Merged
图融合引用文档 #6701
yuwang-Add on 17 days ago
图融合引用文档 #6701
Merged
Merged fromyuwang-/ops-mathmasterintocann/ops-mathmaster
Total: 8 files changed+153-0
| @@ -0,0 +1,87 @@ | |||
| 1 | +# PermuteFusionPass | ||
| 2 | + | ||
| 3 | +<!-- md-trans-meta sourceCommit=unknown translatedAt=2026-10-12T17:03:49.766Z pushedAt=2026-10-13T01:47:29.754Z --> | ||
| 4 | + | ||
| 5 | +## Fusion Pattern | ||
| 6 | + | ||
| 7 | +This fusion replaces the Permute operator in the graph with the TransposeD or Transpose operator. The replacement method depends on the running platform: | ||
| 8 | + | ||
| 9 | +<!-- npu="A3,910b,910,310p,310b" id16 --> | ||
| 10 | +**Use case 1**: | ||
| 11 | + | ||
| 12 | +<!-- npu="A3,910b,910,310p,310b" id14 --> | ||
| 13 | +Permute is replaced with the TransposeD operator. The dimension permutation attribute of Permute is passed as the `perm` attribute of TransposeD. | ||
| 14 | +The supported platforms are as follows: | ||
| 15 | +<!-- end id14 --> | ||
| 16 | +<!-- npu="A3" id7 --> | ||
| 17 | +Atlas A3 products | ||
| 18 | +<!-- end id7 --> | ||
| 19 | +<!-- npu="910b" id8 --> | ||
| 20 | +Atlas A2 products | ||
| 21 | +<!-- end id8 --> | ||
| 22 | +<!-- npu="310b" id9 --> | ||
| 23 | +Atlas 200I/500 A2 inference products | ||
| 24 | +<!-- end id9 --> | ||
| 25 | +<!-- npu="310p" id10 --> | ||
| 26 | +Atlas inference products | ||
| 27 | +<!-- end id10 --> | ||
| 28 | +<!-- npu="910" id11 --> | ||
| 29 | +Atlas training products | ||
| 30 | +<!-- end id11 --> | ||
| 31 | + | ||
| 32 | +The graph structures before and after fusion are as follows: | ||
| 33 | + | ||
| 34 | + | ||
| 35 | +<!-- end id16 --> | ||
| 36 | + | ||
| 37 | +<!-- npu="950" id15 --> | ||
| 38 | +**Use case 2**: | ||
| 39 | + | ||
| 40 | +<!-- npu="950" id13 --> | ||
| 41 | +Replace Permute with the Transpose operator. The dimension permutation attribute of Permute is passed as a Const node to the `perm` input of Transpose. | ||
| 42 | +Supported products are as follows: | ||
| 43 | +<!-- end id13 --> | ||
| 44 | + | ||
| 45 | +<!-- npu="950" id12 --> | ||
| 46 | +Ascend 950PR&950DT products | ||
| 47 | +<!-- end id12 --> | ||
| 48 | + | ||
| 49 | +The graph structures before and after fusion are as follows: | ||
| 50 | + | ||
| 51 | + | ||
| 52 | +<!-- end id15 --> | ||
| 53 | + | ||
| 54 | +The dimension permutation order of the Permute operator is preferentially read from the `perm` attribute. If the attribute does not exist, the `order` attribute is read. The shape and dtype of the output after replacement are the same as those of the original Permute output. | ||
| 55 | + | ||
| 56 | +## Constraints | ||
| 57 | + | ||
| 58 | +- Structural constraints: | ||
| 59 | + - The graph contains a Permute operator, with input x and output y. | ||
| 60 | + - The Permute operator must carry the `perm` or `order` attribute to specify the dimension permutation order. The attribute value must be a valid full permutation of the dimensions of x, without repetition or omission. | ||
