Skip to content

Commit

Permalink
feat(docs): add nms, nonzero add op design docs (Cambricon#9)
Browse files Browse the repository at this point in the history 
* feat(docs): add ops docs.

* feat(bangpy-ops): update add operator computation and the design document.

* fix(docs):add nms design document

* feat(docs):add nonzero design docs.

* fix(docs):fix design docs links.

* fix(docs):rm host.h

* fix(docs):add nms design document.

* fix(docs):add nms design document.

* fix(docs):rename designer name.

* feat(docs): update file link path.

* feat(docs): update file link path.

Co-authored-by: zhourihua <[email protected]>
Co-authored-by: xumingyang <[email protected]>
Co-authored-by: Ai Youngcino <[email protected]>
  • Loading branch information
@withfall @UniqueSquirrel @AIYoungcino
4 people authored Dec 30, 2021
1 parent b5aee29 commit a2b1646
Show file tree
Hide file tree
Showing 13 changed files with 733 additions and 68 deletions.
2 changes: 1 addition & 1 deletion README.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -37,7 +37,7 @@ mlu-ops 提供基于寒武纪机器学习单元（Machine Learning Unit，MLU）

```sh
wget https://www.python.org/ftp/python/3.8.0/Python-3.8.0.tgz
tar -xvf /tmp/Python-3.8.0.tgz
tar -xvf Python-3.8.0.tgz
cd Python-3.8.0
make -j24 && make install
```
Expand Down
2 changes: 1 addition & 1 deletion bangpy-ops/README.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -31,7 +31,7 @@

`注意` : 如果将指定算子列表存放在文件中，其格式是每行一个算子。

编译成功后在 `bangpy-ops/outs` 目录下生成与算子同名的输出文件夹，其中包含 `libmluops.so`，`host.h` 等文件。
编译成功后在 `bangpy-ops/outs` 目录下生成与算子同名的输出文件夹，其中包含 `libmluops.so`等文件。


## 运行测试用例
Expand Down
35 changes: 13 additions & 22 deletions bangpy-ops/ops/add/add.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -28,9 +28,8 @@
from bangpy.platform.bang_config import ALIGN_LENGTH, TARGET
from bangpy.tcp.runtime import TaskType

DTYPES = [bangpy.float16]
DTYPES = [bangpy.float16, bangpy.float32]
TARGET_LIST = ["mlu370-s4", "mlu220-m2", "mlu270", "mlu290"]
SHAPE = (64000,)
KERNEL_NAME = "add"


Expand All @@ -39,42 +38,34 @@ class Add(object):
Add the data in the two buffers.
"""

def __init__(self, shape, dtype, target, task_num):
self.shape = shape
def __init__(self, dtype, target, task_num):
self.dtype = dtype
self.target = target
self.task_num = task_num
self.length = np.prod(shape)
self.bp = tcp.TCP(target)
self.length = self.bp.SizeVar("length")
self.nram_size = TARGET(target).nram_size
self.dtype_sz = dtype.bytes
self.bp = tcp.TCP(target)
self.single_buffer_size = 1024
self.bp.launch_task(self.task_num, 1, 1)

