[arm] fix iOS compiling error (#5040)

PaddlePaddle · Dec 24, 2020 · def0352 · def0352
1 parent 70b7d9b
commit def0352
Show file tree

Hide file tree

Showing 4 changed files with 24 additions and 8 deletions.
diff --git a/lite/backends/arm/math/gemm_prepacked_int8.cc b/lite/backends/arm/math/gemm_prepacked_int8.cc
@@ -5710,8 +5710,9 @@ void gemm_prepack_int8(const int8_t* A_packed,
       alpha[3] = local_alpha;
     }
   }
-#if defined(__aarch64__) && defined(WITH_ARM_DOTPROD)
+#ifdef __aarch64__
   if (ctx->has_dot()) {
+#ifdef WITH_ARM_DOTPROD
     gemm_prepack_sdot_int8<int32_t>(A_packed,
                                     B,
                                     bias,
@@ -5725,6 +5726,7 @@ void gemm_prepack_int8(const int8_t* A_packed,
                                     scale,
                                     alpha,
                                     ctx);
+#endif
   } else {
     gemm_prepack_oth_int8<int32_t>(A_packed,
                                    B,
@@ -5741,7 +5743,10 @@ void gemm_prepack_int8(const int8_t* A_packed,
                                    ctx);
   }
 #else
-  if (ctx->has_dot()) {
+  if (0 && ctx->has_dot()) {
+#ifdef WITH_ARM_DOTPROD
+    // ios build error : no template named 'gemm_prepack_vsdot_int8'
+    // android is ok
     gemm_prepack_vsdot_int8<int32_t>(A_packed,
                                      B,
                                      bias,
@@ -5755,6 +5760,7 @@ void gemm_prepack_int8(const int8_t* A_packed,
                                      scale,
                                      alpha,
                                      ctx);
+#endif
   } else {
     gemm_prepack_oth_int8<int32_t>(A_packed,
                                    B,

diff --git a/lite/core/mir/fusion/quant_dequant_op_fuser.cc b/lite/core/mir/fusion/quant_dequant_op_fuser.cc
@@ -179,14 +179,23 @@ void DequantOpFuser::InsertNewNode(SSAGraph* graph,
   std::vector<float> weight_scale;
   int weight_scale_size = 0;
   if (quantized_op_type_ == "conv2d" ||
-      quantized_op_type_ == "depthwise_conv2d" ||
-      quantized_op_type_ == "conv2d_transpose") {
+      quantized_op_type_ == "depthwise_conv2d") {
     op_desc.SetInput("Input", {quantized_op_input->arg()->name});
     op_desc.SetOutput("Output", {dequant_op_out->arg()->name});
     // Conv weight shape: Cout * Cin/group * kh * hw, the weight_scale_size
     // should
     // be Cout.
     weight_scale_size = quantized_weight_t->dims()[0];
+  } else if (quantized_op_type_ == "conv2d_transpose") {
+    op_desc.SetInput("Input", {quantized_op_input->arg()->name});
+    op_desc.SetOutput("Output", {dequant_op_out->arg()->name});
+
+    auto* conv_op_desc = matched.at("quantized_op")->stmt()->op_info();
+    auto groups = conv_op_desc->GetAttr<int>("groups");
+    // Conv weight shape: Cin * Cout/group * kh * hw, the weight_scale_size
+    // should
+    // be Cout.
+    weight_scale_size = quantized_weight_t->dims()[1] * groups;
   } else if (quantized_op_type_ == "mul" || quantized_op_type_ == "matmul") {
     op_desc.SetInput("X", {quantized_op_input->arg()->name});
     op_desc.SetOutput("Out", {dequant_op_out->arg()->name});

diff --git a/lite/kernels/arm/conv_transpose_compute.cc b/lite/kernels/arm/conv_transpose_compute.cc
@@ -571,6 +571,7 @@ void Conv2DTransposeCompute<PRECISION(kInt8), PRECISION(kInt8)>::Run() {
                                                  &act_param);
   }
 }
+
 }  // namespace arm
 }  // namespace kernels
 }  // namespace lite

diff --git a/lite/tests/cv/image_profiler_test.cc b/lite/tests/cv/image_profiler_test.cc
@@ -363,8 +363,8 @@ void test_flip(const std::vector<int>& cluster_id,
       } else if (srcFormat == ImageFormat::GRAY) {
         size = srch * srcw;
       }
-      uint8_t* src = new uint8_t[size];
-      fill_tensor_host_rand(src, size);
+      uint8_t* src = new uint8_t[2 * size];
+      fill_tensor_host_rand(src, 2 * size);
 
       int out_size = srch * srcw;
       if (dstFormat == ImageFormat::NV12 || dstFormat == ImageFormat::NV21) {
@@ -378,8 +378,8 @@ void test_flip(const std::vector<int>& cluster_id,
       } else if (dstFormat == ImageFormat::GRAY) {
         out_size = srch * srcw;
       }
-      uint8_t* basic_dst = new uint8_t[out_size];
-      uint8_t* lite_dst = new uint8_t[out_size];
+      uint8_t* basic_dst = new uint8_t[2 * out_size];
+      uint8_t* lite_dst = new uint8_t[2 * out_size];
       LOG(INFO) << "basic flip compute";
       Timer t_basic, t_lite;
       for (int i = 0; i < test_iter; i++) {