Merge branch 'remove_eigen_deps' into 'incubate/lite'

[LITE][ARM] Remove Eigen deps of ARM

See merge request inference/paddlelite!87

paddle/fluid/lite/arm/math/CMakeLists.txt

@@ -78,5 +78,5 @@ if (NOT HAS_ARM_MATH_LIB_DIR)
       gemv_arm_int8.cc
       conv3x3s1_direct_int8.cc
       conv3x3s2_direct_int8.cc
-      DEPS ${lite_kernel_deps} eigen3 framework_proto_lite)
+      DEPS ${lite_kernel_deps} framework_proto_lite)
 endif()

paddle/fluid/lite/core/CMakeLists.txt

@@ -29,7 +29,12 @@ endif()
 lite_cc_library(op_registry_lite SRCS op_registry.cc DEPS framework_proto_lite)
 lite_cc_library(scope_lite SRCS scope.cc DEPS ${tensor_lite})
 lite_cc_library(cpu_info_lite SRCS cpu_info.cc)
+
+if (LITE_WITH_ARM)
+lite_cc_library(context_lite SRCS context.cc DEPS ${tensor_lite} any_lite cpu_info_lite CL_DEPS cl_helper)
+else()
 lite_cc_library(context_lite SRCS context.cc DEPS ${tensor_lite} any_lite cpu_info_lite eigen3 CL_DEPS cl_helper)
+endif()
 lite_cc_library(kernel_lite SRCS kernel.cc DEPS context_lite type_system target_wrapper_lite any_lite op_params_lite framework_proto_lite ${tensor_lite})
 lite_cc_library(op_lite SRCS op_lite.cc DEPS scope_lite op_registry_lite target_wrapper_lite kernel_lite
   cpp_op_desc_lite ${tensor_lite})

paddle/fluid/lite/kernels/arm/fc_compute_test.cc

@@ -28,6 +28,35 @@ namespace lite {
 namespace kernels {
 namespace arm {
 
+#define A(i, j) a[i * lda + j]
+#define B(i, j) b[i * ldb + j]
+#define C(i, j) c[i * ldc + j]
+
+template <typename T>
+void gemm_bias(const T* a, const int M, const int K, const T* b, const int K_,
+               const int N, T* biases, T* c) {
+  EXPECT_TRUE(K_ == K && M > 0 && N > 0 && K > 0);
+  EXPECT_TRUE(a && b && c);
+  const int lda = K;
+  const int ldb = N;
+  const int ldc = N;
+  for (int m = 0; m < M; ++m) {
+    for (int n = 0; n < N; ++n) {
+      C(m, n) = 0.0f;
+      for (int k = 0; k < K; ++k) {
+        C(m, n) += A(m, k) * B(k, n);
+      }
+    }
+  }
+  if (biases) {
+    for (int m = 0; m < M; ++m) {
+      for (int n = 0; n < N; ++n) {
+        C(m, n) += biases[n];
+      }
+    }
+  }
+}
+
 template <typename T>
 void FillData(T* a, const int n, const T lower = static_cast<T>(-2.f),
               const T upper = static_cast<T>(2.f)) {
@@ -103,8 +132,8 @@ TEST(fc_arm, compare_test) {
           fc.PrepareForRun();
           fc.Run();
 
-          lite::arm::math::fc_compute_eigen(x_data, m, k, w_data, k, n, b_data,
-                                            ref_data);
+          gemm_bias<T>(x_data, m, k, w_data, k, n, b_data, ref_data);
+
           for (int i = 0; i < out.dims().production(); i++) {
             EXPECT_NEAR(out_data[i], ref_data[i], 1e-3);
           }
@@ -158,8 +187,8 @@ TEST(fc_arm, num_col_dims) {
     fc.PrepareForRun();
     fc.Run();
 
-    lite::arm::math::fc_compute_eigen(x_data, 2, 3, w_data, 3, 4, b_data,
-                                      ref_data);
+    gemm_bias<T>(x_data, 2, 3, w_data, 3, 4, b_data, ref_data);
+
     for (int i = 0; i < out.dims().production(); i++) {
       EXPECT_NEAR(out_data[i], ref_data[i], 1e-3);
     }

paddle/fluid/lite/kernels/arm/mul_compute_test.cc

@@ -28,6 +28,28 @@ namespace lite {
 namespace kernels {
 namespace arm {
 
+#define A(i, j) a[i * lda + j]
+#define B(i, j) b[i * ldb + j]
+#define C(i, j) c[i * ldc + j]
+
+template <typename T>
+void mul_gemm(const T* a, const int M, const int K, const T* b, const int K_,
+              const int N, T* c) {
+  EXPECT_TRUE(K_ == K && M > 0 && N > 0 && K > 0);
+  EXPECT_TRUE(a && b && c);
+  const int lda = K;
+  const int ldb = N;
+  const int ldc = N;
+  for (int m = 0; m < M; ++m) {
+    for (int n = 0; n < N; ++n) {
+      C(m, n) = 0.0f;
+      for (int k = 0; k < K; ++k) {
+        C(m, n) += A(m, k) * B(k, n);
+      }
+    }
+  }
+}
+
 template <typename T>
 void FillData(T* a, const int n, const T lower = static_cast<T>(-2.f),
               const T upper = static_cast<T>(2.f)) {
@@ -91,8 +113,8 @@ TEST(mul_arm, compare_test) {
 
         mul.Run();
 
-        lite::arm::math::mul_compute_eigen(x_data, m, k, y_data, k, n,
-                                           ref_data);
+        mul_gemm<T>(x_data, m, k, y_data, k, n, ref_data);
+
         for (int i = 0; i < out.dims().production(); i++) {
           EXPECT_NEAR(out_data[i], ref_data[i], 1e-3);
         }
@@ -138,7 +160,8 @@ TEST(mul_arm, num_col_dims) {
 
   mul.Run();
 
-  lite::arm::math::mul_compute_eigen(x_data, 2, 12, y_data, 12, 5, ref_data);
+  mul_gemm<T>(x_data, 2, 12, y_data, 12, 5, ref_data);
+
   for (int i = 0; i < out.dims().production(); i++) {
     EXPECT_NEAR(out_data[i], ref_data[i], 1e-3);
   }