ShuriZma
/
dolphin
mirror of https://github.com/dolphin-emu/dolphin.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #12537 from TellowKrinkle/MTLSubgroup 

VideoCommon: More specific subgroup op bugs
This commit is contained in:
Admiral H. Curtiss 2024-05-21 22:17:22 +02:00 committed by GitHub
parent 321dc9b76d b7a451fc87
commit 5c2f73986a
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
8 changed files with 288 additions and 77 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
Source/Core/Core/DolphinAnalytics.h
View File

@ -134,6 +134,9 @@ public:
// to once per game run.
void ReportGameQuirk(GameQuirk quirk);
// Get the base builder for building a report
const Common::AnalyticsReportBuilder& BaseBuilder() const { return m_base_builder; }
struct PerformanceSample
{
double speed_ratio; // See SystemTimers::GetEstimatedEmulationPerformance().

202
Source/Core/VideoBackends/Metal/MTLUtil.mm
View File

@ -91,6 +91,68 @@ void Metal::Util::PopulateBackendInfoAdapters(VideoConfig* config,
}
}
/// For testing driver brokenness
static bool RenderSinglePixel(id<MTLDevice> dev, id<MTLFunction> vs, id<MTLFunction> fs, //
u32 px_in, u32* px_out)
{
auto pdesc = MRCTransfer([MTLRenderPipelineDescriptor new]);
[pdesc setVertexFunction:vs];
[pdesc setFragmentFunction:fs];
[[pdesc colorAttachments][0] setPixelFormat:MTLPixelFormatRGBA8Unorm];
auto pipe = MRCTransfer([dev newRenderPipelineStateWithDescriptor:pdesc error:nil]);
if (!pipe)
return false;
auto buf = MRCTransfer([dev newBufferWithLength:4 options:MTLResourceStorageModeShared]);
memcpy([buf contents], &px_in, sizeof(px_in));
auto tdesc = [MTLTextureDescriptor texture2DDescriptorWithPixelFormat:MTLPixelFormatRGBA8Unorm
width:1
height:1
mipmapped:false];
[tdesc setUsage:MTLTextureUsageRenderTarget];
auto tex = MRCTransfer([dev newTextureWithDescriptor:tdesc]);
auto q = MRCTransfer([dev newCommandQueue]);
id<MTLCommandBuffer> cmdbuf = [q commandBuffer];
id<MTLBlitCommandEncoder> upload_encoder = [cmdbuf blitCommandEncoder];
[upload_encoder copyFromBuffer:buf
sourceOffset:0
sourceBytesPerRow:4
sourceBytesPerImage:4
sourceSize:MTLSizeMake(1, 1, 1)
toTexture:tex
destinationSlice:0
destinationLevel:0
destinationOrigin:MTLOriginMake(0, 0, 0)];
[upload_encoder endEncoding];
auto rpdesc = MRCTransfer([MTLRenderPassDescriptor new]);
[[rpdesc colorAttachments][0] setTexture:tex];
[[rpdesc colorAttachments][0] setLoadAction:MTLLoadActionLoad];
[[rpdesc colorAttachments][0] setStoreAction:MTLStoreActionStore];
id<MTLRenderCommandEncoder> renc = [cmdbuf renderCommandEncoderWithDescriptor:rpdesc];
[renc setRenderPipelineState:pipe];
[renc drawPrimitives:MTLPrimitiveTypeTriangle vertexStart:0 vertexCount:3];
[renc endEncoding];
id<MTLBlitCommandEncoder> download_encoder = [cmdbuf blitCommandEncoder];
[download_encoder copyFromTexture:tex
sourceSlice:0
sourceLevel:0
sourceOrigin:MTLOriginMake(0, 0, 0)
sourceSize:MTLSizeMake(1, 1, 1)
toBuffer:buf
destinationOffset:0
destinationBytesPerRow:4
destinationBytesPerImage:4];
[download_encoder endEncoding];
[cmdbuf commit];
[cmdbuf waitUntilCompleted];
memcpy(px_out, [buf contents], sizeof(*px_out));
return [cmdbuf status] == MTLCommandBufferStatusCompleted;
}
static bool DetectIntelGPUFBFetch(id<MTLDevice> dev)
{
// Even though it's nowhere in the feature set tables, some Intel GPUs support fbfetch!
@ -113,58 +175,14 @@ fragment float4 fbfetch_test(float4 in [[color(0), raster_order_group(0)]]) {
error:nil]);
if (!lib)
return false;
auto pdesc = MRCTransfer([MTLRenderPipelineDescriptor new]);
[pdesc setVertexFunction:MRCTransfer([lib newFunctionWithName:@"fs_triangle"])];
