similar config string thing but for subgroups

Author: keptsecret

23_Arithmetic2UnitTest/app_resources/testSubgroup.comp.hlsl

@@ -5,28 +5,29 @@
 #include "nbl/builtin/hlsl/glsl_compat/core.hlsl"
 #include "nbl/builtin/hlsl/glsl_compat/subgroup_basic.hlsl"
 #include "nbl/builtin/hlsl/subgroup2/arithmetic_portability.hlsl"
+#include "nbl/builtin/hlsl/subgroup2/arithmetic_params.hlsl"
 
 #include "shaderCommon.hlsl"
 #include "nbl/builtin/hlsl/workgroup2/basic.hlsl"
 
-typedef vector<uint32_t, ITEMS_PER_INVOCATION> type_t;
+template<class Binop, class device_capabilities>
+using params_t = SUBGROUP_CONFIG_T;
+
+typedef vector<uint32_t, params_t<typename arithmetic::bit_and<uint32_t>::base_t, device_capabilities>::ItemsPerInvocation> type_t;
 
 uint32_t globalIndex()
 {
     return glsl::gl_WorkGroupID().x*WORKGROUP_SIZE+workgroup::SubgroupContiguousIndex();
 }
 
-template<class Binop, uint32_t N>
+template<class Binop>
 static void subtest(NBL_CONST_REF_ARG(type_t) sourceVal)
 {
-    using config_t = subgroup2::Configuration<SUBGROUP_SIZE_LOG2>;
-    using params_t = subgroup2::ArithmeticParams<config_t, typename Binop::base_t, N, device_capabilities>;
-
     const uint64_t outputBufAddr = pc.pOutputBuf[Binop::BindingIndex];
 
-    assert(glsl::gl_SubgroupSize() == 1u<<SUBGROUP_SIZE_LOG2)
+    assert(glsl::gl_SubgroupSize() == params_t<typename Binop::base_t, device_capabilities>::config_t::Size)
 
-    operation_t<params_t> func;
+    operation_t<params_t<typename Binop::base_t, device_capabilities> > func;
     type_t val = func(sourceVal);
 
     vk::RawBufferStore<type_t>(outputBufAddr + sizeof(type_t) * globalIndex(), val, sizeof(uint32_t));
@@ -37,13 +38,13 @@ type_t test()
     const uint32_t idx = globalIndex();
     type_t sourceVal = vk::RawBufferLoad<type_t>(pc.pInputBuf + idx * sizeof(type_t));
 
-    subtest<arithmetic::bit_and<uint32_t>, ITEMS_PER_INVOCATION>(sourceVal);
-    subtest<arithmetic::bit_xor<uint32_t>, ITEMS_PER_INVOCATION>(sourceVal);
-    subtest<arithmetic::bit_or<uint32_t>, ITEMS_PER_INVOCATION>(sourceVal);
-    subtest<arithmetic::plus<uint32_t>, ITEMS_PER_INVOCATION>(sourceVal);
-    subtest<arithmetic::multiplies<uint32_t>, ITEMS_PER_INVOCATION>(sourceVal);
-    subtest<arithmetic::minimum<uint32_t>, ITEMS_PER_INVOCATION>(sourceVal);
-    subtest<arithmetic::maximum<uint32_t>, ITEMS_PER_INVOCATION>(sourceVal);
+    subtest<arithmetic::bit_and<uint32_t> >(sourceVal);
+    subtest<arithmetic::bit_xor<uint32_t> >(sourceVal);
+    subtest<arithmetic::bit_or<uint32_t> >(sourceVal);
+    subtest<arithmetic::plus<uint32_t> >(sourceVal);
+    subtest<arithmetic::multiplies<uint32_t> >(sourceVal);
+    subtest<arithmetic::minimum<uint32_t> >(sourceVal);
+    subtest<arithmetic::maximum<uint32_t> >(sourceVal);
     return sourceVal;
 }
 

