From c67b92b149eb54252be3adb62bac3030b4879c01 Mon Sep 17 00:00:00 2001
From: HappyAngel <chenjiaobuaa@126.com>
Date: Mon, 29 Jun 2020 21:40:30 -0500
Subject: [PATCH] [arm] update con int8 kernel choose (#3834)

* fix conv int8 kernel choose and sooftmax compute bug

* change axis_size = 4 kernel choose, test=develop

* fix format. test=develop
---
 lite/backends/arm/math/softmax.cc   |  4 ++--
 lite/kernels/arm/conv_compute.cc    | 12 ++++++++++--
 lite/kernels/arm/conv_winograd.cc   |  3 +++
 lite/kernels/arm/conv_winograd.h    |  7 +++++++
 lite/kernels/arm/softmax_compute.cc |  2 +-
 5 files changed, 23 insertions(+), 5 deletions(-)

diff --git a/lite/backends/arm/math/softmax.cc b/lite/backends/arm/math/softmax.cc
index 65d41b0491..b7f82e9f37 100644
--- a/lite/backends/arm/math/softmax.cc
+++ b/lite/backends/arm/math/softmax.cc
@@ -531,7 +531,7 @@ void softmax_inner1_large_axis<float>(const float* din,
     }
     float32x2_t vhmax = vmax_f32(vget_high_f32(vmax), vget_low_f32(vmax));
     float max_data = std::max(vget_lane_f32(vhmax, 0), vget_lane_f32(vhmax, 1));
-    for (j = 4 * j; j < axis_size; ++j) {
+    for (j = 4 * nn; j < axis_size; ++j) {
       max_data = std::max(max_data, din_max_ptr[0]);
       din_max_ptr++;
     }
@@ -557,7 +557,7 @@ void softmax_inner1_large_axis<float>(const float* din,
     float32x2_t vhsum = vadd_f32(vget_high_f32(vsum), vget_low_f32(vsum));
     float sum_data = vget_lane_f32(vhsum, 0) + vget_lane_f32(vhsum, 1);
 
-    for (j = 4 * j; j < axis_size; ++j) {
+    for (j = 4 * nn; j < axis_size; ++j) {
       dout_sum_ptr[0] = expf(din_sum_ptr[0] - max_data);
       sum_data += dout_sum_ptr[0];
       din_sum_ptr++;
diff --git a/lite/kernels/arm/conv_compute.cc b/lite/kernels/arm/conv_compute.cc
index ef174814ce..54e67de5ab 100644
--- a/lite/kernels/arm/conv_compute.cc
+++ b/lite/kernels/arm/conv_compute.cc
@@ -121,10 +121,14 @@ void ConvCompute<PRECISION(kInt8), PRECISION(kFloat)>::PrepareForRun() {
       no_dilation && flag_dw) {
     impl_ = new DepthwiseConv<PRECISION(kInt8), PRECISION(kFloat)>;
     // VLOG(3) << "Run DepthwiseConv Int8";
+  } else if (param.groups == 1 && kw == 3 && sw == 2 && no_dilation &&
+             pads_equal) {
+    impl_ = new DirectConv<PRECISION(kInt8), PRECISION(kFloat)>;
+    // VLOG(3) << "Run DirectConv Int8";
   } else if (param.groups == 1 && kw == 3 && sw == 1 && no_dilation &&
              pads_equal) {
     impl_ = new WinogradConv<PRECISION(kInt8), PRECISION(kFloat)>;
-    // VLOG(3) << "Run DirectConv Int8";
+    // VLOG(3) << "Run WinogradConv Int8";
   } else {
     impl_ = new GemmLikeConv<PRECISION(kInt8), PRECISION(kFloat)>;
     // VLOG(3) << "Run GemmLikeConvInt8";
@@ -168,10 +172,14 @@ void ConvCompute<PRECISION(kInt8), PRECISION(kInt8)>::PrepareForRun() {
       no_dilation && flag_dw) {
     impl_ = new DepthwiseConv<PRECISION(kInt8), PRECISION(kInt8)>;
     // VLOG(3) << "Run DepthwiseConv Int8";
+  } else if (param.groups == 1 && kw == 3 && sw == 2 && no_dilation &&
+             pads_equal) {
+    impl_ = new DirectConv<PRECISION(kInt8), PRECISION(kInt8)>;
+    // VLOG(3) << "Run DirectConv Int8";
   } else if (param.groups == 1 && kw == 3 && sw == 1 && no_dilation &&
              pads_equal) {
     impl_ = new WinogradConv<PRECISION(kInt8), PRECISION(kInt8)>;
-    // VLOG(3) << "Run DirectConv Int8";
+    // VLOG(3) << "Run WinogradConv Int8";
   } else {
     impl_ = new GemmLikeConv<PRECISION(kInt8), PRECISION(kInt8)>;
     // VLOG(3) << "Run GemmLikeConvInt8";
diff --git a/lite/kernels/arm/conv_winograd.cc b/lite/kernels/arm/conv_winograd.cc
index c6e06a243c..f61c6109cd 100644
--- a/lite/kernels/arm/conv_winograd.cc
+++ b/lite/kernels/arm/conv_winograd.cc
@@ -358,6 +358,9 @@ void WinogradConv<PRECISION(kInt8), OutType>::Run() {
                                                       param,
                                                       &ctx);
   }
+#ifdef LITE_WITH_PROFILE
+  kernel_func_name_ = "conv_compute_2x2_3x3_int8";
+#endif
 }
 template class WinogradConv<PRECISION(kInt8), PRECISION(kInt8)>;
 template class WinogradConv<PRECISION(kInt8), PRECISION(kFloat)>;
diff --git a/lite/kernels/arm/conv_winograd.h b/lite/kernels/arm/conv_winograd.h
index 69835a74b4..b93a719f7d 100644
--- a/lite/kernels/arm/conv_winograd.h
+++ b/lite/kernels/arm/conv_winograd.h
@@ -61,6 +61,13 @@ class WinogradConv<PRECISION(kInt8), OutType>
   virtual void PrepareForRun();
   virtual void ReInitWhenNeeded();
   virtual void Run();
+#ifdef LITE_WITH_PROFILE
+  virtual void SetProfileRuntimeKernelInfo(
+      paddle::lite::profile::OpCharacter* ch) {
+    ch->kernel_func_name = kernel_func_name_;
+  }
+  std::string kernel_func_name_{"NotImplForConvWino"};
+#endif
 
  protected:
   using param_t = operators::ConvParam;
diff --git a/lite/kernels/arm/softmax_compute.cc b/lite/kernels/arm/softmax_compute.cc
index 3409d0f5c5..79ea23ab3f 100644
--- a/lite/kernels/arm/softmax_compute.cc
+++ b/lite/kernels/arm/softmax_compute.cc
@@ -34,7 +34,7 @@ void SoftmaxCompute::Run() {
   int inner_num = x_dims.Slice(axis + 1, x_rank).production();
   int axis_size = x_dims[axis];
   if (inner_num == 1) {
-    if (axis_size >= 4) {
+    if (axis_size > 4) {
       lite::arm::math::softmax_inner1_large_axis(
           din, dout, outer_num, axis_size);
     } else {
-- 
GitLab