eden-emu
/
eden
mirror of https://git.eden-emu.dev/eden-emu/eden.git
1
0
Fork 0
Code Releases Wiki Activity
Browse Source

[vulkan] Relaxing Workgroup Memory Explicit Layout flags + INT64 Atomics

vk-experiments5
CamilleLaVey 3 months ago
parent
84b78abe1c
commit
1ba7705279
8 changed files with 62 additions and 48 deletions
Whitespace
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. 7
      src/shader_recompiler/backend/spirv/emit_spirv_atomic.cpp
  2. 27
      src/shader_recompiler/backend/spirv/emit_spirv_shared_memory.cpp
  3. 8
      src/shader_recompiler/backend/spirv/spirv_emit_context.cpp
  4. 4
      src/shader_recompiler/backend/spirv/spirv_emit_context.h
  5. 4
      src/shader_recompiler/profile.h
  6. 5
      src/video_core/renderer_vulkan/vk_pipeline_cache.cpp
  7. 32
      src/video_core/vulkan_common/vulkan_device.cpp
  8. 23
      src/video_core/vulkan_common/vulkan_device.h

7
src/shader_recompiler/backend/spirv/emit_spirv_atomic.cpp
View File

@ -1,3 +1,6 @@
// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
// SPDX-License-Identifier: GPL-3.0-or-later
// SPDX-FileCopyrightText: Copyright 2021 yuzu Emulator Project
// SPDX-License-Identifier: GPL-2.0-or-later
@ -15,7 +18,7 @@ Id SharedPointer(EmitContext& ctx, Id offset, u32 index_offset = 0) {
if (index_offset > 0) {
index = ctx.OpIAdd(ctx.U32[1], index, ctx.Const(index_offset));
}
return ctx.profile.support_explicit_workgroup_layout
return ctx.uses_explicit_workgroup_layout
? ctx.OpAccessChain(ctx.shared_u32, ctx.shared_memory_u32, ctx.u32_zero_value, index)
: ctx.OpAccessChain(ctx.shared_u32, ctx.shared_memory_u32, index);
}
@ -155,7 +158,7 @@ Id EmitSharedAtomicExchange32(EmitContext& ctx, Id offset, Id value) {
}
Id EmitSharedAtomicExchange64(EmitContext& ctx, Id offset, Id value) {
if (ctx.profile.support_int64_atomics && ctx.profile.support_explicit_workgroup_layout) {
if (ctx.profile.support_shared_int64_atomics && ctx.uses_explicit_workgroup_layout) {
const Id shift_id{ctx.Const(3U)};
const Id index{ctx.OpShiftRightArithmetic(ctx.U32[1], offset, shift_id)};
const Id pointer{

