Mojo struct
SM100TensorAccumulatorSS
@register_passable(trivial)
struct SM100TensorAccumulatorSS[operand_type: DType, accum_type: DType, MMA_M: Int, MMA_N: Int, BM: Int, BN: Int, BK: Int, compute_BK: Int, num_softmax_threads: Int, swizzle_a: TensorMapSwizzle = TensorMapSwizzle.SWIZZLE_128B, swizzle_b: TensorMapSwizzle = TensorMapSwizzle.SWIZZLE_128B, *, transpose_b: Bool = True, cta_group: Int = 1, pipeline_stages: Int = 1]
Fields
- mbar (
UnsafePointer[SharedMemBarrier, address_space=AddressSpace.SHARED]): - pipeline (
PipelineState[pipeline_stages]):
Implemented traits
AnyType,
Copyable,
ImplicitlyCopyable,
Movable,
UnknownDestructibility
Aliases
__copyinit__is_trivial
alias __copyinit__is_trivial = True
__del__is_trivial
alias __del__is_trivial = True
__moveinit__is_trivial
alias __moveinit__is_trivial = True
a_offset
alias a_offset = MMAOperandOffsetFn[SM100TensorAccumulatorSS[operand_type, accum_type, MMA_M, MMA_N, BM, BN, BK, compute_BK, num_softmax_threads, swizzle_a, swizzle_b, transpose_b=transpose_b, cta_group=cta_group, pipeline_stages=pipeline_stages].operand_t, BM, BK, swizzle_a, True, MMA_M, 16]()
a_t
alias a_t = MMASmemDescriptor
ab_t
alias ab_t = UMMADescriptorSS[SM100TensorAccumulatorSS[operand_type, accum_type, MMA_M, MMA_N, BM, BN, BK, compute_BK, num_softmax_threads, swizzle_a, swizzle_b, transpose_b=transpose_b, cta_group=cta_group, pipeline_stages=pipeline_stages].operand_t]
accum_t
alias accum_t = accum_type
b_offset
alias b_offset = MMAOperandOffsetFn[SM100TensorAccumulatorSS[operand_type, accum_type, MMA_M, MMA_N, BM, BN, BK, compute_BK, num_softmax_threads, swizzle_a, swizzle_b, transpose_b=transpose_b, cta_group=cta_group, pipeline_stages=pipeline_stages].operand_t, BN, BK, swizzle_b, transpose_b, MMA_N, 16]()
b_t
alias b_t = MMASmemDescriptor
c_t
alias c_t = TMemAccumulator[SM100TensorAccumulatorSS[operand_type, accum_type, MMA_M, MMA_N, BM, BN, BK, compute_BK, num_softmax_threads, swizzle_a, swizzle_b, transpose_b=transpose_b, cta_group=cta_group, pipeline_stages=pipeline_stages].accum_t, (BM // SM100TensorAccumulatorSS[operand_type, accum_type, MMA_M, MMA_N, BM, BN, BK, compute_BK, num_softmax_threads, swizzle_a, swizzle_b, transpose_b=transpose_b, cta_group=cta_group, pipeline_stages=pipeline_stages].num_m_blocks_per_warp), MMA_N, SM100TensorAccumulatorSS[operand_type, accum_type, MMA_M, MMA_N, BM, BN, BK, compute_BK, num_softmax_threads, swizzle_a, swizzle_b, transpose_b=transpose_b, cta_group=cta_group, pipeline_stages=pipeline_stages].num_m_blocks_per_warp, SM100TensorAccumulatorSS[operand_type, accum_type, MMA_M, MMA_N, BM, BN, BK, compute_BK, num_softmax_threads, swizzle_a, swizzle_b, transpose_b=transpose_b, cta_group=cta_group, pipeline_stages=pipeline_stages].num_n_mmas, num_softmax_threads]
idesc
alias idesc = UMMAInsDescriptor.create[UMMAKind.KIND_F16, SM100TensorAccumulatorSS[operand_type, accum_type, MMA_M, MMA_N, BM, BN, BK, compute_BK, num_softmax_threads, swizzle_a, swizzle_b, transpose_b=transpose_b, cta_group=cta_group, pipeline_stages=pipeline_stages].accum_t, SM100TensorAccumulatorSS[operand_type, accum_type, MMA_M, MMA_N, BM, BN, BK, compute_BK, num_softmax_threads, swizzle_a, swizzle_b, transpose_b=transpose_b, cta_group=cta_group, pipeline_stages=pipeline_stages].operand_t, SM100TensorAccumulatorSS[operand_type, accum_type, MMA_M, MMA_N, BM, BN, BK, compute_BK, num_softmax_threads, swizzle_a, swizzle_b, transpose_b=transpose_b, cta_group=cta_group, pipeline_stages=pipeline_stages].operand_t, Index[dtype=DType.uint32](MMA_M, MMA_N), transpose_b=transpose_b]()
MMA_K
alias MMA_K = 16
num_k_mmas
alias num_k_mmas = (compute_BK // 16)
num_m_blocks_per_warp
alias num_m_blocks_per_warp = ((2 * BM) // num_softmax_threads)
num_m_mmas
alias num_m_mmas = (BM // MMA_M)
num_n_mmas
alias num_n_mmas = (BN // MMA_N)
operand_t
alias operand_t = operand_type
smem_ptr_t
alias smem_ptr_t = UnsafePointer[Scalar[SM100TensorAccumulatorSS[operand_type, accum_type, MMA_M, MMA_N, BM, BN, BK, compute_BK, num_softmax_threads, swizzle_a, swizzle_b, transpose_b=transpose_b, cta_group=cta_group, pipeline_stages=pipeline_stages].operand_t], address_space=AddressSpace.SHARED]