def compute_body(self):
# calculate split strategy
# ensure the data size can be divisible by task_num and 128 bytes aligned
assert (self.dtype_sz * self.length) % self.task_num % ALIGN_LENGTH == 0
# gets the data length to be calculated for each task
data_calculated_each_task = self.length // self.task_num
loop_num = np.ceil(
3 * data_calculated_each_task * self.dtype_sz / self.nram_size
)
# ensure the data size is 128 bytes aligned for each calculation
while (
data_calculated_each_task % loop_num != 0
or data_calculated_each_task // loop_num % ALIGN_LENGTH != 0
):
loop_num += 1
data_calculated_each_time = int(data_calculated_each_task // loop_num)
# gets the number of cycles required for each task
loop_num = data_calculated_each_task * self.dtype_sz // self.single_buffer_size
# gets the data length for each calculation
data_calculated_each_time = self.single_buffer_size // self.dtype_sz
# declare I/O buffer
buffer_in0 = self.bp.Buffer(
shape=self.shape, name="INPUT0", dtype=self.dtype, scope="global"
shape=(self.length,), name="INPUT0", dtype=self.dtype, scope="global"
)
buffer_in1 = self.bp.Buffer(
shape=self.shape, name="INPUT1", dtype=self.dtype, scope="global"
shape=(self.length,), name="INPUT1", dtype=self.dtype, scope="global"
)
buffer_out = self.bp.Buffer(
shape=self.shape, name="OUTPUT", dtype=self.dtype, scope="global"
shape=(self.length,), name="OUTPUT", dtype=self.dtype, scope="global"
)
task_id = self.bp.taskId
# declare on-chip buffer
Expand Down Expand Up @@ -117,5 +108,5 @@ def compute_body(self):
def build_add(dtype=None, target=None):
# tasktype fixed in UNION1
task_num = 4
f = Add(SHAPE, dtype, target, task_num).compute_body()
f = Add(dtype, target, task_num).compute_body()
return f
11 changes: 8 additions & 3 deletions bangpy-ops/ops/add/test_add.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -28,11 +28,16 @@
from bangpy.common import utils, load_op_by_type
from bangpy.platform.bang_config import ALIGN_LENGTH, TARGET
from bangpy.tcp.runtime import TaskType
from add import DTYPES, SHAPE, KERNEL_NAME, TARGET_LIST
from add import DTYPES, KERNEL_NAME, TARGET_LIST


@pytest.mark.parametrize(
"shape", [SHAPE],
"shape",
[
(2048,),
(4096,),
(6144,),
],
)
@pytest.mark.parametrize(
"dtype", DTYPES,
Expand All @@ -54,5 +59,5 @@ def test_add(target, shape, dtype):
f1 = load_op_by_type(KERNEL_NAME, dtype.name)
f1(data_in0_dev, data_in1_dev, data_out_dev)
bangpy.assert_allclose(
data_out_dev.numpy(), data_out.astype(dtype.as_numpy_dtype), rtol=1e-3
data_out_dev.numpy(), data_out.astype(dtype.as_numpy_dtype)
)
2 changes: 1 addition & 1 deletion bangpy-ops/ops/nms/nms.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -266,7 +266,7 @@ def nms_compute_body(self, task_num):
with self.tcp.if_scope(self.tcp.taskId == task_num - 1):
self.tcp.assign(self.x1, 0)
self.tcp.memcpy(
self.output[self.output_box_num : self.max_output_size], self.x1
self.output[self.output_box_num : self.max_output_size], self.x1[:]
)

def nms_compute(self):
Expand Down
64 changes: 32 additions & 32 deletions docs/bangpy-docs/BANGPy-OPS算子设计文档模板.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -38,25 +38,25 @@

example:

