matmul-analysis/BMM/BMM_Theory/examples/plans.csv

1	case_id	branch	npu	op	used_core_num	split_b	m_cnt	n_cnt	grid_k	core_map	b_core	merge_b0	single_core_m	single_core_n	single_core_k	k_l1	b_l1	l1_form	base_m	base_n	base_k	l2_policy_in	l2_policy_out	swizzle_w	workspace_bytes	tail_strategy	fixpipe_unitflag	out_dtype_bytes	note
2	merge_demo_1	IterBatch	Ascend950PR	batch_mat_mul_v3	32	32	1	1	1	切B轮转分配(核间零重复读零依赖)	4	1	64	64	512	512	2	b_双batch乒乓	64	64	256	allocate(GM->L1随路驻留L2)	direct_gm(输出仅写一次,直写GM不占L2)	0	0	不涉及(核内不切M/N)	True	2	双batch乒乓: 2*(MK+KN)*dtype=256KB <= L1
3	iter_demo_1	IterBatch	Ascend950PR	batch_mat_mul_v3	32	32	1	1	1	切B轮转分配(核间零重复读零依赖)	4	1	64	64	256	256	2	b_双batch乒乓	64	64	256	allocate(GM->L1随路驻留L2)	direct_gm(输出仅写一次,直写GM不占L2)	0	0	不涉及(核内不切M/N)	True	2	双batch乒乓: 2*(MK+KN)*dtype=128KB <= L1
4	iter_demo_2	IterBatch	Ascend950PR	batch_mat_mul_v3	32	32	1	1	1	切B轮转分配(核间零重复读零依赖)	16	1	128	128	128	128	2	b_双batch乒乓	128	128	128	allocate(GM->L1随路驻留L2)	direct_gm(输出仅写一次,直写GM不占L2)	0	0	不涉及(核内不切M/N)	True	2	双batch乒乓: 2*(MK+KN)*dtype=128KB <= L1
5	iter_demo_3	IterBatch	Ascend950PR	batch_mat_mul_v3	32	32	1	1	1	切B轮转分配(核间零重复读零依赖)	64	1	1024	1024	512	64	1	d_两侧都切K	32	1024	16	allocate(GM->L1随路驻留L2)	direct_gm(输出仅写一次,直写GM不占L2)	0	0	不涉及(核内不切M/N)	True	2	两侧都切K: k_L1=64, K段成对流水, batch边界天然无缝
6	iter_demo_4	IterBatch	Ascend950PR	batch_mat_mul_v3	32	32	1	1	1	切B轮转分配(核间零重复读零依赖)	2	1	64	64	8192	1024	1	d_两侧都切K	64	64	256	allocate(GM->L1随路驻留L2)	direct_gm(输出仅写一次,直写GM不占L2)	0	0	不涉及(核内不切M/N)	True	2	两侧都切K: k_L1=1024, K段成对流水, batch边界天然无缝
7	fp16_out_demo	IterBatch	Ascend950PR	batch_mat_mul_v3	32	32	1	1	1	切B轮转分配(核间零重复读零依赖)	4	1	64	64	512	512	2	b_双batch乒乓	64	64	256	allocate(GM->L1随路驻留L2)	direct_gm(输出仅写一次,直写GM不占L2)	0	0	不涉及(核内不切M/N)	True	2	双batch乒乓: 2*(MK+KN)*dtype=256KB <= L1
8	to_matmul_demo	转Matmul	Ascend950PR	batch_mat_mul_v3	32	1	1	1	1		0	1	0	0	0	0	1		0	0	0			0	0		True	2	BatchA=1或BatchB=1, 折叠转普通Matmul (该分支详实现待后续迭代)
9	special_k1	特殊分支	Ascend950PR	batch_mat_mul_v3	32	1	1	1	1		0	1	0	0	0	0	1		0	0	0			0	0		True	2	K=1, Cube 无用, 走 AIV 向量通路 (该分支详实现待后续迭代)
10	asw_fallback_demo	ASW_Basic	Ascend950PR	batch_mat_mul_v3	32	1	1	1	1		0	1	0	0	0	0	1		0	0	0			0	0		True	2	IterBatch/MergeBatch 进入条件均不满足 (如负载均衡/搬移效率不达标), 回落 ASW_Basic (该分支详实现待后续迭代); IterBatch未过: 3_L1驻留形态(四选一, 核心要求: 单batch核内零重复读); MergeBatch未过: 1_batch关系与每核份额: BatchA==BatchB 且 b_core=B/C>=2*b0; 2_L0C容量: 2*(b0*M)*(b0*N)*4B <= L0C; 5_访存Bound: 2MN/(M+N) < R16/b0