27
src/shader_recompiler/backend/spirv/emit_spirv_shared_memory.cpp
View File

@ -1,3 +1,6 @@
// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
// SPDX-License-Identifier: GPL-3.0-or-later
// SPDX-FileCopyrightText: Copyright 2021 yuzu Emulator Project
// SPDX-License-Identifier: GPL-2.0-or-later
@ -28,7 +31,7 @@ std::pair<Id, Id> ExtractArgs(EmitContext& ctx, Id offset, u32 mask, u32 count)
} // Anonymous namespace
Id EmitLoadSharedU8(EmitContext& ctx, Id offset) {
if (ctx.profile.support_explicit_workgroup_layout) {
if (ctx.uses_explicit_workgroup_layout) {
const Id pointer{
ctx.OpAccessChain(ctx.shared_u8, ctx.shared_memory_u8, ctx.u32_zero_value, offset)};
return ctx.OpUConvert(ctx.U32[1], ctx.OpLoad(ctx.U8, pointer));
@ -39,7 +42,7 @@ Id EmitLoadSharedU8(EmitContext& ctx, Id offset) {
}
Id EmitLoadSharedS8(EmitContext& ctx, Id offset) {
if (ctx.profile.support_explicit_workgroup_layout) {
if (ctx.uses_explicit_workgroup_layout) {
const Id pointer{
ctx.OpAccessChain(ctx.shared_u8, ctx.shared_memory_u8, ctx.u32_zero_value, offset)};
return ctx.OpSConvert(ctx.U32[1], ctx.OpLoad(ctx.U8, pointer));
@ -50,7 +53,7 @@ Id EmitLoadSharedS8(EmitContext& ctx, Id offset) {
}
Id EmitLoadSharedU16(EmitContext& ctx, Id offset) {
if (ctx.profile.support_explicit_workgroup_layout) {
if (ctx.uses_explicit_workgroup_layout) {
const Id pointer{Pointer(ctx, ctx.shared_u16, ctx.shared_memory_u16, offset, 1)};
return ctx.OpUConvert(ctx.U32[1], ctx.OpLoad(ctx.U16, pointer));
} else {
@ -60,7 +63,7 @@ Id EmitLoadSharedU16(EmitContext& ctx, Id offset) {
}
Id EmitLoadSharedS16(EmitContext& ctx, Id offset) {
if (ctx.profile.support_explicit_workgroup_layout) {
if (ctx.uses_explicit_workgroup_layout) {
const Id pointer{Pointer(ctx, ctx.shared_u16, ctx.shared_memory_u16, offset, 1)};
return ctx.OpSConvert(ctx.U32[1], ctx.OpLoad(ctx.U16, pointer));
} else {
@ -70,7 +73,7 @@ Id EmitLoadSharedS16(EmitContext& ctx, Id offset) {
}
Id EmitLoadSharedU32(EmitContext& ctx, Id offset) {
if (ctx.profile.support_explicit_workgroup_layout) {
if (ctx.uses_explicit_workgroup_layout) {
const Id pointer{Pointer(ctx, ctx.shared_u32, ctx.shared_memory_u32, offset, 2)};
return ctx.OpLoad(ctx.U32[1], pointer);
} else {
@ -79,7 +82,7 @@ Id EmitLoadSharedU32(EmitContext& ctx, Id offset) {
}
Id EmitLoadSharedU64(EmitContext& ctx, Id offset) {
if (ctx.profile.support_explicit_workgroup_layout) {
if (ctx.uses_explicit_workgroup_layout) {
const Id pointer{Pointer(ctx, ctx.shared_u32x2, ctx.shared_memory_u32x2, offset, 3)};
return ctx.OpLoad(ctx.U32[2], pointer);
} else {
@ -94,7 +97,7 @@ Id EmitLoadSharedU64(EmitContext& ctx, Id offset) {
}
Id EmitLoadSharedU128(EmitContext& ctx, Id offset) {
if (ctx.profile.support_explicit_workgroup_layout) {
if (ctx.uses_explicit_workgroup_layout) {
const Id pointer{Pointer(ctx, ctx.shared_u32x4, ctx.shared_memory_u32x4, offset, 4)};
return ctx.OpLoad(ctx.U32[4], pointer);
}
@ -110,7 +113,7 @@ Id EmitLoadSharedU128(EmitContext& ctx, Id offset) {
}
void EmitWriteSharedU8(EmitContext& ctx, Id offset, Id value) {
if (ctx.profile.support_explicit_workgroup_layout) {
if (ctx.uses_explicit_workgroup_layout) {
const Id pointer{
ctx.OpAccessChain(ctx.shared_u8, ctx.shared_memory_u8, ctx.u32_zero_value, offset)};
ctx.OpStore(pointer, ctx.OpUConvert(ctx.U8, value));
@ -120,7 +123,7 @@ void EmitWriteSharedU8(EmitContext& ctx, Id offset, Id value) {
}
void EmitWriteSharedU16(EmitContext& ctx, Id offset, Id value) {
if (ctx.profile.support_explicit_workgroup_layout) {
if (ctx.uses_explicit_workgroup_layout) {
const Id pointer{Pointer(ctx, ctx.shared_u16, ctx.shared_memory_u16, offset, 1)};
ctx.OpStore(pointer, ctx.OpUConvert(ctx.U16, value));
} else {
@ -130,7 +133,7 @@ void EmitWriteSharedU16(EmitContext& ctx, Id offset, Id value) {
void EmitWriteSharedU32(EmitContext& ctx, Id offset, Id value) {
Id pointer{};
if (ctx.profile.support_explicit_workgroup_layout) {
if (ctx.uses_explicit_workgroup_layout) {
pointer = Pointer(ctx, ctx.shared_u32, ctx.shared_memory_u32, offset, 2);
} else {
const Id shift{ctx.Const(2U)};
@ -141,7 +144,7 @@ void EmitWriteSharedU32(EmitContext& ctx, Id offset, Id value) {
}
void EmitWriteSharedU64(EmitContext& ctx, Id offset, Id value) {
if (ctx.profile.support_explicit_workgroup_layout) {
if (ctx.uses_explicit_workgroup_layout) {
const Id pointer{Pointer(ctx, ctx.shared_u32x2, ctx.shared_memory_u32x2, offset, 3)};
ctx.OpStore(pointer, value);
return;
@ -156,7 +159,7 @@ void EmitWriteSharedU64(EmitContext& ctx, Id offset, Id value) {
}
void EmitWriteSharedU128(EmitContext& ctx, Id offset, Id value) {
if (ctx.profile.support_explicit_workgroup_layout) {
if (ctx.uses_explicit_workgroup_layout) {
const Id pointer{Pointer(ctx, ctx.shared_u32x4, ctx.shared_memory_u32x4, offset, 4)};
ctx.OpStore(pointer, value);
return;

