Deprecated BF16 support in DepthConvert

* Removed BF16 validation tests for DepthConvert

* Revert back to using inline assembly to convert to/from BF16

* Resolves COMPMID-5800

Change-Id: I803b2ad19ead297417f780c97c5b724cca6b394c
Signed-off-by: Pablo Marquez Tello <pablo.tello@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8929
Reviewed-by: Jakub Sujak <jakub.sujak@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

8 files changed, 94 insertions, 69 deletions

diff --git a/SConstruct b/SConstruct
index 161b3f2c60..8603017f71 100644
--- a/SConstruct
+++ b/SConstruct
@@ -1,6 +1,6 @@
 # -*- coding: utf-8 -*-
 
-# Copyright (c) 2016-2022 Arm Limited.
+# Copyright (c) 2016-2023 Arm Limited.
 #
 # SPDX-License-Identifier: MIT
 #
@@ -302,7 +302,7 @@ if env['multi_isa']:
         if "disable_mmla_fp" not in env['custom_options']:
             env.Append(CPPDEFINES = ['ARM_COMPUTE_ENABLE_SVEF32MM'])
 
-    env.Append(CXXFLAGS = ['-march=armv8.2-a+fp16+bf16']) # explicitly enable fp16 extension otherwise __ARM_FEATURE_FP16_VECTOR_ARITHMETIC is undefined
+    env.Append(CXXFLAGS = ['-march=armv8.2-a+fp16']) # explicitly enable fp16 extension otherwise __ARM_FEATURE_FP16_VECTOR_ARITHMETIC is undefined
 
 else: # NONE "multi_isa" builds
 
diff --git a/filedefs.json b/filedefs.json
index c8a66e3429..3422eeb252 100644
--- a/filedefs.json
+++ b/filedefs.json
@@ -5,7 +5,7 @@
                 "cxxflags": ["-march=armv8.2-a"]
             },
             "armv8.2-a": {
-                "cxxflags": ["-march=armv8.2-a+fp16+bf16"],
+                "cxxflags": ["-march=armv8.2-a+fp16"],
                 "cppdefines": ["ARM_COMPUTE_ENABLE_FP16"]
             },
             "armv8.2-a-sve": {
diff --git a/src/core/NEON/wrapper/intrinsics/cvt.h b/src/core/NEON/wrapper/intrinsics/cvt.h
index c75d43dbf2..1c77a9e9f0 100644
--- a/src/core/NEON/wrapper/intrinsics/cvt.h
+++ b/src/core/NEON/wrapper/intrinsics/cvt.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2020, 2022 Arm Limited.
+ * Copyright (c) 2020, 2022-2023 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -87,6 +87,26 @@ vcvta(const float32x4_t &a)
     return vcvtaq_s32_f32(a);
 }
 #endif //__aarch64__
+
+#if defined(ARM_COMPUTE_ENABLE_BF16)
+/** Convert 2x128-bit floating point vectors into 1x128-bit bfloat16 vector
+ *
+ * @param[in]     inptr  Pointer to the input memory to load values from
+ * @param[in,out] outptr Pointer to the output memory to store values to
+ */
+inline void vcvt_bf16_f32(const float *inptr, uint16_t *outptr)
+{
+    __asm __volatile(
+        "ldp    q0, q1, [%[inptr]]\n"
+        ".inst  0xea16800\n"  // BFCVTN v0, v0
+        ".inst  0x4ea16820\n" // BFCVTN2 v0, v1
+        "str    q0, [%[outptr]]\n"
+        : [inptr] "+r"(inptr)
+        : [outptr] "r"(outptr)
+        : "v0", "v1", "memory");
+}
+#endif /* defined(ARM_COMPUTE_ENABLE_BF16) */
+
 } // namespace wrapper
 } // namespace arm_compute
 #endif /* ARM_COMPUTE_WRAPPER_CVT_H */
diff --git a/src/cpu/kernels/CpuCastKernel.h b/src/cpu/kernels/CpuCastKernel.h
index 95d46fad23..de4ace2140 100644
--- a/src/cpu/kernels/CpuCastKernel.h
+++ b/src/cpu/kernels/CpuCastKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2022 Arm Limited.
+ * Copyright (c) 2016-2023 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -62,6 +62,8 @@ public:
      * @param[in]  src    The src tensor to convert. Data types supported: QASYMM8_SIGNED/QASYMM8/U8/U16/S16/BFLOAT16/F16/F32.
      * @param[out] dst    The dst tensor. Data types supported: QASYMM8_SIGNED/QASYMM8/U8/U16/S16/U32/S32/BFLOAT16/F16/F32.
      * @param[in]  policy Conversion policy.
+     *
+     * @deprecated Support for BFLOAT16 will be removed in 23.05 release
      */
     void configure(const ITensorInfo *src, ITensorInfo *dst, ConvertPolicy policy);
     /** Static function to check if given info will lead to a valid configuration
diff --git a/src/cpu/kernels/cast/generic/neon/bfloat16.cpp b/src/cpu/kernels/cast/generic/neon/bfloat16.cpp
index 942bdfae61..91c15be279 100644
--- a/src/cpu/kernels/cast/generic/neon/bfloat16.cpp
+++ b/src/cpu/kernels/cast/generic/neon/bfloat16.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2022 Arm Limited.
+ * Copyright (c) 2016-2023 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -24,8 +24,10 @@
 #if defined(ARM_COMPUTE_ENABLE_BF16)
 
 #include "arm_compute/core/TensorInfo.h"
+#include "src/core/NEON/wrapper/wrapper.h"
 #include "src/cpu/kernels/CpuCastKernel.h"
 #include "src/cpu/kernels/cast/list.h"
+#include "support/SaturateCast.h"
 
 namespace arm_compute
 {
@@ -36,9 +38,9 @@ void neon_fp32_to_bfloat16_cast(const ITensor *_src, ITensor *_dst, const Thread
     ARM_COMPUTE_UNUSED(info);
     ARM_COMPUTE_UNUSED(_policy);
 
-    const auto    window_start_x = static_cast<int>(window.x().start());
-    const auto    window_end_x   = static_cast<int>(window.x().end());
-    constexpr int window_step_x  = 8;
+    const auto window_start_x = static_cast<int>(window.x().start());
+    const auto window_end_x   = static_cast<int>(window.x().end());
+    const int  window_step_x  = 16;
 
     ARM_COMPUTE_ERROR_ON_NULLPTR(_src, _dst);
     ARM_COMPUTE_ERROR_ON(_src == _dst);
@@ -50,23 +52,25 @@ void neon_fp32_to_bfloat16_cast(const ITensor *_src, ITensor *_dst, const Thread
 
     Iterator src(_src, win);
     Iterator dst(_dst, win);
+
     /* Down-conversion F32 -> BFLOAT16 */
     execute_window_loop(win, [&](const Coordinates &)
     {
-        const auto src_ptr     = reinterpret_cast<const float *>(src.ptr());
-        const auto dst_ptr     = reinterpret_cast<bfloat16_t *>(dst.ptr());
-        int        x           = window_start_x;
-        const int  right_bound = (window_end_x - window_step_x);
-        for(; x <= right_bound; x += window_step_x)
+        const auto src_ptr = reinterpret_cast<const float *>(src.ptr());
+        const auto dst_ptr = reinterpret_cast<bfloat16 *>(dst.ptr());
+
+        int x = window_start_x;
+        for(; x <= (window_end_x - window_step_x); x += window_step_x)
         {
-            const auto vbf16_0 = vcombine_bf16(
-                                     vcvt_bf16_f32(vld1q_f32(src_ptr + x)),
-                                     vcvt_bf16_f32(vld1q_f32(src_ptr + x + 4)));
-            vst1q_bf16(dst_ptr + x, vbf16_0);
+            wrapper::vcvt_bf16_f32(reinterpret_cast<float *>(src.ptr()),
+                                   reinterpret_cast<uint16_t *>(dst.ptr()));
+            wrapper::vcvt_bf16_f32(reinterpret_cast<float *>(src.ptr()) + 8,
+                                   reinterpret_cast<uint16_t *>(dst.ptr()) + 8);
         }
+
         for(; x < window_end_x; ++x)
         {
-            *(reinterpret_cast<bfloat16 *>(dst.ptr()) + x) = *(src_ptr + x);
+            *(dst_ptr + x) = *(src_ptr + x);
         }
     },
     src, dst);
@@ -77,9 +81,9 @@ void neon_bfloat16_to_fp32_cast(const ITensor *_src, ITensor *_dst, const Thread
     ARM_COMPUTE_UNUSED(info);
     ARM_COMPUTE_UNUSED(_policy);
 
-    const auto    window_start_x = static_cast<int>(window.x().start());
-    const auto    window_end_x   = static_cast<int>(window.x().end());
-    constexpr int window_step_x  = 8;
+    const auto window_start_x = static_cast<int>(window.x().start());
+    const auto window_end_x   = static_cast<int>(window.x().end());
+    const int  window_step_x  = 16;
 
     ARM_COMPUTE_ERROR_ON_NULLPTR(_src, _dst);
     ARM_COMPUTE_ERROR_ON(_src == _dst);
@@ -91,26 +95,48 @@ void neon_bfloat16_to_fp32_cast(const ITensor *_src, ITensor *_dst, const Thread
 
     Iterator src(_src, win);
     Iterator dst(_dst, win);
-    /* Up-conversion BFLOAT16 -> F32 */
-    execute_window_loop(win, [&](const Coordinates &)
+    switch(_dst->info()->data_type())
     {
-        const auto src_ptr = reinterpret_cast<const bfloat16_t *>(src.ptr());
-        const auto dst_ptr = reinterpret_cast<float *>(dst.ptr());
-
-        int       x = window_start_x;
-        const int right_bound(window_end_x - window_step_x);
-        for(; x <= right_bound; x += window_step_x)
+        case DataType::F32:
         {
-            const bfloat16x8_t vinput = vld1q_bf16(src_ptr + x);
-            vst1q_f32(dst_ptr + x, vcvt_f32_bf16(vget_low_bf16(vinput)));
-            vst1q_f32(dst_ptr + x + 4, vcvt_f32_bf16(vget_high_bf16(vinput)));
+            /* Up-conversion BFLOAT16 -> F32 */
+            execute_window_loop(win, [&](const Coordinates &)
+            {
+                const auto src_ptr = reinterpret_cast<const bfloat16 *>(src.ptr());
+                const auto dst_ptr = reinterpret_cast<float *>(dst.ptr());
+
+                int x = window_start_x;
+                for(; x <= (window_end_x - window_step_x); x += window_step_x)
+                {
+                    const uint16x8x2_t texels =
+                    {
+                        {
+                            vld1q_u16(reinterpret_cast<uint16_t *>(src.ptr())),
+                            vld1q_u16(reinterpret_cast<uint16_t *>(src.ptr()) + 8)
+                        }
+                    };
+
+                    vst1q_f32(reinterpret_cast<float *>(dst.ptr()),
+                              vreinterpretq_f32_u32(vshlq_n_u32(vmovl_u16(vget_low_u16(texels.val[0])), 16)));
+                    vst1q_f32(reinterpret_cast<float *>(dst.ptr()) + 4,
+                              vreinterpretq_f32_u32(vshlq_n_u32(vmovl_u16(vget_high_u16(texels.val[0])), 16)));
+                    vst1q_f32(reinterpret_cast<float *>(dst.ptr()) + 8,
+                              vreinterpretq_f32_u32(vshlq_n_u32(vmovl_u16(vget_low_u16(texels.val[1])), 16)));
+                    vst1q_f32(reinterpret_cast<float *>(dst.ptr()) + 12,
+                              vreinterpretq_f32_u32(vshlq_n_u32(vmovl_u16(vget_high_u16(texels.val[1])), 16)));
+                }
+
+                for(; x < window_end_x; ++x)
+                {
+                    *(dst_ptr + x) = float(*(src_ptr + x));
+                }
+            },
+            src, dst);
+            break;
         }
-        for(; x < window_end_x; ++x)
-        {
-            *(dst_ptr + x) = float(*(reinterpret_cast<const bfloat16 *>(src_ptr) + x));
-        }
-    },
-    src, dst);
+        default:
+            ARM_COMPUTE_ERROR("dst data type unsupported");
+    }
 }
 
 } // namespace cpu
diff --git a/src/cpu/operators/CpuCast.h b/src/cpu/operators/CpuCast.h
index 5e5f3e0ca6..a8342581cb 100644
--- a/src/cpu/operators/CpuCast.h
+++ b/src/cpu/operators/CpuCast.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2021 Arm Limited.
+ * Copyright (c) 2021, 2023 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -56,6 +56,9 @@ public:
      * @param[in]  src    The source tensor to convert. Data types supported: U8/S8/U16/S16/U32/S32/F16/F32.
      * @param[out] dst    The destination tensor. Data types supported: U8/S8/U16/S16/U32/S32/F16/F32.
      * @param[in]  policy Conversion policy.
+     *
+     * @deprecated Support for BFLOAT16 will be removed in 23.05 release
+     *
      */
     void configure(const ITensorInfo *src, ITensorInfo *dst, ConvertPolicy policy);
     /** Static function to check if given info will lead to a valid configuration
diff --git a/tests/validation/NEON/DepthConvertLayer.cpp b/tests/validation/NEON/DepthConvertLayer.cpp
index 378652c24f..4972708144 100644
--- a/tests/validation/NEON/DepthConvertLayer.cpp
+++ b/tests/validation/NEON/DepthConvertLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2022 Arm Limited.
+ * Copyright (c) 2017-2023 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -56,14 +56,12 @@ const auto DepthConvertLayerU16toU8Dataset      = combine(framework::dataset::ma
 const auto DepthConvertLayerU16toU32Dataset     = combine(framework::dataset::make("DataType", DataType::U16), framework::dataset::make("DataType", DataType::U32));
 const auto DepthConvertLayerS16toU8Dataset      = combine(framework::dataset::make("DataType", DataType::S16), framework::dataset::make("DataType", DataType::U8));
 const auto DepthConvertLayerS16toS32Dataset     = combine(framework::dataset::make("DataType", DataType::S16), framework::dataset::make("DataType", DataType::S32));
-const auto DepthConvertLayerBF16toF32Dataset    = combine(framework::dataset::make("DataType", DataType::BFLOAT16), framework::dataset::make("DataType", DataType::F32));
 const auto DepthConvertLayerF16toU8Dataset      = combine(framework::dataset::make("DataType", DataType::F16), framework::dataset::make("DataType", DataType::U8));
 const auto DepthConvertLayerF16toF32Dataset     = combine(framework::dataset::make("DataType", DataType::F16), framework::dataset::make("DataType", DataType::F32));
 const auto DepthConvertLayerF16toS32Dataset     = combine(framework::dataset::make("DataType", DataType::F16), framework::dataset::make("DataType", DataType::S32));
 const auto DepthConvertLayerF32toF16Dataset     = combine(framework::dataset::make("DataType", DataType::F32), framework::dataset::make("DataType", DataType::F16));
 const auto DepthConvertLayerF32toS32Dataset     = combine(framework::dataset::make("DataType", DataType::F32), framework::dataset::make("DataType", DataType::S32));
 const auto DepthConvertLayerF32toU8Dataset      = combine(framework::dataset::make("DataType", DataType::F32), framework::dataset::make("DataType", DataType::U8));
-const auto DepthConvertLayerF32toBF16Dataset    = combine(framework::dataset::make("DataType", DataType::F32), framework::dataset::make("DataType", DataType::BFLOAT16));
 