[pdesc setFragmentFunction:MRCTransfer([lib newFunctionWithName:@"fbfetch_test"])];
[[pdesc colorAttachments][0] setPixelFormat:MTLPixelFormatRGBA8Unorm];
auto pipe = MRCTransfer([dev newRenderPipelineStateWithDescriptor:pdesc error:nil]);
if (!pipe)
return false;
auto buf = MRCTransfer([dev newBufferWithLength:4 options:MTLResourceStorageModeShared]);
auto tdesc = [MTLTextureDescriptor texture2DDescriptorWithPixelFormat:MTLPixelFormatRGBA8Unorm
width:1
height:1
mipmapped:false];
[tdesc setUsage:MTLTextureUsageRenderTarget];
auto tex = MRCTransfer([dev newTextureWithDescriptor:tdesc]);
auto q = MRCTransfer([dev newCommandQueue]);
u32 px = 0x11223344;
memcpy([buf contents], &px, 4);
id<MTLCommandBuffer> cmdbuf = [q commandBuffer];
id<MTLBlitCommandEncoder> upload_encoder = [cmdbuf blitCommandEncoder];
[upload_encoder copyFromBuffer:buf
sourceOffset:0
sourceBytesPerRow:4
sourceBytesPerImage:4
sourceSize:MTLSizeMake(1, 1, 1)
toTexture:tex
destinationSlice:0
destinationLevel:0
destinationOrigin:MTLOriginMake(0, 0, 0)];
[upload_encoder endEncoding];
auto rpdesc = MRCTransfer([MTLRenderPassDescriptor new]);
[[rpdesc colorAttachments][0] setTexture:tex];
[[rpdesc colorAttachments][0] setLoadAction:MTLLoadActionLoad];
[[rpdesc colorAttachments][0] setStoreAction:MTLStoreActionStore];
id<MTLRenderCommandEncoder> renc = [cmdbuf renderCommandEncoderWithDescriptor:rpdesc];
[renc setRenderPipelineState:pipe];
[renc drawPrimitives:MTLPrimitiveTypeTriangle vertexStart:0 vertexCount:3];
[renc endEncoding];
id<MTLBlitCommandEncoder> download_encoder = [cmdbuf blitCommandEncoder];
[download_encoder copyFromTexture:tex
sourceSlice:0
sourceLevel:0
sourceOrigin:MTLOriginMake(0, 0, 0)
sourceSize:MTLSizeMake(1, 1, 1)
toBuffer:buf
destinationOffset:0
destinationBytesPerRow:4
destinationBytesPerImage:4];
[download_encoder endEncoding];
[cmdbuf commit];
[cmdbuf waitUntilCompleted];
u32 outpx;
memcpy(&outpx, [buf contents], 4);
bool ok = RenderSinglePixel(dev, //
MRCTransfer([lib newFunctionWithName:@"fs_triangle"]), //
MRCTransfer([lib newFunctionWithName:@"fbfetch_test"]), //
0x11223344, &outpx);
if (!ok)
return false;
// Proper fbfetch will double contents, Haswell will return black, and Broadwell will do nothing
if (outpx == 0x22446688)
return true; // Skylake+
@ -174,22 +192,69 @@ fragment float4 fbfetch_test(float4 in [[color(0), raster_order_group(0)]]) {
return false; // Haswell
}
enum class DetectionResult
{
Yes,
No,
Unsure
};
static DetectionResult DetectInvertedIsHelper(id<MTLDevice> dev)
{
static constexpr const char* shader = R"(
vertex float4 fs_triangle(uint vid [[vertex_id]]) {
return float4(vid & 1 ? 3 : -1, vid & 2 ? 3 : -1, 0, 1);
}
fragment float4 is_helper_test() {
float val = metal::simd_is_helper_thread() ? 1 : 0.5;
return float4(val, metal::dfdx(val) + 0.5, metal::dfdy(val) + 0.5, 0);
}
)";
auto lib = MRCTransfer([dev newLibraryWithSource:[NSString stringWithUTF8String:shader]
options:nil
error:nil]);
if (!lib)
return DetectionResult::Unsure;
u32 outpx;
bool ok = RenderSinglePixel(dev, //
MRCTransfer([lib newFunctionWithName:@"fs_triangle"]), //
MRCTransfer([lib newFunctionWithName:@"is_helper_test"]), //
0, &outpx);
// The pixel itself should not be a helper thread (0.5)
// The pixels to its right and below should be helper threads (1.0)
// Correctly working would therefore be 0.5 for the pixel and (0.5 + 0.5) for the derivatives
// Inverted would be 1.0 for the pixel and (-0.5 + 0.5) for the derivatives
if (!ok)
return DetectionResult::Unsure;
if (outpx == 0xffff80)
return DetectionResult::No; // Working correctly
if (outpx == 0x0000ff)
return DetectionResult::Yes; // Inverted
