remove KernelBundle, change signature of [get|is]ValidWorkDiv* (#…

…2349)

* remove `KernelBundle`, change signature of `getValidWorkDiv*`

- revert introduced `KernelBundle` in #2251
- change signature of `getValidWorkDivForKernel`,`isValidWorkDivKernel` and `isValidWorkDivKernel`
- reuse old naming `getValidWorkDiv` and `isValidWorkDiv`

* use new interface for `getValidWorkDiv`

* fix cheat sheet

docs/source/basic/cheatsheet.rst

@@ -180,21 +180,28 @@ Prepare Kernel Bundle
   .. code-block:: c++
 
      HeatEquationKernel heatEqKernel;
-     // Arguments of KernelBundle: The kernel instance and the kernel arguments
-     auto const& bundeledKernel = alpaka::KernelBundle(heatEqKernel, pCurrAcc, pNextAcc, numNodesX, dx, dt);
 
 Automatically select a valid kernel launch configuration
   .. code-block:: c++
 
      Vec<Dim, Idx> const globalThreadExtent = vectorValue;
      Vec<Dim, Idx> const elementsPerThread = vectorValue;
 
-     auto autoWorkDiv = getValidWorkDivForKernel<Acc>(
-       device,
-       bundeledKernel,
-       globalThreadExtent, elementsPerThread,
+     KernelCfg<Acc> const kernelCfg = {
+       globalThreadExtent,
+       elementsPerThread,
        false,
-       GridBlockExtentSubDivRestrictions::Unrestricted);
+       GridBlockExtentSubDivRestrictions::Unrestricted};
+
+     auto autoWorkDiv = getValidWorkDiv(
+       kernelCfg,
+       device,
+       heatEqKernel,
+       pCurrAcc,
+       pNextAcc,
+       numNodesX,
+       dx,
+       dt);
 
 Manually set a kernel launch configuration
   .. code-block:: c++
@@ -204,9 +211,10 @@ Manually set a kernel launch configuration
      Vec<Dim, Idx> const elementsPerThread = vectorValue;
 
      using WorkDiv = WorkDivMembers<Dim, Idx>;
-     auto manualWorkDiv = WorkDiv{blocksPerGrid,
-                                  threadsPerBlock,
-				  elementsPerThread};
+     auto manualWorkDiv = WorkDiv{
+       blocksPerGrid,
+       threadsPerBlock,
+       elementsPerThread};
 
 Instantiate a kernel and create a task that will run it (does not launch it yet)
   .. code-block:: c++

example/bufferCopy/src/bufferCopy.cpp

@@ -118,7 +118,7 @@ auto example(TAccTag const&) -> int
     using Data = std::uint32_t;
     constexpr Idx nElementsPerDim = 2;
 
-    const Vec extents(Vec::all(static_cast<Idx>(nElementsPerDim)));
+    Vec const extents(Vec::all(static_cast<Idx>(nElementsPerDim)));
 
     // Allocate host memory buffers
     //
@@ -164,9 +164,8 @@ auto example(TAccTag const&) -> int
 
     FillBufferKernel fillBufferKernel;
 
-    auto const& bundeledFillBufferKernel = alpaka::KernelBundle(fillBufferKernel, hostViewPlainPtrMdSpan);
-    auto const hostWorkDiv
-        = alpaka::getValidWorkDivForKernel<Host>(devHost, bundeledFillBufferKernel, threadsPerGrid, elementsPerThread);
+    alpaka::KernelCfg<Host> const hostKernelCfg = {threadsPerGrid, elementsPerThread};
+    auto const hostWorkDiv = alpaka::getValidWorkDiv(hostKernelCfg, devHost, fillBufferKernel, hostViewPlainPtrMdSpan);
 
     alpaka::exec<Host>(hostQueue, hostWorkDiv, fillBufferKernel,
                        hostViewPlainPtrMdSpan); // 1st kernel argument
@@ -203,11 +202,10 @@ auto example(TAccTag const&) -> int
     auto deviceBufferMdSpan2 = alpaka::experimental::getMdSpan(deviceBuffer2);
 
