From c2be5da380fbb440d7e80bb57d1ffb3f769d2bd6 Mon Sep 17 00:00:00 2001
From: yuanshuai <ysh329@sina.com>
Date: Mon, 24 Jun 2019 11:05:39 +0000
Subject: [PATCH] [LITE][CI] Enable android armv7 CI, Fix mul, fc bug

---
 paddle/fluid/lite/kernels/arm/fc_compute.cc  |  8 +++++-
 paddle/fluid/lite/kernels/arm/mul_compute.cc | 28 +++++++++++---------
 paddle/fluid/lite/kernels/arm/mul_compute.h  |  3 +++
 paddle/fluid/lite/tools/build.sh             |  9 +------
 4 files changed, 27 insertions(+), 21 deletions(-)

diff --git a/paddle/fluid/lite/kernels/arm/fc_compute.cc b/paddle/fluid/lite/kernels/arm/fc_compute.cc
index 2e6f46a0e07..24619ed9261 100644
--- a/paddle/fluid/lite/kernels/arm/fc_compute.cc
+++ b/paddle/fluid/lite/kernels/arm/fc_compute.cc
@@ -13,10 +13,10 @@
 // limitations under the License.
 
 #include "paddle/fluid/lite/kernels/arm/fc_compute.h"
+#include <vector>
 #include "paddle/fluid/lite/arm/math/funcs.h"
 #include "paddle/fluid/lite/core/op_registry.h"
 #include "paddle/fluid/lite/core/type_system.h"
-
 namespace paddle {
 namespace lite {
 namespace kernels {
@@ -53,6 +53,12 @@ void FcCompute::PrepareForRun() {
       }
     }
   }
+
+  if (m_ > 1) {
+    int hblock = lite::arm::math::get_hblock(ctx.arch());
+    int m_round = hblock * ((m_ + hblock - 1) / hblock);
+    ctx.ExtendWorkspace(DDimLite(std::vector<int64_t>({m_round * k_})));
+  }
 }
 
 void FcCompute::Run() {
diff --git a/paddle/fluid/lite/kernels/arm/mul_compute.cc b/paddle/fluid/lite/kernels/arm/mul_compute.cc
index 57c28e63bbf..c721e8046e7 100644
--- a/paddle/fluid/lite/kernels/arm/mul_compute.cc
+++ b/paddle/fluid/lite/kernels/arm/mul_compute.cc
@@ -13,6 +13,7 @@
 // limitations under the License.
 
 #include "paddle/fluid/lite/kernels/arm/mul_compute.h"
+#include <vector>
 #include "paddle/fluid/lite/arm/math/funcs.h"
 #include "paddle/fluid/lite/core/op_registry.h"
 #include "paddle/fluid/lite/core/type_system.h"
@@ -33,7 +34,7 @@ void MulCompute::Run() {
   const auto* y_data = param.y->data<float>();
   auto* o_data = param.output->mutable_data<float>();
 
-  int m = static_cast<int>(
+  m_ = static_cast<int>(
       param.x->dims().Slice(0, param.x_num_col_dims).production());
   int x_w =
       static_cast<int>(param.x->dims()
@@ -41,26 +42,29 @@ void MulCompute::Run() {
                            .production());
   int y_h = static_cast<int>(
       param.y->dims().Slice(0, param.y_num_col_dims).production());
-  int n =
-      static_cast<int>(param.y->dims()
-                           .Slice(param.y_num_col_dims, param.y->dims().size())
-                           .production());
+  n_ = static_cast<int>(param.y->dims()
+                            .Slice(param.y_num_col_dims, param.y->dims().size())
+                            .production());
 
   CHECK_EQ(x_w, y_h) << "x_w must be equal with y_h";
-  auto k = x_w;
-  if (n == 1) {
-    lite::arm::math::sgemv(x_data, y_data, o_data, false, m, k, false, nullptr,
-                           false);
+  k_ = x_w;
+
+  if (n_ == 1) {
+    lite::arm::math::sgemv(x_data, y_data, o_data, false, m_, k_, false,
+                           nullptr, false);
 
   } else {
     constexpr bool is_tranposed_y = false;
     auto& ctx = this->ctx_->template As<ARMContext>();
+    int hblock = lite::arm::math::get_hblock(ctx.arch());
+    int m_round = hblock * ((m_ + hblock - 1) / hblock);
+    ctx.ExtendWorkspace(DDimLite(std::vector<int64_t>({m_round * k_})));
 
     float* packed_x = static_cast<float*>(ctx.workspace_data<float>()) +
                       ctx.l2_cache_size() / sizeof(float);
-    lite::arm::math::prepackA(packed_x, x_data, k, 0, m, 0, k, false, &ctx);
-    lite::arm::math::sgemm_prepack(packed_x, y_data, nullptr, o_data, m, n, k,
-                                   false, false, is_tranposed_y, &ctx);
+    lite::arm::math::prepackA(packed_x, x_data, k_, 0, m_, 0, k_, false, &ctx);
+    lite::arm::math::sgemm_prepack(packed_x, y_data, nullptr, o_data, m_, n_,
+                                   k_, false, false, is_tranposed_y, &ctx);
   }
 }
 
diff --git a/paddle/fluid/lite/kernels/arm/mul_compute.h b/paddle/fluid/lite/kernels/arm/mul_compute.h
index c18995e5a5c..64c8f813d4e 100644
--- a/paddle/fluid/lite/kernels/arm/mul_compute.h
+++ b/paddle/fluid/lite/kernels/arm/mul_compute.h
@@ -31,6 +31,9 @@ class MulCompute : public KernelLite<TARGET(kARM), PRECISION(kFloat)> {
   void Run() override;
 
   virtual ~MulCompute() = default;
+
+ private:
+  int m_, n_, k_;
 };
 
 }  // namespace arm
diff --git a/paddle/fluid/lite/tools/build.sh b/paddle/fluid/lite/tools/build.sh
index a88e7786c01..49dff53bfa5 100755
--- a/paddle/fluid/lite/tools/build.sh
+++ b/paddle/fluid/lite/tools/build.sh
@@ -246,14 +246,7 @@ function test_arm {
         echo "android do not need armv7hf"
         return 0
     fi
-
-    # TODO(yuanshuai): enable armv7 on android
-    if [[ ${abi} == "armv7" ]]; then
-        echo "skip android v7 test yet"
-        return 0
-    fi
-
-
+   
     echo "test file: ${TESTS_FILE}"
     for _test in $(cat $TESTS_FILE); do
         test_arm_android $_test $port
-- 
GitLab