Zebin: add support for has_dpas

KrystianChmielewski · Compute-Runtime-Automation · commit f20cdac7d364 · 2021-10-28T09:59:21.000+02:00
Signed-off-by: Krystian Chmielewski &lt;krystian.chmielewski@intel.com&gt;
diff --git a/opencl/source/kernel/kernel.cpp b/opencl/source/kernel/kernel.cpp
@@ -2788,7 +2788,7 @@ void Kernel::reconfigureKernel() {
         maxKernelWorkGroupSize >>= 1;
     }
     this->containsStatelessWrites = kernelDescriptor.kernelAttributes.flags.usesStatelessWrites;
-    this->specialPipelineSelectMode = kernelDescriptor.extendedInfo.get() ? kernelDescriptor.extendedInfo->specialPipelineSelectModeRequired() : false;
+    this->specialPipelineSelectMode = kernelDescriptor.kernelAttributes.flags.usesSpecialPipelineSelectMode;
 }
 bool Kernel::requiresCacheFlushCommand(const CommandQueue &commandQueue) const {
     if (false == HwHelper::cacheFlushAfterWalkerSupported(commandQueue.getDevice().getHardwareInfo())) {
diff --git a/shared/source/command_container/command_encoder_xehp_and_later.inl b/shared/source/command_container/command_encoder_xehp_and_later.inl
@@ -70,12 +70,10 @@ void EncodeDispatchKernel<Family>::encode(CommandContainer &container,
         container.allocateNextCommandBuffer();
     }
 
-    if (kernelDescriptor.extendedInfo) {
-        bool specialModeRequired = kernelDescriptor.extendedInfo->specialPipelineSelectModeRequired();
-        if (PreambleHelper<Family>::isSpecialPipelineSelectModeChanged(container.lastPipelineSelectModeRequired, specialModeRequired, hwInfo)) {
-            container.lastPipelineSelectModeRequired = specialModeRequired;
-            EncodeComputeMode<Family>::adjustPipelineSelect(container, kernelDescriptor);
-        }
+    bool specialModeRequired = kernelDescriptor.kernelAttributes.flags.usesSpecialPipelineSelectMode;
+    if (PreambleHelper<Family>::isSpecialPipelineSelectModeChanged(container.lastPipelineSelectModeRequired, specialModeRequired, hwInfo)) {
+        container.lastPipelineSelectModeRequired = specialModeRequired;
+        EncodeComputeMode<Family>::adjustPipelineSelect(container, kernelDescriptor);
     }
 
     WALKER_TYPE walkerCmd = Family::cmdInitGpgpuWalker;
@@ -581,7 +579,7 @@ template <typename Family>
 void EncodeComputeMode<Family>::adjustPipelineSelect(CommandContainer &container, const NEO::KernelDescriptor &kernelDescriptor) {
     using PIPELINE_SELECT = typename Family::PIPELINE_SELECT;
     auto pipelineSelectCmd = Family::cmdInitPipelineSelect;
-    auto isSpecialModeSelected = kernelDescriptor.extendedInfo && kernelDescriptor.extendedInfo->specialPipelineSelectModeRequired();
+    auto isSpecialModeSelected = kernelDescriptor.kernelAttributes.flags.usesSpecialPipelineSelectMode;
 
     PreambleHelper<Family>::appendProgramPipelineSelect(&pipelineSelectCmd, isSpecialModeSelected, container.getDevice()->getHardwareInfo());
 
