Dawn native/wire: Add experimental subgroup limits
This CL add the experimental subgroup limits by introducing a new
structure DawnExperimentalSubgroupLimits chained after SupportedLimits.
The structure can be used for querying adapter or device limits, and the
result would be WGPU_LIMIT_U32_UNDEFINED if AllowUnsafeAPI is not
enabled.
Issue: dawn:464
Change-Id: I2ad591d4c5cf88ad69df48da8c2b01bbc1f38656
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/151282
Reviewed-by: Loko Kung <lokokung@google.com>
Kokoro: Kokoro <noreply+kokoro@google.com>
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
Commit-Queue: Zhaoming Jiang <zhaoming.jiang@intel.com>
diff --git a/dawn.json b/dawn.json
index 8756252..dcb8d10 100644
--- a/dawn.json
+++ b/dawn.json
@@ -1390,6 +1390,15 @@
{"name": "max compute workgroups per dimension", "type": "uint32_t", "default": "WGPU_LIMIT_U32_UNDEFINED"}
]
},
+ "dawn experimental subgroup limits": {
+ "category": "structure",
+ "chained": "out",
+ "chain roots": ["supported limits"],
+ "members": [
+ {"name": "min subgroup size", "type": "uint32_t", "default": "WGPU_LIMIT_U32_UNDEFINED"},
+ {"name": "max subgroup size", "type": "uint32_t", "default": "WGPU_LIMIT_U32_UNDEFINED"}
+ ]
+ },
"required limits": {
"category": "structure",
"extensible": "in",
@@ -3263,6 +3272,7 @@
{"value": 1014, "name": "render pass pixel local storage", "tags": ["dawn"]},
{"value": 1015, "name": "pipeline layout pixel local storage", "tags": ["dawn"]},
{"value": 1016, "name": "buffer host mapped pointer", "tags": ["dawn"]},
+ {"value": 1017, "name": "dawn experimental subgroup limits", "tags": ["dawn"]},
{"value": 1100, "name": "shared texture memory vk image descriptor", "tags": ["dawn", "native"]},
{"value": 1101, "name": "shared texture memory vk dedicated allocation descriptor", "tags": ["dawn", "native"]},
diff --git a/src/dawn/native/Adapter.cpp b/src/dawn/native/Adapter.cpp
index 6c52628..1706458 100644
--- a/src/dawn/native/Adapter.cpp
+++ b/src/dawn/native/Adapter.cpp
@@ -65,7 +65,10 @@
bool AdapterBase::APIGetLimits(SupportedLimits* limits) const {
DAWN_ASSERT(limits != nullptr);
- if (limits->nextInChain != nullptr) {
+ // TODO(dawn:1955): Revisit after deciding how to improve the validation for ChainedStructOut.
+ MaybeError result =
+ ValidateSTypes(limits->nextInChain, {{wgpu::SType::DawnExperimentalSubgroupLimits}});
+ if (mPhysicalDevice->GetInstance()->ConsumedError(std::move(result))) {
return false;
}
if (mUseTieredLimits) {
@@ -73,6 +76,29 @@
} else {
limits->limits = mPhysicalDevice->GetLimits().v1;
}
+ for (auto* chain = limits->nextInChain; chain; chain = chain->nextInChain) {
+ wgpu::ChainedStructOut originalChain = *chain;
+ switch (chain->sType) {
+ case (wgpu::SType::DawnExperimentalSubgroupLimits): {
+ DawnExperimentalSubgroupLimits* subgroupLimits =
+ reinterpret_cast<DawnExperimentalSubgroupLimits*>(chain);
+ if (!mTogglesState.IsEnabled(Toggle::AllowUnsafeAPIs)) {
+ // If AllowUnsafeAPIs is not enabled, return the default-initialized
+ // DawnExperimentalSubgroupLimits object, where minSubgroupSize and
+ // maxSubgroupSize are WGPU_LIMIT_U32_UNDEFINED.
+ *subgroupLimits = DawnExperimentalSubgroupLimits{};
+ } else {
+ *subgroupLimits = mPhysicalDevice->GetLimits().experimentalSubgroupLimits;
+ }
+ break;
+ }
+ default:
+ // ValidateSTypes ensures that all chained sTypes are known.
