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

[spv, vk, qcom] SFC 1 x 2 + SPVE.

Browse Source
This commit is contained in:
CamilleLaVey
2025-12-25 04:44:12 -04:00
committed by Caio Oliveira
parent 2d78333479
commit 81cd6cea66
7 changed files with 269 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

64
src/shader_recompiler/backend/spirv/emit_spirv.cpp
View File

@@ -401,6 +401,69 @@ void SetupSignedNanCapabilities(const Profile& profile, const IR::Program& progr
}
}
void SetupRoundingMode(const Profile& profile, const IR::Program& program, EmitContext& ctx,
Id main_func) {
const auto& info = program.info;
// If no rounding modes used, nothing to do.
if (!info.uses_fp16_rounding_rte && !info.uses_fp16_rounding_rtz &&
!info.uses_fp32_rounding_rte && !info.uses_fp32_rounding_rtz &&
!info.uses_fp64_rounding_rte && !info.uses_fp64_rounding_rtz) {
return;
}
// Capability is required whenever any rounding mode is emitted.
ctx.AddCapability(spv::Capability::FPRoundingMode);
// FP16
if (info.uses_fp16_rounding_rte) {
if (profile.support_shader_rounding_rte_f16) {
ctx.AddExecutionMode(main_func, spv::ExecutionMode::RoundingModeRTE, 16U);
} else {
LOG_DEBUG(Shader_SPIRV, "Fp16 RTE rounding used in shader without host support");
}
}
if (info.uses_fp16_rounding_rtz) {
if (profile.support_shader_rounding_rtz_f16) {
ctx.AddExecutionMode(main_func, spv::ExecutionMode::RoundingModeRTZ, 16U);
} else {
LOG_DEBUG(Shader_SPIRV, "Fp16 RTZ rounding used in shader without host support");
}
}
// FP32
if (info.uses_fp32_rounding_rte) {
if (profile.support_shader_rounding_rte_f32) {
ctx.AddExecutionMode(main_func, spv::ExecutionMode::RoundingModeRTE, 32U);
} else {
LOG_DEBUG(Shader_SPIRV, "Fp32 RTE rounding used in shader without host support");
}
}
if (info.uses_fp32_rounding_rtz) {
if (profile.support_shader_rounding_rtz_f32) {
ctx.AddExecutionMode(main_func, spv::ExecutionMode::RoundingModeRTZ, 32U);
} else {
LOG_DEBUG(Shader_SPIRV, "Fp32 RTZ rounding used in shader without host support");
}
}
// FP64
if (info.uses_fp64_rounding_rte) {
if (profile.support_shader_rounding_rte_f64) {
ctx.AddExecutionMode(main_func, spv::ExecutionMode::RoundingModeRTE, 64U);
} else {
LOG_DEBUG(Shader_SPIRV, "Fp64 RTE rounding used in shader without host support");
}
}
if (info.uses_fp64_rounding_rtz) {
if (profile.support_shader_rounding_rtz_f64) {
ctx.AddExecutionMode(main_func, spv::ExecutionMode::RoundingModeRTZ, 64U);
} else {
LOG_DEBUG(Shader_SPIRV, "Fp64 RTZ rounding used in shader without host support");
}
}
}
void SetupTransformFeedbackCapabilities(EmitContext& ctx, Id main_func) {
if (ctx.runtime_info.xfb_count == 0) {
return;
@@ -519,6 +582,7 @@ std::vector<u32> EmitSPIRV(const Profile& profile, const RuntimeInfo& runtime_in
ctx.AddExtension("SPV_KHR_float_controls");
SetupDenormControl(profile, program, ctx, main);
SetupSignedNanCapabilities(profile, program, ctx, main);
SetupRoundingMode(profile, program, ctx, main);
}
SetupCapabilities(profile, program.info, ctx);
SetupTransformFeedbackCapabilities(ctx, main);