WARN_LOG_FMT(VIDEO, "metal::simd_is_helper_thread might be broken! Test shader returned {:06x}!",
outpx);
return DetectionResult::Unsure;
}
void Metal::Util::PopulateBackendInfoFeatures(VideoConfig* config, id<MTLDevice> device)
{
// Initialize DriverDetails first so we can use it later
DriverDetails::Vendor vendor = DriverDetails::VENDOR_UNKNOWN;
if ([[device name] containsString:@"NVIDIA"])
std::string name = [[device name] UTF8String];
if (name.find("NVIDIA") != std::string::npos)
vendor = DriverDetails::VENDOR_NVIDIA;
else if ([[device name] containsString:@"AMD"])
else if (name.find("AMD") != std::string::npos)
vendor = DriverDetails::VENDOR_ATI;
else if ([[device name] containsString:@"Intel"])
else if (name.find("Intel") != std::string::npos)
vendor = DriverDetails::VENDOR_INTEL;
else if ([[device name] containsString:@"Apple"])
else if (name.find("Apple") != std::string::npos)
vendor = DriverDetails::VENDOR_APPLE;
const NSOperatingSystemVersion cocoa_ver = [[NSProcessInfo processInfo] operatingSystemVersion];
double version = cocoa_ver.majorVersion * 100 + cocoa_ver.minorVersion;
DriverDetails::Init(DriverDetails::API_METAL, vendor, DriverDetails::DRIVER_APPLE, version,
DriverDetails::Family::UNKNOWN);
DriverDetails::Family::UNKNOWN, std::move(name));
#if TARGET_OS_OSX
config->backend_info.bSupportsDepthClamp = true;
@ -247,8 +312,16 @@ void Metal::Util::PopulateBackendInfoFeatures(VideoConfig* config, id<MTLDevice>
[device supportsFamily:MTLGPUFamilyMac2] || [device supportsFamily:MTLGPUFamilyApple6];
config->backend_info.bSupportsFramebufferFetch = [device supportsFamily:MTLGPUFamilyApple1];
}
if (DriverDetails::HasBug(DriverDetails::BUG_BROKEN_SUBGROUP_OPS))
g_features.subgroup_ops = false;
if (g_features.subgroup_ops)
{
DetectionResult result = DetectInvertedIsHelper(device);
if (result != DetectionResult::Unsure)
{
bool is_helper_inverted = result == DetectionResult::Yes;
if (is_helper_inverted != DriverDetails::HasBug(DriverDetails::BUG_INVERTED_IS_HELPER))
DriverDetails::OverrideBug(DriverDetails::BUG_INVERTED_IS_HELPER, is_helper_inverted);
}
}
#if TARGET_OS_OSX
if (@available(macOS 11, *))
if (vendor == DriverDetails::VENDOR_INTEL)
@ -436,22 +509,29 @@ std::optional<std::string> Metal::Util::TranslateShaderToMSL(ShaderStage stage,
full_source.append(header);
if (Metal::g_features.subgroup_ops)
full_source.append(SUBGROUP_HELPER_HEADER);
if (DriverDetails::HasBug(DriverDetails::BUG_INVERTED_IS_HELPER))
{
full_source.append("#define gl_HelperInvocation !gl_HelperInvocation "
"// Work around broken AMD Metal driver\n");
}
if (DriverDetails::HasBug(DriverDetails::BUG_BROKEN_SUBGROUP_OPS_WITH_DISCARD))
full_source.append("#define BROKEN_SUBGROUP_WITH_DISCARD 1\n");
full_source.append(source);
std::optional<SPIRV::CodeVector> code;
switch (stage)
{
case ShaderStage::Vertex:
code = SPIRV::CompileVertexShader(full_source, APIType::Metal, glslang::EShTargetSpv_1_3);
code = SPIRV::CompileVertexShader(full_source, APIType::Metal, glslang::EShTargetSpv_1_5);
break;
case ShaderStage::Geometry:
PanicAlertFmt("Tried to compile geometry shader for Metal, but Metal doesn't support them!");
break;
case ShaderStage::Pixel:
code = SPIRV::CompileFragmentShader(full_source, APIType::Metal, glslang::EShTargetSpv_1_3);
code = SPIRV::CompileFragmentShader(full_source, APIType::Metal, glslang::EShTargetSpv_1_5);
break;
case ShaderStage::Compute:
code = SPIRV::CompileComputeShader(full_source, APIType::Metal, glslang::EShTargetSpv_1_3);
code = SPIRV::CompileComputeShader(full_source, APIType::Metal, glslang::EShTargetSpv_1_5);
break;
}
if (!code.has_value())