+ DAWN_UNREACHABLE();
+ }
+ // Recover the original chain
+ *chain = originalChain;
+ }
return true;
}
@@ -181,15 +207,17 @@
}
if (descriptor->requiredLimits != nullptr) {
+ // Only consider limits in RequiredLimits structure, and currently no chained structure
+ // supported.
+ DAWN_INVALID_IF(descriptor->requiredLimits->nextInChain != nullptr,
+ "can not chain after requiredLimits.");
+
SupportedLimits supportedLimits;
bool success = APIGetLimits(&supportedLimits);
DAWN_ASSERT(success);
DAWN_TRY_CONTEXT(ValidateLimits(supportedLimits.limits, descriptor->requiredLimits->limits),
"validating required limits");
-
- DAWN_INVALID_IF(descriptor->requiredLimits->nextInChain != nullptr,
- "nextInChain is not nullptr.");
}
return mPhysicalDevice->CreateDevice(this, descriptor, deviceToggles);
diff --git a/src/dawn/native/Device.cpp b/src/dawn/native/Device.cpp
index 808f453..7a9bd34c 100644
--- a/src/dawn/native/Device.cpp
+++ b/src/dawn/native/Device.cpp
@@ -219,6 +219,9 @@
} else {
GetDefaultLimits(&mLimits.v1, adapter->GetFeatureLevel());
}
+ // Get experimentalSubgroupLimits from physical device
+ mLimits.experimentalSubgroupLimits =
+ GetPhysicalDevice()->GetLimits().experimentalSubgroupLimits;
mFormatTable = BuildFormatTable(this);
@@ -1505,10 +1508,37 @@
bool DeviceBase::APIGetLimits(SupportedLimits* limits) const {
DAWN_ASSERT(limits != nullptr);
- if (limits->nextInChain != nullptr) {
+ // TODO(dawn:1955): Revisit after deciding how to improve the validation for ChainedStructOut.
+ MaybeError result =
+ ValidateSTypes(limits->nextInChain, {{wgpu::SType::DawnExperimentalSubgroupLimits}});
+ if (GetPhysicalDevice()->GetInstance()->ConsumedError(std::move(result))) {
return false;
}
+
limits->limits = mLimits.v1;
+
+ for (auto* chain = limits->nextInChain; chain; chain = chain->nextInChain) {
+ wgpu::ChainedStructOut originalChain = *chain;
+ switch (chain->sType) {
+ case (wgpu::SType::DawnExperimentalSubgroupLimits): {
+ DawnExperimentalSubgroupLimits* subgroupLimits =
+ reinterpret_cast<DawnExperimentalSubgroupLimits*>(chain);
+ if (!mToggles.IsEnabled(Toggle::AllowUnsafeAPIs)) {
+ // If AllowUnsafeAPIs is not enabled, return the default-initialized
+ // DawnExperimentalSubgroupLimits object, where minSubgroupSize and
+ // maxSubgroupSize are WGPU_LIMIT_U32_UNDEFINED.
+ *subgroupLimits = DawnExperimentalSubgroupLimits{};
+ } else {
+ *subgroupLimits = mLimits.experimentalSubgroupLimits;
+ }
+ break;
+ }
+ default:
+ DAWN_UNREACHABLE();
+ }
+ // Recover the original chain
+ *chain = originalChain;
+ }
return true;
}
diff --git a/src/dawn/native/Limits.h b/src/dawn/native/Limits.h
index 83ae805..798b1da 100644
--- a/src/dawn/native/Limits.h
+++ b/src/dawn/native/Limits.h
@@ -24,6 +24,7 @@
struct CombinedLimits {
Limits v1;
+ DawnExperimentalSubgroupLimits experimentalSubgroupLimits;
};
// Populate |limits| with the default limits.