| 算子功能简介 | 简要填写算子功能，详细描述在 1.2 中进行说明 |
| ---------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
| 需求来源 | PyTorch/Tensorflow/... |
| 应用网络 | resnet50/... |
| 输入数据类型 | half, float |
| 输入 Shape | input1: [batches, hi, wi, channels]; input2: [batches, 4] |
| 输入 Layout | input1: NHWC; input2: ARRAY |
| 输出数据类型 | half, float... |
| 输出 Shape | [batches, ho, wo, channels] |
| 输出 Layout | NHWC |
| 模式(可选） | |
| 是否含有 dim/axis 等类似语义的参数且该参数支持负数/其他特殊处理 | 有 dim/axis 参数且需要支持负数 / 不含带 dim/axis 语义的参数等<br>(例如 scatter 算子接口中的 dim 参数支持为负，当 dim=-1 时，实际在最低维上做计算) |
| 算子功能简介 | 简要填写算子功能，详细描述在 1.2 中进行说明 |
| ------------ | ----------------------------------------------------------- |
| 需求来源 | PyTorch/Tensorflow/... |
| 应用网络 | resnet50/... |
| 输入数据类型 | half, float |
| 输入 Shape | input1: [batches, hi, wi, channels]; input2: [batches, 4] |
| 输入 Layout | input1: NHWC; input2: ARRAY |
| 输出数据类型 | half, float... |
| 输出 Shape | [batches, ho, wo, channels] |
| 输出 Layout | NHWC |
| 模式(可选） | |
| 是否含有 dim/axis 等类似语义的参数且该参数支持负数/其他特殊处理 | 有 dim/axis 参数且需要支持负数 / 不含带 dim/axis 语义的参数等<br>(例如 scatter 算子接口中的 dim 参数支持为负，当 dim=-1 时，实际在最低维上做计算) |
| 是否含有 labels/index 等类似语义的参数且该参数支持负数/界外情况/其他特殊处理 | 有 labels 参数，labels 取值在 dim 范围之外时输出为 Nan / 有 index 参数，要求 index 支持负数 / 不含带 labels/index 语义的参数等<br>(例如 sparse_softmax_ce_logits 算子支持 label 取值在 dim 范围之外，advanced_index 算子支持负数 index 等) |
| 是否需要支持原位 | 是/否 |
| 是否需要支持 stride 机制 | 是/否 |
| 是否需要支持广播 | 是/否 (若是，列清楚具体哪些参数要支持) |
| 0 元素检查是否直接返回 | 是/否 |
| 其他特殊需求(在线量化，融合，转数提前等，可选) | |
| 本次开发优先支持的规模/模式 | 优先支持 xxx 模式/NHWC 的 layout |
| 是否需要支持原位 | 是/否 |
| 是否需要支持 stride 机制 | 是/否 |
| 是否需要支持广播 | 是/否 (若是，列清楚具体哪些参数要支持) |
| 0 元素检查是否直接返回 | 是/否 |
| 其他特殊需求(在线量化，融合，转数提前等，可选) | |
| 本次开发优先支持的规模/模式 | 优先支持 xxx 模式/NHWC 的 layout |

### 1.2 算子功能和应用场景描述

Expand Down Expand Up @@ -87,17 +87,17 @@ example:

example:

| 限制类型 | 详细说明 |
| ------------ | --------------------------------------------------------------------------------------------------------------- |
| 数据类型限制 | 不支持 input 和 output 同时为 half |
| 限制类型 | 详细说明 |
| ------------ | ------------------------------------------------------------------------------------------------- |
| 数据类型限制 | 不支持 input 和 output 同时为 half |
| 布局限制 | 不支持 NCHW 的 Layout, 需要明确感知 layout 的算子才需要说明限制, 对于不需要感知 layout 的算子, 不要加额外的防呆 |
| 规模限制 | output_size <= 384KB |
| 功能限制 | 不支持 xxx 模式/不支持 xx 和 xxx 的模式组合 |
| 数据范围限制 | half 类型下，数据需满足[xx, xx]范围，否则有精度问题 |
| 原位限制 | 不支持原位 |
| stride 限制 | 不支持 stride 机制 |
| 广播限制 | xxx 参数不支持广播 |
| xx 限制 | xxx |
| 规模限制 | output_size <= 384KB |
| 功能限制 | 不支持 xxx 模式/不支持 xx 和 xxx 的模式组合 |
| 数据范围限制 | half 类型下，数据需满足[xx, xx]范围，否则有精度问题 |
| 原位限制 | 不支持原位 |
| stride 限制 | 不支持 stride 机制 |
| 广播限制 | xxx 参数不支持广播 |
| xx 限制 | xxx |

### 1.5 验收标准

Expand Down Expand Up @@ -160,12 +160,12 @@ example:

1、资源分配

| 表项 | 分配策略 |
| --------------- | ---------------------------------------------------------------------- |
| 表项 | 分配策略 |
| --------------- | ------------------------------------------------------------------- |
| NRAM | 举例: 保存神经元数据 |
| WRAM | 举例: 保存权值数据 |
| SRAM | 举例：暂时存储 IO 的输入输出数据，权值最先 load 在 SM 上然后 broadcast |
| DRAM(workspace) | 举例：存储计算中的临时数据，比如 ci 拆分情况下的部分和 |
| SRAM | 举例：暂时存储 IO 的输入输出数据，权值最先 load 在 SM 上然后 broadcast |
| DRAM(workspace) | 举例：存储计算中的临时数据，比如 ci 拆分情况下的部分和 |

2、流水设计

Expand Down
Loading

0 comments on commit a2b1646

Please sign in to comment.