Commit Graph

7 Commits

Author SHA1 Message Date
PanZezhong1725 7f6aec6c17
针对bert和gpt2模型分布式推理的优化 (#221)
* fix(dist): 改善分布式脚本,只打印绝对误差

* feat(dist): 增加可导出onnx的pytorch运行脚本

* feat(front): 增加对Y值为-inf的where算子的图优化

* feat(kernel): 对b为常数的pow和div算子进行特判优化

* fix(front): 消除前端对global output形状信息的依赖,分布式脚本删除不必要的shape infer

* feat(kernel): 针对matmul中bias为行向量时的expand操作的特化优化

* fix(kernel): 删除div pow const中不必要的同步

* Update expand.cu

* fix: fix comments

---------

Co-authored-by: Haojie Wang <haojie0429@gmail.com>
Co-authored-by: Derui Yang <ydrml@hotmail.com>
2024-04-01 14:04:28 +08:00
Chenjie Duan 51086d2b8d
Modify kernel registration & support fp16 (#205)
* - Remove dataType from the kernel registration.

* - support fp16 for conv

* - cpu kernel: adapt the new registration mechanism

* modified all register kernel

* add where fp16

* add layernorm fp16

* add split_concat fp16

* - element_wise support fp16

* feat: support transpose fp16

* feat: support sliceOp fp16

* - unary support fp16

* - feat: support reduceOp fp16

* feat: support matmulOp/expandOp fp16

* feat: support powOp int8

* add cuda cast & support half-precision for gather

* style: fix style

* feat:support int8 for gather

* style:fix style

* modified test_cuda_conv_transposed

* fix: fix dist code to support fp16

* fix(graph.cc): fix topo_sort

* fix: fix recv and send kernel registration

* feat: add field tensors for stub

* refactor(frontend): 先排序后构图

Signed-off-by: YdrMaster <ydrml@hotmail.com>

* fix: 为中间结果提供tensor到node的mapping

* fix (slice): add guard for area out of range

* fix: fix matmul fp16

* fix: fix re-dataMalloc for weight tensor and use of naive allocator

* feat: add dataType filter for cuda kernel

* feat: bang kernel adapt the new registration mechanism

* fix: fix some error on mlu

* feat: intelcpu kernel adapt the new registration mechanism

* feat: modify kernel registration on kunlun

* fix intelcpu compiler bug

* feat: bang reshape support all dataType

* fix: fix bang reduce

* fix(all_reduce.cc): fix as reviewer suggessted

* fix: fix style and restore unary test codes

---------

Signed-off-by: YdrMaster <ydrml@hotmail.com>
Co-authored-by: xgqdut2016 <kenan_gewei@163.com>
Co-authored-by: xgqdut2016 <140036308+xgqdut2016@users.noreply.github.com>
Co-authored-by: zhangyunze <z13785159769@163.com>
Co-authored-by: OdinaryWord <sx-hz@163.com>
Co-authored-by: YdrMaster <ydrml@hotmail.com>
Co-authored-by: panzezhong <panzezhong@qiyuanlab.com>
2024-01-15 11:02:13 +08:00
zhangyunze 331f7ab2b8
support Dynamic tensor infer shape and fix memory pool (#176)
* feat: support dynamic tensor part1

* feat: support dynamic-tensor part2

* feat: support dynamic tensor part 3

* fix: fix some ..

* - add kvcache example

* feat: support concat to identity kernel

* add a simple mempory pool for allocator

* fix: rebase to master

* fix bug after merging

* - remove outdated script

* fix: fix as review

---------

Co-authored-by: kilinchange <kilinchange@163.com>
Co-authored-by: Haojie Wang <haojie0429@gmail.com>
2023-11-23 13:11:50 +08:00
YdrMaster 26f0d13c26
Dev for 202303ddl (#66)
* add activation operatiopn relu, tanh, sigmoid on mlu

