[mlir][sparse][gpu] cleanup GPUDataTransferStrategy (#71615)

The flag seems to be doing practically the same thing for zero cost and
pinned dma. In addition, the register host is not truly the right zero
cost mechanism according to Thomas. So we are simplifying the setup for
now, until we have a better definition for what to implement and test.
    
#64316

Author: aartbik

mlir/include/mlir/Dialect/SparseTensor/Pipelines/Passes.h

@@ -52,21 +52,6 @@ struct SparseCompilerOptions
               mlir::SparseParallelizationStrategy::kAnyStorageAnyLoop,
               "any-storage-any-loop",
               "Enable sparse parallelization for any storage and loop."))};
-  PassOptions::Option<mlir::GPUDataTransferStrategy> gpuDataTransfer{
-      *this, "gpu-data-transfer-strategy",
-      ::llvm::cl::desc(
-          "Set the data transfer strategy between the host and the GPUs"),
-      ::llvm::cl::init(mlir::GPUDataTransferStrategy::kRegularDMA),
-      llvm::cl::values(
-          clEnumValN(mlir::GPUDataTransferStrategy::kRegularDMA, "regular-dma",
-                     "Default option: malloc on host without additional "
-                     "options or care and then use DMA to copy the data"),
-          clEnumValN(mlir::GPUDataTransferStrategy::kPinnedDMA, "pinned-dma",
-                     "Based on the default option, pin the host memory to "
-                     "accelerate the data transfer"),
-          clEnumValN(mlir::GPUDataTransferStrategy::kZeroCopy, "zero-copy",
-                     "Use zero-copy to perform the data transfer from the host "
-                     "to the GPU"))};
 
   PassOptions::Option<bool> enableIndexReduction{
       *this, "enable-index-reduction",
@@ -166,9 +151,8 @@ struct SparseCompilerOptions
 
   /// Projects out the options for `createSparsificationPass`.
   SparsificationOptions sparsificationOptions() const {
-    return SparsificationOptions(parallelization, gpuDataTransfer,
-                                 enableIndexReduction, enableGPULibgen,
-                                 enableRuntimeLibrary);
+    return SparsificationOptions(parallelization, enableIndexReduction,
+                                 enableGPULibgen, enableRuntimeLibrary);
   }
 
   /// Projects out the options for `createConvertVectorToLLVMPass`.

mlir/include/mlir/Dialect/SparseTensor/Transforms/Passes.h

@@ -47,10 +47,6 @@ enum class ReinterpretMapScope {
   kExceptGeneric, // reinterprets operation other than linalg.generic
 };
 
-/// Defines data movement strategy between host and device for GPU.
-// TODO : Zero copy is disabled due to correctness bugs (tracker #64316)
-enum class GPUDataTransferStrategy { kRegularDMA, kZeroCopy, kPinnedDMA };
-
 #define GEN_PASS_DECL
 #include "mlir/Dialect/SparseTensor/Transforms/Passes.h.inc"
 
@@ -78,18 +74,14 @@ std::unique_ptr<Pass> createPreSparsificationRewritePass();
 
 /// Options for the Sparsification pass.
 struct SparsificationOptions {
-  SparsificationOptions(SparseParallelizationStrategy p,
-                        GPUDataTransferStrategy t, bool idxReduc,
+  SparsificationOptions(SparseParallelizationStrategy p, bool idxReduc,
                         bool gpuLibgen, bool enableRT)
-      : parallelizationStrategy(p), gpuDataTransferStrategy(t),
-        enableIndexReduction(idxReduc), enableGPULibgen(gpuLibgen),
-        enableRuntimeLibrary(enableRT) {}
+      : parallelizationStrategy(p), enableIndexReduction(idxReduc),
+        enableGPULibgen(gpuLibgen), enableRuntimeLibrary(enableRT) {}
   SparsificationOptions()
-      : SparsificationOptions(SparseParallelizationStrategy::kNone,
-                              GPUDataTransferStrategy::kRegularDMA, false,
+      : SparsificationOptions(SparseParallelizationStrategy::kNone, false,
                               false, true) {}
   SparseParallelizationStrategy parallelizationStrategy;
-  GPUDataTransferStrategy gpuDataTransferStrategy;
   bool enableIndexReduction;
   bool enableGPULibgen;
   bool enableRuntimeLibrary;
@@ -201,8 +193,8 @@ std::unique_ptr<Pass> createSparseVectorizationPass(unsigned vectorLength,
 void populateSparseGPUCodegenPatterns(RewritePatternSet &patterns,
                                       unsigned numThreads);
 
-void populateSparseGPULibgenPatterns(RewritePatternSet &patterns, bool enableRT,
-                                     GPUDataTransferStrategy gpuDataTransfer);
+void populateSparseGPULibgenPatterns(RewritePatternSet &patterns,
+                                     bool enableRT);
 
 std::unique_ptr<Pass> createSparseGPUCodegenPass();
 std::unique_ptr<Pass> createSparseGPUCodegenPass(unsigned numThreads);

mlir/include/mlir/Dialect/SparseTensor/Transforms/Passes.td

@@ -134,19 +134,6 @@ def SparsificationPass : Pass<"sparsification", "ModuleOp"> {
              clEnumValN(mlir::SparseParallelizationStrategy::kAnyStorageAnyLoop,
                         "any-storage-any-loop",
                         "Enable sparse parallelization for any storage and loop."))}]>,
-    Option<"gpuDataTransfer", "gpu-data-transfer-strategy", "mlir::GPUDataTransferStrategy",
-            "mlir::GPUDataTransferStrategy::kRegularDMA",
-            "Set the data transfer strategy", [{llvm::cl::values(
-               clEnumValN(mlir::GPUDataTransferStrategy::kRegularDMA,
-                     "regular-dma",
-                     "Default option: malloc on host without additional "
-                     "options or care and then use DMA to copy the data"),
-          clEnumValN(mlir::GPUDataTransferStrategy::kPinnedDMA, "pinned-dma",
-                     "Based on the default option, pin the host memory to "
-                     "accelerate the data transfer"),
-          clEnumValN(mlir::GPUDataTransferStrategy::kZeroCopy, "zero-copy",
-                     "Use zero-copy to perform the data transfer from the host "
-                     "to the GPU"))}]>,
     Option<"enableGPULibgen", "enable-gpu-libgen", "bool",
            "false",
            "Enable GPU acceleration by means of direct library calls (like cuSPARSE)">,