openai / blocksparse - Conditional Complexity

Intro

Conditional complexity (also called cyclomatic complexity) is a term used to measure the complexity of software. The term refers to the number of possible paths through a program function. A higher value ofter means higher maintenance and testing costs (infosecinstitute.com).
Conditional complexity is calculated by counting all conditions in the program that can affect the execution path (e.g. if statement, loops, switches, and/or operators, try and catch blocks...).
Conditional complexity is measured at the unit level (methods, functions...).
Units are classified in four categories based on the measured McCabe index: 1-5 (simple units), 6-10 (medium complex units), 11-25 (complex units), 26+ (very complex units).

Learn more...

Conditional Complexity Overall

There are 842 units with 8,715 lines of code in units (25.8% of code).

1 very complex units (71 lines of code)
3 complex units (232 lines of code)
38 medium complex units (1,843 lines of code)
68 simple units (1,500 lines of code)
732 very simple units (5,069 lines of code)

Legend:

51+

26-50

11-25

6-10

1-5

Alternative Visuals

3D view of all units

Conditional Complexity per Extension

51+

26-50

11-25

6-10

1-5

Conditional Complexity per Logical Component

primary logical decomposition

51+

26-50

11-25

6-10

1-5

Most Complex Units

Top 50 most complex units

Unit	# lines	McCabe index	# params
def gradients() in blocksparse/grads.py	71	51	6
def group_param_grads() in blocksparse/matmul.py	83	31	2
Status Compute_Xprop() in src/blocksparse_matmul_op.cc	81	29	2
def xprop_lut() in blocksparse/matmul.py	68	26	8
Status Compute_Updat() in src/blocksparse_matmul_op.cc	73	23	1
def spatial_grid() in blocksparse/conv.py	33	22	6
def __init__() in blocksparse/matmul.py	55	16	6
def serialize_allreduce_ops() in blocksparse/nccl.py	53	16	3
def _PendingCount() in blocksparse/grads.py	34	16	2
def gpu_fprop_tune() in src/dev/spatial_conv.py	60	16	1
def get_kernel() in generate_kernels.py	49	15	1
def block_reduced_full_dw() in blocksparse/matmul.py	31	15	4
def bst_conv_layout() in blocksparse/utils.py	47	15	7
def bst_deconv_layout() in blocksparse/utils.py	47	15	7
void Compute() in src/layer_norm_op.cc	73	15	1
def group_allreduce() in blocksparse/nccl.py	46	14	7
def __init__() in blocksparse/conv.py	43	14	9
def group_lstm_grads() in blocksparse/lstm.py	51	14	4
Status BlocksparseConvShape() in src/blocksparse_conv_op.cc	115	14	1
virtual void Compute() in src/blocksparse_conv_op.cc	94	14	1
__device__ __forceinline__ void mma_sync() in src/gpu_hmma.h	14	14	2
def __init__() in blocksparse/transformer.py	57	13	6
def cwise_linear() in blocksparse/conv.py	18	13	6
void Compute_XN() in src/bst_op.cc	55	13	3
void Compute() in src/quantize_op.cc	61	13	1
void Compute() in src/quantize_op.cc	59	13	1
void Compute() in src/layer_norm_op.cc	68	13	1
void Compute() in src/ew_op.cc	48	13	1
void Compute() in src/ew_op.cc	43	13	1
def masked_softmax_test() in blocksparse/transformer.py	35	12	4
def _build_edge_lut() in blocksparse/conv.py	35	12	3
void NcclCommThread() in src/nccl_op.cc	48	12	0
void Compute() in src/blocksparse_matmul_op.cc	82	12	1
def group_dg_grads() in blocksparse/matmul.py	29	11	3
def quantize() in blocksparse/quantize.py	40	11	4
def broadcast_check() in blocksparse/ewops.py	15	11	6
def bias_relu() in blocksparse/ewops.py	16	11	8
def layer_norm() in blocksparse/norms.py	24	11	10
explicit BlocksparseConvOp() in src/blocksparse_conv_op.cc	69	11	1
void Compute_NT() in src/bst_op.cc	53	11	1
void Compute() in src/transformer_op.cc	34	11	1
void Compute() in src/transformer_op.cc	36	11	1
def apply() in blocksparse/optimize.py	23	10	3
void Compute() in src/ew_op.cc	29	10	1
void magicu64() in src/gpu_types.cc	23	10	3
def get_ptx_file() in generate_kernels.py	34	9	5
def _create_slots() in blocksparse/optimize.py	9	9	2
def fprop_test() in blocksparse/matmul.py	22	9	4
def bprop_test() in blocksparse/matmul.py	22	9	4
def log_stats() in blocksparse/quantize.py	19	9	8