| 61 | +- Version constraint: The fusion takes effect when the GE compiler version is 9.1.0 or later. | ||
| 62 | + | ||
| 63 | +## Applicable Products | ||
| 64 | + | ||
| 65 | +<!-- npu="310b" id1 --> | ||
| 66 | +Atlas 200I/500 A2 inference products | ||
| 67 | +<!-- end id1 --> | ||
| 68 | + | ||
| 69 | +<!-- npu="310p" id2 --> | ||
| 70 | +Atlas inference products | ||
| 71 | +<!-- end id2 --> | ||
| 72 | + | ||
| 73 | +<!-- npu="910" id3 --> | ||
| 74 | +Atlas training products | ||
| 75 | +<!-- end id3 --> | ||
| 76 | + | ||
| 77 | +<!-- npu="910b" id4 --> | ||
| 78 | +Atlas A2 products | ||
| 79 | +<!-- end id4 --> | ||
| 80 | + | ||
| 81 | +<!-- npu="A3" id5 --> | ||
| 82 | +Atlas A3 products | ||
| 83 | +<!-- end id5 --> | ||
| 84 | + | ||
| 85 | +<!-- npu="950" id6 --> | ||
| 86 | +Ascend 950PR&950DT products | ||
| 87 | +<!-- end id6 --> | ||
| @@ -0,0 +1,3 @@ | |||
| 1 | +version https://git-lfs.github.com/spec/v1 | ||
| 2 | +oid sha256:dd43e69a6b30987d5fe9c4fd0b97c56f6c285fba583d697d48f9de512fea08dc | ||
| 3 | +size 18794 | ||
| @@ -0,0 +1,3 @@ | |||
| 1 | +version https://git-lfs.github.com/spec/v1 | ||
| 2 | +oid sha256:d739524c54e0d17c3573d3c45ab7a6825af904fcae5689893c04bd1863909941 | ||
| 3 | +size 17445 | ||
| @@ -0,0 +1,3 @@ | |||
| 1 | +version https://git-lfs.github.com/spec/v1 | ||
| 2 | +oid sha256:a0360e2060f9f8ca37d3389776e85cb60edc54403c23ef6a61fd7eb82f12ccfd | ||
| 3 | +size 19963 | ||
| @@ -0,0 +1,3 @@ | |||
| 1 | +version https://git-lfs.github.com/spec/v1 | ||
| 2 | +oid sha256:a32bfe75be6af51d87f9ca29a1f8ee4ba354a697e168f4c2f64e071ac3e7aa24 | ||
| 3 | +size 22466 | ||
| @@ -0,0 +1,3 @@ | |||
| 1 | +version https://git-lfs.github.com/spec/v1 | ||
| 2 | +oid sha256:451d1c0c02a0a02ca2fdc9b1e7f20d0ddf02436a318cb7029e492b2e5d6a75e8 | ||
| 3 | +size 26092 | ||
| @@ -0,0 +1,21 @@ | |||
| 1 | +# Globalavgpoolpass | ||
| 2 | + | ||
| 3 | +<!-- md-trans-meta sourceCommit=unknown translatedAt=2026-10-12T16:56:50.046Z pushedAt=2026-10-13T01:47:29.733Z --> | ||
| 4 | + | ||
| 5 | +## Fusion Pattern | ||
| 6 | + | ||
| 7 | +This fusion replaces the `GlobalAveragePool` operator in the graph with the `ReduceMean` operator, which computes the mean along the spatial dimensions (all dimensions from the second one onward), as shown in the following figure. | ||
 | |||
| 8 | + | ||
| 9 | + | ||
| 10 | + | ||
| 11 | +After the replacement, the attributes of `ReduceMean` are set to `keep_dims=true` and `noop_with_empty_axes=true`. | ||
| 12 | + | ||
| 13 | +## Constraints | ||
| 14 | + | ||
| 15 | +- Version constraint: The fusion takes effect when the GE compiler version is not lower than 9.1.0. | ||
| 16 | +- Input dimension constraint: The input `x` of `GlobalAveragePool` must be 3D, 4D, or 5D. Inputs of other dimensions are not fused. | ||
| 17 | +- Data types and data formats follow the constraints of the underlying `GlobalAveragePool`/`ReduceMean` operators. | ||
| 18 | + | ||
| 19 | +## Applicable Products | ||
| 20 | + | ||
