COMPMID-2789: Add support for QASYMM8_SIGNED in CLGEMMDeconvolutionLayer

Change-Id: I7e3bcb01025e827f6f62491749c691c205ee7481
Signed-off-by: Sheri Zhang <sheri.zhang@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/2844
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

2 files changed, 56 insertions, 23 deletions

diff --git a/arm_compute/runtime/CL/functions/CLGEMMDeconvolutionLayer.h b/arm_compute/runtime/CL/functions/CLGEMMDeconvolutionLayer.h
index 3df9205f48..01687b69ec 100644
--- a/arm_compute/runtime/CL/functions/CLGEMMDeconvolutionLayer.h
+++ b/arm_compute/runtime/CL/functions/CLGEMMDeconvolutionLayer.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019 ARM Limited.
+ * Copyright (c) 2019-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -68,7 +68,7 @@ class ICLTensor;
  * This function calls the following OpenCL kernels/functions:
  *
  * -# @ref CLGEMMLowpMatrixMultiplyCore
- * -# @ref CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint
+ * -# @ref CLGEMMLowpOutputStage
  * -# @ref CLPermute
  * -# @ref CLPermute
  * -# @ref CLReshapeLayer
@@ -91,7 +91,8 @@ public:
     CLGEMMDeconvolutionLayer &operator=(CLGEMMDeconvolutionLayer &&) = default;
     /** Set the input, weights, biases and output tensors.
      *
-     * @param[in,out] input       Input tensor. 3 lower dimensions represent a single input, and an optional 4th dimension for batch of inputs. Data types supported: F16/F32. Data layout supported: NHWC
+     * @param[in,out] input       Input tensor. 3 lower dimensions represent a single input, and an optional 4th dimension for batch of inputs.
+     *                            Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. Data layout supported: NHWC
      * @param[in]     weights     The 4d weights with dimensions [width, height, IFM, OFM]. Data type supported: Same as @p input. Data layout supported: same as @p input.
      * @param[in]     bias        (Optional) The biases have one dimension. Data type supported: Same as @p input. Data layout supported: same as @p input.
      * @param[out]    output      Output tensor. The output has the same number of dimensions as the @p input. Data layout supported: same as @p input.
@@ -100,7 +101,8 @@ public:
     void configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *bias, ICLTensor *output, const PadStrideInfo &deconv_info);
     /** Static function to check if given info will lead to a valid configuration of @ref CLDeconvolutionLayer
      *
-     * @param[in] input       Input tensor info. 3 lower dimensions represent a single input, and an optional 4th dimension for batch of inputs. Data types supported: F16/F32. Data layout supported: NHWC
+     * @param[in] input       Input tensor info. 3 lower dimensions represent a single input, and an optional 4th dimension for batch of inputs.
+     *                        Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. Data layout supported: NHWC
      * @param[in] weights     The 4d weights info with dimensions [width, height, IFM, OFM]. Data type supported: Same as @p input. Data layout supported: same as @p input.
      * @param[in] bias        (Optional) The biases have one dimension. Data type supported: Same as @p input. Data layout supported: same as @p input.
      * @param[in] output      Output tensor info. The output has the same number of dimensions as the @p input. Data layout supported: same as @p input.
@@ -117,15 +119,15 @@ public:
 private:
     MemoryGroup _memory_group;
 
-    CLGEMM                                              _mm_gemm;
-    CLGEMMLowpMatrixMultiplyCore                        _mm_gemmlowp;
-    CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint _gemmlowp_output_stage;
-    CLPermute                                           _permute_input_to_nhwc;
-    CLPermute                                           _permute_weights_to_nhwc;
-    CLReshapeLayer                                      _reshape_weights;
-    CLTranspose                                         _transpose_weights;
-    CLDeconvolutionReshapeOutputKernel                  _deconv_reshape;
-    CLSlice                                             _slice_gemm;
+    CLGEMM                             _mm_gemm;
+    CLGEMMLowpMatrixMultiplyCore       _mm_gemmlowp;
+    CLGEMMLowpOutputStage              _gemmlowp_output_stage;
+    CLPermute                          _permute_input_to_nhwc;
+    CLPermute                          _permute_weights_to_nhwc;
+    CLReshapeLayer                     _reshape_weights;
+    CLTranspose                        _transpose_weights;
+    CLDeconvolutionReshapeOutputKernel _deconv_reshape;
+    CLSlice                            _slice_gemm;
 
