COMPMID-2805: Add QASYMM8_SIGNED support in NEGEMMLowpOutputStage

Add support from requantizing down from S32 to Int8 with fixed point
requantization. This involves the following:
- Compute fixed point multiplication between each entry of input by
  result_fixedpoint_multiplier
- Add bias to final result if bias tensor is not a nullptr
- Round to nearest division by a power-of-two using result_shift
- Add offset to each result
- Clamp the value between the specified min and max bounds
- Cast to int8 data type

Change-Id: I641b3fac0833c568d8565ccb859bbc561a24c17d
Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-on: https://review.mlplatform.org/c/2340
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

5 files changed, 241 insertions, 2 deletions

diff --git a/arm_compute/core/NEON/NEAsymm.h b/arm_compute/core/NEON/NEAsymm.h
index c75a58046b..40bdd0f5bf 100644
--- a/arm_compute/core/NEON/NEAsymm.h
+++ b/arm_compute/core/NEON/NEAsymm.h
@@ -115,6 +115,66 @@ uint8x16_t finalize_quantization(int32x4x4_t &in_s32,
     return out_u8;
 }
 
+/** Performs final quantization step on 16 elements
+ *
+ * @tparam is_bounded_relu Specified if a fused bounded relu should be applied
+ *
+ * @param in_s32                        Input to be quantized.
+ * @param result_fixedpoint_multiplier  Result multiplier parameter
+ * @param result_shift                  Result shift parameter
+ * @param result_offset_after_shift_s32 Result offset parameter
+ * @param min_s8                        Relu lower bound
+ * @param max_s8                        Relu upper bound
+ *
+ * @return Quantized values
+ */
+template <bool is_bounded_relu>
+int8x16_t finalize_quantization(int32x4x4_t &in_s32,
+                                int          result_fixedpoint_multiplier,
+                                int32_t      result_shift,
+                                int32x4_t    result_offset_after_shift_s32,
+                                int8x16_t    min_s8,
+                                int8x16_t    max_s8)
+{
+    // Fixed point multiplication with vector saturating rounding doubling multiply high with scalar
+    in_s32.val[0] = vqrdmulhq_n_s32(in_s32.val[0], result_fixedpoint_multiplier);
+    in_s32.val[1] = vqrdmulhq_n_s32(in_s32.val[1], result_fixedpoint_multiplier);
+    in_s32.val[2] = vqrdmulhq_n_s32(in_s32.val[2], result_fixedpoint_multiplier);
+    in_s32.val[3] = vqrdmulhq_n_s32(in_s32.val[3], result_fixedpoint_multiplier);
+
+    // Round to the nearest division by a power-of-two using result_shift_s32
+    in_s32.val[0] = rounding_divide_by_pow2(in_s32.val[0], result_shift);
+    in_s32.val[1] = rounding_divide_by_pow2(in_s32.val[1], result_shift);
+    in_s32.val[2] = rounding_divide_by_pow2(in_s32.val[2], result_shift);
+    in_s32.val[3] = rounding_divide_by_pow2(in_s32.val[3], result_shift);
+
+    // Add the offset terms
+    in_s32.val[0] = vaddq_s32(in_s32.val[0], result_offset_after_shift_s32);
+    in_s32.val[1] = vaddq_s32(in_s32.val[1], result_offset_after_shift_s32);
+    in_s32.val[2] = vaddq_s32(in_s32.val[2], result_offset_after_shift_s32);
+    in_s32.val[3] = vaddq_s32(in_s32.val[3], result_offset_after_shift_s32);
+
+    // Convert S32 to S16
+    const int16x8x2_t in_s16 =
+    {
+        {
+            vcombine_s16(vqmovn_s32(in_s32.val[0]), vqmovn_s32(in_s32.val[1])),
+            vcombine_s16(vqmovn_s32(in_s32.val[2]), vqmovn_s32(in_s32.val[3]))
+        }
+    };
+
+    // Convert S16 to S8
+    int8x16_t out_s8 = vcombine_s8(vqmovn_s16(in_s16.val[0]), vqmovn_s16(in_s16.val[1]));
+
+    if(is_bounded_relu)
+    {
+        out_s8 = vmaxq_s8(out_s8, min_s8);
+        out_s8 = vminq_s8(out_s8, max_s8);
+    }
+
+    return out_s8;
+}
+
 /** Performs final quantization step on 16 elements for symmetric quantization
  *
  * @tparam is_bounded_relu Specified if a fused bounded relu should be applied
diff --git a/arm_compute/core/NEON/NEKernels.h b/arm_compute/core/NEON/NEKernels.h
index aa46a346e9..05485d847a 100644
--- a/arm_compute/core/NEON/NEKernels.h
+++ b/arm_compute/core/NEON/NEKernels.h
@@ -80,6 +80,7 @@
 #include "arm_compute/core/NEON/kernels/NEGEMMLowpOffsetContributionKernel.h"
 #include "arm_compute/core/NEON/kernels/NEGEMMLowpOffsetContributionOutputStageKernel.h"
 #include "arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.h"
+#include "arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.h"
 #include "arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.h"
 #include "arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.h"
 #include "arm_compute/core/NEON/kernels/NEGEMMLowpReductionKernel.h"
