COMPMID-3160: Add Bfloat16 support in NEGEMMConvolutionLayer

Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Change-Id: I0e449306c138a562ffc1455e76ec44b2fd059d85
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/2860
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

6 files changed, 40 insertions, 30 deletions

diff --git a/arm_compute/core/NEON/kernels/NEIm2ColKernel.h b/arm_compute/core/NEON/kernels/NEIm2ColKernel.h
index 292e941272..1c358b379d 100644
--- a/arm_compute/core/NEON/kernels/NEIm2ColKernel.h
+++ b/arm_compute/core/NEON/kernels/NEIm2ColKernel.h
@@ -77,7 +77,8 @@ public:
     /** Set the input and output of the kernel.
      *
      * @param[in]  input       The input tensor to convert. 3 lower dimensions represent a single input [width, height, IFM],
-     *                         while every optional dimension from 4 and above represent a batch of inputs. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32
+     *                         while every optional dimension from 4 and above represent a batch of inputs.
+     *                         Data types supported: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32
      *                         Note: QASYMM8 works only for has_bias = false
      * @param[out] output      The output tensor. Data types supported: Same as @p input
      * @param[in]  kernel_dims The kernel dimensions (width and height).
@@ -91,7 +92,8 @@ public:
     /** Static function to check if given info will lead to a valid configuration of @ref NEIm2ColKernel
      *
      * @param[in] input       The input tensor to convert. 3 lower dimensions represent a single input [width, height, IFM],
-     *                        while every optional dimension from 4 and above represent a batch of inputs. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32
+     *                        while every optional dimension from 4 and above represent a batch of inputs.
+     *                        Data types supported: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32
      *                        Note: QASYMM8 works only for has_bias = false
      * @param[in] output      The output tensor. Data types supported: Same as @p input
      * @param[in] kernel_dims The kernel dimensions (width and height).
diff --git a/arm_compute/core/NEON/kernels/NEWeightsReshapeKernel.h b/arm_compute/core/NEON/kernels/NEWeightsReshapeKernel.h
index 46be5233eb..b68cb50c7b 100644
--- a/arm_compute/core/NEON/kernels/NEWeightsReshapeKernel.h
+++ b/arm_compute/core/NEON/kernels/NEWeightsReshapeKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2019 ARM Limited.
+ * Copyright (c) 2017-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -76,7 +76,7 @@ public:
      *
      * @param[in]  input  The input tensor to convert. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM] if shared,
      *                    and 5D tensor with dimensions [kernel_x, kernel_y, IFM, OFM, num_patches] if unshared.
-     *                    Data types supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/FP16/F32
+     *                    Data types supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/BFLOAT16/FP16/F32
      * @param[in]  bias   The shared biases tensor to append.  Bias is 1D tensor with dimensions [OFM] if shared and 2D tensor with
      *                    dimensions [OFM, num_patches] if unshared. Data types supported: Same as @p input
      *                    @warning Appending biases to weights reshaped matrix is not supported for quantized asymmetric types.
@@ -87,7 +87,7 @@ public:
      *
      * @param[in] input  The input tensor to convert. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM] if shared,
      *                   and 5D tensor with dimensions [kernel_x, kernel_y, IFM, OFM,  num_patches] if unshared.
-     *                   Data types supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/F16/F32
+     *                   Data types supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/BFLOAT16/F16/F32
      * @param[in] biases The shared biases tensor to append.  Bias is 1D tensor with dimensions [OFM] if shared and 2D tensor with
      *                   dimensions [OFM, num_patches] if unshared. Data types supported: Same as @p input
      *                   @warning Appending biases to weights reshaped matrix is not supported for quantized asymmetric types.
diff --git a/arm_compute/core/Utils.h b/arm_compute/core/Utils.h
index 8577046af0..eff6157b1f 100644
--- a/arm_compute/core/Utils.h
+++ b/arm_compute/core/Utils.h
@@ -1342,7 +1342,7 @@ void print_consecutive_elements_impl(std::ostream &s, const T *ptr, unsigned int
         }
         else if(std::is_same<typename std::decay<T>::type, bfloat16>::value)
         {
-            // We use T instead of print_type here is because the std::is_floating_point<bfloat> returns false and then the print_type becomes int.
+            // We use T instead of print_type here is because the std::is_floating_point<bfloat16> returns false and then the print_type becomes int.
             s << std::right << float(ptr[i]) << element_delim;
         }
         else
diff --git a/arm_compute/runtime/NEON/functions/NEGEMM.h b/arm_compute/runtime/NEON/functions/NEGEMM.h
index f08bd9fac5..c87e806d0c 100644
--- a/arm_compute/runtime/NEON/functions/NEGEMM.h
+++ b/arm_compute/runtime/NEON/functions/NEGEMM.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2019 ARM Limited.
