set maxDynSharedBytes for CPUs and some cleaning

alpaka-group · Jun 19, 2024 · b7f84a5 · b7f84a5
1 parent 1386694
commit b7f84a5
Show file tree

Hide file tree

Showing 6 changed files with 25 additions and 6 deletions.
diff --git a/include/alpaka/kernel/TaskKernelCpuOmp2Blocks.hpp b/include/alpaka/kernel/TaskKernelCpuOmp2Blocks.hpp
@@ -975,6 +975,8 @@ namespace alpaka
                 // properties function.
                 auto const& props = alpaka::getAccDevProps<AccCpuOmp2Blocks<TDim, TIdx>>(dev);
                 kernelFunctionAttributes.maxThreadsPerBlock = static_cast<int>(props.m_blockThreadCountMax);
+                kernelFunctionAttributes.maxDynamicSharedSizeBytes
+                    = static_cast<int>(alpaka::BlockSharedDynMemberAllocKiB * 1024);
                 return kernelFunctionAttributes;
             }
         };

diff --git a/include/alpaka/kernel/TaskKernelCpuOmp2Threads.hpp b/include/alpaka/kernel/TaskKernelCpuOmp2Threads.hpp
@@ -221,6 +221,8 @@ namespace alpaka
                 // properties function.
                 auto const& props = alpaka::getAccDevProps<AccCpuOmp2Threads<TDim, TIdx>>(dev);
                 kernelFunctionAttributes.maxThreadsPerBlock = static_cast<int>(props.m_blockThreadCountMax);
+                kernelFunctionAttributes.maxDynamicSharedSizeBytes
+                    = static_cast<int>(alpaka::BlockSharedDynMemberAllocKiB * 1024);
                 return kernelFunctionAttributes;
             }
         };

diff --git a/include/alpaka/kernel/TaskKernelCpuSerial.hpp b/include/alpaka/kernel/TaskKernelCpuSerial.hpp
@@ -166,6 +166,8 @@ namespace alpaka
                 // properties function.
                 auto const& props = alpaka::getAccDevProps<AccCpuSerial<TDim, TIdx>>(dev);
                 kernelFunctionAttributes.maxThreadsPerBlock = static_cast<int>(props.m_blockThreadCountMax);
+                kernelFunctionAttributes.maxDynamicSharedSizeBytes
+                    = static_cast<int>(alpaka::BlockSharedDynMemberAllocKiB * 1024);
                 return kernelFunctionAttributes;
             }
         };

diff --git a/include/alpaka/kernel/TaskKernelCpuTbbBlocks.hpp b/include/alpaka/kernel/TaskKernelCpuTbbBlocks.hpp
@@ -178,6 +178,8 @@ namespace alpaka
                 // properties function.
                 auto const& props = alpaka::getAccDevProps<AccCpuTbbBlocks<TDim, TIdx>>(dev);
                 kernelFunctionAttributes.maxThreadsPerBlock = static_cast<int>(props.m_blockThreadCountMax);
+                kernelFunctionAttributes.maxDynamicSharedSizeBytes
+                    = static_cast<int>(alpaka::BlockSharedDynMemberAllocKiB * 1024);
                 return kernelFunctionAttributes;
             }
         };

diff --git a/include/alpaka/kernel/TaskKernelCpuThreads.hpp b/include/alpaka/kernel/TaskKernelCpuThreads.hpp
@@ -229,6 +229,8 @@ namespace alpaka
                 // properties function.
                 auto const& props = alpaka::getAccDevProps<AccCpuThreads<TDim, TIdx>>(dev);
                 kernelFunctionAttributes.maxThreadsPerBlock = static_cast<int>(props.m_blockThreadCountMax);
+                kernelFunctionAttributes.maxDynamicSharedSizeBytes
+                    = static_cast<int>(alpaka::BlockSharedDynMemberAllocKiB * 1024);
                 return kernelFunctionAttributes;
             }
         };

diff --git a/test/unit/workDiv/src/WorkDivForKernelTest.cpp b/test/unit/workDiv/src/WorkDivForKernelTest.cpp
@@ -86,10 +86,10 @@ TEMPLATE_LIST_TEST_CASE("getValidWorkDivForKernel.1D", "[workDivKernel]", TestAc
     auto const dev = alpaka::getDevByIdx(platform, 0);
 