diff --git a/shared/source/device_binary_format/elf/zebin_elf.h b/shared/source/device_binary_format/elf/zebin_elf.h
@@ -171,6 +171,7 @@ static constexpr ConstStringRef disableMidThreadPreemption("disable_mid_thread_p
 static constexpr ConstStringRef grfCount("grf_count");
 static constexpr ConstStringRef has4gbBuffers("has_4gb_buffers");
 static constexpr ConstStringRef hasDeviceEnqueue("has_device_enqueue");
+static constexpr ConstStringRef hasDpas("has_dpas");
 static constexpr ConstStringRef hasFenceForImageAccess("has_fence_for_image_access");
 static constexpr ConstStringRef hasGlobalAtomics("has_global_atomics");
 static constexpr ConstStringRef hasMultiScratchSpaces("has_multi_scratch_spaces");
@@ -287,6 +288,7 @@ using DisableMidThreadPreemptionT = bool;
 using GrfCountT = int32_t;
 using Has4GBBuffersT = bool;
 using HasDeviceEnqueueT = bool;
+using HasDpasT = bool;
 using HasFenceForImageAccessT = bool;
 using HasGlobalAtomicsT = bool;
 using HasMultiScratchSpacesT = bool;
@@ -309,6 +311,7 @@ static constexpr BarrierCountT barrierCount = 0;
 static constexpr DisableMidThreadPreemptionT disableMidThreadPreemption = false;
 static constexpr Has4GBBuffersT has4GBBuffers = false;
 static constexpr HasDeviceEnqueueT hasDeviceEnqueue = false;
+static constexpr HasDpasT hasDpas = false;
 static constexpr HasFenceForImageAccessT hasFenceForImageAccess = false;
 static constexpr HasGlobalAtomicsT hasGlobalAtomics = false;
 static constexpr HasMultiScratchSpacesT hasMultiScratchSpaces = false;
@@ -338,6 +341,7 @@ struct ExecutionEnvBaseT {
     GrfCountT grfCount = -1;
     Has4GBBuffersT has4GBBuffers = Defaults::has4GBBuffers;
     HasDeviceEnqueueT hasDeviceEnqueue = Defaults::hasDeviceEnqueue;
+    HasDpasT hasDpas = Defaults::hasDpas;
     HasFenceForImageAccessT hasFenceForImageAccess = Defaults::hasFenceForImageAccess;
     HasGlobalAtomicsT hasGlobalAtomics = Defaults::hasGlobalAtomics;
     HasMultiScratchSpacesT hasMultiScratchSpaces = Defaults::hasMultiScratchSpaces;
diff --git a/shared/source/device_binary_format/zebin_decoder.cpp b/shared/source/device_binary_format/zebin_decoder.cpp
@@ -253,6 +253,8 @@ DecodeError readZeInfoExecutionEnvironment(const NEO::Yaml::YamlParser &parser,
             validExecEnv = validExecEnv & readZeInfoValueChecked(parser, execEnvMetadataNd, outExecEnv.has4GBBuffers, context, outErrReason);
         } else if (NEO::Elf::ZebinKernelMetadata::Tags::Kernel::ExecutionEnv::hasDeviceEnqueue == key) {
             validExecEnv = validExecEnv & readZeInfoValueChecked(parser, execEnvMetadataNd, outExecEnv.hasDeviceEnqueue, context, outErrReason);
+        } else if (NEO::Elf::ZebinKernelMetadata::Tags::Kernel::ExecutionEnv::hasDpas == key) {
+            validExecEnv = validExecEnv & readZeInfoValueChecked(parser, execEnvMetadataNd, outExecEnv.hasDpas, context, outErrReason);
         } else if (NEO::Elf::ZebinKernelMetadata::Tags::Kernel::ExecutionEnv::hasFenceForImageAccess == key) {
             validExecEnv = validExecEnv & readZeInfoValueChecked(parser, execEnvMetadataNd, outExecEnv.hasFenceForImageAccess, context, outErrReason);
         } else if (NEO::Elf::ZebinKernelMetadata::Tags::Kernel::ExecutionEnv::hasGlobalAtomics == key) {
@@ -1005,6 +1007,7 @@ NEO::DecodeError populateKernelDescriptor(NEO::ProgramInfo &dst, NEO::Elf::Elf<N
     kernelDescriptor.kernelAttributes.flags.usesFencesForReadWriteImages = execEnv.hasFenceForImageAccess;
     kernelDescriptor.kernelAttributes.flags.useGlobalAtomics = execEnv.hasGlobalAtomics;
     kernelDescriptor.kernelAttributes.flags.usesStatelessWrites = (false == execEnv.hasNoStatelessWrite);
+    kernelDescriptor.kernelAttributes.flags.usesSpecialPipelineSelectMode = execEnv.hasDpas;
     kernelDescriptor.entryPoints.skipPerThreadDataLoad = execEnv.offsetToSkipPerThreadDataLoad;
     kernelDescriptor.entryPoints.skipSetFFIDGP = execEnv.offsetToSkipSetFfidGp;
     kernelDescriptor.kernelMetadata.requiredSubGroupSize = execEnv.requiredSubGroupSize;
diff --git a/shared/source/kernel/kernel_descriptor.h b/shared/source/kernel/kernel_descriptor.h
@@ -28,7 +28,6 @@ using InstructionsSegmentOffset = uint16_t;
 
 struct ExtendedInfoBase {
     virtual ~ExtendedInfoBase() = default;
-    virtual bool specialPipelineSelectModeRequired() const { return false; }
 };
 
 struct KernelDescriptor {
@@ -84,6 +83,7 @@ struct KernelDescriptor {
 
         union {
             struct {
+                bool usesSpecialPipelineSelectMode : 1;
                 bool usesStringMapForPrintf : 1;
                 bool usesPrintf : 1;
                 bool usesFencesForReadWriteImages : 1;
diff --git a/shared/source/kernel/kernel_descriptor_from_patchtokens.cpp b/shared/source/kernel/kernel_descriptor_from_patchtokens.cpp
@@ -55,11 +55,12 @@ void populateKernelDescriptor(KernelDescriptor &dst, const SPatchExecutionEnviro
     dst.kernelAttributes.simdSize = execEnv.LargestCompiledSIMDSize;
     dst.kernelAttributes.barrierCount = execEnv.HasBarriers;
 
