Mirrors/yuzu
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Query Cache: Fix Prefix Sums

Browse source
This commit is contained in:
Fernando Sahmkow 2023-08-24 03:58:59 +02:00
parent bf0d6b8806
commit 57d8cd6c40
5 changed files with 174 additions and 131 deletions
Download patch file Download diff file Expand all files Collapse all files

99
src/video_core/renderer_vulkan/vk_compute_pass.cpp
View file

@ -179,8 +179,10 @@ struct AstcPushConstants {
};
struct QueriesPrefixScanPushConstants {
u32 min_accumulation_base;
u32 max_accumulation_base;
u32 accumulation_limit;
u32 buffer_offset;
};
} // Anonymous namespace
@ -416,56 +418,65 @@ QueriesPrefixScanPass::QueriesPrefixScanPass(
device_.IsSubgroupFeatureSupported(VK_SUBGROUP_FEATURE_SHUFFLE_BIT) &&
device_.IsSubgroupFeatureSupported(VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT)
? std::span<const u32>(QUERIES_PREFIX_SCAN_SUM_COMP_SPV)
: std::span<const u32>(QUERIES_PREFIX_SCAN_SUM_NOSUBGROUPS_COMP_SPV),
{32}),
: std::span<const u32>(QUERIES_PREFIX_SCAN_SUM_NOSUBGROUPS_COMP_SPV)),
scheduler{scheduler_}, compute_pass_descriptor_queue{compute_pass_descriptor_queue_} {}
void QueriesPrefixScanPass::Run(VkBuffer accumulation_buffer, VkBuffer dst_buffer,
VkBuffer src_buffer, size_t number_of_sums,
size_t max_accumulation_limit) {
size_t aligned_runs = Common::AlignUp(number_of_sums, 32);
size_t min_accumulation_limit, size_t max_accumulation_limit) {
size_t current_runs = number_of_sums;
size_t offset = 0;
while (current_runs != 0) {
static constexpr size_t DISPATCH_SIZE = 2048U;
size_t runs_to_do = std::min<size_t>(current_runs, DISPATCH_SIZE);
current_runs -= runs_to_do;
compute_pass_descriptor_queue.Acquire();
compute_pass_descriptor_queue.AddBuffer(src_buffer, 0, number_of_sums * sizeof(u64));
compute_pass_descriptor_queue.AddBuffer(dst_buffer, 0, number_of_sums * sizeof(u64));
compute_pass_descriptor_queue.AddBuffer(accumulation_buffer, 0, sizeof(u64));
const void* const descriptor_data{compute_pass_descriptor_queue.UpdateData()};
size_t used_offset = offset;
offset += runs_to_do;
compute_pass_descriptor_queue.Acquire();
compute_pass_descriptor_queue.AddBuffer(src_buffer, 0, aligned_runs * sizeof(u64));
compute_pass_descriptor_queue.AddBuffer(dst_buffer, 0, aligned_runs * sizeof(u64));
compute_pass_descriptor_queue.AddBuffer(accumulation_buffer, 0, sizeof(u64));
const void* const descriptor_data{compute_pass_descriptor_queue.UpdateData()};
scheduler.RequestOutsideRenderPassOperationContext();
scheduler.Record([this, descriptor_data, min_accumulation_limit, max_accumulation_limit,
runs_to_do, used_offset](vk::CommandBuffer cmdbuf) {
static constexpr VkMemoryBarrier read_barrier{
.sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER,
.pNext = nullptr,
.srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
.dstAccessMask = VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT,
};
static constexpr VkMemoryBarrier write_barrier{
.sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER,
.pNext = nullptr,
.srcAccessMask = VK_ACCESS_SHADER_WRITE_BIT,
.dstAccessMask = VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_TRANSFER_READ_BIT |
VK_ACCESS_VERTEX_ATTRIBUTE_READ_BIT |
VK_ACCESS_INDIRECT_COMMAND_READ_BIT | VK_ACCESS_INDEX_READ_BIT |
VK_ACCESS_UNIFORM_READ_BIT |
VK_ACCESS_CONDITIONAL_RENDERING_READ_BIT_EXT,
};
const QueriesPrefixScanPushConstants uniforms{
.min_accumulation_base = static_cast<u32>(min_accumulation_limit),
.max_accumulation_base = static_cast<u32>(max_accumulation_limit),
.accumulation_limit = static_cast<u32>(runs_to_do - 1),
.buffer_offset = static_cast<u32>(used_offset),
};
const VkDescriptorSet set = descriptor_allocator.Commit();
device.GetLogical().UpdateDescriptorSet(set, *descriptor_template, descriptor_data);
scheduler.RequestOutsideRenderPassOperationContext();
scheduler.Record([this, descriptor_data, max_accumulation_limit, number_of_sums,
aligned_runs](vk::CommandBuffer cmdbuf) {
static constexpr VkMemoryBarrier read_barrier{
.sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER,
.pNext = nullptr,
.srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
.dstAccessMask = VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT,
};
static constexpr VkMemoryBarrier write_barrier{
.sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER,
.pNext = nullptr,
.srcAccessMask = VK_ACCESS_SHADER_WRITE_BIT,
.dstAccessMask = VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_TRANSFER_READ_BIT |
VK_ACCESS_VERTEX_ATTRIBUTE_READ_BIT |
VK_ACCESS_INDIRECT_COMMAND_READ_BIT | VK_ACCESS_INDEX_READ_BIT |
VK_ACCESS_UNIFORM_READ_BIT |
VK_ACCESS_CONDITIONAL_RENDERING_READ_BIT_EXT,
};
const QueriesPrefixScanPushConstants uniforms{
.max_accumulation_base = static_cast<u32>(max_accumulation_limit),
.accumulation_limit = static_cast<u32>(number_of_sums - 1),
};
const VkDescriptorSet set = descriptor_allocator.Commit();
device.GetLogical().UpdateDescriptorSet(set, *descriptor_template, descriptor_data);
cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, 0, read_barrier);
cmdbuf.BindPipeline(VK_PIPELINE_BIND_POINT_COMPUTE, *pipeline);
cmdbuf.BindDescriptorSets(VK_PIPELINE_BIND_POINT_COMPUTE, *layout, 0, set, {});
cmdbuf.PushConstants(*layout, VK_SHADER_STAGE_COMPUTE_BIT, uniforms);
cmdbuf.Dispatch(static_cast<u32>(aligned_runs / 32U), 1, 1);
cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
VK_PIPELINE_STAGE_CONDITIONAL_RENDERING_BIT_EXT, 0, write_barrier);
});
cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, 0, read_barrier);
cmdbuf.BindPipeline(VK_PIPELINE_BIND_POINT_COMPUTE, *pipeline);
cmdbuf.BindDescriptorSets(VK_PIPELINE_BIND_POINT_COMPUTE, *layout, 0, set, {});
cmdbuf.PushConstants(*layout, VK_SHADER_STAGE_COMPUTE_BIT, uniforms);
cmdbuf.Dispatch(1, 1, 1);
cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
VK_PIPELINE_STAGE_CONDITIONAL_RENDERING_BIT_EXT, 0,
write_barrier);
});
}
}
ASTCDecoderPass::ASTCDecoderPass(const Device& device_, Scheduler& scheduler_,