3
Source/Core/VideoBackends/OGL/OGLConfig.cpp
View File

@ -211,7 +211,8 @@ void InitDriverInfo()
default:
break;
}
DriverDetails::Init(DriverDetails::API_OPENGL, vendor, driver, version, family);
DriverDetails::Init(DriverDetails::API_OPENGL, vendor, driver, version, family,
std::string(srenderer));
}
bool PopulateConfig(GLContext* m_main_gl_context)

8
Source/Core/VideoBackends/Vulkan/ShaderCompiler.cpp
View File

@ -7,6 +7,7 @@
#include <string>
#include "VideoBackends/Vulkan/VulkanContext.h"
#include "VideoCommon/DriverDetails.h"
#include "VideoCommon/Spirv.h"
namespace Vulkan::ShaderCompiler
@ -98,6 +99,13 @@ static std::string GetShaderCode(std::string_view source, std::string_view heade
full_source_code.append(header);
if (g_vulkan_context->SupportsShaderSubgroupOperations())
full_source_code.append(SUBGROUP_HELPER_HEADER, subgroup_helper_header_length);
if (DriverDetails::HasBug(DriverDetails::BUG_INVERTED_IS_HELPER))
{
full_source_code.append("#define gl_HelperInvocation !gl_HelperInvocation "
"// Work around broken AMD Metal driver\n");
}
if (DriverDetails::HasBug(DriverDetails::BUG_BROKEN_SUBGROUP_OPS_WITH_DISCARD))
full_source_code.append("#define BROKEN_SUBGROUP_WITH_DISCARD 1\n");
full_source_code.append(source);
}

