#include <sdfg/targets/rocm/rocm_arch.h>
|
| | RocmMmaSupport (bool base_support, bool f32_support, uint16_t threads, uint16_t int8_block_k=0) |
| MmaBlockSize | int8_block_size () const |
| | The block size Int8 matmuls use by default, at the full Int8 rate.
|
| bool | supported_types (types::PrimitiveType input_type, types::PrimitiveType output_type) const override |
| types::PrimitiveType | get_accumulator_type (types::PrimitiveType input_type, types::PrimitiveType output_type, types::PrimitiveType desired_acc_type) const override |
| bool | is_valid_block_size (const MmaBlockSize &block_size, types::PrimitiveType input_type, types::PrimitiveType acc_type) const override |
| std::optional< GpuMmaTiling > | try_get_mma_tiling (const MmaBlockSize &block_size, const BlockCounts *wave_tile_hint, const symbolic::MultiExpression &res_shape, types::PrimitiveType acc_type, bool ceil_counts=false) const |
| std::optional< GpuMmaTiling > | get_mma_tiling (const symbolic::MultiExpression &res_shape, types::PrimitiveType input_type, types::PrimitiveType acc_type, const MmaBlockSize *block_size_hint, const BlockCounts *wave_tile_hint) const override |
| std::optional< GpuMmaTiling > | get_ragged_mma_tiling (const symbolic::MultiExpression &padded_shape, types::PrimitiveType input_type, types::PrimitiveType output_type, types::PrimitiveType acc_type, const MmaBlockSize *block_size_hint, const BlockCounts *wave_tile_hint) const override |
| | Tiling for a ragged problem: padded_shape is a constant upper bound of each extent that need not be a multiple of the MMA block.
|
| void | set_mma_fragment_storage_type (types::StorageType &storage_type, const GpuMmaTiling &size, MmaFragmentType type, MmaFragmentLayout layout) const override |
| data_flow::ImplementationType | get_mma_impl_type (const GpuMmaTiling &tiling) const override |
| void | emit_block_frag_type (std::ostream &os, const types::StorageType &storage_type, types::PrimitiveType element_type) const |
| std::vector< MmaBlockSize > | get_supported_block_sizes (types::PrimitiveType input_type, types::PrimitiveType acc_type) const override |
| virtual | ~GpuMmaSupport ()=default |
| types::PrimitiveType | get_accumulator_type (types::PrimitiveType input_type, types::PrimitiveType output_type) const |
| | GpuMmaSupport ()=default |
|
| static bool | is_mma_type (const types::StorageType &storage) |
| static RocmMmaVariant | select_variant (const GpuMmaTiling &tiling) |
| static void | emit_bare_fragment_decl_type (std::ostream &os, MmaFragmentType type, const MmaBlockSize &dims, MmaFragmentLayout layout, types::PrimitiveType scalar_type, std::optional< std::pair< int, int > > coop_dims) |
| static void | emit_block_frag_type (std::ostream &os, MmaFragmentType type, const MmaBlockSize &dims, MmaFragmentLayout layout, types::PrimitiveType scalar_type, std::optional< std::pair< int, int > > coop_dims) |
| static int | get_integer_block_count (const symbolic::Expression &size, uint16_t block_size) |
| static int | get_ceil_block_count (const symbolic::Expression &size, uint16_t block_size) |
| | ceil(size / block_size) for a positive integer size, otherwise 0.
|
|
| const bool | base_support |
| const bool | f32_support |
| const uint16_t | threads_per_mma_block |
| const uint16_t | int8_block_k |
| | K of the native Int8 x Int8 -> Int32 instruction of 16x16 blocks (16 on CDNA2 and RDNA, 32 on CDNA3, where rocWMMA pads 16x16x16 to 16x16x32 at half the throughput); 0 without Int8 matrix support.
|
◆ RocmMmaSupport()
| sdfg::gpu::rocm::RocmMmaSupport::RocmMmaSupport |
( |
bool | base_support, |
|
|
bool | f32_support, |
|
|
uint16_t | threads, |
|
|
uint16_t | int8_block_k = 0 ) |
|
inline |
◆ emit_bare_fragment_decl_type()
◆ emit_block_frag_type() [1/2]
◆ emit_block_frag_type() [2/2]
◆ get_accumulator_type()
◆ get_mma_impl_type()
◆ get_mma_tiling()
◆ get_ragged_mma_tiling()
Tiling for a ragged problem: padded_shape is a constant upper bound of each extent that need not be a multiple of the MMA block.
Fragments crossing the exact extent are clipped (zero-filled loads, masked stores; see GpuMmaFromMemoryLayout::rows/cols). Targets without masked fragment transfers return nullopt.
Reimplemented from sdfg::gpu::GpuMmaSupport.
◆ get_supported_block_sizes()
◆ int8_block_size()
| MmaBlockSize sdfg::gpu::rocm::RocmMmaSupport::int8_block_size |
( |
| ) |
const |
|
inline |
The block size Int8 matmuls use by default, at the full Int8 rate.
◆ is_mma_type()
◆ is_valid_block_size()
◆ select_variant()
◆ set_mma_fragment_storage_type()
◆ supported_types()
◆ try_get_mma_tiling()
◆ base_support
| const bool sdfg::gpu::rocm::RocmMmaSupport::base_support |
◆ CDNA_BLOCK_SIZE
| MmaBlockSize sdfg::gpu::rocm::RocmMmaSupport::CDNA_BLOCK_SIZE = {32, 32, 8} |
|
staticconstexprprotected |
◆ DEFAULT_BLOCK_SIZE
| MmaBlockSize sdfg::gpu::rocm::RocmMmaSupport::DEFAULT_BLOCK_SIZE = {16, 16, 16} |
|
staticconstexprprotected |
◆ f32_support
| const bool sdfg::gpu::rocm::RocmMmaSupport::f32_support |
◆ int8_block_k
| const uint16_t sdfg::gpu::rocm::RocmMmaSupport::int8_block_k |
K of the native Int8 x Int8 -> Int32 instruction of 16x16 blocks (16 on CDNA2 and RDNA, 32 on CDNA3, where rocWMMA pads 16x16x16 to 16x16x32 at half the throughput); 0 without Int8 matrix support.
◆ MMA_FRAGMENT_STORAGE_TYPE
| const char* sdfg::gpu::rocm::RocmMmaSupport::MMA_FRAGMENT_STORAGE_TYPE = "ROCM_MMA" |
|
staticconstexpr |
◆ MMA_REGS_STORAGE_TYPE
| const char* sdfg::gpu::rocm::RocmMmaSupport::MMA_REGS_STORAGE_TYPE = "ROCM_MMA_REGS" |
|
staticconstexpr |
◆ threads_per_mma_block
| const uint16_t sdfg::gpu::rocm::RocmMmaSupport::threads_per_mma_block |
The documentation for this struct was generated from the following file:
1.15.0