fix fp16 bug

mace/core/tensor.h

@@ -46,7 +46,7 @@ namespace mace {
     break;                            \
   }
 
-#if defined(MACE_ENABLE_NEON) && defined(__ANDROID__) && defined(ANDROID)
+#if defined(MACE_ENABLE_NEON) && defined(__ANDROID__)
 #define MACE_TYPE_ENUM_SWITCH_CASE_NEON(STATEMENTS)             \
   MACE_CASE(float16_t, MACE_SINGLE_ARG(STATEMENTS))
 #else

mace/core/types.cc

@@ -47,7 +47,7 @@ size_t GetEnumTypeSize(const DataType dt) {
       return sizeof(float);
     case DT_HALF:
       return sizeof(half);
-#if defined(MACE_ENABLE_NEON) && defined(__ANDROID__) && defined(ANDROID)
+#if defined(MACE_ENABLE_NEON) && defined(__ANDROID__)
     case DT_FLOAT16:
       return sizeof(float16_t);
 #endif

mace/core/types.h

@@ -17,7 +17,7 @@
 
 #include <cstdint>
 #include <string>
-#if defined(MACE_ENABLE_NEON) && defined(__ANDROID__) && defined(ANDROID)
+#if defined(MACE_ENABLE_NEON) && defined(__ANDROID__)
 #include <arm_neon.h>
 #endif
 
@@ -54,7 +54,7 @@ struct EnumToDataType;
   };
 
 MACE_MAPPING_DATA_TYPE_AND_ENUM(half, DT_HALF);
-#if defined(MACE_ENABLE_NEON) && defined(__ANDROID__) && defined(ANDROID)
+#if defined(MACE_ENABLE_NEON) && defined(__ANDROID__)
 MACE_MAPPING_DATA_TYPE_AND_ENUM(float16_t, DT_FLOAT16);
 #endif
 MACE_MAPPING_DATA_TYPE_AND_ENUM(float, DT_FLOAT);

mace/ops/BUILD.bazel

@@ -38,7 +38,7 @@ cc_library(
     ]) + if_neon_enabled([
         "-DMACE_ENABLE_NEON",
     ]) + if_android_armv7([
-        "-mfpu=neon",
+        "-mfpu=neon-fp16",
         "-mfloat-abi=softfp",
     ]) + if_opencl_enabled([
         "-DMACE_ENABLE_OPENCL",
@@ -61,7 +61,20 @@ cc_library(
         "-Werror",
         "-Wextra",
         "-Wno-missing-field-initializers",
-    ],
+    ] + if_openmp_enabled([
+        "-fopenmp",
+    ]) + if_neon_enabled([
+        "-DMACE_ENABLE_NEON",
+    ]) + if_android_armv7([
+        "-mfpu=neon-fp16",
+        "-mfloat-abi=softfp",
+    ]) + if_opencl_enabled([
+        "-DMACE_ENABLE_OPENCL",
+    ]) + if_quantize_enabled([
+        "-DMACE_ENABLE_QUANTIZE",
+    ]) + if_hexagon_enabled([
+        "-DMACE_ENABLE_HEXAGON",
+    ]),
     deps = [
         "//mace/core",
         "@gtest",
@@ -89,7 +102,7 @@ cc_library(
     ]) + if_neon_enabled([
         "-DMACE_ENABLE_NEON",
     ]) + if_android_armv7([
-        "-mfpu=neon",
+        "-mfpu=neon-fp16",
         "-mfloat-abi=softfp",
     ]) + if_opencl_enabled([
         "-DMACE_ENABLE_OPENCL",
@@ -141,7 +154,7 @@ cc_library(
     ]) + if_neon_enabled([
         "-DMACE_ENABLE_NEON",
     ]) + if_android_armv7([
-        "-mfpu=neon",
+        "-mfpu=neon-fp16",
         "-mfloat-abi=softfp",
     ]) + if_opencl_enabled([
         "-DMACE_ENABLE_OPENCL",
@@ -186,7 +199,7 @@ cc_library(
     ]) + if_neon_enabled([
         "-DMACE_ENABLE_NEON",
     ]) + if_android_armv7([
-        "-mfpu=neon",
+        "-mfpu=neon-fp16",
         "-mfloat-abi=softfp",
     ]) + if_opencl_enabled([
         "-DMACE_ENABLE_OPENCL",
@@ -221,7 +234,7 @@ cc_library(
     ]) + if_neon_enabled([
         "-DMACE_ENABLE_NEON",
     ]) + if_android_armv7([
-        "-mfpu=neon",
+        "-mfpu=neon-fp16",
         "-mfloat-abi=softfp",
     ]) + if_opencl_enabled([
         "-DMACE_ENABLE_OPENCL",
@@ -256,7 +269,7 @@ cc_library(
     ]) + if_neon_enabled([
         "-DMACE_ENABLE_NEON",
     ]) + if_android_armv7([
-        "-mfpu=neon",
+        "-mfpu=neon-fp16",
         "-mfloat-abi=softfp",
     ]) + if_opencl_enabled([
         "-DMACE_ENABLE_OPENCL",
@@ -325,7 +338,6 @@ cc_library(
         "-DMACE_ENABLE_NEON",
     ]) + if_android_armv7([
         "-mfpu=neon-fp16",
-    ]) + if_android_armv7([
         "-mfloat-abi=softfp",
     ]) + if_opencl_enabled([
         "-DMACE_ENABLE_OPENCL",
@@ -370,7 +382,6 @@ cc_library(
         "-DMACE_ENABLE_NEON",
     ]) + if_android_armv7([
         "-mfpu=neon-fp16",
-    ]) + if_android_armv7([
         "-mfloat-abi=softfp",
     ]) + if_opencl_enabled([
         "-DMACE_ENABLE_OPENCL",

mace/ops/arm/fp16_gemm.h

@@ -17,15 +17,13 @@
 
 #include "mace/core/types.h"
 