+ * Copyright (c) 2017-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -74,7 +74,7 @@ public:
      * @note GEMM: General Matrix Multiply - [alpha * A * B + beta * C].
      * @note GEMM: The tensors a, b, c, d must have the same data type. You should not mix data types when calling this function.
      *
-     * @param[in]  a         First input tensor  (Matrix A or Vector A). Data type supported: F16/F32
+     * @param[in]  a         First input tensor  (Matrix A or Vector A). Data type supported: BLOAT16/F16/F32
      * @param[in]  b         Second input tensor (Matrix B). Data type supported: same as @p a
      * @param[in]  c         Third input tensor  (Matrix C). It can be a nullptr if just the multiplication between @p a and @p b is needed. Data type supported: same as @p a
      * @param[out] d         Output tensor. Data type supported: same as @p a
@@ -86,7 +86,7 @@ public:
     void configure(const ITensor *a, const ITensor *b, const ITensor *c, ITensor *d, float alpha, float beta, const GEMMInfo &gemm_info = GEMMInfo());
     /** Static function to check if given info will lead to a valid configuration of @ref NEGEMM.
      *
-     * @param[in]  a         First input tensor info  (Matrix or Vector A). Data types supported: F16/F32
+     * @param[in]  a         First input tensor info  (Matrix or Vector A). Data types supported: BLOAT16/F16/F32
      * @param[in]  b         Second input tensor info (Matrix B). Data type supported: same as @p a.
      * @param[in]  c         Third input tensor info  (Matrix C). It can be a nullptr if just the multiplication between @p a and @p b is needed. Data type supported: same as @p a.
      * @param[out] output    Output tensor info. Data type supported: same as @p a
diff --git a/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h b/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h
index 660f55953e..5368384b19 100644
--- a/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2019 ARM Limited.
+ * Copyright (c) 2017-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -63,18 +63,22 @@ public:
     NEConvolutionLayerReshapeWeights &operator=(NEConvolutionLayerReshapeWeights &&) = default;
     /** Set the input and output tensors.
      *
-     * @param[in]  weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/F16/F32.
+     * @param[in]  weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM].
+     *                     Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/BFLOAT16/F16/F32.
      * @param[in]  biases  Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
-     *                     Data type supported: Same as @p weights, S32 if @p weights is QASYMM8/QASYMM8_SIGNED.
-     * @param[out] output  Destination tensor. Data types supported: Same as @p weights.
+     *                     Data type supported: Same as @p weights, S32 if @p weights is QASYMM8/QASYMM8_SIGNED, FP32 if @p weights is BLOAT16
+     * @param[out] output  Destination tensor.
+     *                     Data types supported: Same as @p weights, FP32 if @p weights is BLOAT16
      */
     void configure(const ITensor *weights, const ITensor *biases, ITensor *output);
     /** Static function to check if given info will lead to a valid configuration of @ref NEConvolutionLayerReshapeWeights
      *
-     * @param[in] weights Weights tensor info. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/F16/F32.
+     * @param[in] weights Weights tensor info. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM].
+     *                    Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/BFLOAT16/F16/F32.
      * @param[in] biases  Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
-     *                    Data type supported: Same as @p weights, S32 if @p weights is QASYMM8/QASYMM8_SIGNED.
-     * @param[in] output  Destination tensor. Data types supported: Same as @p weights.
+     *                    Data type supported: Same as @p weights, S32 if @p weights is QASYMM8/QASYMM8_SIGNED, FP32 if @p weights is BLOAT16
+     * @param[in] output  Destination tensor.
+     *                    Data types supported: Same as @p weights FP32 if @p weights is BLOAT16
      *
      * @return an error status
      */
@@ -136,7 +140,7 @@ private:
 /** Basic function to compute the convolution layer. This function calls the following NEON kernels/functions:
  *
  * -# @ref NEIm2ColKernel
- * -# @ref NEGEMM (if the data type is FP32 or FP16)
+ * -# @ref NEGEMM (if the data type is BLOAT16/FP16/FP32)
  * -# @ref NEGEMMLowpMatrixMultiplyCore (if the data type is QASYMM8/QASYMM8_SIGNED)
  * -# @ref NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint (if the data type is QASYMM8/QASYMM8_SIGNED)
  * -# @ref NEArithmeticAdditionKernel (if biases != nullptr and we have a 1x1 convolution with the NHWC data layout)
@@ -160,8 +164,9 @@ public:
      *
      * @param[in]  input        Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
      *                          while every optional dimension from 4 and above represent a batch of inputs.
