[x86] conv_depthwise_5x5 optimization #6979
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Collaborator
lxwlaq
commented
Sep 16, 2021
•
lxwlaq
commented
| conv | old | new | rate |
|---|---|---|---|
| 1×240×14×14 s1 | 0.3609 | 0.0802 | 300% |
| 1×96×28×28 s2 | 0.1745 | 0.04215 | 300% |
| 1×120×28×28 s1 | 0.7046 | 0.1082 | 600% |
| 1×72×56×56 s2 | 0.3437 | 0.0982 | 200% |
| model | old | new | rate |
|---|---|---|---|
| MobileNetV3 large | 17.49 | 12.48 | 28.6% |
| MobileNetV3 small | 9.32 | 5.34 | 42.7% |
…into depthwise_conv2d
|
Thanks for your contribution! |
| namespace math { | ||
| #define Min(a, b) (a < b ? a : b) | ||
| #define ROUNDUP(a, b) ((((a) + (b)-1) / (b)) * (b)) | ||
| void conv_depthwise_5x5s1s2(const float* din, |
There was a problem hiding this comment.
写成两个函数实现：conv_depthwise_5x5s1 和 conv_depthwise_5x5s2
| __m256 _bias = flag_bias ? _mm256_maskload_ps(bias + c, bias_mask) | ||
| : _mm256_set1_ps(0.f); | ||
|
|
||
| if (stride == 1) { |
There was a problem hiding this comment.
分开两个函数实现，减少if-else 判断
| impl_ = new DepthwiseConv<PRECISION(kFloat), PRECISION(kFloat)>; | ||
| } | ||
| if (dw_kernel && kps_equal && no_dilation && flag_dw && | ||
| (flag_dw_5x5 || paddings[0] == 1)) { |
There was a problem hiding this comment.
paddings 只支持1，是指3x3_dw 吗？我看原有实现是都支持呀
| @@ -89,11 +88,11 @@ void Conv2dCompute<PRECISION(kFloat), PRECISION(kFloat)>::PrepareForRun() { | |||
| const int iw = param.x->dims()[3]; | |||
|
|
|||
| //! select conv impl | |||
| if (this->device_ctx->avx_level() != AVXType::AVX_NONE) { | |||
lite/kernels/x86/conv_depthwise.cc
Outdated
| lite::x86::math::unpack8_m256(&output_pack_, param.output); | ||
| } else if (pack_size == 4) { | ||
| lite::x86::math::unpack4_m128(&output_pack_, param.output); | ||
| LOG(FATAL) << "weights scale size must equal to filter size"; |
There was a problem hiding this comment.
这个报错信息不对吧，应该是conv_depthwise 只支持3x3 5x5 吧
| for (auto stride : {1, 2}) { | ||
| for (auto pad : {0, 1}) { | ||
| for (auto bias : {false, true}) { | ||
| for (auto act : {"relu", "leaky_relu"}) { |
| r = _mm256_blendv_ps(r, | ||
| _mm256_mul_ps(negative_slope, r), | ||
| _mm256_cmp_ps(r, zero, 2)); | ||
| } else { |
| @@ -418,27 +417,32 @@ void TestConvDepthwise(Place place, float abs_error = 2e-5) { | |||
| // time-consuming | |||
| for (int64_t n : {1, 3, 4}) { | |||
| for (auto win : {3, 4, 7, 16, 30}) { | |||
| std::vector<int64_t> dims{n, 32, win, win}; | |||
There was a problem hiding this comment.
TestConvDepthwise 加入单测测试中，当前只有这个测试：
| @@ -74,6 +74,18 @@ void pack_padding8_m256(lite::Tensor* input, | |||
| const int channel_num, | |||
| const std::vector<int>& paddings); | |||
|
|
|||
| void packC8_with_Cleft(const float* din, | |||
There was a problem hiding this comment.
下个PR加上注视，介绍函数作用。类似：
| r3 = _mm256_blendv_ps(r3, | ||
| _mm256_mul_ps(negative_slope, r3), | ||
| _mm256_cmp_ps(r3, zero, 2)); | ||
| } else { |
There was a problem hiding this comment.
下个PR加上hard_siwsh 激活判断实现
| _mm256_mul_ps(negative_slope, r3), | ||
| _mm256_cmp_ps(r3, zero, 2)); | ||
| } else { | ||
| LOG(FATAL) << "[X86] activation type not supported"; |
There was a problem hiding this comment.
这句话修改下： LOG(FATAL) << "[X86] activation type" << static_cast(act_type) << "not supported";
| _mm256_mul_ps(negative_slope, r), | ||
| _mm256_cmp_ps(r, zero, 2)); | ||
| } else { | ||
| LOG(FATAL) << "[X86] activation type not supported"; |
|
下个PR完善下上述评论信息 |
lxwlaq
added a commit
to lxwlaq/Paddle-Lite
that referenced
this pull request
Nov 17, 2021
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.