sdfg::gpu::GpuMmaEinsumReplacer::EinsumMmaAnalysis Struct Reference

#include <sdfg/targets/gpu/gpu_mma_einsum_transform.h>

Inheritance diagram for sdfg::gpu::GpuMmaEinsumReplacer::EinsumMmaAnalysis:

Public Attributes
types::PrimitiveType	acc_type
bool	ragged = false
	Some extent is not a constant multiple of the MMA block: the tiling covers the padded extents and the fragments crossing the exact extents are clipped.
Public Attributes inherited from sdfg::einsum::Einsum2MatMul::MatMulAnalysis
symbolic::Symbol	indvar_outer_1 = SymEngine::null
symbolic::Symbol	indvar_outer_2 = SymEngine::null
symbolic::Symbol	indvar_inner = SymEngine::null
symbolic::Expression	m = SymEngine::null
symbolic::Expression	n = SymEngine::null
symbolic::Expression	k = SymEngine::null
symbolic::Expression	m_init = SymEngine::null
symbolic::Expression	n_init = SymEngine::null
symbolic::Expression	k_init = SymEngine::null
int	a_idx = -1
int	b_idx = -1
std::optional< math::tensor::TensorLayout >	layout_a
std::optional< math::tensor::TensorLayout >	layout_b
std::optional< math::tensor::TensorLayout >	layout_y
types::PrimitiveType	input_type = types::PrimitiveType::Void
	whether the indices in our contributors array are swapped to how we canonically name them this is not "transposed" as that will simply depend on the strides associated with the respective indices.
types::PrimitiveType	output_type = types::PrimitiveType::Void

Member Data Documentation

acc_type

types::PrimitiveType sdfg::gpu::GpuMmaEinsumReplacer::EinsumMmaAnalysis::acc_type

ragged

bool sdfg::gpu::GpuMmaEinsumReplacer::EinsumMmaAnalysis::ragged = false

Some extent is not a constant multiple of the MMA block: the tiling covers the padded extents and the fragments crossing the exact extents are clipped.

---

The documentation for this struct was generated from the following file:

• opt/include/sdfg/targets/gpu/gpu_mma_einsum_transform.h