8
src/shader_recompiler/backend/spirv/spirv_emit_context.cpp
View File

@ -371,7 +371,7 @@ Id CasFunction(EmitContext& ctx, Operation operation, Id value_type) {
Id CasLoop(EmitContext& ctx, Operation operation, Id array_pointer, Id element_pointer,
Id value_type, Id memory_type, spv::Scope scope) {
const bool is_shared{scope == spv::Scope::Workgroup};
const bool is_struct{!is_shared || ctx.profile.support_explicit_workgroup_layout};
const bool is_struct{!is_shared || ctx.uses_explicit_workgroup_layout};
const Id cas_func{CasFunction(ctx, operation, value_type)};
const Id zero{ctx.u32_zero_value};
const Id scope_id{ctx.Const(static_cast<u32>(scope))};
@ -620,7 +620,11 @@ void EmitContext::DefineSharedMemory(const IR::Program& program) {
return std::make_tuple(variable, element_pointer, pointer);
}};
if (profile.support_explicit_workgroup_layout) {
uses_explicit_workgroup_layout =
profile.support_explicit_workgroup_layout &&
(!program.info.uses_int8 || profile.support_workgroup_layout_8bit_access) &&
(!program.info.uses_int16 || profile.support_workgroup_layout_16bit_access);
if (uses_explicit_workgroup_layout) {
AddExtension("SPV_KHR_workgroup_memory_explicit_layout");
AddCapability(spv::Capability::WorkgroupMemoryExplicitLayoutKHR);
if (program.info.uses_int8) {

4
src/shader_recompiler/backend/spirv/spirv_emit_context.h
View File

@ -310,6 +310,10 @@ public:
Id local_memory{};
/// True when this shader's shared memory uses SPV_KHR_workgroup_memory_explicit_layout.
/// False when the host lacks the extension or a width this shader accesses natively.
bool uses_explicit_workgroup_layout{};
Id shared_memory_u8{};
Id shared_memory_u16{};
Id shared_memory_u32{};

4
src/shader_recompiler/profile.h
View File

@ -31,14 +31,16 @@ struct Profile {
bool support_fp32_signed_zero_nan_preserve{};
bool support_fp64_signed_zero_nan_preserve{};
bool support_explicit_workgroup_layout{};
bool support_workgroup_layout_8bit_access{};
bool support_workgroup_layout_16bit_access{};
bool support_vote{};
/// Bitmask over Shader::Stage of stages where the host supports subgroup operations
u32 supported_subgroup_stages{0x7F};
bool support_viewport_index_layer_non_geometry{};
bool support_viewport_mask{};
bool support_typeless_image_loads{};
bool support_demote_to_helper_invocation{};
bool support_int64_atomics{};
bool support_shared_int64_atomics{};
bool support_derivative_control{};
bool support_geometry_shader_passthrough{};
bool support_native_ndc{};

5
src/video_core/renderer_vulkan/vk_pipeline_cache.cpp
View File

@ -420,6 +420,10 @@ PipelineCache::PipelineCache(Tegra::MaxwellDeviceMemoryManager& device_memory_,
float_control.shaderSignedZeroInfNanPreserveFloat64 != VK_FALSE &&
!has_broken_sz_inf_nan,
.support_explicit_workgroup_layout = device.IsKhrWorkgroupMemoryExplicitLayoutSupported(),
.support_workgroup_layout_8bit_access =
device.IsWorkgroupMemoryExplicitLayout8BitSupported(),
.support_workgroup_layout_16bit_access =
device.IsWorkgroupMemoryExplicitLayout16BitSupported(),
.support_vote = device.IsSubgroupFeatureSupported(VK_SUBGROUP_FEATURE_VOTE_BIT),
.supported_subgroup_stages = supported_subgroup_stages,
.support_viewport_index_layer_non_geometry =
@ -429,6 +433,7 @@ PipelineCache::PipelineCache(Tegra::MaxwellDeviceMemoryManager& device_memory_,
.support_demote_to_helper_invocation =
device.IsExtShaderDemoteToHelperInvocationSupported(),
.support_int64_atomics = device.IsExtShaderAtomicInt64Supported(),
.support_shared_int64_atomics = device.IsSharedInt64AtomicsSupported(),
.support_derivative_control = true,
.support_geometry_shader_passthrough = device.IsNvGeometryShaderPassthroughSupported(),
.support_native_ndc = device.IsExtDepthClipControlSupported(),