125
src/shader_recompiler/ir_opt/collect_shader_info_pass.cpp
View File

@@ -790,8 +790,133 @@ void VisitFpModifiers(Info& info, IR::Inst& inst) {
info.uses_fp16_denorms_preserve = true;
break;
}
switch (control.rounding) {
case IR::FpRounding::DontCare:
break;
case IR::FpRounding::RN:
info.uses_fp16_rounding_rte = true;
break;
case IR::FpRounding::RZ:
info.uses_fp16_rounding_rtz = true;
break;
default:
break;
}
break;
}
...
case IR::Opcode::FPAdd32:
case IR::Opcode::FPFma32:
case IR::Opcode::FPMul32:
case IR::Opcode::FPRoundEven32:
case IR::Opcode::FPFloor32:
case IR::Opcode::FPCeil32:
case IR::Opcode::FPTrunc32:
case IR::Opcode::FPOrdEqual32:
case IR::Opcode::FPUnordEqual32:
case IR::Opcode::FPOrdNotEqual32:
case IR::Opcode::FPUnordNotEqual32:
case IR::Opcode::FPOrdLessThan32:
case IR::Opcode::FPUnordLessThan32:
case IR::Opcode::FPOrdGreaterThan32:
case IR::Opcode::FPUnordGreaterThan32:
case IR::Opcode::FPOrdLessThanEqual32:
case IR::Opcode::FPUnordLessThanEqual32:
case IR::Opcode::FPOrdGreaterThanEqual32:
case IR::Opcode::FPUnordGreaterThanEqual32:
case IR::Opcode::ConvertF16F32:
case IR::Opcode::ConvertF64F32: {
const auto control{inst.Flags<IR::FpControl>()};
switch (control.fmz_mode) {
case IR::FmzMode::DontCare:
break;
case IR::FmzMode::FTZ:
case IR::FmzMode::FMZ:
info.uses_fp32_denorms_flush = true;
break;
case IR::FmzMode::None:
info.uses_fp32_denorms_preserve = true;
break;
}
switch (control.rounding) {
case IR::FpRounding::DontCare:
break;
case IR::FpRounding::RN:
info.uses_fp32_rounding_rte = true;
break;
case IR::FpRounding::RZ:
info.uses_fp32_rounding_rtz = true;
break;
default:
break;
}
break;
}
...
case IR::Opcode::FPAdd64:
case IR::Opcode::FPCeil64:
case IR::Opcode::FPFloor64:
case IR::Opcode::FPFma64:
case IR::Opcode::FPMax64:
case IR::Opcode::FPMin64:
case IR::Opcode::FPMul64:
case IR::Opcode::FPNeg64:
case IR::Opcode::FPRecip64:
case IR::Opcode::FPRecipSqrt64:
case IR::Opcode::FPRoundEven64:
case IR::Opcode::FPSaturate64:
case IR::Opcode::FPClamp64:
case IR::Opcode::FPTrunc64:
case IR::Opcode::FPOrdEqual64:
case IR::Opcode::FPUnordEqual64:
case IR::Opcode::FPOrdNotEqual64:
case IR::Opcode::FPUnordNotEqual64:
case IR::Opcode::FPOrdLessThan64:
case IR::Opcode::FPUnordLessThan64:
case IR::Opcode::FPOrdGreaterThan64:
case IR::Opcode::FPUnordGreaterThan64:
case IR::Opcode::FPOrdLessThanEqual64:
case IR::Opcode::FPUnordLessThanEqual64:
case IR::Opcode::FPOrdGreaterThanEqual64:
case IR::Opcode::FPUnordGreaterThanEqual64:
case IR::Opcode::FPIsNan64:
case IR::Opcode::ConvertS16F64:
case IR::Opcode::ConvertS32F64:
case IR::Opcode::ConvertS64F64:
case IR::Opcode::ConvertU16F64:
case IR::Opcode::ConvertU32F64:
case IR::Opcode::ConvertU64F64:
case IR::Opcode::ConvertF32F64:
case IR::Opcode::ConvertF64F32:
case IR::Opcode::ConvertF64S8:
case IR::Opcode::ConvertF64S16:
case IR::Opcode::ConvertF64S32:
case IR::Opcode::ConvertF64S64:
case IR::Opcode::ConvertF64U8:
case IR::Opcode::ConvertF64U16:
case IR::Opcode::ConvertF64U32:
case IR::Opcode::ConvertF64U64:
info.uses_fp64 = true;
{
const auto control{inst.Flags<IR::FpControl>()};
switch (control.rounding) {
case IR::FpRounding::DontCare:
break;
case IR::FpRounding::RN:
info.uses_fp64_rounding_rte = true;
break;
case IR::FpRounding::RZ:
info.uses_fp64_rounding_rtz = true;
break;
default:
break;
}
}
break;
case IR::Opcode::FPAdd32:
case IR::Opcode::FPFma32:
case IR::Opcode::FPMul32:

