COMPMID-2790: Add support for QASYMM8_SIGNED in CLGEMMLowpMatrixMultiplyCore

Change-Id: Ifdaeb53c512ba697f174649c026075010f54f628
Signed-off-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/2472
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Sang-Hoon Park <sang-hoon.park@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Giuseppe Rossini <giuseppe.rossini@arm.com>

5 files changed, 24 insertions, 22 deletions

diff --git a/arm_compute/core/CL/kernels/CLGEMMLowpMatrixMultiplyNativeKernel.h b/arm_compute/core/CL/kernels/CLGEMMLowpMatrixMultiplyNativeKernel.h
index db4bf3664a..e1191f265e 100644
--- a/arm_compute/core/CL/kernels/CLGEMMLowpMatrixMultiplyNativeKernel.h
+++ b/arm_compute/core/CL/kernels/CLGEMMLowpMatrixMultiplyNativeKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019 ARM Limited.
+ * Copyright (c) 2019-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -30,7 +30,7 @@ namespace arm_compute
 {
 class ICLTensor;
 
-/** OpenCL kernel to multiply matrices with QASYMM8 data type */
+/** OpenCL kernel to multiply matrices with QASYMM8/QASYMM8_SIGNED data type */
 class CLGEMMLowpMatrixMultiplyNativeKernel : public ICLKernel
 {
 public:
@@ -46,7 +46,7 @@ public:
     CLGEMMLowpMatrixMultiplyNativeKernel &operator=(CLGEMMLowpMatrixMultiplyNativeKernel &&) = default;
     /** Initialise the kernel's input and output.
      *
-     * @param[in]  input0    Input tensor containing the LHS matrix. Data type supported: QASYMM8
+     * @param[in]  input0    Input tensor containing the LHS matrix. Data type supported: QASYMM8/QASYMM8_SIGNED
      * @param[in]  input1    Input tensor containing the RHS matrix. Data type supported: same as @p input0
      * @param[out] output    Output tensor to store the result of matrix multiplication. Data type supported: S32
      * @param[in]  lhs_info  LHS matrix information used to retrieve the number of rows to be processed by each thread
@@ -60,7 +60,7 @@ public:
     void configure(const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, const GEMMReshapeInfo &gemm_info);
     /** Static function to check if given info will lead to a valid configuration of @ref CLGEMMLowpMatrixMultiplyNativeKernel
      *
-     * @param[in] input0    Input tensor info for the LHS matrix. Data type supported: QASYMM8
+     * @param[in] input0    Input tensor info for the LHS matrix. Data type supported: QASYMM8/QASYMM8_SIGNED
      * @param[in] input1    Input tensor info for the RHS matrix. Data type supported: same as @p input0
      * @param[in] output    Output tensor info. Data type supported: S32
      * @param[in] lhs_info  LHS matrix information used to retrieve the number of rows to be processed by each thread
diff --git a/arm_compute/core/CL/kernels/CLGEMMLowpOffsetContributionOutputStageKernel.h b/arm_compute/core/CL/kernels/CLGEMMLowpOffsetContributionOutputStageKernel.h
index 44a91fef18..4094bc681e 100644
--- a/arm_compute/core/CL/kernels/CLGEMMLowpOffsetContributionOutputStageKernel.h
+++ b/arm_compute/core/CL/kernels/CLGEMMLowpOffsetContributionOutputStageKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2019 ARM Limited.
+ * Copyright (c) 2018-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -35,6 +35,7 @@ class ICLTensor;
  * This kernel takes a final int32 accumulator value (the output of @ref CLGEMMLowpMatrixMultiplyKernel), adds to it the offset contribution
  * of matrix A and matrix B and performs the output stage defined by the output_stage argument
  *
+ * @note For quantized computations the output data type for auto-initialization must be passed as part of the @ref GEMMLowpOutputStageInfo.
  */
 class CLGEMMLowpOffsetContributionOutputStageKernel : public ICLKernel
 {
@@ -58,7 +59,7 @@ public:
      *                                Note: vector_sum_row can be a nullptr in case b_offset = 0. Data type supported: same as @p mm_result
      * @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: QASYMM8.
+     * @param[out] output             Output tensor. Data type supported: QASYMM8/QASYMM8_SIGNED.
      * @param[in]  k                  Number of matrix A columns or Matrix B rows
      * @param[in]  a_offset           Offset to be added to each element of the matrix A.
      * @param[in]  b_offset           Offset to be added to each element of the matrix B.
@@ -72,14 +73,14 @@ public:
                    const GEMMLowpOutputStageInfo &output_stage, const ICLTensor *output_multipliers, const ICLTensor *output_shifts);
     /** Static function to check if given info will lead to a valid configuration of @ref CLGEMMLowpOffsetContributionKernel
      *
-     * @param[in] mm_result          Input tensor containing the result of @ref CLGEMMLowpOffsetContributionKernel. Data type supported: S32 or QASYMM8 if output_stage != NONE
+     * @param[in] mm_result          Input tensor containing the result of @ref CLGEMMLowpOffsetContributionKernel. Data type supported: S32
      * @param[in] vector_sum_col     Input row-vector of sums of all the entries in each column of matrix B.
      *                               Note: vector_sum_col can be a nullptr in case a_offset = 0. Data type supported: same as @p mm_result
      * @param[in] vector_sum_row     Input row-vector of sums of all the entries in each row of matrix A.
      *                               Note: vector_sum_row can be a nullptr in case b_offset = 0. Data type supported: same as @p mm_result
      * @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.
+     * @param[in] output             Output tensor. Data type supported: QASYMM8/QASYMM8_SIGNED.
      * @param[in] a_offset           Offset to be added to each element of the matrix A.
      * @param[in] b_offset           Offset to be added to each element of the matrix B.
      * @param[in] output_stage       GEMMLowp output stage info
diff --git a/arm_compute/core/CL/kernels/CLGEMMLowpReductionKernel.h b/arm_compute/core/CL/kernels/CLGEMMLowpReductionKernel.h
index c42b218dc1..4e52a8029e 100644
--- a/arm_compute/core/CL/kernels/CLGEMMLowpReductionKernel.h
+++ b/arm_compute/core/CL/kernels/CLGEMMLowpReductionKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2019 ARM Limited.
+ * Copyright (c) 2017-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -67,13 +67,13 @@ class CLGEMMLowpMatrixAReductionKernel : public ICLGEMMLowpReductionKernel
 public:
     /** Initialise the kernel's input and output.
      *
-     * @param[in]  mtx_a          Input tensor. Data type supported: QASYMM8
+     * @param[in]  mtx_a          Input tensor. Data type supported: QASYMM8/QASYMM8_SIGNED
      * @param[out] vector_sum_row Output row-vector of sums of all the entries in each row of mtx_a. Data type supported: S32
      */
     void configure(const ICLTensor *mtx_a, ICLTensor *vector_sum_row) override;
     /** Static function to check if given info will lead to a valid configuration of @ref CLGEMMLowpMatrixAReductionKernel
      *
-     * @param[in] mtx_a          Input tensor. Data type supported: QASYMM8
+     * @param[in] mtx_a          Input tensor. Data type supported: QASYMM8/QASYMM8_SIGNED
      * @param[in] vector_sum_row Output row-vector of sums of all the entries in each row of mtx_a. Data type supported: S32
      *
      * @return a status
@@ -94,13 +94,13 @@ class CLGEMMLowpMatrixBReductionKernel : public ICLGEMMLowpReductionKernel
 public:
     /** Initialise the kernel's input and output.
      *
-     * @param[in]  mtx_b          Input tensor. Data type supported: Data type supported: QASYMM8
+     * @param[in]  mtx_b          Input tensor. Data type supported: Data type supported: QASYMM8/QASYMM8_SIGNED
      * @param[out] vector_sum_col Output row-vector of sums of all the entries in each column of mtx_b. Data type supported: S32
      */
     void configure(const ICLTensor *mtx_b, ICLTensor *vector_sum_col) override;
     /** Static function to check if given info will lead to a valid configuration of @ref CLGEMMLowpMatrixBReductionKernel
      *
-     * @param[in] mtx_b          Input tensor. Data type supported: Data type supported: QASYMM8
+     * @param[in] mtx_b          Input tensor. Data type supported: Data type supported: QASYMM8/QASYMM8_SIGNED
      * @param[in] vector_sum_col Output row-vector of sums of all the entries in each column of mtx_b. Data type supported: S32
      *
      * @return a status
diff --git a/arm_compute/core/Types.h b/arm_compute/core/Types.h
index 901d080b0e..cbcd3fa783 100644
--- a/arm_compute/core/Types.h
+++ b/arm_compute/core/Types.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2019 ARM Limited.
+ * Copyright (c) 2016-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -1866,10 +1866,10 @@ struct DepthwiseConvolutionReshapeInfo
 /** GEMMLowp output stage type */
 enum class GEMMLowpOutputStageType
 {
-    NONE,                     /**< No quantization to uint8 */
-    QUANTIZE_DOWN,            /**< Quantize to uint8 using an integer multiplication */
-    QUANTIZE_DOWN_FIXEDPOINT, /**< Quantize to uint8 using a fixed point multiplication */
-    QUANTIZE_DOWN_FLOAT       /**< Quantize to uint8 using a floating point multiplication */
+    NONE,                     /**< No quantization */
+    QUANTIZE_DOWN,            /**< Quantize using an integer multiplication */
+    QUANTIZE_DOWN_FIXEDPOINT, /**< Quantize using a fixed point multiplication */
+    QUANTIZE_DOWN_FLOAT       /**< Quantize using a floating point multiplication */
 };
 