32
src/video_core/vulkan_common/vulkan_device.cpp
View File

@ -501,13 +501,6 @@ Device::Device(VkInstance instance_, vk::PhysicalDevice physical_, VkSurfaceKHR
LOG_WARNING(Render_Vulkan,
"Qualcomm drivers have slow push descriptor implementation");
RemoveExtension(extensions.push_descriptor, VK_KHR_PUSH_DESCRIPTOR_EXTENSION_NAME);
LOG_WARNING(Render_Vulkan,
"Disabling 64-bit integer features on Qualcomm proprietary drivers");
RemoveExtensionFeature(extensions.shader_atomic_int64, features.shader_atomic_int64,
VK_KHR_SHADER_ATOMIC_INT64_EXTENSION_NAME);
features.shader_atomic_int64.shaderBufferInt64Atomics = false;
features.shader_atomic_int64.shaderSharedInt64Atomics = false;
features.features.shaderInt64 = false;
#if defined(__ANDROID__) && defined(ARCHITECTURE_arm64)
// BCn patching only safe on Android 9+ (API 28+). Older versions crash on driver load.
@ -560,7 +553,6 @@ Device::Device(VkInstance instance_, vk::PhysicalDevice physical_, VkSurfaceKHR
features.shader_float16_int8.shaderFloat16 = false;
}
// Mali/ NVIDIA proprietary drivers: Shader stencil export not supported
// Use hardware depth/stencil blits instead when available
if (!extensions.shader_stencil_export) {
LOG_INFO(Render_Vulkan,
@ -573,8 +565,8 @@ Device::Device(VkInstance instance_, vk::PhysicalDevice physical_, VkSurfaceKHR
if (!is_blit_depth24_stencil8_supported && !is_blit_depth32_stencil8_supported) {
LOG_WARNING(Render_Vulkan,
"NVIDIA: Neither shader export nor hardware blits available for "
"depth/stencil. Performance may be degraded.");
"Neither shader export nor hardware blits available for "
"depth/stencil.");
}
}
}
@ -656,17 +648,8 @@ Device::Device(VkInstance instance_, vk::PhysicalDevice physical_, VkSurfaceKHR
}
const auto dyna_state = Settings::values.dyna_state.GetValue();
// Base dynamic states (VIEWPORT, SCISSOR, DEPTH_BIAS, etc.) are ALWAYS active in vk_graphics_pipeline.cpp
// This slider controls EXTENDED dynamic states with accumulative levels per Vulkan specs:
// Level 0 = Core Dynamic States only (Vulkan 1.0)
// Level 1 = Core + VK_EXT_extended_dynamic_state
// Level 2 = Core + VK_EXT_extended_dynamic_state + VK_EXT_extended_dynamic_state2
// Level 3 = Core + VK_EXT_extended_dynamic_state + VK_EXT_extended_dynamic_state2 + VK_EXT_extended_dynamic_state3
switch (dyna_state) {
case Settings::ExtendedDynamicState::Disabled:
// Level 0: Disable all extended dynamic state extensions
RemoveExtensionFeature(extensions.extended_dynamic_state, features.extended_dynamic_state,
VK_EXT_EXTENDED_DYNAMIC_STATE_EXTENSION_NAME);
RemoveExtensionFeature(extensions.extended_dynamic_state2, features.extended_dynamic_state2,
@ -677,7 +660,6 @@ Device::Device(VkInstance instance_, vk::PhysicalDevice physical_, VkSurfaceKHR
dynamic_state3_enables = false;
break;
case Settings::ExtendedDynamicState::EDS1:
// Level 1: Enable EDS1, disable EDS2 and EDS3
RemoveExtensionFeature(extensions.extended_dynamic_state2, features.extended_dynamic_state2,
VK_EXT_EXTENDED_DYNAMIC_STATE_2_EXTENSION_NAME);
RemoveExtensionFeature(extensions.extended_dynamic_state3, features.extended_dynamic_state3,
@ -686,7 +668,6 @@ Device::Device(VkInstance instance_, vk::PhysicalDevice physical_, VkSurfaceKHR
dynamic_state3_enables = false;
break;
case Settings::ExtendedDynamicState::EDS2:
// Level 2: Enable EDS1 + EDS2, disable EDS3
RemoveExtensionFeature(extensions.extended_dynamic_state3, features.extended_dynamic_state3,
VK_EXT_EXTENDED_DYNAMIC_STATE_3_EXTENSION_NAME);
dynamic_state3_blending = false;
@ -694,12 +675,9 @@ Device::Device(VkInstance instance_, vk::PhysicalDevice physical_, VkSurfaceKHR
break;
case Settings::ExtendedDynamicState::EDS3:
default:
// Level 3: Enable all (EDS1 + EDS2 + EDS3)
break;
}
// VK_EXT_vertex_input_dynamic_state is independent from EDS
// It can be enabled even without extended_dynamic_state
if (!Settings::values.vertex_input_dynamic_state.GetValue()) {
RemoveExtensionFeature(extensions.vertex_input_dynamic_state, features.vertex_input_dynamic_state, VK_EXT_VERTEX_INPUT_DYNAMIC_STATE_EXTENSION_NAME);
}
@ -1281,8 +1259,7 @@ void Device::RemoveUnsuitableExtensions() {
VK_EXT_IMAGE_ROBUSTNESS_EXTENSION_NAME);
// VK_KHR_shader_atomic_int64
extensions.shader_atomic_int64 = features.shader_atomic_int64.shaderBufferInt64Atomics &&
features.shader_atomic_int64.shaderSharedInt64Atomics;
extensions.shader_atomic_int64 = features.shader_atomic_int64.shaderBufferInt64Atomics;
RemoveExtensionFeatureIfUnsuitable(extensions.shader_atomic_int64, features.shader_atomic_int64,
VK_KHR_SHADER_ATOMIC_INT64_EXTENSION_NAME);
@ -1331,10 +1308,7 @@ void Device::RemoveUnsuitableExtensions() {
// VK_KHR_workgroup_memory_explicit_layout
extensions.workgroup_memory_explicit_layout =
features.features.shaderInt16 &&
features.workgroup_memory_explicit_layout.workgroupMemoryExplicitLayout &&
features.workgroup_memory_explicit_layout.workgroupMemoryExplicitLayout8BitAccess &&
features.workgroup_memory_explicit_layout.workgroupMemoryExplicitLayout16BitAccess &&
features.workgroup_memory_explicit_layout.workgroupMemoryExplicitLayoutScalarBlockLayout;
RemoveExtensionFeatureIfUnsuitable(extensions.workgroup_memory_explicit_layout,
features.workgroup_memory_explicit_layout,