| 21 | +This fusion rule is platform-independent. For the specific supported models, see the Applicable Products of the underlying `GlobalAveragePool` and `ReduceMean` operators. | ||
| @@ -0,0 +1,30 @@ | |||
| 1 | +# ReduceMeanWithCastFusionPass | ||
| 2 | + | ||
| 3 | +<!-- md-trans-meta sourceCommit=unknown translatedAt=2026-10-12T17:09:24.202Z pushedAt=2026-10-13T01:47:29.768Z --> | ||
| 4 | + | ||
| 5 | +## Fusion Pattern | ||
| 6 | + | ||
| 7 | +This fusion decomposes the `ReduceMeanWithCast` operator in the graph into `Cast` + `ReduceMean`. When `ReduceMeanWithCast` carries a valid `dtype` attribute (not `DT_UNDEFINED`), a `Cast` node is inserted first to convert the input to the target data type, and then `ReduceMean` is connected. When `dtype` is `DT_UNDEFINED`, no `Cast` is inserted, and the operator is directly replaced with `ReduceMean`. | ||
| 8 | + | ||
| 9 | +**Use case 1: dtype is valid (Cast is inserted)** | ||
| 10 | + | ||
| 11 | + | ||
| 12 | + | ||
| 13 | +**Use case 2: dtype is DT_UNDEFINED (Cast is not inserted)** | ||
| 14 | + | ||
| 15 | + | ||
| 16 | + | ||
| 17 | +After the replacement, the `keep_dims` and `noop_with_empty_axes` attributes of `ReduceMean` are inherited from the original `ReduceMeanWithCast` node, and `dst_type` of `Cast` takes the value of the `dtype` attribute of the original node. | ||
| 18 | + | ||
| 19 | +## Constraints | ||
| 20 | + | ||
| 21 | +- Version constraint: The fusion takes effect when the GE compiler version is 9.1.0 or later. | ||
| 22 | +- Attribute constraints: | ||
| 23 | + - `keep_dims`: Boolean. Default value: `false`. It is inherited as-is by `ReduceMean`. | ||
| 24 | + - `noop_with_empty_axes`: Boolean, defaults to `true`, inherited as-is by `ReduceMean`. | ||
| 25 | + - `dtype`: target data type, defaults to `DT_UNDEFINED`. If the value is `DT_UNDEFINED`, no `Cast` is inserted; otherwise, `Cast` is inserted and the value is used as `dst_type`. | ||
| 26 | +- Input shape and data type are not validated. Relevant constraints follow those of the underlying `Cast` and `ReduceMean` operators. | ||
| 27 | + | ||
| 28 | +## Applicable Products | ||
| 29 | + | ||
| 30 | +This fusion rule is platform-independent. The supported models depend on the applicable products of the underlying `Cast` and `ReduceMean` operators. | ||
🟡 Medium Priority
变更行：math/reduce_mean/docs/GlobalavgpoolFusionPass_en.md:7 新增「computes the mean along the spatial dimensions (all dimensions from the second one onward)」。
问题链路：GlobalAveragePool 的语义是沿空间维度（NCHW 下的 H、W）求均值，输出仍保留 batch(N) 与 channel(C) 维，即归约轴为 {2,…,n-1}；而该句把归约范围写成「从第 2 维起的所有维度」= {1,…,n-1}，等于把 channel 维也一起平均掉，描述出的计算与 GlobalAveragePool 标准语义（输出 N×C×1×1）不一致。
失败模式：读者按此文档理解融合后 ReduceMean 的 axes，会认为 C 维被压平；若下游按此理解做 shape 推导、对齐适配层或编写校验用例，就会出现 shape/数值预期错误（例如期望输出为 N×1×1 而实际为 N×C×1×1）。触发条件：任何 4D NCHW 输入阅读该文档并据此推导输出形状。由于本环境无法读取 pass 实现核对实际 axes，置信度已下调，但即便是实现侧真的传 axes={1..n-1}，该实现同样与 GlobalAveragePool 语义不符，仍需修正。
建议：核对 pass 实际传入的 axes 后修正该句：GlobalAveragePool 只沿空间维（NCHW 的 H、W，即从第三维起）求均值并保留 batch/channel 维；建议直接写出 axes 取值，避免「from the second one onward」这种会吞掉 channel 维的表述。