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

1	case_id	batch_a	batch_b	m	n	k	dtype_a	dtype_b	dtype_c	trans_a	trans_b	has_bias	out_nd	deterministic_level	plan_case_id	plan_branch	plan_npu	plan_op	plan_used_core_num	plan_split_b	plan_m_cnt	plan_n_cnt	plan_grid_k	plan_core_map	plan_b_core	plan_merge_b0	plan_single_core_m	plan_single_core_n	plan_single_core_k	plan_k_l1	plan_b_l1	plan_l1_form	plan_base_m	plan_base_n	plan_base_k	plan_l2_policy_in	plan_l2_policy_out	plan_swizzle_w	plan_workspace_bytes	plan_tail_strategy	plan_fixpipe_unitflag	plan_out_dtype_bytes	plan_note	gm_read_bytes	l2_read_bytes	t_mte2_gm	t_mte2_l2	t_mte2	dma_cmd_count	t_dma_cmd	cube_flops	t_mmad	fixpipe_bytes	t_fixpipe	t_reduce	t_steady	t_drain	t_total	bottleneck	feasible	violations	bound_type	advice
2	merge_demo_1	128	128	64	64	512	bf16	bf16	bf16	False	False	False	True	0	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	524288	0.0	1.048576e-05	0.0	1.0685759999999999e-05	4	2.0000000000000002e-07	16777216.0	1.104672658436214e-06	32768	6.5536e-07	0.0	1.0685759999999999e-05	4.400081646090535e-07	1.1125768164609053e-05	MTE2_GM	True		访存Bound(GM)	瓶颈在 GM 搬入: 可考虑增大 tile 提升 dValue/单核搬移量, 或利用 L2 驻留吸收重复读 (MergeBatch/ASW swizzle 方向)
3	iter_demo_1	128	128	64	64	256	bf16	bf16	bf16	False	False	False	True	0	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	262144	0.0	5.24288e-06	0.0	5.4428799999999995e-06	4	2.0000000000000002e-07	8388608.0	5.52336329218107e-07	32768	6.5536e-07	0.0	5.4428799999999995e-06	3.0192408230452674e-07	5.744804082304526e-06	MTE2_GM	True		访存Bound(GM)	瓶颈在 GM 搬入: 可考虑增大 tile 提升 dValue/单核搬移量, 或利用 L2 驻留吸收重复读 (MergeBatch/ASW swizzle 方向)
4	iter_demo_2	512	512	128	128	128	bf16	bf16	bf16	False	False	False	True	0	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	1048576	0.0	2.097152e-05	0.0	2.1771519999999998e-05	16	8.000000000000001e-07	67108864.0	4.418690633744856e-06	524288	1.048576e-05	0.0	2.1771519999999998e-05	9.315281646090535e-07	2.270304816460905e-05	MTE2_GM	True		访存Bound(GM)	瓶颈在 GM 搬入: 可考虑增大 tile 提升 dValue/单核搬移量, 或利用 L2 驻留吸收重复读 (MergeBatch/ASW swizzle 方向)
5	iter_demo_3	2048	2048	1024	1024	512	bf16	bf16	bf16	False	False	False	True	0	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	134217728	0.0	0.00268435456	0.0	0.00270995456	512	2.5600000000000002e-05	68719476736.0	0.004524739208954733	134217728	0.00268435456	0.0	0.004524739208954733	5.078042126748971e-05	0.004575519630222223	MMAD	True		计算Bound	瓶颈在 Cube 计算: 已接近理论算力上限, 检查是否有冗余计算 (MergeBatch 交叉项) 可消除
6	iter_demo_4	64	64	64	64	8192	bf16	bf16	bf16	False	False	False	True	0	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	4194304	0.0	8.388608e-05	0.0	8.468608e-05	16	8.000000000000001e-07	134217728.0	8.837381267489712e-06	16384	3.2768e-07	0.0	8.468608e-05	7.16176329218107e-07	8.540225632921811e-05	MTE2_GM	True		访存Bound(GM)	瓶颈在 GM 搬入: 可考虑增大 tile 提升 dValue/单核搬移量, 或利用 L2 驻留吸收重复读 (MergeBatch/ASW swizzle 方向)
7	fp16_out_demo	128	128	64	64	512	bf16	bf16	fp16	False	False	False	True	0	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	524288	0.0	1.048576e-05	0.0	1.0685759999999999e-05	4	2.0000000000000002e-07	16777216.0	1.104672658436214e-06	32768	6.5536e-07	0.0	1.0685759999999999e-05	4.400081646090535e-07	1.1125768164609053e-05	MTE2_GM	True		访存Bound(GM)	瓶颈在 GM 搬入: 可考虑增大 tile 提升 dValue/单核搬移量, 或利用 L2 驻留吸收重复读 (MergeBatch/ASW swizzle 方向)
8	to_matmul_demo	1	1	2048	2048	2048	bf16	bf16	bf16	False	False	False	True	0	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)	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0		False	dValue=0B < 下限 128B, 搬移带宽利用率崩塌		分支 '转Matmul' 的评估模型待后续迭代; 当前支持: ['IterBatch', 'MergeBatch']
9	special_k1	128	128	256	256	1	bf16	bf16	bf16	False	False	False	True	0	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)	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0		False	dValue=0B < 下限 128B, 搬移带宽利用率崩塌		分支 '特殊分支' 的评估模型待后续迭代; 当前支持: ['IterBatch', 'MergeBatch']
10	asw_fallback_demo	32	32	4096	4096	4096	bf16	bf16	bf16	False	False	False	True	0	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	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0		False	dValue=0B < 下限 128B, 搬移带宽利用率崩塌		分支 'ASW_Basic' 的评估模型待后续迭代; 当前支持: ['IterBatch', 'MergeBatch']