diff --git a/arm_compute/core/NEON/kernels/NEGEMMLowpOffsetContributionOutputStageKernel.h b/arm_compute/core/NEON/kernels/NEGEMMLowpOffsetContributionOutputStageKernel.h
index c284ca5c5f..dadc5c221b 100644
--- a/arm_compute/core/NEON/kernels/NEGEMMLowpOffsetContributionOutputStageKernel.h
+++ b/arm_compute/core/NEON/kernels/NEGEMMLowpOffsetContributionOutputStageKernel.h
@@ -83,7 +83,7 @@ public:
      * @param[in]  vector_sum_row Input row-vector of sums of all the entries in each row of matrix A.
      * @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 mm_result.
-     * @param[out] output         Output tensor containing the final quantized result. Data type supported: QASYMM8
+     * @param[out] output         Output tensor containing the final quantized result. 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.
@@ -100,7 +100,7 @@ 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 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 mm_result.
-     * @param[in] output         Output tensor info containing the final quantized result. Data type supported: QASYMM8
+     * @param[in] output         Output tensor info containing the final quantized result. 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, providing the type of quantization and the necessary parameters.
diff --git a/arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.h b/arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.h
new file mode 100644
index 0000000000..2b3657c728
--- /dev/null
+++ b/arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.h
@@ -0,0 +1,119 @@
+/*
+ * Copyright (c) 2019 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef ARM_COMPUTE_NEGEMMLOWPQUANTIZEDOWNINT32TOINT8SCALEBYFIXEDPOINTKERNEL_H
+#define ARM_COMPUTE_NEGEMMLOWPQUANTIZEDOWNINT32TOINT8SCALEBYFIXEDPOINTKERNEL_H
+
+#include "arm_compute/core/NEON/INEKernel.h"
+
+namespace arm_compute
+{
+class ITensor;
+
+/** NEON kernel used to quantize down the int32 accumulator values of GEMMLowp to QASYMM8_SIGNED
+ *
+ * This kernel takes a final int32 accumulator value (the output of @ref NEGEMMLowpMatrixMultiplyKernel), and processes it to obtain the final 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
+ *  -# Add bias to final result if bias tensor is not a nullptr
+ *  -# Round to nearest division by a power-of-two using result_shift
+ *  -# Add offset to each result
+ *  -# Clamp the value between the specified min and max bounds
+ *  -# Clamp the resulting int32 values to the [-128..127] range and cast to QASYMM8_SIGNED.
+ *
+ */
+class NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel : public INEKernel
+{
+public:
+    const char *name() const override
+    {
+        return "NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel";
+    }
+    /** Constructor */
+    NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers)*/
+    NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel(const NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers)*/
+    NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel &operator=(const NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel(NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel &operator=(NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel &&) = 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_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                 Integer value used to round to nearest division by a power-of-two 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
+     * @param[in]  min                          (Optional) Min value used to saturate down the output result before converting back to QASYMM8_SIGNED
+     * @param[in]  max                          (Optional) Max value used to saturate up the output result before converting back to QASYMM8_SIGNED,
+     *                                          Along with @p min, this value can be used to implement "rectified linear unit" activation functions
+     */
+    void configure(const ITensor *input, const ITensor *bias, ITensor *output, int result_fixedpoint_multiplier, int result_shift, int result_offset_after_shift, int min = 0, int max = 0);
+    /** Static function to check if given info will lead to a valid configuration of @ref NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel
+     *
+     * @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_SIGNED
+     * @param[in] min    (Optional) Min value used to saturate down the output result before converting back to QASYMM8_SIGNED
+     * @param[in] max    (Optional) Max value used to saturate up the output result before converting back to QASYMM8_SIGNED,
+     *                            Along with @p min, this value can be used to implement "rectified linear unit" activation functions