5
Source/Core/VideoBackends/Vulkan/VulkanContext.cpp
View File

@ -967,7 +967,7 @@ void VulkanContext::InitDriverDetails()
DriverDetails::Init(DriverDetails::API_VULKAN, vendor, driver,
static_cast<double>(m_device_properties.driverVersion),
DriverDetails::Family::UNKNOWN);
DriverDetails::Family::UNKNOWN, std::move(device_name));
}
void VulkanContext::PopulateShaderSubgroupSupport()
@ -1000,8 +1000,7 @@ void VulkanContext::PopulateShaderSubgroupSupport()
VK_SUBGROUP_FEATURE_BALLOT_BIT | VK_SUBGROUP_FEATURE_SHUFFLE_BIT;
m_supports_shader_subgroup_operations =
(subgroup_properties.supportedOperations & required_operations) == required_operations &&
subgroup_properties.supportedStages & VK_SHADER_STAGE_FRAGMENT_BIT &&
!DriverDetails::HasBug(DriverDetails::BUG_BROKEN_SUBGROUP_OPS);
subgroup_properties.supportedStages & VK_SHADER_STAGE_FRAGMENT_BIT;
}
bool VulkanContext::SupportsExclusiveFullscreen(const WindowSystemInfo& wsi, VkSurfaceKHR surface)

