Module reduction 

Reduction kernels — each CTA handles one output element (one “row” of the flattened [outer, inner] view). The caller is responsible for reshaping the input to [n_outer, n_inner] before invoking these kernels.

Grid: [n_outer, 1, 1] Block: [BLOCK_INNER, 1, 1]

Structs

CumProdForward

Forward: y = cumprod(x, axis=0) over a 1-D block

CumSumForward

Forward: y = cumsum(x, axis=0) over a 1-D block Each CTA handles one complete row (n_inner elements).

GlobalAvgPoolForward

Forward: y[row] = mean(x[row, :]) (same as ReduceMean)

GlobalMaxPoolForward

Forward: y[row] = max(x[row, :]) (same as ReduceMax)

ReduceL1Forward

Forward: y[row] = sum(|x[row, :]|)

ReduceL2Forward

Forward: y[row] = sqrt(sum(x[row, :]^2))

ReduceLogSumExpForward

Forward: y[row] = log(sum(exp(x[row, :]))) — numerically stable via max subtraction

ReduceLogSumForward

Forward: y[row] = log(sum(x[row, :])) (numerically unsafe; use ReduceLogSumExp for stable)

ReduceMaxForward

Forward: y[row] = max(x[row, :])

ReduceMeanForward

Forward: y[row] = mean(x[row, :])

ReduceMinForward

Forward: y[row] = min(x[row, :])

ReduceProdForward

Forward: y[row] = prod(x[row, :]) Note: implemented as exp(sum(log(x))) — only valid for positive x. For general use this is a placeholder.

ReduceSumForward

Forward: y[row] = sum(x[row, :])

ReduceSumSquareForward

Forward: y[row] = sum(x[row, :]^2)

Functions

cum_prod_forward

Forward: y = cumprod(x, axis=0) over a 1-D block

cum_sum_forward

Forward: y = cumsum(x, axis=0) over a 1-D block Each CTA handles one complete row (n_inner elements).

global_avg_pool_forward

Forward: y[row] = mean(x[row, :]) (same as ReduceMean)

global_max_pool_forward

Forward: y[row] = max(x[row, :]) (same as ReduceMax)

reduce_l1_forward

Forward: y[row] = sum(|x[row, :]|)

reduce_l2_forward

Forward: y[row] = sqrt(sum(x[row, :]^2))

reduce_log_sum_exp_forward

Forward: y[row] = log(sum(exp(x[row, :]))) — numerically stable via max subtraction

reduce_log_sum_forward

Forward: y[row] = log(sum(x[row, :])) (numerically unsafe; use ReduceLogSumExp for stable)

reduce_max_forward

Forward: y[row] = max(x[row, :])

reduce_mean_forward

Forward: y[row] = mean(x[row, :])

reduce_min_forward

Forward: y[row] = min(x[row, :])

reduce_prod_forward

Forward: y[row] = prod(x[row, :]) Note: implemented as exp(sum(log(x))) — only valid for positive x. For general use this is a placeholder.

reduce_sum_forward

Forward: y[row] = sum(x[row, :])

reduce_sum_square_forward

Forward: y[row] = sum(x[row, :]^2)