 /** GEMMLowp output stage info */
@@ -1884,6 +1884,7 @@ struct GEMMLowpOutputStageInfo
     std::vector<int32_t>    gemmlowp_multipliers{};                /**< GEMMLowp output stage multiplier used for quantizing to QASYMM8 */
     std::vector<int32_t>    gemmlowp_shifts{};                     /**< GEMMLowp output stage multiplier used for quantizing to QASYMM8 */
     bool                    is_quantized_per_channel{ false };     /**< GEMMLowp quantized per-channel flag */
+    DataType                output_data_type{ DataType::UNKNOWN }; /**< Output tensor data type to use if the output is not initialized */
 };
 
 /** GEMM LHS (Left Hand Side) matrix information */
diff --git a/arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h b/arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h
index 770ef0b287..66c5e9ee46 100644
--- a/arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h
+++ b/arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2019 ARM Limited.
+ * Copyright (c) 2017-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -76,10 +76,10 @@ public:
      *  -# Compute the matrix product of the resulting a * b in int32.
      *  -# Quantize to uint8 if gemm_info.gemmlowp_output_stage != NONE
      *
-     * @param[in]  a         First input tensor  (Matrix A). Data type supported: QASYMM8.
+     * @param[in]  a         First input tensor  (Matrix A). Data type supported: QASYMM8/QASYMM8_SIGNED.
      * @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. Data type supported: S32
-     * @param[out] output    Output tensor. Data type supported: S32 or QASYMM8 if gemm_info.gemmlowp_output_stage != NONE
+     * @param[out] output    Output tensor. Data type supported: S32 or QASYMM8/QASYMM8_SIGNED if gemm_info.gemmlowp_output_stage != NONE
      * @param[in]  gemm_info (Optional) Specifies if the matrix A and/or matrix B have been reshaped and
      *                       if the reshape of matrix B should be executed only for the first run
      */
@@ -89,7 +89,7 @@ public:
      * @param[in] a         First input tensor info (Matrix A). Data type supported: QASYMM8.
      * @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. Data type supported: S32
-     * @param[in] output    Output tensor info. Data type supported: S32 or QASYMM8 if gemm_info.gemmlowp_output_stage != NONE
+     * @param[in] output    Output tensor info. Data type supported: S32 or QASYMM8/QASYMM8_SIGNED if gemm_info.gemmlowp_output_stage != NONE
      * @param[in] gemm_info (Optional) Specifies if the matrix A and/or matrix B have been reshaped and
      *                      if the reshape of matrix B should be executed only for the first run
      *