2
src/video_core/renderer_vulkan/vk_compute_pass.h
View file

@ -104,7 +104,7 @@ public:
ComputePassDescriptorQueue& compute_pass_descriptor_queue_);
void Run(VkBuffer accumulation_buffer, VkBuffer dst_buffer, VkBuffer src_buffer,
size_t number_of_sums, size_t max_accumulation_limit);
size_t number_of_sums, size_t min_accumulation_limit, size_t max_accumulation_limit);
private:
Scheduler& scheduler;

13
src/video_core/renderer_vulkan/vk_query_cache.cpp
View file

@ -181,7 +181,8 @@ public:
});
rasterizer->SyncOperation(std::move(func));
accumulation_since_last_sync = false;
last_accumulation_checkpoint = std::min(last_accumulation_checkpoint, num_slots_used);
first_accumulation_checkpoint = std::min(first_accumulation_checkpoint, num_slots_used);
last_accumulation_checkpoint = std::max(last_accumulation_checkpoint, num_slots_used);
}
void CloseCounter() override {
@ -285,7 +286,9 @@ public:
resolve_buffers.push_back(intermediary_buffer_index);
queries_prefix_scan_pass->Run(*accumulation_buffer, *buffers[intermediary_buffer_index],
*buffers[resolve_buffer_index], num_slots_used,
std::min(last_accumulation_checkpoint, num_slots_used));
std::min(first_accumulation_checkpoint, num_slots_used),
last_accumulation_checkpoint);
} else {
scheduler.RequestOutsideRenderPassOperationContext();
scheduler.Record([buffer = *accumulation_buffer](vk::CommandBuffer cmdbuf) {
@ -298,7 +301,8 @@ public:
rasterizer->SyncOperation(std::move(func));
AbandonCurrentQuery();
num_slots_used = 0;
last_accumulation_checkpoint = std::numeric_limits<size_t>::max();
first_accumulation_checkpoint = std::numeric_limits<size_t>::max();
last_accumulation_checkpoint = 0;
accumulation_since_last_sync = has_multi_queries;
pending_sync.clear();
}
@ -506,7 +510,7 @@ private:
template <bool is_resolve>
size_t ObtainBuffer(size_t num_needed) {
const size_t log_2 = std::max<size_t>(6U, Common::Log2Ceil64(num_needed));
const size_t log_2 = std::max<size_t>(11U, Common::Log2Ceil64(num_needed));
if constexpr (is_resolve) {
if (resolve_table[log_2] != 0) {
return resolve_table[log_2] - 1;
@ -563,6 +567,7 @@ private:
VkQueryPool current_query_pool;
size_t current_query_id;
size_t num_slots_used{};
size_t first_accumulation_checkpoint{};
size_t last_accumulation_checkpoint{};
bool accumulation_since_last_sync{};
VideoCommon::HostQueryBase* current_query;