     TestBufferKernel testBufferKernel;
-    auto const& bundeledTestBufferKernel = alpaka::KernelBundle(testBufferKernel, deviceBufferMdSpan1);
 
     // Let alpaka calculate good block and grid sizes given our full problem extent
-    auto const devWorkDiv
-        = alpaka::getValidWorkDivForKernel<Acc>(devAcc, bundeledTestBufferKernel, threadsPerGrid, elementsPerThread);
+    alpaka::KernelCfg<Acc> const devKernelCfg = {threadsPerGrid, elementsPerThread};
+    auto const devWorkDiv = alpaka::getValidWorkDiv(devKernelCfg, devAcc, testBufferKernel, deviceBufferMdSpan1);
 
     alpaka::exec<Acc>(devQueue, devWorkDiv, testBufferKernel, deviceBufferMdSpan1);
     alpaka::exec<Acc>(devQueue, devWorkDiv, testBufferKernel, deviceBufferMdSpan2);

example/complex/src/complex.cpp

@@ -58,10 +58,10 @@ auto example(TAccTag const&) -> int
 
     ComplexKernel complexKernel;
 
-    auto const& bundeledKernel = alpaka::KernelBundle(complexKernel);
+    alpaka::KernelCfg<Acc> const kernelCfg = {threadsPerGrid, elementsPerThread};
+
     // Let alpaka calculate good block and grid sizes given our full problem extent
-    auto const workDiv
-        = alpaka::getValidWorkDivForKernel<Acc>(devAcc, bundeledKernel, threadsPerGrid, elementsPerThread);
+    auto const workDiv = alpaka::getValidWorkDiv(kernelCfg, devAcc, complexKernel);
 
     // Run the kernel
     alpaka::exec<Acc>(queue, workDiv, complexKernel);

example/conv2DWithMdspan/src/conv2DWithMdspan.cpp

@@ -147,16 +147,16 @@ auto example(TAccTag const&) -> int
     //  Construct kernel object
     ConvolutionKernelMdspan2D convolutionKernel2D;
 
-    // Make a bundle
-    auto const& bundeledKernel = alpaka::KernelBundle(
+    //   Let alpaka calculate good block and grid sizes given our full problem extent.
+    alpaka::KernelCfg<DevAcc> const kernelCfg = {extent, Vec::ones()};
+    auto const workDiv = alpaka::getValidWorkDiv(
+        kernelCfg,
+        devAcc,
         convolutionKernel2D,
         alpaka::experimental::getMdSpan(bufInputAcc),
         alpaka::experimental::getMdSpan(outputDeviceMemory),
         alpaka::experimental::getMdSpan(bufFilterAcc));
 
-    //   Let alpaka calculate good block and grid sizes given our full problem extent.
-    auto const workDiv = alpaka::getValidWorkDivForKernel<DevAcc>(devAcc, bundeledKernel, extent, Vec::ones());
-
 
     // Run the kernel, pass 3 arrays as 2D mdspans
     alpaka::exec<DevAcc>(

example/convolution1D/src/convolution1D.cpp

@@ -37,8 +37,8 @@ struct ConvolutionKernel
         TElem const* const input,
         TElem const* const filter,
         TElem* const output,
-        const std::size_t inputSize,
-        const std::size_t filterSize) const -> void
+        std::size_t const inputSize,
+        std::size_t const filterSize) const -> void
     {
         auto const globalThreadIdxX = alpaka::getIdx<alpaka::Grid, alpaka::Threads>(acc)[0];
 
@@ -140,17 +140,19 @@ auto example(TAccTag const&) -> int
     DataType* nativeInputDeviceMemory = std::data(inputDeviceMemory);
     DataType* nativeOutputDeviceMemory = std::data(outputDeviceMemory);
 
-    auto const& bundeledKernel = alpaka::KernelBundle(
+    alpaka::KernelCfg<DevAcc> const kernelCfg = {threadsPerGrid, elementsPerThread};
+
+    // Let alpaka calculate good block and grid sizes given our full problem extent
+    auto const workDiv = alpaka::getValidWorkDiv(
+        kernelCfg,
+        devAcc,
         convolutionKernel,
         nativeInputDeviceMemory,
         nativeFilterDeviceMemory,
         nativeOutputDeviceMemory,
         inputSize,
         filterSize);
 