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min = 0, int max = 0);
+
+    // Inherited methods overridden:
+    void run(const Window &window, const ThreadInfo &info) override;
+
+private:
+    /** Template function to run the NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel
+     *
+     * @param[in] window Region on which to execute the kernel. (Must be a valid region of the window returned by window()).
+     */
+    template <bool is_bounded_relu>
+    void run(const Window &window);
+
+    /** Common signature for all the specialised NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel functions
+     *
+     * @param[in] window Region on which to execute the kernel.
+     */
+    using QuantizeDownFunctionPtr = void (NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::*)(const Window &window);
+
+    QuantizeDownFunctionPtr _func;
+    const ITensor          *_input;
+    const ITensor          *_bias;
+    ITensor                *_output;
+    int                     _result_fixedpoint_multiplier;
+    int                     _result_shift;
+    int                     _result_offset_after_shift;
+    int                     _min;
+    int                     _max;
+};
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_NEGEMMLOWPQUANTIZEDOWNINT32TOINT8SCALEBYFIXEDPOINTKERNEL_H */
diff --git a/arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h b/arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h
index 5ece753660..1a65f3b6ce 100644
--- a/arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h
+++ b/arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h
@@ -147,6 +147,65 @@ public:
      */
     static Status validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min = 0, int max = 0);
 };
+/** Basic function to execute NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint on NEON.
+ *
+ *  NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint depends on 3 parameters:
+ *
+ *  result_fixedpoint_multiplier, result_shift, result_offset_after_shift
+ *
+ * The final result is:
+ *
+ * (FixedPointMul(input[i][k], result_fixedpoint_multiplier) >> result_shift) + result_offset_after_shift
+ *
+ * where FixedPointMul(x, y) is the nearest integer to the following
+ * mathematical expression, evaluated without overflow or intermediate rounding:
+ *
+ * (x * y) / 2^31
+ *
+ * For more information: https://github.com/google/gemmlowp/blob/master/public/output_stages.h#L68
+ *
+ * In case the bias tensor is provided, the final result is:
+ *
+ * ((FixedPointMul(input[i][k] + bias[k], result_fixedpoint_multiplier)) >> result_shift) + result_offset_after_shift
+ *
+ *  This function calls the following NEON kernels:
+ *
+ * -# @ref NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel
+ *
+ * @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
+*/
+class NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint : public INESimpleFunctionNoBorder
+{
+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: 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
+     * @param[in]  min                          (Optional) Min value used to saturate down the output result before converting back to QASYMM8_SIGNED
+     * @param[in]  max                          (Optional) Max value used to saturate up the output result before converting back to QASYMM8_SIGNED,
+     *                                          Along with @p min, this value can be used to implement "rectified linear unit" activation functions
+     */
+    void configure(const ITensor *input, const ITensor *bias, ITensor *output, int result_fixedpoint_multiplier, int result_shift, int result_offset_after_shift, int min = 0, int max = 0);
+    /** Static function to check if given info will lead to a valid configuration of @ref NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint
+     *
+     * @param[in] input  Input tensor. It is the output of @ref NEGEMMLowpMatrixMultiplyCore 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] min    (Optional) Min value used to saturate down the output result before converting back to QASYMM8_SIGNED
+     * @param[in] max    (Optional) Max value used to saturate up the output result before converting back to QASYMM8_SIGNED,
+     *                            Along with @p min, this value can be used to implement "rectified linear unit" activation functions
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min = 0, int max = 0);
+};
 /** Basic function to execute NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint on NEON.
  *
  *  NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint depends on 2 parameters: