From 97b3f11a1655c05bedaf378f85f94cdccb1536ba Mon Sep 17 00:00:00 2001
From: Sheri Zhang <sheri.zhang@arm.com>
Date: Mon, 4 Jan 2021 17:14:23 +0000
Subject: COMPMID-4076: ArmNN unittest failure with memory access voilation in
 FuseReLUIntoBatchNormFloat32CpuAccTest

1. Fix fusable and non-fusable configuration issue
2. Fix FP16 issue

Signed-off-by: Sheri Zhang <sheri.zhang@arm.com>
Change-Id: I6d0eacca7ac437f236ad403ddb283c10c8f419a6
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4761
Reviewed-by: Pablo Marquez Tello <pablo.tello@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
---
 src/core/NEON/kernels/NEBatchNormalizationLayerKernel.cpp   |  6 ++++--
 src/core/NEON/kernels/batchnormalization/impl/NEON/fp16.cpp | 11 +++++++++--
 src/core/NEON/kernels/batchnormalization/impl/NEON/fp32.cpp | 11 +++++++++--
 3 files changed, 22 insertions(+), 6 deletions(-)

diff --git a/src/core/NEON/kernels/NEBatchNormalizationLayerKernel.cpp b/src/core/NEON/kernels/NEBatchNormalizationLayerKernel.cpp
index b4cac74dc4..6f2c72a970 100644
--- a/src/core/NEON/kernels/NEBatchNormalizationLayerKernel.cpp
+++ b/src/core/NEON/kernels/NEBatchNormalizationLayerKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2020 Arm Limited.
+ * Copyright (c) 2017-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -74,12 +74,14 @@ static const BatchNormalizationKernel available_kernels[] =
         [](const BatchNormalizationSelectorData & data) { return data.dt == DataType::F32; },
         REGISTER_FP32_SVE(arm_compute::cpu::fp32_sve_batch_normalization)
     },
-#else  /* !defined(__ARM_FEATURE_SVE) */
+#else /* !defined(__ARM_FEATURE_SVE) */
+#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
     {
         "fp16_neon_batch_normalization",
         [](const BatchNormalizationSelectorData & data) { return data.dt == DataType::F16; },
         REGISTER_FP16_NEON(arm_compute::cpu::fp16_neon_batch_normalization)
     },
+#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
     {
         "f32_neon_batch_normalization",
         [](const BatchNormalizationSelectorData & data) { return data.dt == DataType::F32; },
diff --git a/src/core/NEON/kernels/batchnormalization/impl/NEON/fp16.cpp b/src/core/NEON/kernels/batchnormalization/impl/NEON/fp16.cpp
index dfadef34f7..fd17b98f7b 100644
--- a/src/core/NEON/kernels/batchnormalization/impl/NEON/fp16.cpp
+++ b/src/core/NEON/kernels/batchnormalization/impl/NEON/fp16.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2020 Arm Limited.
+ * Copyright (c) 2020-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -138,7 +138,14 @@ namespace cpu
 void fp16_neon_batch_normalization(ITensor *src, ITensor *dst, const ITensor *mean, const ITensor *var, const ITensor *beta, const ITensor *gamma,
                                    float epsilon, ActivationLayerInfo &act_info, const Window &window)
 {
-    fused_map[act_info.activation()](src, dst, mean, var, beta, gamma, epsilon, act_info, window);
+    if(act_info.enabled())
+    {
+        fused_map[act_info.activation()](src, dst, mean, var, beta, gamma, epsilon, act_info, window);
+    }
+    else
+    {
+        batch_normalization<detail::dummy<float16_t, 8>>(src, dst, mean, var, beta, gamma, epsilon, act_info, window);
+    }
 }
 } // namespace cpu
 } // namespace arm_compute
diff --git a/src/core/NEON/kernels/batchnormalization/impl/NEON/fp32.cpp b/src/core/NEON/kernels/batchnormalization/impl/NEON/fp32.cpp
index a24f7f624a..5b375e5d4d 100644
--- a/src/core/NEON/kernels/batchnormalization/impl/NEON/fp32.cpp
+++ b/src/core/NEON/kernels/batchnormalization/impl/NEON/fp32.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2020 Arm Limited.
+ * Copyright (c) 2020-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -137,7 +137,14 @@ namespace cpu
 void fp32_neon_batch_normalization(ITensor *src, ITensor *dst, const ITensor *mean, const ITensor *var, const ITensor *beta, const ITensor *gamma,
                                    float epsilon, ActivationLayerInfo &act_info, const Window &window)
 {
-    fused_map[act_info.activation()](src, dst, mean, var, beta, gamma, epsilon, act_info, window);
+    if(act_info.enabled())
+    {
+        fused_map[act_info.activation()](src, dst, mean, var, beta, gamma, epsilon, act_info, window);
+    }
+    else
+    {
+        batch_normalization<detail::dummy<float, 4>>(src, dst, mean, var, beta, gamma, epsilon, act_info, window);
+    }
 }
 } // namespace cpu
 } // namespace arm_compute
-- 
cgit v1.2.1