-     *                          Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
-     * @param[in]  weights      Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/F16/F32.
+     *                          Data types supported: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32.
+     * @param[in]  weights      Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM].
+     *                          Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/BFLOAT16/F16/F32.
      * @param[in]  biases       Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
      *                          Data type supported: Should match @p input data type, except for input of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type.
      * @param[out] output       Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
@@ -179,8 +184,9 @@ public:
      *
      * @param[in] input        Source tensor info. 3 lower dimensions represent a single input [width, height, IFM],
      *                         while every optional dimension from 4 and above represent a batch of inputs.
-     *                         Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
-     * @param[in] weights      Weights tensor info. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/F16/F32.
+     *                         Data types supported: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32.
+     * @param[in] weights      Weights tensor info. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM].
+     *                         Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/BFLOAT16/F16/F32.
      * @param[in] biases       Biases tensor info. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
      *                         Data type supported: Should match @p input data type, except for input of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type.
      * @param[in] output       Destination tensor info. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
@@ -204,8 +210,8 @@ public:
 private:
     /** Configures the appropriate matrix multiply routine
      *
-     * @param[in]  input         Input tensor. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
-     * @param[in]  weights       Weights tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/F16/F32.
+     * @param[in]  input         Input tensor. Data types supported: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32.
+     * @param[in]  weights       Weights tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/BFLOAT16/F16/F32.
      * @param[in]  biases        Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
      *                           Data type supported: Should match @p input data type, except for input of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type.
      * @param[out] output        Output tensor. Data types supported: Same as @p input,
@@ -216,8 +222,8 @@ private:
     void configure_mm(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const ActivationLayerInfo &act_info = ActivationLayerInfo(), int gemm_3d_depth = 1);
     /** Static function to check if given info will lead to a valid configuration of @ref NEGEMMConvolutionLayer matrix multiply routines
      *
-     * @param[in] input         Input tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
-     * @param[in] weights       Weights tensor info. Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/F16/F32.
+     * @param[in] input         Input tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32.
+     * @param[in] weights       Weights tensor info. Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/BFLOAT16/F16/F32.
      * @param[in] biases        Biases tensor info. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
      *                          Data type supported: Should match @p input data type, except for input of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type.
      * @param[in] output        Output tensor info. Data types supported: Same as @p input,
@@ -232,8 +238,8 @@ private:
                               int gemm_3d_depth = 1, bool skip_im2col = false);
     /** Static function to check if GEMM3D is supported in @ref NEGEMM or in @ref NEGEMMLowpMatrixMultiplyCore
      *
-     * @param[in] input_info    Input tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
-     * @param[in] weights_info  Weights tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
+     * @param[in] input_info    Input tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32.
+     * @param[in] weights_info  Weights tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32.
      * @param[in] act_info      Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU supported.
      * @param[in] gemm_3d_depth Depth of GEMM 3D
      * @param[in] skip_im2col   Flag which specifies if im2col has to be skipped. i.e. 1x1 convolution with NHWC data layout
diff --git a/arm_compute/runtime/NEON/functions/NEIm2Col.h b/arm_compute/runtime/NEON/functions/NEIm2Col.h
index ac1b2674a5..cb905a3652 100644
--- a/arm_compute/runtime/NEON/functions/NEIm2Col.h
+++ b/arm_compute/runtime/NEON/functions/NEIm2Col.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2019 ARM Limited.
+ * Copyright (c) 2017-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -44,7 +44,8 @@ public:
     /** Configure the im2col NEON kernel
      *
      * @param[in]  input       The input tensor to convert. 3 lower dimensions represent a single input [width, height, IFM],
-     *                         while every optional dimension from 4 and above represent a batch of inputs. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32
+     *                         while every optional dimension from 4 and above represent a batch of inputs.
+     *                         Data types supported: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32
      *                         Note: QASYMM8 works only for has_bias = false
      * @param[out] output      The output tensor. Data types supported: Same as @p input
      * @param[in]  kernel_dims The kernel dimensions (width and height).
@@ -58,7 +59,8 @@ public:
     /** Static function to check if given info will lead to a valid configuration of @ref NEIm2Col
      *
      * @param[in] input       The input tensor to convert. 3 lower dimensions represent a single input [width, height, IFM],
-     *                        while every optional dimension from 4 and above represent a batch of inputs. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32
+     *                        while every optional dimension from 4 and above represent a batch of inputs.
+     *                        Data types supported: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32
      *                        Note: QASYMM8 works only for has_bias = false
      * @param[in] output      The output tensor. Data types supported: Same as @p input
      * @param[in] kernel_dims The kernel dimensions (width and height).