11
src/shader_recompiler/profile.h
View File

@@ -22,8 +22,19 @@ struct Profile {
bool support_int64{};
bool support_vertex_instance_id{};
bool support_float_controls{};
bool support_float_controls2{};
bool support_separate_denorm_behavior{};
bool support_separate_rounding_mode{};
/* Shader rounding mode support (SPV_KHR_float_controls floating rounding modes).
* These are set from VkPhysicalDeviceFloatControlsProperties/KHR/2KHR fields. */
bool support_shader_rounding_rte_f16{};
bool support_shader_rounding_rte_f32{};
bool support_shader_rounding_rte_f64{};
bool support_shader_rounding_rtz_f16{};
bool support_shader_rounding_rtz_f32{};
bool support_shader_rounding_rtz_f64{};
bool support_fp16_denorm_preserve{};
bool support_fp32_denorm_preserve{};
bool support_fp16_denorm_flush{};

8
src/shader_recompiler/shader_info.h
View File

@@ -290,6 +290,14 @@ struct Info {
bool uses_fp16_denorms_preserve{};
bool uses_fp32_denorms_flush{};
bool uses_fp32_denorms_preserve{};
/* Rounding mode usage detected in shader IR (set during analysis). */
bool uses_fp16_rounding_rte{};
bool uses_fp16_rounding_rtz{};
bool uses_fp32_rounding_rte{};
bool uses_fp32_rounding_rtz{};
bool uses_fp64_rounding_rte{};
bool uses_fp64_rounding_rtz{};
bool uses_int8{};
bool uses_int16{};
bool uses_int64{};

14
src/video_core/renderer_vulkan/vk_pipeline_cache.cpp
View File

@@ -377,10 +377,24 @@ PipelineCache::PipelineCache(Tegra::MaxwellDeviceMemoryManager& device_memory_,
.support_int64 = device.IsShaderInt64Supported(),
.support_vertex_instance_id = false,
.support_float_controls = float_controls_supported,
.support_float_controls2 = device.IsKhrShaderFloatControls2Supported(),
.support_separate_denorm_behavior = float_controls_supported &&
float_control.denormBehaviorIndependence == VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_ALL,
.support_separate_rounding_mode = float_controls_supported &&
float_control.roundingModeIndependence == VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_ALL,
/* Rounding mode support (RTE/RTZ) per precision */
.support_shader_rounding_rte_f16 = float_controls_supported &&
float_control.shaderRoundingModeRTEFloat16 != VK_FALSE,
.support_shader_rounding_rte_f32 = float_controls_supported &&
float_control.shaderRoundingModeRTEFloat32 != VK_FALSE,
.support_shader_rounding_rte_f64 = float_controls_supported &&
float_control.shaderRoundingModeRTEFloat64 != VK_FALSE,
.support_shader_rounding_rtz_f16 = float_controls_supported &&
float_control.shaderRoundingModeRTZFloat16 != VK_FALSE,
.support_shader_rounding_rtz_f32 = float_controls_supported &&
float_control.shaderRoundingModeRTZFloat32 != VK_FALSE,
.support_shader_rounding_rtz_f64 = float_controls_supported &&
float_control.shaderRoundingModeRTZFloat64 != VK_FALSE,
.support_fp16_denorm_preserve = float_controls_supported &&
float_control.shaderDenormPreserveFloat16 != VK_FALSE,
.support_fp32_denorm_preserve = float_controls_supported &&