120
Source/Core/VideoCommon/DriverDetails.cpp
View File

@ -6,6 +6,7 @@
#include <map>
#include "Common/Logging/LogManager.h"
#include "Core/DolphinAnalytics.h"
namespace DriverDetails
{
@ -46,6 +47,7 @@ static Vendor m_vendor = VENDOR_UNKNOWN;
static Driver m_driver = DRIVER_UNKNOWN;
static Family m_family = Family::UNKNOWN;
static double m_version = 0.0;
static std::string m_name;
// This is a list of all known bugs for each vendor
// We use this to check if the device and driver has a issue
@ -130,14 +132,14 @@ constexpr BugInfo m_known_bugs[] = {
-1.0, -1.0, true},
{API_VULKAN, OS_ALL, VENDOR_ARM, DRIVER_ARM, Family::UNKNOWN, BUG_BROKEN_VECTOR_BITWISE_AND,
-1.0, -1.0, true},
{API_VULKAN, OS_OSX, VENDOR_ATI, DRIVER_PORTABILITY, Family::UNKNOWN, BUG_BROKEN_SUBGROUP_OPS,
{API_VULKAN, OS_OSX, VENDOR_ATI, DRIVER_PORTABILITY, Family::UNKNOWN, BUG_INVERTED_IS_HELPER,
-1.0, -1.0, true},
{API_VULKAN, OS_OSX, VENDOR_INTEL, DRIVER_PORTABILITY, Family::UNKNOWN, BUG_BROKEN_SUBGROUP_OPS,
-1.0, -1.0, true},
{API_METAL, OS_OSX, VENDOR_ATI, DRIVER_APPLE, Family::UNKNOWN, BUG_BROKEN_SUBGROUP_OPS, -1.0,
-1.0, true},
{API_METAL, OS_OSX, VENDOR_INTEL, DRIVER_APPLE, Family::UNKNOWN, BUG_BROKEN_SUBGROUP_OPS, -1.0,
{API_METAL, OS_OSX, VENDOR_ATI, DRIVER_APPLE, Family::UNKNOWN, BUG_INVERTED_IS_HELPER, -1.0,
-1.0, true},
{API_VULKAN, OS_OSX, VENDOR_INTEL, DRIVER_PORTABILITY, Family::UNKNOWN,
BUG_BROKEN_SUBGROUP_OPS_WITH_DISCARD, -1.0, -1.0, true},
{API_METAL, OS_OSX, VENDOR_INTEL, DRIVER_APPLE, Family::UNKNOWN,
BUG_BROKEN_SUBGROUP_OPS_WITH_DISCARD, -1.0, -1.0, true},
{API_OPENGL, OS_ANDROID, VENDOR_ALL, DRIVER_ALL, Family::UNKNOWN,
BUG_BROKEN_MULTITHREADED_SHADER_PRECOMPILATION, -1.0, -1.0, true},
{API_VULKAN, OS_ANDROID, VENDOR_ALL, DRIVER_ALL, Family::UNKNOWN,
@ -160,13 +162,15 @@ constexpr BugInfo m_known_bugs[] = {
static std::map<Bug, BugInfo> m_bugs;
void Init(API api, Vendor vendor, Driver driver, const double version, const Family family)
void Init(API api, Vendor vendor, Driver driver, const double version, const Family family,
std::string name)
{
m_api = api;
m_vendor = vendor;
m_driver = driver;
m_version = version;
m_family = family;
m_name = std::move(name);
if (driver == DRIVER_UNKNOWN)
{
@ -217,4 +221,106 @@ bool HasBug(Bug bug)
return false;
return it->second.m_hasbug;
}
#ifdef __clang__
// Make sure we handle all these switch cases
#pragma clang diagnostic error "-Wswitch"
#pragma clang diagnostic error "-Wcovered-switch-default"
#endif
// clang-format off
static const char* to_string(API api)
{
switch (api)
{
case API_OPENGL: return "OpenGL";
case API_VULKAN: return "Vulkan";
case API_METAL: return "Metal";
}
return "Unknown";
}
static const char* to_string(Driver driver)
{
switch (driver)
{
case DRIVER_ALL: return "All";
case DRIVER_NVIDIA: return "Nvidia";
case DRIVER_NOUVEAU: return "Nouveau";
case DRIVER_ATI: return "ATI";
case DRIVER_R600: return "R600";
case DRIVER_INTEL: return "Intel";
case DRIVER_I965: return "I965";
case DRIVER_ARM: return "ARM";
case DRIVER_LIMA: return "Lima";
case DRIVER_QUALCOMM: return "Qualcomm";
case DRIVER_FREEDRENO: return "Freedreno";
case DRIVER_IMGTEC: return "Imgtech";
case DRIVER_VIVANTE: return "Vivante";
case DRIVER_PORTABILITY: return "Portability";
case DRIVER_APPLE: return "Apple";
case DRIVER_UNKNOWN: return "Unknown";
}
return "Unknown";
}
static const char* to_string(Bug bug)
{
switch (bug)
{
case BUG_BROKEN_UBO: return "broken-ubo";
case BUG_BROKEN_PINNED_MEMORY: return "broken-pinned-memory";
case BUG_BROKEN_BUFFER_STREAM: return "broken-buffer-stream";
case BUG_BROKEN_BUFFER_STORAGE: return "broken-buffer-storage";
case BUG_PRIMITIVE_RESTART: return "primitive-restart";
case BUG_BROKEN_UNSYNC_MAPPING: return "broken-unsync-mapping";
case BUG_INTEL_BROKEN_BUFFER_STORAGE: return "intel-broken-buffer-storage";
case BUG_BROKEN_NEGATED_BOOLEAN: return "broken-negated-boolean";
case BUG_BROKEN_COPYIMAGE: return "broken-copyimage";
case BUG_BROKEN_VSYNC: return "broken-vsync";
case BUG_BROKEN_GEOMETRY_SHADERS: return "broken-geometry-shaders";
case BUG_SLOW_GETBUFFERSUBDATA: return "slow-getBufferSubData";
case BUG_BROKEN_CLIP_DISTANCE: return "broken-clip-distance";
case BUG_BROKEN_DUAL_SOURCE_BLENDING: return "broken-dual-source-blending";
case BUG_BROKEN_BITWISE_OP_NEGATION: return "broken-bitwise-op-negation";
case BUG_SHARED_CONTEXT_SHADER_COMPILATION: return "shared-context-shader-compilation";
case BUG_BROKEN_MSAA_CLEAR: return "broken-msaa-clear";
case BUG_BROKEN_CLEAR_LOADOP_RENDERPASS: return "broken-clear-loadop-renderpass";
case BUG_BROKEN_D32F_CLEAR: return "broken-d32f-clear";
case BUG_BROKEN_REVERSED_DEPTH_RANGE: return "broken-reversed-depth-range";
case BUG_SLOW_CACHED_READBACK_MEMORY: return "slow-cached-readback-memory";
case BUG_BROKEN_VECTOR_BITWISE_AND: return "broken-vector-bitwise-and";
case BUG_BROKEN_SUBGROUP_OPS_WITH_DISCARD: return "broken-subgroup-ops-with-discard";
case BUG_INVERTED_IS_HELPER: return "inverted-is-helper";
case BUG_BROKEN_MULTITHREADED_SHADER_PRECOMPILATION: return "broken-multithreaded-shader-precompilation";
case BUG_BROKEN_DISCARD_WITH_EARLY_Z: return "broken-discard-with-early-z";
case BUG_BROKEN_DYNAMIC_SAMPLER_INDEXING: return "broken-dynamic-sampler-indexing";
case BUG_SLOW_OPTIMAL_IMAGE_TO_BUFFER_COPY: return "slow-optimal-image-to-buffer-copy";
}
return "Unknown";
}
// clang-format on
void OverrideBug(Bug bug, bool new_value)
{
const auto [it, added] = m_bugs.try_emplace(
bug, BugInfo{m_api, m_os, m_vendor, m_driver, m_family, bug, -1, -1, false});
if (it->second.m_hasbug != new_value)
{
DolphinAnalytics& analytics = DolphinAnalytics::Instance();
Common::AnalyticsReportBuilder builder(analytics.BaseBuilder());
builder.AddData("type", "gpu-bug-override");
builder.AddData("bug", to_string(bug));
builder.AddData("value", new_value);
builder.AddData("gpu", m_name);
builder.AddData("api", to_string(m_api));
builder.AddData("driver", to_string(m_driver));
builder.AddData("version", std::to_string(m_version));
analytics.Send(builder);
it->second.m_hasbug = new_value;
}
}
} // namespace DriverDetails