Methods
__init__
__init__(smem: UnsafePointer[SharedMemBarrier, address_space=AddressSpace.SHARED]) -> Self
check_constraints
static check_constraints()
init
init(self)
mma_descriptors
static mma_descriptors[dtype_a: DType, dtype_b: DType](p_a: UnsafePointer[Scalar[dtype_a], address_space=AddressSpace.SHARED], p_b: UnsafePointer[Scalar[dtype_b], address_space=AddressSpace.SHARED]) -> UMMADescriptorSS[SM100TensorAccumulatorSS[operand_type, accum_type, MMA_M, MMA_N, BM, BN, BK, compute_BK, num_softmax_threads, swizzle_a, swizzle_b, transpose_b=transpose_b, cta_group=cta_group, pipeline_stages=pipeline_stages].operand_t]
Returns:
UMMADescriptorSS
mma
mma(mut self, a: MMASmemDescriptor, b: MMASmemDescriptor, c_base: TMemAccumulator[SM100TensorAccumulatorSS[operand_type, accum_type, MMA_M, MMA_N, BM, BN, BK, compute_BK, num_softmax_threads, swizzle_a, swizzle_b, transpose_b=transpose_b, cta_group=cta_group, pipeline_stages=pipeline_stages].accum_t, (BM // SM100TensorAccumulatorSS[operand_type, accum_type, MMA_M, MMA_N, BM, BN, BK, compute_BK, num_softmax_threads, swizzle_a, swizzle_b, transpose_b=transpose_b, cta_group=cta_group, pipeline_stages=pipeline_stages].num_m_blocks_per_warp), MMA_N, SM100TensorAccumulatorSS[operand_type, accum_type, MMA_M, MMA_N, BM, BN, BK, compute_BK, num_softmax_threads, swizzle_a, swizzle_b, transpose_b=transpose_b, cta_group=cta_group, pipeline_stages=pipeline_stages].num_m_blocks_per_warp, SM100TensorAccumulatorSS[operand_type, accum_type, MMA_M, MMA_N, BM, BN, BK, compute_BK, num_softmax_threads, swizzle_a, swizzle_b, transpose_b=transpose_b, cta_group=cta_group, pipeline_stages=pipeline_stages].num_n_mmas, num_softmax_threads], scale_c: UInt32)
wait_for_tmem
wait_for_tmem(self)
Wait for the accumulator tmem to finish being read.
wait_for_mma
wait_for_mma(self, c_base: TMemAccumulator[SM100TensorAccumulatorSS[operand_type, accum_type, MMA_M, MMA_N, BM, BN, BK, compute_BK, num_softmax_threads, swizzle_a, swizzle_b, transpose_b=transpose_b, cta_group=cta_group, pipeline_stages=pipeline_stages].accum_t, (BM // SM100TensorAccumulatorSS[operand_type, accum_type, MMA_M, MMA_N, BM, BN, BK, compute_BK, num_softmax_threads, swizzle_a, swizzle_b, transpose_b=transpose_b, cta_group=cta_group, pipeline_stages=pipeline_stages].num_m_blocks_per_warp), MMA_N, SM100TensorAccumulatorSS[operand_type, accum_type, MMA_M, MMA_N, BM, BN, BK, compute_BK, num_softmax_threads, swizzle_a, swizzle_b, transpose_b=transpose_b, cta_group=cta_group, pipeline_stages=pipeline_stages].num_m_blocks_per_warp, SM100TensorAccumulatorSS[operand_type, accum_type, MMA_M, MMA_N, BM, BN, BK, compute_BK, num_softmax_threads, swizzle_a, swizzle_b, transpose_b=transpose_b, cta_group=cta_group, pipeline_stages=pipeline_stages].num_n_mmas, num_softmax_threads]) -> TMemAccumulator[SM100TensorAccumulatorSS[operand_type, accum_type, MMA_M, MMA_N, BM, BN, BK, compute_BK, num_softmax_threads, swizzle_a, swizzle_b, transpose_b=transpose_b, cta_group=cta_group, pipeline_stages=pipeline_stages].accum_t, (BM // SM100TensorAccumulatorSS[operand_type, accum_type, MMA_M, MMA_N, BM, BN, BK, compute_BK, num_softmax_threads, swizzle_a, swizzle_b, transpose_b=transpose_b, cta_group=cta_group, pipeline_stages=pipeline_stages].num_m_blocks_per_warp), MMA_N, SM100TensorAccumulatorSS[operand_type, accum_type, MMA_M, MMA_N, BM, BN, BK, compute_BK, num_softmax_threads, swizzle_a, swizzle_b, transpose_b=transpose_b, cta_group=cta_group, pipeline_stages=pipeline_stages].num_m_blocks_per_warp, SM100TensorAccumulatorSS[operand_type, accum_type, MMA_M, MMA_N, BM, BN, BK, compute_BK, num_softmax_threads, swizzle_a, swizzle_b, transpose_b=transpose_b, cta_group=cta_group, pipeline_stages=pipeline_stages].num_n_mmas, num_softmax_threads]
Wait for the accumulator tmem to finish being read.
Returns:
tmem_arrive_init
tmem_arrive_init(self)
tmem_arrive
tmem_arrive(mut self)
Indicate that the accumulator is ready to be updated.
Was this page helpful?
Thank you! We'll create more content like this.
Thank you for helping us improve!