23
src/video_core/vulkan_common/vulkan_device.h
View File

@ -488,6 +488,18 @@ FN_MAX_LIMIT_LIST
return extensions.workgroup_memory_explicit_layout;
}
/// Returns true if the device supports 8-bit accesses to workgroup explicit layout memory.
bool IsWorkgroupMemoryExplicitLayout8BitSupported() const {
return extensions.workgroup_memory_explicit_layout &&
features.workgroup_memory_explicit_layout.workgroupMemoryExplicitLayout8BitAccess;
}
/// Returns true if the device supports 16-bit accesses to workgroup explicit layout memory.
bool IsWorkgroupMemoryExplicitLayout16BitSupported() const {
return extensions.workgroup_memory_explicit_layout && features.features.shaderInt16 &&
features.workgroup_memory_explicit_layout.workgroupMemoryExplicitLayout16BitAccess;
}
/// Returns true if the device supports VK_KHR_image_format_list.
bool IsKhrImageFormatListSupported() const {
return extensions.image_format_list || instance_version >= VK_API_VERSION_1_2;
@ -716,9 +728,16 @@ FN_MAX_LIMIT_LIST
features.provoking_vertex.transformFeedbackPreservesProvokingVertex;
}
/// Returns true if the device supports VK_KHR_shader_atomic_int64.
/// Returns true if the device supports int64 atomics on storage buffers.
bool IsExtShaderAtomicInt64Supported() const {
return extensions.shader_atomic_int64;
return extensions.shader_atomic_int64 &&
features.shader_atomic_int64.shaderBufferInt64Atomics;
}
/// Returns true if the device supports int64 atomics on workgroup (shared) memory.
bool IsSharedInt64AtomicsSupported() const {
return extensions.shader_atomic_int64 &&
features.shader_atomic_int64.shaderSharedInt64Atomics;
}
bool IsExtConditionalRendering() const {

Write Preview
Loading…
Cancel
Save