_and_wait(ctx);
bool cs_state_saved = false;
d3d12_compute_transform_save_restore save;
for (unsigned i = 0; i < ctx->gfx_pipeline_state.num_so_targets; ++i) {
struct d3d12_stream_output_target *target = (struct d3d12_stream_output_target *)ctx->so_targets[i];
struct d3d12_stream_output_target *fake_target = (struct d3d12_stream_output_target *)ctx->fake_so_targets[i];
uint64_t filled_size = 0;
struct pipe_transfer *src_transfer, *dst_transfer;
uint8_t *src, *dst;
if (fake_target == NULL)
continue;
pipe_buffer_read(&ctx->base, fake_target->fill_buffer,
fake_target->fill_buffer_offset, sizeof(uint64_t),
&filled_size);
src = (uint8_t *)pipe_buffer_map_range(&ctx->base, fake_target->base.buffer,
fake_target->base.buffer_offset,
fake_target->base.buffer_size,
PIPE_MAP_READ, &src_transfer);
dst = (uint8_t *)pipe_buffer_map_range(&ctx->base, target->base.buffer,
target->base.buffer_offset,
target->base.buffer_size,
PIPE_MAP_READ, &dst_transfer);
/* Note: This will break once support for gl_SkipComponents is added */
uint32_t stride = ctx->gfx_pipeline_state.so_info.stride[i] * 4;
uint64_t src_offset = 0, dst_offset = fake_target->cached_filled_size;
while (src_offset < filled_size) {
memcpy(dst + dst_offset, src + src_offset, stride);
src_offset += stride * ctx->fake_so_buffer_factor;
dst_offset += stride;
if (!cs_state_saved) {
cs_state_saved = true;
d3d12_save_compute_transform_state(ctx, &save);
}
pipe_buffer_unmap(&ctx->base, src_transfer);
pipe_buffer_unmap(&ctx->base, dst_transfer);
d3d12_compute_transform_key key;
memset(&key, 0, sizeof(key));
key.type = d3d12_compute_transform_type::fake_so_buffer_vertex_count;
ctx->base.bind_compute_state(&ctx->base, d3d12_get_compute_transform(ctx, &key));
ctx->transform_state_vars[0] = ctx->gfx_pipeline_state.so_info.stride[i];
ctx->transform_state_vars[1] = ctx->fake_so_buffer_factor;
pipe_shader_buffer new_cs_ssbos[3];
new_cs_ssbos[0].buffer = fake_target->fill_buffer;
new_cs_ssbos[0].buffer_offset = fake_target->fill_buffer_offset;
new_cs_ssbos[0].buffer_size = fake_target->fill_buffer->width0 - fake_target->fill_buffer_offset;
new_cs_ssbos[1].buffer = target->fill_buffer;
new_cs_ssbos[1].buffer_offset = target->fill_buffer_offset;
new_cs_ssbos[1].buffer_size = target->fill_buffer->width0 - target->fill_buffer_offset;
ctx->base.set_shader_buffers(&ctx->base, PIPE_SHADER_COMPUTE, 0, 2, new_cs_ssbos, 2);
pipe_grid_info grid = {};
grid.block[0] = grid.block[1] = grid.block[2] = 1;
grid.grid[0] = grid.grid[1] = grid.grid[2] = 1;
ctx->base.launch_grid(&ctx->base, &grid);
key.type = d3d12_compute_transform_type::fake_so_buffer_copy_back;
key.fake_so_buffer_copy_back.stride = ctx->gfx_pipeline_state.so_info.stride[i];
for (unsigned j = 0; j < ctx->gfx_pipeline_state.so_info.num_outputs; ++j) {
auto& output = ctx->gfx_pipeline_state.so_info.output[j];
if (output.output_buffer != i)
continue;
if (key.fake_so_buffer_copy_back.num_ranges > 0) {
auto& last_range = key.fake_so_buffer_copy_back.ranges[key.fake_so_buffer_copy_back.num_ranges - 1];
if (output.dst_offset * 4 == last_range.offset + last_range.size) {
last_range.size += output.num_components * 4;
continue;
}
}
auto& new_range = key.fake_so_buffer_copy_back.ranges[key.fake_so_buffer_copy_back.num_ranges++];
new_range.offset = output.dst_offset * 4;
new_range.size = output.num_components * 4;
}
ctx->base.bind_compute_state(&ctx->base, d3d12_get_compute_transform(ctx, &key));
ctx->transform_state_vars[0] = ctx->fake_so_buffer_factor;
new_cs_ssbos[0].buffer = target->base.buffer;
new_cs_ssbos[0].buffer_offset = target->base.buffer_offset;
new_cs_ssbos[0].buffer_size = target->base.buffer_size;
new_cs_ssbos[1].buffer = fake_target->base.buffer;
new_cs_ssbos[1].buffer_offset = fake_target->base.buffer_offset;
new_cs_ssbos[1].buffer_size = fake_target->base.buffer_size;
ctx->base.set_shader_buffers(&ctx->base, PIPE_SHADER_COMPUTE, 0, 2, new_cs_ssbos, 2);
pipe_constant_buffer cbuf = {};
cbuf.buffer = fake_target->fill_buffer;
cbuf.buffer_offset = fake_target->fill_buffer_offset;
cbuf.buffer_size = fake_target->fill_buffer->width0 - cbuf.buffer_offset;
ctx->base.set_constant_buffer(&ctx->base, PIPE_SHADER_COMPUTE, 0, true, &cbuf);
grid.indirect = fake_target->fill_buffer;
grid.indirect_offset = fake_target->fill_buffer_offset + 4;
ctx->base.launch_grid(&ctx->base, &grid);
pipe_so_target_reference(&ctx->fake_so_targets[i], NULL);
ctx->fake_so_buffer_views[i].SizeInBytes = 0;
@@ -1856,6 +1897,9 @@ d3d12_disable_fake_so_buffers(struct d3d12_context *ctx)
ctx->fake_so_buffer_factor = 0;
ctx->cmdlist_dirty |= D3D12_DIRTY_STREAM_OUTPUT;
if (cs_state_saved)
d3d12_restore_compute_transform_state(ctx, &save);
return true;
}
src/gallium/drivers/d3d12/d3d12_context.h
-1
View File
@@ -142,7 +142,6 @@ struct d3d12_stream_output_target {
struct pipe_stream_output_target base;
struct pipe_resource *fill_buffer;
unsigned fill_buffer_offset;
uint32_t cached_filled_size;
};
struct d3d12_shader_state {