-#if defined(MACE_ENABLE_NEON) && defined(__ANDROID__) && defined(ANDROID)
+#if defined(MACE_ENABLE_NEON) && defined(__ANDROID__)
 #include <arm_neon.h>
 #endif
 
-#if defined(MACE_ENABLE_NEON) && !defined(__aarch64__)
-#if defined(__ANDROID__) && defined(ANDROID)
+#if defined(MACE_ENABLE_NEON) && !defined(__aarch64__) && defined(__ANDROID__)
 #define vaddvq_f32(v) ((v)[0] + (v)[1] + (v)[2] + (v)[3])
 #endif
-#endif
 
 namespace mace {
 namespace ops {
@@ -39,7 +37,7 @@ void FP16Gemv(const INPUT_TYPE_LEFT *m_ptr,
               const index_t width,
               OUTPUT_TYPE *result);
 
-#if defined(MACE_ENABLE_NEON) && defined(__ANDROID__) && defined(ANDROID)
+#if defined(MACE_ENABLE_NEON) && defined(__ANDROID__)
 template<>
 void FP16Gemv<float16_t, float, float>(const float16_t *m_ptr,
                                        const float *v_ptr,

mace/ops/matmul.cc

@@ -513,7 +513,7 @@ class MatMulOp<DeviceType::GPU, T> : public MatMulOpBase {
 };
 #endif  // MACE_ENABLE_OPENCL
 
-#if defined(MACE_ENABLE_NEON) && defined(__ANDROID__) && defined(ANDROID)
+#if defined(MACE_ENABLE_NEON) && defined(__ANDROID__)
 template <>
 class MatMulOp<CPU, float16_t> : public MatMulOpBase {
  public:
@@ -610,7 +610,7 @@ void RegisterMatMul(OpRegistryBase *op_registry) {
                    DeviceType::GPU, half);
 #endif  // MACE_ENABLE_OPENCL
 
-#if defined(MACE_ENABLE_NEON) && defined(__ANDROID__) && defined(ANDROID)
+#if defined(MACE_ENABLE_NEON) && defined(__ANDROID__)
   MACE_REGISTER_OP(op_registry, "MatMul", MatMulOp,
                    DeviceType::CPU, float16_t);
 #endif  // MACE_ENABLE_NEON

mace/ops/matmul_benchmark.cc

@@ -263,13 +263,17 @@ void MatMulBenchmark(
   OpsTestNet net;
 
   // Add input data
+#if defined(MACE_ENABLE_NEON) && defined(__ANDROID__)
   if (DataTypeToEnum<T>::value == DT_FLOAT16) {
     net.AddRandomInput<D, float16_t>("A", {batch, height, channels});
     net.AddRandomInput<D, float>("B", {batch, channels, out_width});
   } else {
+#endif
     net.AddRandomInput<D, T>("A", {batch, height, channels});
     net.AddRandomInput<D, T>("B", {batch, channels, out_width});
+#if defined(MACE_ENABLE_NEON) && defined(__ANDROID__)
   }
+#endif
   net.GetTensor("A")->SetIsWeight(true);
   net.GetTensor("B")->SetIsWeight(true);
   if (DataTypeToEnum<T>::value == DT_UINT8) {
@@ -310,13 +314,17 @@ void MatMulTransposeBenchmark(
   OpsTestNet net;
 
   // Add input data
+#if defined(MACE_ENABLE_NEON) && defined(__ANDROID__)
   if (DataTypeToEnum<T>::value == DT_FLOAT16) {
     net.AddRandomInput<D, float>("A", {batch, height, channels});
     net.AddRandomInput<D, float16_t>("B", {batch, out_width, channels});
   } else {
+#endif
     net.AddRandomInput<D, T>("A", {batch, height, channels});
     net.AddRandomInput<D, float>("B", {batch, out_width, channels});
+#if defined(MACE_ENABLE_NEON) && defined(__ANDROID__)
   }
+#endif
   net.GetTensor("A")->SetIsWeight(true);
   net.GetTensor("B")->SetIsWeight(true);
   if (DataTypeToEnum<T>::value == DT_UINT8) {
@@ -388,11 +396,16 @@ void MatMulTransposeBenchmark(
   }                                                                            \
   MACE_BENCHMARK(MACE_BM_MATMUL_##T_##N##_##H##_##C##_##W##_##TYPE##_##DEVICE)
 
-#ifdef MACE_ENABLE_QUANTIZE
+#if defined(MACE_ENABLE_NEON) && defined(__ANDROID__)
 #define MACE_BM_MATMUL_TRANPOSE(N, H, C, W)                   \
   MACE_BM_MATMUL_TRANSPOSE_MACRO(N, H, C, W, float, CPU);     \
   MACE_BM_MATMUL_TRANSPOSE_MACRO(N, H, C, W, float16_t, CPU);     \
   MACE_BM_MATMUL_TRANSPOSE_MACRO(N, H, C, W, uint8_t, CPU);
+#else
+#define MACE_BM_MATMUL_TRANPOSE(N, H, C, W)                   \
+  MACE_BM_MATMUL_TRANSPOSE_MACRO(N, H, C, W, float, CPU);     \
+  MACE_BM_MATMUL_TRANSPOSE_MACRO(N, H, C, W, uint8_t, CPU);
+#endif
 
 MACE_BM_MATMUL_OP(1, 30000, 256, 1);
 MACE_BM_MATMUL_OP(1, 128, 256, 128);