diff --git a/src/dawn/native/d3d12/PhysicalDeviceD3D12.cpp b/src/dawn/native/d3d12/PhysicalDeviceD3D12.cpp
index 2424039..d93b1fe 100644
--- a/src/dawn/native/d3d12/PhysicalDeviceD3D12.cpp
+++ b/src/dawn/native/d3d12/PhysicalDeviceD3D12.cpp
@@ -348,6 +348,13 @@
// TODO(crbug.com/dawn/1448):
// - maxInterStageShaderVariables
+ // Experimental limits for subgroups
+ limits->experimentalSubgroupLimits.minSubgroupSize = mDeviceInfo.waveLaneCountMin;
+ // Currently the WaveLaneCountMax queried from D3D12 API is not reliable and the meaning is
+ // unclear. Use 128 instead, which is the largest possible size. Reference:
+ // https://github.com/Microsoft/DirectXShaderCompiler/wiki/Wave-Intrinsics#:~:text=UINT%20WaveLaneCountMax
+ limits->experimentalSubgroupLimits.maxSubgroupSize = 128u;
+
return {};
}
diff --git a/src/dawn/native/metal/BackendMTL.mm b/src/dawn/native/metal/BackendMTL.mm
index 434f8bb..69cf2d4 100644
--- a/src/dawn/native/metal/BackendMTL.mm
+++ b/src/dawn/native/metal/BackendMTL.mm
@@ -805,6 +805,10 @@
// TODO(crbug.com/dawn/1448):
// - maxInterStageShaderVariables
+ // Experimental limits for subgroups
+ limits->experimentalSubgroupLimits.minSubgroupSize = 4;
+ limits->experimentalSubgroupLimits.maxSubgroupSize = 64;
+
return {};
}
diff --git a/src/dawn/native/vulkan/PhysicalDeviceVk.cpp b/src/dawn/native/vulkan/PhysicalDeviceVk.cpp
index e0ddd4c..bb92142 100644
--- a/src/dawn/native/vulkan/PhysicalDeviceVk.cpp
+++ b/src/dawn/native/vulkan/PhysicalDeviceVk.cpp
@@ -518,6 +518,12 @@
// TODO(crbug.com/dawn/1448):
// - maxInterStageShaderVariables
+ // Experimental limits for subgroups
+ limits->experimentalSubgroupLimits.minSubgroupSize =
+ mDeviceInfo.subgroupSizeControlProperties.minSubgroupSize;
+ limits->experimentalSubgroupLimits.maxSubgroupSize =
+ mDeviceInfo.subgroupSizeControlProperties.maxSubgroupSize;
+
return {};
}
diff --git a/src/dawn/wire/client/LimitsAndFeatures.cpp b/src/dawn/wire/client/LimitsAndFeatures.cpp
index 6983b34..9cea113 100644
--- a/src/dawn/wire/client/LimitsAndFeatures.cpp
+++ b/src/dawn/wire/client/LimitsAndFeatures.cpp
@@ -25,10 +25,28 @@
bool LimitsAndFeatures::GetLimits(WGPUSupportedLimits* limits) const {
DAWN_ASSERT(limits != nullptr);
- if (limits->nextInChain != nullptr) {
- return false;
- }
+ auto* originalNextInChain = limits->nextInChain;
*limits = mLimits;
+ limits->nextInChain = originalNextInChain;
+ // Handle other requiring limits that chained after WGPUSupportedLimits
+ for (auto* chain = limits->nextInChain; chain; chain = chain->next) {
+ // Store the WGPUChainedStructOut to restore the chain after assignment.
+ WGPUChainedStructOut originalChainedStructOut = *chain;
+ switch (chain->sType) {
+ case (WGPUSType_DawnExperimentalSubgroupLimits): {
+ auto* experimentalSubgroupLimits =
+ reinterpret_cast<WGPUDawnExperimentalSubgroupLimits*>(chain);
+ // This assignment break the next field of WGPUChainedStructOut head.
+ *experimentalSubgroupLimits = mExperimentalSubgroupLimits;
+ break;
+ }
+ default:
+ // Fail if unknown sType found.
+ return false;
+ }
+ // Restore the chain.
