[MLIR] Fix TMA descriptors

[guraypp]

Summary

Fix unused inputs when building TMA descriptor

[nicolasvasilache]

Can the commit message be improved ?
Are you fixing a runtime out of bounds error ?

mlir/lib/ExecutionEngine/CudaRuntimeWrappers.cpp

352

why is globalStrides not also updated ?

mlir/lib/ExecutionEngine/CudaRuntimeWrappers.cpp

extern "C" MLIR_CUDA_WRAPPERS_EXPORT void *mgpuTensorMapEncodeTiledMemref(
     CUtensorMapSwizzle swizzle,               // Bank swizzling pattern
     CUtensorMapL2promotion l2Promotion,       // L2 promotion size
     CUtensorMapFloatOOBfill oobFill,          // Padding zfill or NaN fill
     int64_t *inputBoxDims // Tensor size (number of elements)
 ) {
   CUtensorMap tensorMap;
 
   auto *globalAddress = descriptor->data;
-  uint32_t boxDim[5] = {0}, elementStrides[5] = {0};
-  uint64_t globalDim[5] = {0}, globalStrides[5] = {0};
+  uint32_t boxDim[5] = {1, 1, 1, 1, 1}, elementStrides[5] = {1, 1, 1, 1, 1};
+  uint64_t globalDim[5] = {1, 1, 1, 1, 1}, globalStrides[5] = {0};
   uint32_t tensorRank32 = uint32_t(tensorRank);
 
   static const int elementSizeInBytes[] = {1, 2, 4, 4, 8, 8, 2,
                                            4, 8, 2, 4, 4, 4};
   for (int64_t r = 0; r < tensorRank; ++r) {
     elementStrides[r] = uint32_t(1);
     boxDim[r] = static_cast<uint32_t>(inputBoxDims[tensorRank - r - 1]);
     globalDim[r] = static_cast<uint64_t>(descriptor->sizes[tensorRank - r - 1]);