OptimizeSharedMemory.cpp File Reference

#include "mlir/Dialect/NVGPU/Transforms/Passes.h"
#include "mlir/Dialect/Arith/IR/Arith.h"
#include "mlir/Dialect/GPU/IR/GPUDialect.h"
#include "mlir/Dialect/MemRef/IR/MemRef.h"
#include "mlir/Dialect/NVGPU/IR/NVGPUDialect.h"
#include "mlir/Dialect/NVGPU/Transforms/Transforms.h"
#include "mlir/Dialect/NVGPU/Transforms/Utils.h"
#include "mlir/Dialect/Vector/IR/VectorOps.h"
#include "mlir/Interfaces/SideEffectInterfaces.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/Support/MathExtras.h"
#include "mlir/Dialect/NVGPU/Transforms/Passes.h.inc"

Go to the source code of this file.

Namespaces
	mlir
	Include the generated interface declarations.
	mlir::nvgpu

Macros
#define	GEN_PASS_DEF_OPTIMIZESHAREDMEMORY

Functions
static Value	permuteVectorOffset (OpBuilder &b, Location loc, ArrayRef< Value > indices, MemRefType memrefTy, int64_t srcDim, int64_t tgtDim)
	Uses srcIndexValue to permute tgtIndexValue via `result = xor(floordiv(srcIdxVal,permuteEveryN), floordiv(tgtIdxVal,vectorSize))) More...
static void	transformIndices (OpBuilder &builder, Location loc, SmallVector< Value, 4 > &indices, MemRefType memrefTy, int64_t srcDim, int64_t tgtDim)
static LogicalResult	getShmReadAndWriteOps (Operation *parentOp, Value shmMemRef, SmallVector< Operation *, 16 > &readOps, SmallVector< Operation *, 16 > &writeOps)
	Return all operations within parentOp that read from or write to shmMemRef. More...

Variables
constexpr int64_t	kSharedMemoryLineSizeBytes = 128
	The size of a shared memory line according to NV documentation. More...
constexpr int64_t	kDefaultVectorSizeBits = 128
	We optimize for 128bit accesses, but this can be made an argument in the future. More...

Macro Definition Documentation

GEN_PASS_DEF_OPTIMIZESHAREDMEMORY

#define GEN_PASS_DEF_OPTIMIZESHAREDMEMORY

Definition at line 28 of file OptimizeSharedMemory.cpp.

Function Documentation

getShmReadAndWriteOps()

static LogicalResult getShmReadAndWriteOps ( Operation *  parentOp, Value  shmMemRef, SmallVector< Operation *, 16 > &  readOps, SmallVector< Operation *, 16 > &  writeOps  )

static

Return all operations within parentOp that read from or write to shmMemRef.

Definition at line 113 of file OptimizeSharedMemory.cpp.

References mlir::nvgpu::getIndices(), and mlir::Operation::walk().

Referenced by mlir::nvgpu::optimizeSharedMemoryReadsAndWrites().

permuteVectorOffset()

static Value permuteVectorOffset ( OpBuilder &  b, Location  loc, ArrayRef< Value >  indices, MemRefType  memrefTy, int64_t  srcDim, int64_t  tgtDim  )

static

Uses srcIndexValue to permute tgtIndexValue via `result = xor(floordiv(srcIdxVal,permuteEveryN), floordiv(tgtIdxVal,vectorSize)))

• tgtIdxVal % vectorSize This is done using an optimized sequence ofarith` operations.

Definition at line 47 of file OptimizeSharedMemory.cpp.

References kDefaultVectorSizeBits, and kSharedMemoryLineSizeBytes.

Referenced by transformIndices().

transformIndices()

static void transformIndices ( OpBuilder &  builder, Location  loc, SmallVector< Value, 4 > &  indices, MemRefType  memrefTy, int64_t  srcDim, int64_t  tgtDim  )

static

Definition at line 102 of file OptimizeSharedMemory.cpp.

References permuteVectorOffset().

Referenced by mlir::nvgpu::optimizeSharedMemoryReadsAndWrites().

Variable Documentation

kDefaultVectorSizeBits

constexpr int64_t kDefaultVectorSizeBits = 128

constexpr

We optimize for 128bit accesses, but this can be made an argument in the future.

Definition at line 40 of file OptimizeSharedMemory.cpp.

Referenced by mlir::nvgpu::optimizeSharedMemoryReadsAndWrites(), and permuteVectorOffset().

kSharedMemoryLineSizeBytes

constexpr int64_t kSharedMemoryLineSizeBytes = 128

constexpr

The size of a shared memory line according to NV documentation.

Definition at line 37 of file OptimizeSharedMemory.cpp.

Referenced by mlir::nvgpu::optimizeSharedMemoryReadsAndWrites(), and permuteVectorOffset().