+ *chain = originalChainedStructOut;
+ }
return true;
}
@@ -50,6 +68,20 @@
DAWN_ASSERT(limits != nullptr);
mLimits = *limits;
mLimits.nextInChain = nullptr;
+ // Handle other limits that chained after WGPUSupportedLimits
+ for (auto* chain = limits->nextInChain; chain; chain = chain->next) {
+ switch (chain->sType) {
+ case (WGPUSType_DawnExperimentalSubgroupLimits): {
+ auto* experimentalSubgroupLimits =
+ reinterpret_cast<WGPUDawnExperimentalSubgroupLimits*>(chain);
+ mExperimentalSubgroupLimits = *experimentalSubgroupLimits;
+ mExperimentalSubgroupLimits.chain.next = nullptr;
+ break;
+ }
+ default:
+ DAWN_UNREACHABLE();
+ }
+ }
}
void LimitsAndFeatures::SetFeatures(const WGPUFeatureName* features, uint32_t featuresCount) {
diff --git a/src/dawn/wire/client/LimitsAndFeatures.h b/src/dawn/wire/client/LimitsAndFeatures.h
index cfe9353..379e7e6 100644
--- a/src/dawn/wire/client/LimitsAndFeatures.h
+++ b/src/dawn/wire/client/LimitsAndFeatures.h
@@ -35,6 +35,7 @@
private:
WGPUSupportedLimits mLimits;
+ WGPUDawnExperimentalSubgroupLimits mExperimentalSubgroupLimits;
std::unordered_set<WGPUFeatureName> mFeatures;
};
diff --git a/src/dawn/wire/server/ServerAdapter.cpp b/src/dawn/wire/server/ServerAdapter.cpp
index 3bac8a1..dbee88e 100644
--- a/src/dawn/wire/server/ServerAdapter.cpp
+++ b/src/dawn/wire/server/ServerAdapter.cpp
@@ -83,6 +83,10 @@
cmd.features = features.data();
WGPUSupportedLimits limits = {};
+ // Also query the DawnExperimentalSubgroupLimits and report to client.
+ WGPUDawnExperimentalSubgroupLimits experimentalSubgroupLimits = {};
+ experimentalSubgroupLimits.chain.sType = WGPUSType_DawnExperimentalSubgroupLimits;
+ limits.nextInChain = &experimentalSubgroupLimits.chain;
mProcs.deviceGetLimits(device, &limits);
cmd.limits = &limits;
diff --git a/src/dawn/wire/server/ServerInstance.cpp b/src/dawn/wire/server/ServerInstance.cpp
index 9d7cf94..8f0d5ea 100644
--- a/src/dawn/wire/server/ServerInstance.cpp
+++ b/src/dawn/wire/server/ServerInstance.cpp
@@ -56,13 +56,12 @@
return;
}
- WGPUAdapterProperties properties = {};
- WGPUSupportedLimits limits = {};
- std::vector<WGPUFeatureName> features;
-
// Assign the handle and allocated status if the adapter is created successfully.
AdapterObjects().FillReservation(data->adapterObjectId, adapter);
+ // Query and report the adapter supported features.
+ std::vector<WGPUFeatureName> features;
+
size_t featuresCount = mProcs.adapterEnumerateFeatures(adapter, nullptr);
features.resize(featuresCount);
mProcs.adapterEnumerateFeatures(adapter, features.data());
@@ -73,9 +72,19 @@
cmd.featuresCount = std::distance(features.begin(), it);
cmd.features = features.data();
+ // Query and report the adapter properties.
+ WGPUAdapterProperties properties = {};
mProcs.adapterGetProperties(adapter, &properties);
- mProcs.adapterGetLimits(adapter, &limits);
cmd.properties = &properties;
+
+ // Query and report the adapter limits, including DawnExperimentalSubgroupLimits.
+ WGPUSupportedLimits limits = {};
+
+ WGPUDawnExperimentalSubgroupLimits experimentalSubgroupLimits = {};
+ experimentalSubgroupLimits.chain.sType = WGPUSType_DawnExperimentalSubgroupLimits;
+ limits.nextInChain = &experimentalSubgroupLimits.chain;
+
+ mProcs.adapterGetLimits(adapter, &limits);
cmd.limits = &limits;
SerializeCommand(cmd);