23_Arithmetic2UnitTest/main.cpp

@@ -2,6 +2,7 @@
 #include "nbl/application_templates/MonoAssetManagerAndBuiltinResourceApplication.hpp"
 #include "app_resources/common.hlsl"
 #include "nbl/builtin/hlsl/workgroup2/arithmetic_config.hlsl"
+#include "nbl/builtin/hlsl/subgroup2/arithmetic_params.hlsl"
 
 using namespace nbl;
 using namespace core;
@@ -186,7 +187,7 @@ class Workgroup2ScanTestApp final : public application_templates::BasicMultiQueu
 			for (auto subgroupSize = MinSubgroupSize; subgroupSize <= MaxSubgroupSize; subgroupSize *= 2u)
 			{
 				const uint8_t subgroupSizeLog2 = hlsl::findMSB(subgroupSize);
-				for (uint32_t workgroupSize = 64; workgroupSize <= MaxWorkgroupSize; workgroupSize *= 2u)
+				for (uint32_t workgroupSize = subgroupSize; workgroupSize <= MaxWorkgroupSize; workgroupSize *= 2u)
 				{
 					// make sure renderdoc captures everything for debugging
 					m_api->startCapture();
@@ -198,12 +199,12 @@ class Workgroup2ScanTestApp final : public application_templates::BasicMultiQueu
 						uint32_t itemsPerWG = workgroupSize * itemsPerInvocation;
 						m_logger->log("Testing Items per Invocation %u", ILogger::ELL_INFO, itemsPerInvocation);
 						bool passed = true;
-						//passed = runTest<emulatedReduction, false>(subgroupTestSource, elementCount, subgroupSizeLog2, workgroupSize, bool(useNative), itemsPerWG, itemsPerInvocation) && passed;
-						//logTestOutcome(passed, itemsPerWG);
-						//passed = runTest<emulatedScanInclusive, false>(subgroupTestSource, elementCount, subgroupSizeLog2, workgroupSize, bool(useNative), itemsPerWG, itemsPerInvocation) && passed;
-						//logTestOutcome(passed, itemsPerWG);
-						//passed = runTest<emulatedScanExclusive, false>(subgroupTestSource, elementCount, subgroupSizeLog2, workgroupSize, bool(useNative), itemsPerWG, itemsPerInvocation) && passed;
-						//logTestOutcome(passed, itemsPerWG);
+						passed = runTest<emulatedReduction, false>(subgroupTestSource, elementCount, subgroupSizeLog2, workgroupSize, bool(useNative), itemsPerWG, itemsPerInvocation) && passed;
+						logTestOutcome(passed, itemsPerWG);
+						passed = runTest<emulatedScanInclusive, false>(subgroupTestSource, elementCount, subgroupSizeLog2, workgroupSize, bool(useNative), itemsPerWG, itemsPerInvocation) && passed;
+						logTestOutcome(passed, itemsPerWG);
+						passed = runTest<emulatedScanExclusive, false>(subgroupTestSource, elementCount, subgroupSizeLog2, workgroupSize, bool(useNative), itemsPerWG, itemsPerInvocation) && passed;
+						logTestOutcome(passed, itemsPerWG);
 
 						hlsl::workgroup2::SArithmeticConfiguration wgConfig;
 					    wgConfig.init(hlsl::findMSB(workgroupSize), subgroupSizeLog2, itemsPerInvocation);
@@ -331,24 +332,25 @@ class Workgroup2ScanTestApp final : public application_templates::BasicMultiQueu
 		}
 		else
 		{
-			const std::string definitions[4] = { 
+			hlsl::subgroup2::SArithmeticParams sgParams;
+			sgParams.init(subgroupSizeLog2, itemsPerInvoc);
+
+			const std::string definitions[3] = { 
 				"subgroup2::" + arith_name,
 				std::to_string(workgroupSize),
-				std::to_string(itemsPerInvoc),
-				std::to_string(subgroupSizeLog2)
+				sgParams.getParamTemplateStructString()
 			};
 
-			const IShaderCompiler::SMacroDefinition defines[5] = {
+			const IShaderCompiler::SMacroDefinition defines[4] = {
 				{ "OPERATION", definitions[0] },
 				{ "WORKGROUP_SIZE", definitions[1] },
-				{ "ITEMS_PER_INVOCATION", definitions[2] },
-				{ "SUBGROUP_SIZE_LOG2", definitions[3] },
+				{ "SUBGROUP_CONFIG_T", definitions[2] },
 				{ "TEST_NATIVE", "1" }
 			};
 			if (useNative)
-				options.preprocessorOptions.extraDefines = { defines, defines + 5 };
-			else
 				options.preprocessorOptions.extraDefines = { defines, defines + 4 };
+			else
+				options.preprocessorOptions.extraDefines = { defines, defines + 3 };
 
 			overriddenUnspecialized = compiler->compileToSPIRV((const char*)source->getContent()->getPointer(), options);
 		}