40
src/video_core/vulkan_common/vulkan_device.cpp
View File

@@ -530,15 +530,30 @@ Device::Device(VkInstance instance_, vk::PhysicalDevice physical_, VkSurfaceKHR
"Qualcomm drivers require scaled vertex format emulation; forcing fallback");
if (extensions.shader_float_controls) {
LOG_WARNING(Render_Vulkan,
"Qualcomm drivers: VK_KHR_shader_float_controls is unstable; disabling usage");
RemoveExtension(extensions.shader_float_controls,
VK_KHR_SHADER_FLOAT_CONTROLS_EXTENSION_NAME);
// Allowlist: enable float controls when VK_KHR_shader_float_controls2 is present
// (more precise properties), or when driver >= 512.849.0 where the driver fix
// was introduced.
const bool adreno_safe = extensions.shader_float_controls2 ||
(properties.properties.driverVersion >= VK_MAKE_VERSION(512, 849, 0));
if (!adreno_safe) {
LOG_WARNING(Render_Vulkan,
"Qualcomm drivers: VK_KHR_shader_float_controls is unstable; disabling usage");
RemoveExtension(extensions.shader_float_controls,
VK_KHR_SHADER_FLOAT_CONTROLS_EXTENSION_NAME);
disable_shader_float_controls_usage = true;
} else {
LOG_INFO(Render_Vulkan,
"Qualcomm drivers: enabling VK_KHR_shader_float_controls (driver version {}.{}.{} or shader_float_controls2 present)",
VK_VERSION_MAJOR(properties.properties.driverVersion),
VK_VERSION_MINOR(properties.properties.driverVersion),
VK_VERSION_PATCH(properties.properties.driverVersion));
}
} else {
LOG_INFO(Render_Vulkan,
"Qualcomm drivers: VK_KHR_shader_float_controls already unavailable");
}
disable_shader_float_controls_usage = true;
// Other Qualcomm workarounds continue to apply.
RemoveExtensionFeature(extensions.shader_atomic_int64, features.shader_atomic_int64,
VK_KHR_SHADER_ATOMIC_INT64_EXTENSION_NAME);
features.shader_atomic_int64.shaderBufferInt64Atomics = false;
@@ -1172,6 +1187,11 @@ bool Device::GetSuitability(bool requires_swapchain) {
VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT_CONTROLS_PROPERTIES;
SetNext(next, properties.float_controls);
}
if (extensions.shader_float_controls2) {
properties.float_controls2.sType =
VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT_CONTROLS_PROPERTIES_2_KHR;
SetNext(next, properties.float_controls2);
}
if (extensions.push_descriptor) {
properties.push_descriptor.sType =
VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR;
@@ -1201,6 +1221,16 @@ bool Device::GetSuitability(bool requires_swapchain) {
// Perform the property fetch.
physical.GetProperties2(properties2);
// If VK_KHR_shader_float_controls2 is available, copy its inner properties for
// backwards compatibility with code that expects VkPhysicalDeviceFloatControlsProperties.
if (extensions.shader_float_controls2) {
properties.float_controls = properties.float_controls2.floatControls;
LOG_INFO(Render_Vulkan,
"VK_KHR_shader_float_controls2 supported: denormBehaviorIndependence={}, roundingModeIndependence={}",
properties.float_controls.denormBehaviorIndependence,
properties.float_controls.roundingModeIndependence);
}
// Store base properties
properties.properties = properties2.properties;

12
src/video_core/vulkan_common/vulkan_device.h
View File

@@ -90,6 +90,7 @@ VK_DEFINE_HANDLE(VmaAllocator)
EXTENSION(KHR, PUSH_DESCRIPTOR, push_descriptor) \
EXTENSION(KHR, SAMPLER_MIRROR_CLAMP_TO_EDGE, sampler_mirror_clamp_to_edge) \
EXTENSION(KHR, SHADER_FLOAT_CONTROLS, shader_float_controls) \
EXTENSION(KHR, SHADER_FLOAT_CONTROLS_2, shader_float_controls2) \
EXTENSION(KHR, SPIRV_1_4, spirv_1_4) \
EXTENSION(KHR, SWAPCHAIN, swapchain) \
EXTENSION(KHR, SWAPCHAIN_MUTABLE_FORMAT, swapchain_mutable_format) \
@@ -344,6 +345,11 @@ public:
return properties.float_controls;
}
/// Returns float control properties (KHR v2) of the device.
const VkPhysicalDeviceFloatControlsProperties2KHR& FloatControlProperties2() const {
return properties.float_controls2;
}
/// Returns true if ASTC is natively supported.
bool IsOptimalAstcSupported() const {
@@ -500,6 +506,11 @@ public:
return extensions.shader_float_controls && !disable_shader_float_controls_usage;
}
/// Returns true if VK_KHR_shader_float_controls2 is enabled.
bool IsKhrShaderFloatControls2Supported() const {
return extensions.shader_float_controls2 && !disable_shader_float_controls_usage;
}
/// Returns true if the device supports VK_KHR_workgroup_memory_explicit_layout.
bool IsKhrWorkgroupMemoryExplicitLayoutSupported() const {
return extensions.workgroup_memory_explicit_layout;
@@ -1044,6 +1055,7 @@ private:
VkPhysicalDeviceDriverProperties driver{};
VkPhysicalDeviceSubgroupProperties subgroup_properties{};
VkPhysicalDeviceFloatControlsProperties float_controls{};
VkPhysicalDeviceFloatControlsProperties2KHR float_controls2{};
VkPhysicalDevicePushDescriptorPropertiesKHR push_descriptor{};
VkPhysicalDeviceSubgroupSizeControlProperties subgroup_size_control{};
VkPhysicalDeviceTransformFeedbackPropertiesEXT transform_feedback{};