22
Source/Core/VideoCommon/DriverDetails.h
View File

@ -1,6 +1,7 @@
// Copyright 2013 Dolphin Emulator Project
// SPDX-License-Identifier: GPL-2.0-or-later
#pragma once
#include <string>
#include "Common/CommonTypes.h"
#undef OS // CURL defines that, nobody uses it...
@ -299,14 +300,23 @@ enum Bug
// BUG: Accessing gl_SubgroupInvocationID causes the Metal shader compiler to crash.
// Affected devices: AMD (older macOS)
// BUG: gl_HelperInvocation always returns true, even for non-helper invocations
// Affected devices: AMD (newer macOS)
// Started version: ???
// Ended version: ???
// (Workaround currently disabled, will put it back when someone hits the issue and we can
// find out what devices and OSes it actually affects)
// BUG: Using subgroupMax in a shader that can discard results in garbage data
// (For some reason, this only happens at 4x+ IR on Metal, but 2x+ IR on MoltenVK)
// Affected devices: Intel (macOS)
// Started version: -1
// Ended version: -1
BUG_BROKEN_SUBGROUP_OPS,
BUG_BROKEN_SUBGROUP_OPS_WITH_DISCARD,
// BUG: gl_HelperInvocation is actually !gl_HelperInvocation
// Affected devices: AMD (macOS)
// Started version: -1
// Ended version: -1
BUG_INVERTED_IS_HELPER,
// BUG: Multi-threaded shader pre-compilation sometimes crashes
// Used primarily in Videoconfig.cpp's GetNumAutoShaderPreCompilerThreads()
@ -336,9 +346,13 @@ enum Bug
};
// Initializes our internal vendor, device family, and driver version
void Init(API api, Vendor vendor, Driver driver, const double version, const Family family);
void Init(API api, Vendor vendor, Driver driver, const double version, const Family family,
std::string name);
// Once Vendor and driver version is set, this will return if it has the applicable bug passed to
// it.
bool HasBug(Bug bug);
// Overrides the current state of a bug
void OverrideBug(Bug bug, bool new_value);
} // namespace DriverDetails

2
Source/Core/VideoCommon/PixelShaderGen.cpp
View File

@ -485,7 +485,7 @@ void UpdateBoundingBox(float2 rawpos) {{
int2 pos_tl = pos & ~1; // round down to even
int2 pos_br = pos | 1; // round up to odd
#ifdef SUPPORTS_SUBGROUP_REDUCTION
#if defined(SUPPORTS_SUBGROUP_REDUCTION) && !defined(BROKEN_SUBGROUP_WITH_DISCARD)
if (!IS_HELPER_INVOCATION)
{{
SUBGROUP_MIN(pos_tl);