JDLOBrew/eden
16
0
Fork 0
Code Pull Requests Actions Packages Releases 4 Activity

[maxwell, vk] Allocate data transfer from Maxwell to VK using Sample Locations as channel for MSAA

Browse Source
This commit is contained in:
CamilleLaVey
2025-12-04 02:44:06 -04:00
committed by Caio Oliveira
parent e96972c62e
commit 8bd4c76b22
2 changed files with 39 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp
View File

@@ -830,8 +830,14 @@ void GraphicsPipeline::MakePipeline(VkRenderPass render_pass) {
const bool alpha_to_one_supported = device.SupportsAlphaToOne();
const auto msaa_mode = key.state.msaa_mode.Value();
const VkSampleCountFlagBits vk_samples = MaxwellToVK::MsaaMode(msaa_mode);
const bool supports_sample_locations =
device.IsExtSampleLocationsSupported() && device.SupportsSampleLocationsFor(vk_samples);
const auto [grid_width, grid_height] = VideoCommon::SampleLocationGridSize(msaa_mode);
const auto& sample_location_props = device.SampleLocationProperties();
const bool grid_within_limits = grid_width <= sample_location_props.maxSampleLocationGridSize.width &&
grid_height <= sample_location_props.maxSampleLocationGridSize.height;
const bool supports_sample_locations = device.IsExtSampleLocationsSupported() &&
device.SupportsSampleLocationsFor(vk_samples) &&
sample_location_props.variableSampleLocations == VK_TRUE &&
grid_within_limits;
VkPipelineMultisampleStateCreateInfo multisample_ci{
.sType = VK_STRUCTURE_TYPE_PIPELINE_MULTISAMPLE_STATE_CREATE_INFO,
@@ -856,7 +862,6 @@ void GraphicsPipeline::MakePipeline(VkRenderPass render_pass) {
if (supports_sample_locations) {
sample_locations_chain.emplace();
auto& chain = *sample_locations_chain;
const auto [grid_width, grid_height] = VideoCommon::SampleLocationGridSize(msaa_mode);
const u32 samples_per_pixel = static_cast<u32>(VideoCommon::NumSamples(msaa_mode));
const u32 sample_locations_count = grid_width * grid_height * samples_per_pixel;
chain.locations.fill(VkSampleLocationEXT{0.5f, 0.5f});

33
src/video_core/renderer_vulkan/vk_rasterizer.cpp
View File

@@ -6,6 +6,7 @@
#include <algorithm>
#include <array>
#include <cmath>
#include <memory>
#include <mutex>
@@ -1702,9 +1703,22 @@ void RasterizerVulkan::UpdateSampleLocations(Maxwell& regs) {
return;
}
const auto& sample_props = device.SampleLocationProperties();
if (sample_props.variableSampleLocations == VK_FALSE) {
return;
}
const auto [grid_width_u32, grid_height_u32] = VideoCommon::SampleLocationGridSize(msaa_mode);
const u32 grid_width = grid_width_u32;
const u32 grid_height = grid_height_u32;
if (grid_width > sample_props.maxSampleLocationGridSize.width ||
grid_height > sample_props.maxSampleLocationGridSize.height) {
LOG_WARNING(Render_Vulkan,
"Sample location grid {}x{} exceeds device limit {}x{}, falling back to fixed pattern",
grid_width, grid_height, sample_props.maxSampleLocationGridSize.width,
sample_props.maxSampleLocationGridSize.height);
return;
}
const u32 samples_per_pixel = static_cast<u32>(VideoCommon::NumSamples(msaa_mode));
const u32 grid_cells = grid_width * grid_height;
const u32 sample_locations_count = grid_cells * samples_per_pixel;
@@ -1712,15 +1726,30 @@ void RasterizerVulkan::UpdateSampleLocations(Maxwell& regs) {
const auto raw_locations = DecodeSampleLocationRegisters(regs);
std::array<VkSampleLocationEXT, VideoCommon::MaxSampleLocationSlots> resolved{};
const float coord_min = sample_props.sampleLocationCoordinateRange[0];
const float coord_max = sample_props.sampleLocationCoordinateRange[1];
const float host_unit = sample_props.sampleLocationSubPixelBits > 0
? 1.0f / static_cast<float>(1u << sample_props.sampleLocationSubPixelBits)
: 0.0f;
const auto sanitize_coord = [&](float value) {
float clamped = std::clamp(value, coord_min, coord_max);
if (host_unit > 0.0f) {
clamped = coord_min + std::round((clamped - coord_min) / host_unit) * host_unit;
clamped = std::clamp(clamped, coord_min, coord_max);
}
return clamped;
};
for (u32 cell = 0; cell < grid_cells; ++cell) {
const u32 slot_base = cell * samples_per_pixel;
const u32 cell_x = cell % grid_width;
const u32 cell_y = cell / grid_width;
for (u32 sample = 0; sample < samples_per_pixel; ++sample) {
const VkSampleLocationEXT raw = raw_locations[slot_base + sample];
const float sample_x = static_cast<float>(cell_x) + raw.x;
const float sample_y = static_cast<float>(cell_y) + raw.y;
resolved[slot_base + sample] = VkSampleLocationEXT{
.x = static_cast<float>(cell_x) + raw.x,
.y = static_cast<float>(cell_y) + raw.y,
.x = sanitize_coord(sample_x),
.y = sanitize_coord(sample_y),
};
}
}