     CLTensor _gemmlowp_final;
     CLTensor _reshaped_weights;
diff --git a/arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h b/arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h
index 564135eed8..b6619da5d2 100644
--- a/arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h
+++ b/arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h
@@ -64,7 +64,7 @@ public:
      * @param[in]  input           Input tensor. It is the output of @ref CLGEMMLowpMatrixMultiplyCore function. Data type supported: S32
      * @param[in]  bias            Biases tensor. Only shared biases supported and it can be a nullptr if the addition of biases is not required.
      *                             Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input.
-     * @param[out] output          Output tensor. Data type supported: Data type supported: QASYMM8
+     * @param[out] output          Output tensor. Data type supported: QASYMM8
      * @param[in]  result_offset   Offset to be added to each element of the input matrix
      * @param[in]  result_mult_int Value to be multiplied to each element of the input matrix when once the result_offset has been add
      * @param[in]  result_shift    Number of bits to shift right the result before converting back to QASYMM8
@@ -79,7 +79,7 @@ public:
      * @param[in] input  Input tensor. It is the output of @ref CLGEMMLowpMatrixMultiplyCore function. Data type supported: S32
      * @param[in] bias   Biases tensor. Only shared biases supported and it can be a nullptr if the addition of biases is not required.
      *                   Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input.
-     * @param[in] output Output tensor. Data type supported: Data type supported: QASYMM8
+     * @param[in] output Output tensor. Data type supported: QASYMM8
      * @param[in] min    (Optional) Min value used to saturate down the output result before converting back to QASYMM8. Defaults to the minimum possible 32-bit signed integer.
      * @param[in] max    (Optional) Max value used to saturate up the output result before converting back to QASYMM8,
      *                   Along with @p min, this value can be used to implement "rectified linear unit" activation functions. Defaults to the maximum possible 32-bit signed integer.
@@ -125,7 +125,7 @@ public:
      * @param[in]  input                        Input tensor. Data type supported: S32
      * @param[in]  bias                         Biases tensor. Only shared biases supported and it can be a nullptr if the biases addition is not required.
      *                                          Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input.
-     * @param[out] output                       Output tensor. Data type supported: Data type supported: QASYMM8
+     * @param[out] output                       Output tensor. Data type supported: QASYMM8
      * @param[in]  result_fixedpoint_multiplier Fixed point value to be multiplied to each element of the input matrix when once the result_offset has been add
      * @param[in]  result_shift                 Number of bits to shift right the result after the fixed point multiplication
      * @param[in]  result_offset_after_shift    Offset to be applied to result before converting it back to QASYMM8
@@ -140,7 +140,7 @@ public:
      * @param[in] input  Input tensor. It is the output of @ref CLGEMMLowpMatrixMultiplyCore function. Data type supported: S32
      * @param[in] bias   Biases tensor. Only shared biases supported and it can be a nullptr if the addition of biases is not required.
      *                   Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input.
-     * @param[in] output Output tensor. Data type supported: Data type supported: QASYMM8
+     * @param[in] output Output tensor. Data type supported: QASYMM8
      * @param[in] min    (Optional) Min value used to saturate down the output result before converting back to QASYMM8. Defaults to the minimum possible 32-bit signed integer.
      * @param[in] max    (Optional) Max value used to saturate up the output result before converting back to QASYMM8,
      *                            Along with @p min, this value can be used to implement "rectified linear unit" activation functions. Defaults to the maximum possible 32-bit signed integer.
@@ -186,7 +186,7 @@ public:
      * @param[in]  input                        Input tensor. Data type supported: S32
      * @param[in]  bias                         Biases tensor. Only shared biases supported and it can be a nullptr if the biases addition is not required.
      *                                          Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input.
-     * @param[out] output                       Output tensor. Data type supported: Data type supported: QASYMM8_SIGNED
+     * @param[out] output                       Output tensor. Data type supported: QASYMM8_SIGNED
      * @param[in]  result_fixedpoint_multiplier Fixed point value to be multiplied to each element of the input matrix when once the result_offset has been add
      * @param[in]  result_shift                 Number of bits to shift right the result after the fixed point multiplication
      * @param[in]  result_offset_after_shift    Offset to be applied to result before converting it back to QASYMM8_SIGNED
@@ -201,7 +201,7 @@ public:
      * @param[in] input  Input tensor. It is the output of @ref CLGEMMLowpMatrixMultiplyCore function. Data type supported: S32
      * @param[in] bias   Biases tensor. Only shared biases supported and it can be a nullptr if the addition of biases is not required.
      *                   Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input.
-     * @param[in] output Output tensor. Data type supported: Data type supported: QASYMM8_SIGNED
+     * @param[in] output Output tensor. Data type supported: QASYMM8_SIGNED
      * @param[in] min    (Optional) Min value used to saturate down the output result before converting back to QASYMM8_SIGNED. Defaults to the minimum possible 32-bit signed integer.
      * @param[in] max    (Optional) Max value used to saturate up the output result before converting back to QASYMM8_SIGNED. Defaults to 0