     TestKernelWithManyRegisters kernel;
-    auto const& bundeledKernel = alpaka::KernelBundle(kernel, 200ul);
+    auto const bundeledKernel = alpaka::KernelBundle(kernel, 200ul);
 
     // Get hard limits for test
-    auto const& props = alpaka::getAccDevProps<Acc, decltype(dev)>(dev);
+    auto const props = alpaka::getAccDevProps<Acc, decltype(dev)>(dev);
     Idx const threadsPerGridTestValue = props.m_blockThreadCountMax * props.m_gridBlockCountMax;
 
     // Test getValidWorkDivForKernel for threadsPerGridTestValue threads per grid
@@ -136,10 +136,14 @@ TEMPLATE_LIST_TEST_CASE("getValidWorkDivForKernel.1D", "[workDivKernel]", TestAc
         // CPU must have only 1 thread per block. In other words, number of blocks is equal to number of threads.
         CHECK(workDiv == WorkDiv{Vec{threadsPerGridTestValue}, Vec{1}, Vec{1}});
         // Test a new 1D workdiv. Threads per block can not be larger than 1 for CPU. Hence 2 is not valid.
-        auto const& workDiv1DUsingInitList = WorkDiv{Vec{threadsPerGridTestValue / 2}, Vec{2}, Vec{1}};
+        auto const workDiv1DUsingInitList = WorkDiv{Vec{threadsPerGridTestValue / 2}, Vec{2}, Vec{1}};
         auto const isWorkDivValidForCPU
             = alpaka::isValidWorkDivKernel<Acc>(dev, bundeledKernel, workDiv1DUsingInitList);
         CHECK(isWorkDivValidForCPU == false);
+        // Check maxDynamicSharedSizeBytes for CPU backends
+        auto const funcAttributes = alpaka::getFunctionAttributes<Acc>(dev, bundeledKernel);
+        CHECK(
+            funcAttributes.maxDynamicSharedSizeBytes == static_cast<int>(alpaka::BlockSharedDynMemberAllocKiB * 1024));
     }
     else
     {
@@ -162,10 +166,10 @@ TEMPLATE_LIST_TEST_CASE("getValidWorkDivForKernel.2D", "[workDivKernel]", TestAc
     TestKernelWithManyRegisters kernel;
     // A random value
     size_t val(200ul);
-    auto const& bundeledKernel = alpaka::KernelBundle(kernel, val);
+    auto const bundeledKernel = alpaka::KernelBundle(kernel, val);
 
     // Get hard limits for test
-    auto const& props = alpaka::getAccDevProps<Acc>(dev);
+    auto const props = alpaka::getAccDevProps<Acc>(dev);
     Idx const threadsPerGridTestValue = props.m_blockThreadCountMax * props.m_gridBlockCountMax;
 
     // Test getValidWorkDivForKernel function for threadsPerGridTestValue threads per grid.
@@ -241,9 +245,14 @@ TEMPLATE_LIST_TEST_CASE("getValidWorkDivForKernel.2D", "[workDivKernel]", TestAc
         // CPU must have only 1 thread per block. In other words, number of blocks is equal to number of threads.
         CHECK(workDiv == WorkDiv{Vec{8, threadsPerGridTestValue / 8}, Vec{1, 1}, Vec{1, 1}});
         // Test a new 2D workdiv. Threads per block can not be larger than 1 for CPU. Hence 2x1 threads is not valid.
-        auto const& invalidWorkDiv2D = WorkDiv{Vec{1, 2048}, Vec{1, 2}, Vec{1, 1}};
+        auto const invalidWorkDiv2D = WorkDiv{Vec{1, 2048}, Vec{1, 2}, Vec{1, 1}};
         auto const isWorkDivValidForCpu = alpaka::isValidWorkDivKernel<Acc>(dev, bundeledKernel, invalidWorkDiv2D);
         CHECK(isWorkDivValidForCpu == false);
+
+        // Check maxDynamicSharedSizeBytes for CPU backends
+        CHECK(
+            alpaka::getFunctionAttributes<Acc>(dev, bundeledKernel).maxDynamicSharedSizeBytes
+            == static_cast<int>(alpaka::BlockSharedDynMemberAllocKiB * 1024));
     }
     else
     {