SMAUG
Simulating Machine Learning Applications on gem5-Aladdin
|
Tiling optimizer for SMV inner product kernel. More...
#include <smv_inner_product_tiling.h>
Tiling optimizer for SMV inner product kernel.
Definition at line 19 of file smv_inner_product_tiling.h.
Static Public Member Functions | |
static std::array< TiledTensor, 3 > | doTiling (SmvInnerProductOp *op) |
static TilingConfig | computeBasicTileShapes (SmvInnerProductOp *op) |
Determine the best basic tiling shape for this fc layer without bias. More... | |
Static Protected Member Functions | |
static std::array< TilingDims, 3 > | determineBestTilingDims (Tensor *inputs, Tensor *weights, Tensor *outputs, int maxTileSize) |
Determine the best tiling dimensions for running inner product on SMV. More... | |
Static Protected Member Functions inherited from smaug::smv::TilingOptimizerBase | |
static TilingDims | findBestTilingDims (const TensorShape &shape, int maxTileSize, const std::vector< int > &minShape) |
Find the best set of dimensions to tile a given tensor shape. More... | |
static void | enum2DTensorTilingConfigs (TensorShape shape, int maxTileSize, const std::vector< int > &minShape, const std::vector< int > &strides, std::vector< TensorShape > &configs) |
Enumerates all tiling configs for a two dimensional Tensor. More... | |
static void | enum4DTensorTilingConfigs (TensorShape shape, int maxTileSize, const std::vector< int > &minShape, const std::vector< int > &strides, std::vector< TensorShape > &configs) |
Enumerates all tiling configs for a four dimensional Tensor. More... | |
|
static |
Determine the best basic tiling shape for this fc layer without bias.
The algorithm first determines the dimensions along which the inputs, weights, and outputs will be tiled. Then based on those dimensions, we enumerate all possible basic tile shapes for inputs, weights, and outputs. A basic shape is the shape that all but potentially the last tile along a set of dimensions will use. This triplet of tile shapes defines a TilingConfig. The TilingConfig that maximizes the total combined size of input, weights, and output tiles is chosen as the best.
To limit the number of possibilities, we only enumerate each dimension in certain increments. For example, input channels are only enumerated in multiples of kNumMaccsPerPE, and output channels are only enumerated in multiples in kNumPEs.
This algorithm assumes that the maximum tile size for weights, inputs, and outputs are all the same and that they will reside in separate scratchpads (no sharing).
op | The SMV inner product operator. All tensors must have been created with createAllTensors() prior to calling this function. |
Definition at line 35 of file smv_inner_product_tiling.cpp.
|
staticprotected |
Determine the best tiling dimensions for running inner product on SMV.
This function imposes some additional constraints on the tiling dimensions, in that certain combinations of input/weight/output tiling dimensions are not allowed in the interest of tiling code complexity.
Definition at line 13 of file smv_inner_product_tiling.cpp.