[Mlir-commits] [mlir] [mlir][gpu] Generate multiple rank-specializations for tensor map cre… (PR #74082)

[Adam Paszke]

https://github.com/apaszke updated https://github.com/llvm/llvm-project/pull/74082

>From 35b2bcd093a2f6194c30d19845632379f3e815a8 Mon Sep 17 00:00:00 2001
From: Adam Paszke <apaszke at google.com>
Date: Fri, 1 Dec 2023 14:22:09 +0000
Subject: [PATCH] [mlir][gpu] Generate multiple rank-specializations for tensor
 map creation

The previous code was technically incorrect in that the type indicated
that the memref only has 1 dimension, while the code below was happily
dereferencing the size array out of bounds. Now, if the compiler doesn't
get too smart about optimizations, this code *might even work*. But, if
the compiler realizes that the array has 1 element it might starrt doing
silly things. This generates a specialization per each supported rank,
making sure we don't do any UB.
---
 .../ExecutionEngine/CudaRuntimeWrappers.cpp   | 43 +++++++++++++++++--
 1 file changed, 40 insertions(+), 3 deletions(-)

diff --git a/mlir/lib/ExecutionEngine/CudaRuntimeWrappers.cpp b/mlir/lib/ExecutionEngine/CudaRuntimeWrappers.cpp
index b8ac9ab90a9f3b8..db9957e3ed86682 100644
--- a/mlir/lib/ExecutionEngine/CudaRuntimeWrappers.cpp
+++ b/mlir/lib/ExecutionEngine/CudaRuntimeWrappers.cpp
@@ -423,9 +423,24 @@ extern "C" MLIR_CUDA_WRAPPERS_EXPORT void mgpuTensorMapEncodeTiled(
               elementStrides[4], interleave, swizzle, l2Promotion, oobFill);
 }
 
+namespace {
+
+template <int rank>
+void mgpuGetMemRefDataAndShape(void *raw_descriptor, char **addr,
+                               uint64_t *globalDim) {
+  auto descriptor =
+      reinterpret_cast<StridedMemRefType<char, rank> *>(raw_descriptor);
+  *addr = descriptor->data;
+  for (int i = 0; i < rank; ++i) {
+    globalDim[i] = static_cast<uint64_t>(descriptor->sizes[rank - i - 1]);
+  }
+}
+
+} // namespace
+
 extern "C" MLIR_CUDA_WRAPPERS_EXPORT void *mgpuTensorMapEncodeTiledMemref(
     int64_t tensorRank,                       // Dimensionality of tensor
-    StridedMemRefType<char, 1> *descriptor,   // Starting address
+    void *ranked_descriptor,                  // Starting address
     const CUtensorMapDataType tensorDataType, // Stride size (in bytes)
     CUtensorMapInterleave interleave,         // Type of interleaved layout
     CUtensorMapSwizzle swizzle,               // Bank swizzling pattern
@@ -435,17 +450,39 @@ extern "C" MLIR_CUDA_WRAPPERS_EXPORT void *mgpuTensorMapEncodeTiledMemref(
 ) {
   CUtensorMap tensorMap;
 
-  auto *globalAddress = descriptor->data;
   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);
 
+  char *globalAddress = nullptr;
+  switch (tensorRank) {
+  case 1:
+    mgpuGetMemRefDataAndShape<1>(ranked_descriptor, &globalAddress, globalDim);
+    break;
+  case 2:
+    mgpuGetMemRefDataAndShape<2>(ranked_descriptor, &globalAddress, globalDim);
+    break;
+  case 3:
+    mgpuGetMemRefDataAndShape<3>(ranked_descriptor, &globalAddress, globalDim);
+    break;
+  case 4:
+    mgpuGetMemRefDataAndShape<4>(ranked_descriptor, &globalAddress, globalDim);
+    break;
+  case 5:
+    mgpuGetMemRefDataAndShape<5>(ranked_descriptor, &globalAddress, globalDim);
+    break;
+  default:
+    fprintf(
+        stderr,
+        "'mgpuTensorMapEncodeTiledMemref' failed with 'rank is too high'\n");
+    return NULL;
+  }
+
   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]);
   }
 
   globalStrides[0] = globalDim[0] * elementSizeInBytes[tensorDataType];