      *                            Along with @p min, this value can be used to implement "rectified linear unit" activation functions. Defaults to the maximum possible 32-bit signed integer.
@@ -228,7 +228,7 @@ public:
      * @param[in]  input      Input tensor. Data type supported: S32
      * @param[in]  bias       Biases tensor. Only shared biases supported and it can be a nullptr if the biases addition is not required.
      *                        Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input.
-     * @param[out] output     Output tensor. Data type supported: Data type supported: QASYMM8
+     * @param[out] output     Output tensor. Data type supported: QASYMM8
      * @param[in]  multiplier Float multiplier to be multiplied to each element of the input matrix
      * @param[in]  offset     Offset to be applied to result before converting it back to QASYMM8
      * @param[in]  min        (Optional) Min value used to saturate down the output result before converting back to QASYMM8. Defaults to the minimum possible 32-bit signed integer.
@@ -242,7 +242,7 @@ public:
      * @param[in] input  Input tensor. It is the output of @ref CLGEMMLowpMatrixMultiplyCore function. Data type supported: S32
      * @param[in] bias   Biases tensor. Only shared biases supported and it can be a nullptr if the addition of biases is not required.
      *                   Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input.
-     * @param[in] output Output tensor. Data type supported: Data type supported: QASYMM8
+     * @param[in] output Output tensor. Data type supported: QASYMM8
      * @param[in] min    (Optional) Min value used to saturate down the output result before converting back to QASYMM8. Defaults to the minimum possible 32-bit signed integer.
      * @param[in] max    (Optional) Max value used to saturate up the output result before converting back to QASYMM8,
      *                   Along with @p min, this value can be used to implement "rectified linear unit" activation functions. Defaults to the maximum possible 32-bit signed integer.
@@ -287,7 +287,7 @@ public:
      * @param[in]  input                        Input tensor. Data type supported: S32
      * @param[in]  bias                         Biases tensor. Only shared biases supported and it can be a nullptr if the biases addition is not required.
      *                                          Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input.
-     * @param[out] output                       Output tensor. Data type supported: Data type supported: QSYMM16
+     * @param[out] output                       Output tensor. Data type supported: QSYMM16
      * @param[in]  result_fixedpoint_multiplier Fixed point value to be multiplied to each element of the input matrix when once the result_offset has been add
      * @param[in]  result_shift                 Number of bits to shift right the result after the fixed point multiplication
      * @param[in]  min                          (Optional) Min value used to saturate down the output result before converting back to QSYMM16. Defaults to the minimum possible 32-bit signed integer.
@@ -301,7 +301,7 @@ public:
      * @param[in] input  Input tensor info. It is the output of @ref CLGEMMLowpMatrixMultiplyCore function. Data type supported: S32
      * @param[in] bias   Biases tensor info. Only shared biases supported and it can be a nullptr if the addition of biases is not required.
      *                   Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input.
-     * @param[in] output Output tensor info. Data type supported: Data type supported: QSYMM16
+     * @param[in] output Output tensor info. Data type supported: QSYMM16
      * @param[in] min    (Optional) Min value used to saturate down the output result before converting back to QSYMM16. Defaults to the minimum possible 32-bit signed integer.
      * @param[in] max    (Optional) Max value used to saturate up the output result before converting back to QSYMM16,
      *                            Along with @p min, this value can be used to implement "rectified linear unit" activation functions. Defaults to the maximum possible 32-bit signed integer.
@@ -310,5 +310,36 @@ public:
      */
     static Status validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min = std::numeric_limits<int32_t>::lowest(), int max = std::numeric_limits<int32_t>::max());
 };
+/** Basic function to execute GEMMLowpQuantizeDown kernels on CL.
+ *
+ *  This function calls the following CL kernels:
+ *
+ * -# @ref CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel
+ * -# @ref CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel
+*/
+class CLGEMMLowpOutputStage : public ICLSimpleFunction
+{
+public:
+    /** Initialise the kernel's inputs, output
+     *
+     * @param[in]  input  Input tensor. Data type supported: S32
+     * @param[in]  bias   Biases tensor. Only shared biases supported and it can be a nullptr if the biases addition is not required.
+     *                    Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input.
+     * @param[out] output Output tensor. Data type supported: QASYMM8/QASYMM8_SIGNED
+     * @param[in]  info   GEMMLowp output stage metadata.
+     */
+    void configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output, const GEMMLowpOutputStageInfo &info);
+    /** Static function to check if given info will lead to a valid configuration of @ref CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel
+     *
+     * @param[in] input  Input tensor. It is the output of @ref CLGEMMLowpMatrixMultiplyCore function. Data type supported: S32
+     * @param[in] bias   Biases tensor. Only shared biases supported and it can be a nullptr if the addition of biases is not required.
+     *                   Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input.
+     * @param[in] output Output tensor. Data type supported: QASYMM8/QASYMM8_SIGNED
+     * @param[in] info   GEMMLowp output stage metadata.
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const GEMMLowpOutputStageInfo &info);
+};
 } // namespace arm_compute
 #endif /*ARM_COMPUTE_CLGEMMLOWPOUTPUTSTAGE_H */
\ No newline at end of file