-    dst.kernelAttributes.flags.usesDeviceSideEnqueue = (0 != execEnv.HasDeviceEnqueue);
     dst.kernelAttributes.flags.requiresDisabledMidThreadPreemption = (0 != execEnv.DisableMidThreadPreemption);
-    dst.kernelAttributes.flags.usesFencesForReadWriteImages = (0 != execEnv.UsesFencesForReadWriteImages);
     dst.kernelAttributes.flags.requiresSubgroupIndependentForwardProgress = (0 != execEnv.SubgroupIndependentForwardProgressRequired);
     dst.kernelAttributes.flags.useGlobalAtomics = (0 != execEnv.HasGlobalAtomics);
+    dst.kernelAttributes.flags.usesDeviceSideEnqueue = (0 != execEnv.HasDeviceEnqueue);
+    dst.kernelAttributes.flags.usesFencesForReadWriteImages = (0 != execEnv.UsesFencesForReadWriteImages);
+    dst.kernelAttributes.flags.usesSpecialPipelineSelectMode = (0 != execEnv.HasDPAS);
     dst.kernelAttributes.flags.usesStatelessWrites = (0 != execEnv.StatelessWritesCount);
 
     dst.kernelMetadata.compiledSubGroupsNumber = execEnv.CompiledSubGroupsNumber;
diff --git a/shared/test/unit_test/device_binary_format/zebin_decoder_tests.cpp b/shared/test/unit_test/device_binary_format/zebin_decoder_tests.cpp
@@ -946,6 +946,7 @@ TEST(ReadZeInfoExecutionEnvironment, GivenValidYamlEntriesThenSetProperMembers)
         grf_count : 13
         has_4gb_buffers : true
         has_device_enqueue : true
+        has_dpas : true
         has_fence_for_image_access : true
         has_global_atomics : true
         has_multi_scratch_spaces : true
@@ -987,6 +988,7 @@ TEST(ReadZeInfoExecutionEnvironment, GivenValidYamlEntriesThenSetProperMembers)
     EXPECT_EQ(13, execEnv.grfCount);
     EXPECT_TRUE(execEnv.has4GBBuffers);
     EXPECT_TRUE(execEnv.hasDeviceEnqueue);
+    EXPECT_TRUE(execEnv.hasDpas);
     EXPECT_TRUE(execEnv.hasFenceForImageAccess);
     EXPECT_TRUE(execEnv.hasGlobalAtomics);
     EXPECT_TRUE(execEnv.hasMultiScratchSpaces);
diff --git a/shared/test/unit_test/kernel/kernel_descriptor_tests.cpp b/shared/test/unit_test/kernel/kernel_descriptor_tests.cpp
@@ -109,8 +109,3 @@ TEST(KernelDescriptorAttributesSupportsBuffersBiggerThan4Gb, GivenStatefulBuffer
     desc.kernelAttributes.bufferAddressingMode = NEO::KernelDescriptor::BindlessAndStateless;
     EXPECT_FALSE(desc.kernelAttributes.supportsBuffersBiggerThan4Gb());
 }
-
-TEST(KernelDescriptorTest, givenExtendedInfoWhenAskingForSpecialPipelineSelectModeThenReturnFalse) {
-    NEO::ExtendedInfoBase extendedInfo;
-    EXPECT_FALSE(extendedInfo.specialPipelineSelectModeRequired());
-}

Original file line number	Diff line number	Diff line change
`@@ -2788,7 +2788,7 @@ void Kernel::reconfigureKernel() {`
`2788`	`2788`	`maxKernelWorkGroupSize >>= 1;`
`2789`	`2789`	`}`
`2790`	`2790`	`this->containsStatelessWrites = kernelDescriptor.kernelAttributes.flags.usesStatelessWrites;`
`2791`		`- this->specialPipelineSelectMode = kernelDescriptor.extendedInfo.get() ? kernelDescriptor.extendedInfo->specialPipelineSelectModeRequired() : false;`
	`2791`	`+ this->specialPipelineSelectMode = kernelDescriptor.kernelAttributes.flags.usesSpecialPipelineSelectMode;`
`2792`	`2792`	`}`
`2793`	`2793`	`bool Kernel::requiresCacheFlushCommand(const CommandQueue &commandQueue) const {`
`2794`	`2794`	`if (false == HwHelper::cacheFlushAfterWalkerSupported(commandQueue.getDevice().getHardwareInfo())) {`
Original file line number	Diff line number	Diff line change
`@@ -109,8 +109,3 @@ TEST(KernelDescriptorAttributesSupportsBuffersBiggerThan4Gb, GivenStatefulBuffer`
`109`	`109`	`desc.kernelAttributes.bufferAddressingMode = NEO::KernelDescriptor::BindlessAndStateless;`
`110`	`110`	`EXPECT_FALSE(desc.kernelAttributes.supportsBuffersBiggerThan4Gb());`
`111`	`111`	`}`
`112`		`-`
`113`		`-TEST(KernelDescriptorTest, givenExtendedInfoWhenAskingForSpecialPipelineSelectModeThenReturnFalse) {`
`114`		`- NEO::ExtendedInfoBase extendedInfo;`
`115`		`- EXPECT_FALSE(extendedInfo.specialPipelineSelectModeRequired());`
`116`		`-}`