* commit for format

* add activation backward operation

* add test for activation_backward

* add test

* add convbpfilter

* fix

* add transpsoe code and test

* add trigon function operation on mlu: sin,cos,tan,asin,sinh,asinh

* add copy operation on mlu

* add ceil operation and floor operation

* add operation clip

* add operation cnnl div, test and test for divdemo bangc kernel

* add divnonan operation and test

* add erf operation

* add exp operation

* add operation fill

* add log operation

* add log1p operation

* add l2loss operation

* add maximum and minimum operation

* add mseloss operation

* add negTensor operation

* add power operation

* add reciprocal operation

* add sqrt and rsqrt operation

* add transform operation

* add addn operation

* add muln operation

* cherrry pick some operation

* add floordiv operation and floordivtrunc operation

* add floormod operation

* add cumsum operation

* add det operation

* add pad operation

* format

* add concat operation

* format

* add split operation

* fix concat and split operation

* add round operation

* add pooling operation

* add square operation

* add squaredDifference operation

* code format fix

* add flip operation

* code format fix

* add hardtanh operation

* add logic operation

* add addcdiv and addcmul operation

* add arange operation

* add bitcompute operation

* add net test

* fmt

Signed-off-by: YdrMaster <ydrml@hotmail.com>

* style: rename

Signed-off-by: YdrMaster <ydrml@hotmail.com>

* fix: 用 NativeCpuRuntime 替换 CpuRuntime

Signed-off-by: YdrMaster <ydrml@hotmail.com>

* fix code

* fix code

* fix code by review suggestion

* remove operation which is not the onnx operation

* fix format

* clang format

* refactor: tensor 的 print 加一层模板的 dataToString

Signed-off-by: YdrMaster <ydrml@hotmail.com>

* fix: onnx 导出

Signed-off-by: YdrMaster <ydrml@hotmail.com>

* feat: 增加计算图优化接口

Signed-off-by: YdrMaster <ydrml@hotmail.com>

* add clip operation

* feat: 支持导入 clip

Signed-off-by: YdrMaster <ydrml@hotmail.com>

* test: 导入导出测试加入 ci

Signed-off-by: YdrMaster <ydrml@hotmail.com>

* fix batch norm

* feat: 增加 Shape 算子

Signed-off-by: YdrMaster <ydrml@hotmail.com>

* feat: 支持导入 unsqueeze

Signed-off-by: YdrMaster <ydrml@hotmail.com>

* fix: 修正 clip 接口

feat: 支持导入 transpose
Signed-off-by: YdrMaster <ydrml@hotmail.com>

* add broadcast operation

* fix elementwise-broadcast

* fix elementwise broadcast

* add broadcast for gpu elementsie

* feat: pad 支持 axes 负数

feat: 不支持的 padding 导出为独立的 pad 算子

feat: 支持导入 onnxsim 过的 inception
Signed-off-by: YdrMaster <ydrml@hotmail.com>

* fix: 修正池化的测试

Signed-off-by: YdrMaster <ydrml@hotmail.com>

* feat: 导出 pads,支持 inception 导入导出,已加入 ci

Signed-off-by: YdrMaster <ydrml@hotmail.com>

* feat: 支持 densenet 导入导出,并加入 ci

Signed-off-by: YdrMaster <ydrml@hotmail.com>

* feat: 导入 squeeze

Signed-off-by: YdrMaster <ydrml@hotmail.com>

* fix softmax

* feat: 导出 clip 和 transpose

Signed-off-by: YdrMaster <ydrml@hotmail.com>

* feat: 支持 Conv 的 bias

Signed-off-by: YdrMaster <ydrml@hotmail.com>

* fix: bias of conv

Signed-off-by: YdrMaster <ydrml@hotmail.com>

* fix: bias of conv

Signed-off-by: YdrMaster <ydrml@hotmail.com>

* feat: 导入 split

Signed-off-by: YdrMaster <ydrml@hotmail.com>

* feat: 导出 split

Signed-off-by: YdrMaster <ydrml@hotmail.com>

* fix: conv

Signed-off-by: YdrMaster <ydrml@hotmail.com>

* fix: conv group

Signed-off-by: YdrMaster <ydrml@hotmail.com>

* fix: matmul 的 bias 没有放在输入里,修正