 const auto DepthConvertLayerS32toF32Dataset     = combine(framework::dataset::make("DataType", DataType::S32), framework::dataset::make("DataType", DataType::F32));
 const auto DepthConvertLayerS32toQASYMM8Dataset = combine(framework::dataset::make("DataType", DataType::S32), framework::dataset::make("DataType", DataType::QASYMM8));
@@ -127,8 +125,6 @@ using NEDepthConvertLayerToU8Fixture = DepthConvertLayerValidationFixture<Tensor
 template <typename T>
 using NEDepthConvertLayerToU32Fixture = DepthConvertLayerValidationFixture<Tensor, Accessor, NEDepthConvertLayer, T, uint32_t>;
 template <typename T>
-using NEDepthConvertLayerToBF16Fixture = DepthConvertLayerValidationFixture<Tensor, Accessor, NEDepthConvertLayer, T, bfloat16>;
-template <typename T>
 using NEDepthConvertLayerToF16Fixture = DepthConvertLayerValidationFixture<Tensor, Accessor, NEDepthConvertLayer, T, half>;
 template <typename T>
 using NEDepthConvertLayerToF32Fixture = DepthConvertLayerValidationFixture<Tensor, Accessor, NEDepthConvertLayer, T, float>;
@@ -342,28 +338,6 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NEDepthConvertLayerToS32Fixture<int16_t>, frame
 }
 TEST_SUITE_END() // S16_to_S32
 
-#if defined(ARM_COMPUTE_ENABLE_BF16)
-TEST_SUITE(BFLOAT16_to_F32)
-FIXTURE_DATA_TEST_CASE(RunSmall, NEDepthConvertLayerToF32Fixture<bfloat16>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), DepthConvertLayerBF16toF32Dataset),
-                                                                                                                       framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })),
-                                                                                                                       DepthConvertLayerZeroShiftDataset))
-{
-    // Validate output
-    validate(Accessor(_target), _reference);
-}
-TEST_SUITE_END() // BFLOAT16_to_F32
-
-TEST_SUITE(F32_to_BFLOAT16)
-FIXTURE_DATA_TEST_CASE(RunSmall, NEDepthConvertLayerToBF16Fixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), DepthConvertLayerF32toBF16Dataset),
-                                                                                                                     framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })),
-                                                                                                                     DepthConvertLayerZeroShiftDataset))
-{
-    // Validate output
-    validate(Accessor(_target), _reference);
-}
-TEST_SUITE_END() // F32_to_BFLOAT16
-#endif           /* defined(ARM_COMPUTE_ENABLE_BF16) */
-
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
 TEST_SUITE(F16_to_QASYMM8)
 FIXTURE_DATA_TEST_CASE(RunSmall, NEDepthConvertLayerToQASYMM8Fixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallShapes(),
diff --git a/utils/TypePrinter.h b/utils/TypePrinter.h
index a31a0e559b..8f9c24955e 100644
--- a/utils/TypePrinter.h
+++ b/utils/TypePrinter.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2022 Arm Limited.
+ * Copyright (c) 2017-2023 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -482,7 +482,7 @@ inline ::std::ostream &operator<<(::std::ostream &os, const BoundingBoxTransform
 inline ::std::ostream &operator<<(::std::ostream &os, const bfloat16 &v)
 {
     std::stringstream str;
-    str << static_cast<float>(v);
+    str << v;
     os << str.str();
     return os;
 }