COMPMID-2968: Add support for QASYMM8_SIGNED in CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel

Signed-off-by: Sheri Zhang <sheri.zhang@arm.com>
Change-Id: I37e6e76dbd5546c0eaedfacd01ea905c37148e8a
Signed-off-by: Sheri Zhang <sheri.zhang@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/2861
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

4 files changed, 31 insertions, 31 deletions

diff --git a/arm_compute/core/CL/CLKernels.h b/arm_compute/core/CL/CLKernels.h
index f2e16ca139..b265aa2fe7 100644
--- a/arm_compute/core/CL/CLKernels.h
+++ b/arm_compute/core/CL/CLKernels.h
@@ -79,11 +79,11 @@
 #include "arm_compute/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMLowpOffsetContributionKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMLowpOffsetContributionOutputStageKernel.h"
+#include "arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ScaleKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.h"
-#include "arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMLowpReductionKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMMatrixAccumulateBiasesKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyKernel.h"
diff --git a/arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel.h b/arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel.h
index 900a8c3b5d..439f569d07 100644
--- a/arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel.h
+++ b/arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2019 ARM Limited.
+ * Copyright (c) 2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -21,8 +21,8 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
-#ifndef ARM_COMPUTE_CLGEMMLOWPQUANTIZEDOWNINT32TOUINT8SCALEBYFLOATKERNEL_H
-#define ARM_COMPUTE_CLGEMMLOWPQUANTIZEDOWNINT32TOUINT8SCALEBYFLOATKERNEL_H
+#ifndef ARM_COMPUTE_CLGEMMLOWPQUANTIZEDOWNINT32SCALEBYFLOATKERNEL_H
+#define ARM_COMPUTE_CLGEMMLOWPQUANTIZEDOWNINT32SCALEBYFLOATKERNEL_H
 
 #include "arm_compute/core/CL/ICLKernel.h"
 
@@ -31,9 +31,9 @@ namespace arm_compute
 // Forward declarations
 class ICLTensor;
 
-/** OpenCL kernel used to quantize down the int32 accumulator values of GEMMLowp to QASYMM8
+/** OpenCL kernel used to quantize down the int32 accumulator values of GEMMLowp to QASYMM8/QASYMM8_SIGNED
  *
- * This kernel takes a final int32 accumulator value (the output of @ref CLGEMMLowpMatrixMultiplyKernel), and processes it to obtain the final QASYMM8 value.
+ * This kernel takes a final int32 accumulator value (the output of @ref CLGEMMLowpMatrixMultiplyKernel), and processes it to obtain the final QASYMM8/QASYMM8_SIGNED value.
  * The following computations will be performed by the kernel:
  *
  *  -# Compute fixed point multiplication between each entry of input by result_fixedpoint_multiplier
@@ -41,47 +41,43 @@ class ICLTensor;
  *  -# Requantize
  *  -# Add offset to each result
  *  -# Clamp the value between the specified min and max bounds
- *  -# Clamp the resulting int32 values to the [0..255] range and cast to QASYMM8.
+ *  -# Clamp the resulting int32 values to
+ *      - to the [0..255] range and cast to QASYMM8.
+ *      - to the [-128..127] range and cast to QASYMM8_SIGNED.
  */
-class CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel : public ICLKernel
+class CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel : public ICLKernel
 {
 public:
     /** Constructor */
-    CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel();
+    CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel();
     /** Prevent instances of this class from being copied (As this class contains pointers)*/
-    CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel(const CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel &) = delete;
+    CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel(const CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel &) = delete;
     /** Prevent instances of this class from being copied (As this class contains pointers)*/
-    CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel &operator=(const CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel &) = delete;
+    CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel &operator=(const CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel &) = delete;
     /** Allow instances of this class to be moved */
-    CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel(CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel &&) = default;
+    CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel(CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel &&) = default;
     /** Allow instances of this class to be moved */
-    CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel &operator=(CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel &&) = default;
+    CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel &operator=(CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel &&) = default;
     /** Initialise the kernel's input and 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: 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
-     * @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
+     * @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/QASYMM8_SIGNED
+     * @param[in]  info   Output stage info. Used to pass the quantized output data type
      */
-    void configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output, float multiplier, int offset, int min = 0, int max = 0);
-    /** Static function to check if given info will lead to a valid configuration of @ref CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel
+    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 CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel
      *
      * @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[in] output Output tensor. Data type supported: Data type supported: QASYMM8
-     * @param[in] min    (Optional) Min value used to saturate down the output result before converting back to QASYMM8
-     * @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
+     * @param[in] output Output tensor. Data type supported: Data type supported: QASYMM8/QASYMM8_SIGNED
+     * @param[in] info   Output stage info. Used to pass the quantized output data type
      *
      * @return a status
      */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min = 0, int max = 0);
+    static Status validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const GEMMLowpOutputStageInfo *info);
 
     // Inherited methods overridden:
     void run(const Window &window, cl::CommandQueue &queue) override;
@@ -92,4 +88,4 @@ private:
     ICLTensor       *_output;
 };
 } // namespace arm_compute
-#endif /* ARM_COMPUTE_CLGEMMLOWPQUANTIZEDOWNINT32TOUINT8SCALEBYFLOATKERNEL_H */
+#endif /* ARM_COMPUTE_CLGEMMLOWPQUANTIZEDOWNINT32SCALEBYFLOATKERNEL_H */
diff --git a/arm_compute/core/Types.h b/arm_compute/core/Types.h
index 711b68f236..37a9679a21 100644
--- a/arm_compute/core/Types.h
+++ b/arm_compute/core/Types.h
@@ -1956,6 +1956,7 @@ struct GEMMLowpOutputStageInfo
     int32_t                 gemmlowp_max_bound{ std::numeric_limits<int32_t>::max() };    /**< GEMMLowp max value used to saturate down the output result before converting back to QASYMM8 */
     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 */
+    float                   gemmlowp_real_multiplier{ 0 };                                /**< GEMMLowp output stage real 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 */
 };
diff --git a/arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h b/arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h
index 184d827d4b..05cffa6680 100644
--- a/arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h
+++ b/arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h
@@ -217,7 +217,7 @@ public:
  *
  *  This function calls the following OpenCL kernels:
  *
- * -# @ref CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel
+ * -# @ref CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel
  *
  * @note The function accepts also 2 optional input arguments (min and max) which can be used to implement "rectified linear unit" activation functions
  *       after the result is shifted right by result_shift
@@ -237,6 +237,7 @@ public:
      * @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.
      */
+    ARM_COMPUTE_DEPRECATED_REL(20.05)
     void configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output, float multiplier, int offset, int min = std::numeric_limits<int32_t>::lowest(),
                    int max = std::numeric_limits<int32_t>::max());
     /** Static function to check if given info will lead to a valid configuration of @ref CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint
@@ -251,6 +252,7 @@ public:
      *
      * @return a status
      */
+    ARM_COMPUTE_DEPRECATED_REL(20.05)
     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 CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint on OpenCL.
@@ -317,6 +319,7 @@ public:
  *  This function calls the following CL kernels:
  *
  * -# @ref CLGEMMLowpQuantizeDownInt32ScaleKernel
+ * -# @ref CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel
  * -# @ref CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel
  * -# @ref CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel
 */