XeTLA GEMM

This is a gemm API to compute matAcc = matA * matB. More...

Classes
class	gpu::xetla::group::gemm_t< compute_policy_int4_dequantize_xmx< compute_attr_, perf_tuning_knob_, dtype_scale_, dtype_zero_pt_, dequant_s_, gpu_arch::Xe >, tile_shape_, mem_desc_a_t_, mem_desc_b_t_, pre_processing_t_ >
	Is the gemm functor for Xe architecture and matrix engine. More...
struct	gpu::xetla::kernel::dispatch_policy_int4_dequantize_kslicing< group_swizzle_policy_, num_global_kslicing_, num_local_kslicing_ >
	4bit kslicing GEMM implementation. More...
class	gpu::xetla::kernel::gemm_universal_t< dispatch_policy_int4_dequantize_kslicing< group_swizzle_, num_global_kslicing_, num_local_kslicing_ >, gemm_t_, epilogue_t_ >
	Is the GEMM functor, specialized in bit4 matB kslicing dispatch policy and Xe architecture. More...
struct	gpu::xetla::group::pre_processing_default_t< tile_shape, arch_tag, enable >
	Gemm default pre_processing functor. More...
struct	gpu::xetla::group::pre_processing_matA_neg_filter_t< tile_shape, arch_tag, enable >
	Gemm pre_processing functor with applying relu op to matA. More...
class	gpu::xetla::group::gemm_t< compute_policy, tile_shape, mem_desc_a, mem_desc_b, pre_processing, enable >
	Gemm functor. More...
class	gpu::xetla::group::gemm_selector_t< dtype_a, dtype_b, mem_layout_a, mem_layout_b, mem_space_a, mem_space_b, alignment_a, alignment_b, dtype_acc, tile_shape, k_stride, engine, arch_tag, stages, sync_freq, enable >
	Gemm selection functor. More...
struct	gpu::xetla::group::compute_policy_default_xmx< compute_attr_, perf_tuning_knob_, arch_tag_ >
	Compute policy for xmx engine. More...
struct	gpu::xetla::group::compute_policy_default_xmx< compute_attr_, perf_tuning_knob_, gpu_arch::Xe >
	Specialized for Xe architecture. More...
struct	gpu::xetla::group::compute_policy_unaligned_xmx< compute_attr_, perf_tuning_knob_, arch_tag_ >
	Compute policy for unaligned shape and xmx engine. More...
struct	gpu::xetla::group::compute_policy_unaligned_xmx< compute_attr_, perf_tuning_knob_, gpu_arch::Xe >
	Specialized for Xe architecture. More...
struct	gpu::xetla::group::compute_policy_default_fpu< compute_attr_, perf_tuning_knob_, arch_tag_ >
	Compute policy for fpu engine. More...
struct	gpu::xetla::group::compute_policy_default_fpu< compute_attr_, perf_tuning_knob_, gpu_arch::Xe >
	Specialized for Xe architecture. More...
class	gpu::xetla::group::gemm_t< compute_policy_default_fpu< compute_attr_, perf_tuning_knob_, arch_tag_ >, tile_shape_, mem_desc_a_t_, mem_desc_b_t_, pre_processing_t_, std::enable_if_t<(arch_tag_==gpu_arch::Xe)> >
	Is the gemm functor for Xe architecture and vector engine. More...
class	gpu::xetla::group::gemm_t< compute_policy_default_xmx< compute_attr_, perf_tuning_knob_, arch_tag_ >, tile_shape_, mem_desc_a_t_, mem_desc_b_t_, pre_processing_t_, std::enable_if_t<(arch_tag_==gpu_arch::Xe)> >
	Is the gemm functor for Xe architecture and matrix engine. More...
class	gpu::xetla::group::pre_processing_default_t< tile_shape_, arch_tag, std::enable_if_t<(arch_tag==gpu_arch::Xe)> >
	gemm default pre_processing functor. Specialized for Xe architecture. More...
class	gpu::xetla::group::pre_processing_matA_neg_filter_t< tile_shape_, arch_tag, std::enable_if_t<(arch_tag==gpu_arch::Xe)> >
	gemm pre_processing functor with applying relu op to matA. Specialized for Xe architecture. More...
class	gpu::xetla::group::gemm_selector_t< dtype_a, dtype_b, mem_layout_a, mem_layout_b, mem_space_a, mem_space_b, alignment_a, alignment_b, dtype_acc, tile_shape, k_stride, mma_engine::xmx, arch_tag, stages, sync_freq, std::enable_if_t< detail::check_2d_block_pitch_alignment< dtype_a, dtype_b, alignment_a, alignment_b, arch_tag >::value > >
	Selects 2d block && xmx based gemm. More...
class	gpu::xetla::group::gemm_selector_t< dtype_a, dtype_b, mem_layout_a, mem_layout_b, mem_space_a, mem_space_b, alignment_a, alignment_b, dtype_acc, tile_shape, k_stride, mma_engine::xmx, arch_tag, stages, sync_freq, std::enable_if_t<!detail::check_2d_block_pitch_alignment< dtype_a, dtype_b, alignment_a, alignment_b, arch_tag >::value > >
	Selects scatter && xmx based brgemm. More...
class	gpu::xetla::group::gemm_selector_t< dtype_a, dtype_b, mem_layout_a, mem_layout_b, mem_space_a, mem_space_b, alignment_a, alignment_b, dtype_acc, tile_shape, k_stride, mma_engine::fpu, arch_tag, stages, sync_freq, std::enable_if_t< detail::check_2d_block_pitch_alignment< dtype_a, dtype_b, alignment_a, alignment_b, arch_tag >::value > >
	Selects 2d block && fpu based gemm. More...
class	gpu::xetla::group::gemm_t< compute_policy_unaligned_xmx< compute_attr_, perf_tuning_knob_, arch_tag_ >, tile_shape_, mem_desc_a_t_, mem_desc_b_t_, pre_processing_t_, std::enable_if_t<(arch_tag_==gpu_arch::Xe)> >
	Is the gemm functor for unaligned input, Xe architecture and matrix engine. More...

Detailed Description

This is a gemm API to compute matAcc = matA * matB.