Signed-off-by: YdrMaster <ydrml@hotmail.com>

* fix exmaple

* fix: 改正 reduce_mean 导出

Signed-off-by: YdrMaster <ydrml@hotmail.com>

* refactor: 修改 slice 实现与 onnx 一致

Signed-off-by: YdrMaster <ydrml@hotmail.com>

* style: 不导出两个 runtime 函数

Signed-off-by: YdrMaster <ydrml@hotmail.com>

* doc: 中文使用指南

Signed-off-by: YdrMaster <ydrml@hotmail.com>

* doc: 补全指南

Signed-off-by: YdrMaster <ydrml@hotmail.com>

* fix: 修复导入数据的问题

Signed-off-by: YdrMaster <ydrml@hotmail.com>

* fmt

Signed-off-by: YdrMaster <ydrml@hotmail.com>

* feat: 添加 Dropout 基本结构,但不支持两个输出是不同的类型

Signed-off-by: YdrMaster <ydrml@hotmail.com>

* feat: 重新导出优化接口

feat: dropout 导入
Signed-off-by: YdrMaster <ydrml@hotmail.com>

* build: BANG 选项加入 Makefile

Signed-off-by: YdrMaster <ydrml@hotmail.com>

* fxi code, change of test/kernels/bang/test* is use NativeCpuRuntime.
chaneg of include/bang/bang_runtime is for the cntoolkit upgrade.

* feat: 导出 bang runtime

Signed-off-by: YdrMaster <ydrml@hotmail.com>

* add USE_BANG=1

* fix matmul

* fix reshape

* fix

* fix activation

* fix transpose

* format

* format

* update Makefile

Signed-off-by: YdrMaster <ydrml@hotmail.com>

* feat: 支持导入导出 ConvTranspose

Signed-off-by: YdrMaster <ydrml@hotmail.com>

* add prelu on mlu

* fix: ConvTranspose

Signed-off-by: YdrMaster <ydrml@hotmail.com>

* feat: 支持导入导出 PRelu

Signed-off-by: YdrMaster <ydrml@hotmail.com>

* add convtrans on mlu

* fmt

Signed-off-by: YdrMaster <ydrml@hotmail.com>

* docs: 更新 README_CN.md

Signed-off-by: YdrMaster <ydrml@hotmail.com>

* fix code by review suggestions

* style

Signed-off-by: YdrMaster <ydrml@hotmail.com>

* fix: Softmax 的 axis 可以用默认值?感觉是 onnx 不标准

Signed-off-by: YdrMaster <ydrml@hotmail.com>

* fix cuda & intelcpu bugs after merging

---------

Signed-off-by: YdrMaster <ydrml@hotmail.com>
Co-authored-by: wanghailu <wanghailu0717@163.com>
Co-authored-by: wanghailu <wanghailu@qiyuanlab.com>
Co-authored-by: whjthu <haojie0429@gmail.com>
2023-04-18 15:10:33 +08:00
wendy12022 a4d6426589
ADD: batch norm operator and cuda kernel. (#44)
fix numInputs of batchNorm, add new line in file ending.

ADD: batch norm operator and cuda kernel.

add training

remove comments.

fix compile error.

add batch norm operator and cuda kernel.
2022-10-15 16:29:28 +08:00
wendy12022 3c6e208f42
ADD:concat/split operator and cuda kernels (#29)
* ADD:concat/split operator and cuda kernels

refector

minor change comment

ADD:concat/split operator and cuda kernels

merge split_kernel and concat_kernel to split_concat_kernel.

Revert "fix"

This reverts commit 459926be09a838658ec55f1e0a72b3cf17037d5c.

fix

ADD:concat/split operator and cuda kernels

change whole tensor name to composed tensor

fix some

remove unused header.

rebase

add CudaKernel

add test for split.

ADD split operator and cuda kernel.

modify test.

ADD:concat operator and cuda kernel.

ADD:concat/split operator and cuda kernels

fix some

remove unused header.

rebase

add CudaKernel

ADD:concat/split operator and cuda kernels

add test for split.

ADD split operator and cuda kernel.

modify test.

ADD:concat operator and cuda kernel.

* remove extra comment; typo fix.

Co-authored-by: Haojie Wang <haojie0429@gmail.com>
2022-09-29 11:01:30 +08:00
wendy12022 13b7a2604b
ADD add/mul/sub/div/pow operators and CPU/CUDA kernels (#26)
Fix some

remove useless code.

add div/pow kernel

Add add/mul/sub operators.

fix cpu kernel.

add element wise kenerl for cuda.

ADD element wise operator.
2022-09-09 13:43:59 +08:00