-    // Let alpaka calculate good block and grid sizes given our full problem extent
-    auto const workDiv
-        = alpaka::getValidWorkDivForKernel<DevAcc>(devAcc, bundeledKernel, threadsPerGrid, elementsPerThread);
     // Run the kernel
     alpaka::exec<DevAcc>(
         queue,

example/convolution2D/src/convolution2D.cpp

@@ -265,7 +265,7 @@ auto example(TAccTag const&) -> int
     alpaka::wait(queueAcc);
 
     // Calculate the allocated width, due to padding it might be larger then the matrix width
-    auto const intputWidthAllocated = [&]() -> const Idx
+    auto const intputWidthAllocated = [&]() -> Idx const
     {
         // Calculate pitch: The size of one line in bytes including padding.
         auto const rowPitchInput{alpaka::getPitchesInBytes(bufInputAcc)[0]};
@@ -294,7 +294,7 @@ auto example(TAccTag const&) -> int
     alpaka::wait(queueAcc);
 
     // Calculate the allocated width, due to padding it might be larger then the matrix width
-    auto const filterWidthAllocated = [&]() -> const Idx
+    auto const filterWidthAllocated = [&]() -> Idx const
     {
         // Calculate pitch: The size of one line in bytes including padding.
         auto const rowPitchFilter{alpaka::getPitchesInBytes(bufFilterAcc)[0]};
@@ -305,20 +305,22 @@ auto example(TAccTag const&) -> int
     //  ConvolutionKernel2DSharedMemory
     ConvolutionKernel2DSharedMemory convolutionKernel2D;
 
-    auto const& bundeledKernel = alpaka::KernelBundle(
+    alpaka::KernelCfg<DevAcc> kernelCfg = {extent, Vec::ones()};
+
+    //   Let alpaka calculate good block and grid sizes given our full problem extent.
+    auto const workDiv = alpaka::getValidWorkDiv(
+        kernelCfg,
+        devAcc,
         convolutionKernel2D,
-        alpaka::getPtrNative(bufInputAcc),
-        alpaka::getPtrNative(outputDeviceMemory),
+        std::data(bufInputAcc),
+        std::data(outputDeviceMemory),
         matrixWidth,
         matrixHeight,
-        alpaka::getPtrNative(bufFilterAcc),
+        std::data(bufFilterAcc),
         filterWidth,
         intputWidthAllocated,
         filterWidthAllocated);
 
-    //   Let alpaka calculate good block and grid sizes given our full problem extent.
-    auto const workDiv = alpaka::getValidWorkDivForKernel<DevAcc>(devAcc, bundeledKernel, extent, Vec::ones());
-
     // Run the kernel
     alpaka::exec<DevAcc>(
         queueAcc,

example/counterBasedRng/src/counterBasedRng.cpp

@@ -147,22 +147,31 @@ auto example(TAccTag const&) -> int
     BufAcc bufAcc(alpaka::allocBuf<Data, Idx>(devAcc, extent));
 
     CounterBasedRngKernel counterBasedRngKernel;
-    auto const& bundeledKernel
-        = alpaka::KernelBundle(counterBasedRngKernel, alpaka::experimental::getMdSpan(bufAcc), key);
-    auto const& bundeledKernel2
-        = alpaka::KernelBundle(counterBasedRngKernel, alpaka::experimental::getMdSpan(bufHost), key);
 
     // Let alpaka calculate good block and grid sizes given our full problem extent
-    auto const workDivAcc = alpaka::getValidWorkDivForKernel<Acc>(devAcc, bundeledKernel, extent, elementsPerThread);
-    auto const workDivHost
-        = alpaka::getValidWorkDivForKernel<AccHost>(devHost, bundeledKernel2, extent, elementsPerThreadHost);
+    alpaka::KernelCfg<Acc> kernerlCfgAccDev = {extent, elementsPerThread};
+    auto const workDivAcc = alpaka::getValidWorkDiv(
+        kernerlCfgAccDev,
+        devAcc,
+        counterBasedRngKernel,
+        alpaka::experimental::getMdSpan(bufAcc),
+        key);
 
     // Create the kernel execution task.
     auto const taskKernelAcc = alpaka::createTaskKernel<Acc>(
         workDivAcc,
         CounterBasedRngKernel(),
         alpaka::experimental::getMdSpan(bufAcc),
         key);
+
+    alpaka::KernelCfg<AccHost> kernerlCfgAccHost = {extent, elementsPerThreadHost};
+    auto const workDivHost = alpaka::getValidWorkDiv(
+        kernerlCfgAccHost,
+        devHost,
+        counterBasedRngKernel,
+        alpaka::experimental::getMdSpan(bufHost),
+        key);
+
     auto const taskKernelHost = alpaka::createTaskKernel<AccHost>(
         workDivHost,
         CounterBasedRngKernel(),

example/heatEquation/src/heatEquation.cpp

@@ -134,9 +134,11 @@ auto example(TAccTag const&) -> int
 
     HeatEquationKernel heatEqKernel;
 
-    auto const& bundeledKernel = alpaka::KernelBundle(heatEqKernel, pCurrAcc, pNextAcc, numNodesX, dx, dt);
+    alpaka::KernelCfg<Acc> const kernelCfg = {extent, elemPerThread};
+
     // Let alpaka calculate good block and grid sizes given our full problem extent
-    auto const workDiv = alpaka::getValidWorkDivForKernel<Acc>(devAcc, bundeledKernel, extent, elemPerThread);
+    auto const workDiv
+        = alpaka::getValidWorkDiv(kernelCfg, devAcc, heatEqKernel, pCurrAcc, pNextAcc, numNodesX, dx, dt);
 
     // Copy host -> device
     alpaka::memcpy(queue, uCurrBufAcc, uCurrBufHost);

example/helloWorld/src/helloWorld.cpp

@@ -135,10 +135,10 @@ auto example(TAccTag const&) -> int
     // argument. So a kernel can be a class or struct, a lambda, etc.
     HelloWorldKernel helloWorldKernel;
 
-    auto const& bundeledKernel = alpaka::KernelBundle(helloWorldKernel);
+    alpaka::KernelCfg<Acc> const kernelCfg = {threadsPerGrid, elementsPerThread};
+
     // Let alpaka calculate good block and grid sizes given our full problem extent
-    auto const workDiv
-        = alpaka::getValidWorkDivForKernel<Acc>(devAcc, bundeledKernel, threadsPerGrid, elementsPerThread);
+    auto const workDiv = alpaka::getValidWorkDiv(kernelCfg, devAcc, helloWorldKernel);
 
     // Run the kernel
     //

example/helloWorldLambda/src/helloWorldLambda.cpp

@@ -78,7 +78,7 @@ auto example(TAccTag const&) -> int
     auto const threadsPerGrid = Vec{4, 2, 4};
 
 
-    const size_t nExclamationMarks = 10;
+    size_t const nExclamationMarks = 10;
 
     // Run "Hello World" kernel with a lambda function
     //
@@ -117,10 +117,10 @@ auto example(TAccTag const&) -> int
         printf("\n");
     };
 
-    auto const& bundeledKernel = alpaka::KernelBundle(kernelLambda, nExclamationMarks);
+    alpaka::KernelCfg<Acc> const kernelCfg = {threadsPerGrid, elementsPerThread};
+
     // Let alpaka calculate good block and grid sizes given our full problem extent
-    auto const workDiv
-        = alpaka::getValidWorkDivForKernel<Acc>(devAcc, bundeledKernel, threadsPerGrid, elementsPerThread);
+    auto const workDiv = alpaka::getValidWorkDiv(kernelCfg, devAcc, kernelLambda, nExclamationMarks);
 
     alpaka::exec<Acc>(queue, workDiv, kernelLambda, nExclamationMarks);
     alpaka::wait(queue);

example/kernelSpecialization/src/kernelSpecialization.cpp

@@ -81,10 +81,10 @@ auto example(TAccTag const&) -> int
     std::size_t const elementsPerThread = 1u;
     Kernel kernel;
 
-    auto const& bundeledKernel = alpaka::KernelBundle(kernel);
+    alpaka::KernelCfg<Acc> const kernelCfg = {threadsPerGrid, elementsPerThread};
+
     // Let alpaka calculate good block and grid sizes given our full problem extent
-    auto const workDiv
-        = alpaka::getValidWorkDivForKernel<Acc>(devAcc, bundeledKernel, threadsPerGrid, elementsPerThread);
+    auto const workDiv = alpaka::getValidWorkDiv(kernelCfg, devAcc, kernel);
 
     // Run the kernel
     alpaka::exec<Acc>(queue, workDiv, kernel);

example/matrixMulWithMdspan/src/matrixMulMdSpan.cpp

@@ -147,19 +147,14 @@ auto example(TAccTag const&) -> int
     auto mdDevC = alpaka::experimental::getMdSpan(bufDevC);
 
     MatrixMulKernel kernel;
-    auto const& bundeledKernel = alpaka::KernelBundle(kernel, mdDevA, mdDevB, mdDevC);
 
     // Let alpaka calculate good block and grid sizes given our full problem extent
-    auto const workDiv = alpaka::getValidWorkDivForKernel<Acc>(
-        devAcc,
-        bundeledKernel,
-        extentC,
-        Vec::ones(),
-        false,
-        alpaka::GridBlockExtentSubDivRestrictions::Unrestricted);
+    alpaka::KernelCfg<Acc> const kernelCfg
+        = {extentC, Vec::ones(), false, alpaka::GridBlockExtentSubDivRestrictions::Unrestricted};
+    auto const workDiv = alpaka::getValidWorkDiv<Acc>(kernelCfg, devAcc, kernel, mdDevA, mdDevB, mdDevC);
 
     // Execute the kernel
-    alpaka::exec<Acc>(queue, workDiv, MatrixMulKernel{}, mdDevA, mdDevB, mdDevC);
+    alpaka::exec<Acc>(queue, workDiv, kernel, mdDevA, mdDevB, mdDevC);
 
     // Copy result back to host
     alpaka::memcpy(queue, bufHostC, bufDevC);

example/monteCarloIntegration/src/monteCarloIntegration.cpp

@@ -112,14 +112,11 @@ auto example(TAccTag const&) -> int
     bufHost[0] = 0.0f;
     alpaka::memcpy(queue, bufAcc, bufHost);
 
+    alpaka::KernelCfg<Acc> const kernelCfg = {Vec(numThreads), Vec(numAlpakaElementsPerThread)};
     Kernel kernel;
-    auto const& bundeledKernel = alpaka::KernelBundle(kernel, numPoints, ptrBufAcc, Function{});
+
     // Let alpaka calculate good block and grid sizes given our full problem extent
-    auto const workDiv = alpaka::getValidWorkDivForKernel<Acc>(
-        devAcc,
-        bundeledKernel,
-        Vec(numThreads),
-        Vec(numAlpakaElementsPerThread));
+    auto const workDiv = alpaka::getValidWorkDiv(kernelCfg, devAcc, kernel, numPoints, ptrBufAcc, Function{});
 
     alpaka::exec<Acc>(queue, workDiv, kernel, numPoints, ptrBufAcc, Function{});
     alpaka::memcpy(queue, bufHost, bufAcc);

example/openMPSchedule/src/openMPSchedule.cpp

@@ -108,10 +108,10 @@ auto main() -> int
     Idx const elementsPerThread = 1u;
 
     OpenMPScheduleDefaultKernel openMPScheduleDefaultKernel;
-    auto const& bundeledKernel = alpaka::KernelBundle(openMPScheduleDefaultKernel);
+
     // Let alpaka calculate good block and grid sizes given our full problem extent
-    auto const workDiv
-        = alpaka::getValidWorkDivForKernel<Acc>(devAcc, bundeledKernel, threadsPerGrid, elementsPerThread);
+    alpaka::KernelCfg<Acc> kernelCfg = {threadsPerGrid, elementsPerThread};
+    auto const workDiv = alpaka::getValidWorkDiv(kernelCfg, devAcc, openMPScheduleDefaultKernel);
 
     // Run the kernel setting no schedule explicitly.
     std::cout << "OpenMPScheduleDefaultKernel setting no schedule explicitly:\n";