sdfg::gpu::GpuMmaTiling Struct Reference

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

Public Member Functions
int	get_mma_blocks_m_per_wavefront () const
int	get_mma_blocks_n_per_wavefront () const

Static Public Member Functions
static bool	wave_tile_divides (const BlockCounts &wave_tiling, const BlockCounts &macro_block)
	Check that we can use the given wave tiling for the given macro block tiling, i.e.

Public Attributes
MmaBlockSize	mma_block_size
	The MMA block size (m, n, k) in elements.
int	threads_per_mma_wavefront = 0
	How many threads will work together on 1 MMA block.
BlockCounts	wave_tiling
	How many MMA blocks will be grouped together in a single wavefront / warp.
BlockCounts	macro_tiling
	How many MMA blocks will be grouped together in a macro block / on a SM or CU This controls how much data one block needs to handle and will typically need to be stored in block-shared memory.
types::PrimitiveType	acc_type
	The type used by the MMA hardware to accumulate the results of the MMA operation.
int	impl_details = 0
	implementation-specific details, e.g. for ROCm MFMA variant selection

Member Function Documentation

get_mma_blocks_m_per_wavefront()

int sdfg::gpu::GpuMmaTiling::get_mma_blocks_m_per_wavefront

(

)

const

get_mma_blocks_n_per_wavefront()

int sdfg::gpu::GpuMmaTiling::get_mma_blocks_n_per_wavefront

(

)

const

wave_tile_divides()

bool sdfg::gpu::GpuMmaTiling::wave_tile_divides ( const BlockCounts & wave_tiling, const BlockCounts & macro_block )

static

Check that we can use the given wave tiling for the given macro block tiling, i.e.

that the wave tiling divides the macro block tiling and that the number of threads per wave does not exceed 1024.

Parameters

wave_tiling
macro_block

Returns

Member Data Documentation

acc_type

types::PrimitiveType sdfg::gpu::GpuMmaTiling::acc_type

The type used by the MMA hardware to accumulate the results of the MMA operation.

This may be different from the input and output types, e.g. for FP16 inputs and FP32 outputs. For a MatmulNode this would not be externally visible other in the precision of the results

impl_details

int sdfg::gpu::GpuMmaTiling::impl_details = 0

implementation-specific details, e.g. for ROCm MFMA variant selection

macro_tiling

BlockCounts sdfg::gpu::GpuMmaTiling::macro_tiling

How many MMA blocks will be grouped together in a macro block / on a SM or CU This controls how much data one block needs to handle and will typically need to be stored in block-shared memory.

mma_block_size

MmaBlockSize sdfg::gpu::GpuMmaTiling::mma_block_size

The MMA block size (m, n, k) in elements.

While every underlying storage will only use 2 dimensions of this, this qualifies all the possible operations on it

threads_per_mma_wavefront

int sdfg::gpu::GpuMmaTiling::threads_per_mma_wavefront = 0

How many threads will work together on 1 MMA block.

wave_tiling

BlockCounts sdfg::gpu::GpuMmaTiling::wave_tiling

How many MMA blocks will be grouped together in a single wavefront / warp.

MMA BlockSize * wave_tiling != macro_tiling

---

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

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