COMPMID-2406: Create a new GEMMLowpQuantizeDownInt32ToInt16ScaleKernel for NEON

Change-Id: I3f3e247728fd6dafca066e41835f0ef9442d9b7a
Signed-off-by: giuros01 <giuseppe.rossini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1379
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

4 files changed, 283 insertions, 2 deletions

diff --git a/arm_compute/core/NEON/NEKernels.h b/arm_compute/core/NEON/NEKernels.h
index 206ff06f8a..4023d82107 100644
--- a/arm_compute/core/NEON/NEKernels.h
+++ b/arm_compute/core/NEON/NEKernels.h
@@ -78,6 +78,7 @@
 #include "arm_compute/core/NEON/kernels/NEGEMMLowpMatrixMultiplyKernel.h"
 #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/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/NESymm.h b/arm_compute/core/NEON/NESymm.h
new file mode 100644
index 0000000000..0479753426
--- /dev/null
+++ b/arm_compute/core/NEON/NESymm.h
@@ -0,0 +1,106 @@
+/*
+ * 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_NESYMM_H__
+#define __ARM_COMPUTE_NESYMM_H__
+
+#include "NEAsymm.h"
+#include <arm_neon.h>
+
+namespace arm_compute
+{
+/** Performs final quantization step on 8 signed 16-bit elements
+ *
+ * @tparam is_bounded_relu Specified if a fused bounded relu should be applied
+ *
+ * @param[in] in_s32                       Input to be quantized.
+ * @param[in] result_fixedpoint_multiplier Result multiplier parameter
+ * @param[in] result_shift                 Result shift parameter
+ * @param[in] min_s16                      Relu lower bound
+ * @param[in] max_s16                      Relu upper bound
+ *
+ * @return Quantized values
+ */
+template <bool is_bounded_relu>
+int16x8_t finalize_quantization_int16(int32x4x2_t &in_s32,
+                                      int          result_fixedpoint_multiplier,
+                                      int32_t      result_shift,
+                                      int16x8_t    min_s16,
+                                      int16x8_t    max_s16)
+{
+    // 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);
+
+    // 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);
+
+    // Convert S32 to S16
+    int16x8_t out_s16 = vcombine_s16(vqmovn_s32(in_s32.val[0]), vqmovn_s32(in_s32.val[1]));
+
+    if(is_bounded_relu)
+    {
+        out_s16 = vmaxq_s16(out_s16, min_s16);
+        out_s16 = vminq_s16(out_s16, max_s16);
+    }
+
+    return out_s16;
+}
+
+/** Performs final quantization step on single signed 16-bit element
+ *
+ * @tparam is_bounded_relu Specified if a fused bounded relu should be applied
+ *
+ * @param[in] in_value                     Input to be quantized.
+ * @param[in] result_fixedpoint_multiplier Result multiplier parameter
+ * @param[in] result_shift                 Result shift parameter
+ * @param[in] min_s16                      Relu lower bound
+ * @param[in] max_s16                      Relu upper bound
+ *
+ * @return Quantized values
+ */
+template <bool is_bounded_relu>
+inline int16_t finalize_quantization_int16(int32_t in_value, int result_fixedpoint_multiplier,
+                                           int32_t result_shift, int16_t min_s16, int16_t max_s16)
+{
+    int32x4_t in_s32 = vdupq_n_s32(in_value);
+
+    // Fixed point multiplication with vector saturating rounding doubling multiply high with scalar
+    in_value = vgetq_lane_s32(vqrdmulhq_n_s32(in_s32, result_fixedpoint_multiplier), 0);
+
+    // Shift value by result_shift_s32
+    in_value = rounding_divide_by_pow2(in_value, result_shift);
+
+    // Bound the result
+    int16_t out_s16 = static_cast<int16_t>(std::max<int32_t>(-32768, std::min<int32_t>(32767, in_value)));
+
+    if(is_bounded_relu)
+    {
+        out_s16 = static_cast<int16_t>(std::max(min_s16, std::min(max_s16, out_s16)));
+    }
+
+    return out_s16;
+}
+} // namespace arm_compute
+#endif // __ARM_COMPUTE_NESYMM_H__
diff --git a/arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.h b/arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.h
new file mode 100644
index 0000000000..806c0e4c2f
--- /dev/null
+++ b/arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.h
@@ -0,0 +1,116 @@
+/*
+ * 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_NEGEMMLOWPQUANTIZEDOWNINT32TOINT16SCALEBYFIXEDPOINTKERNEL_H__
+#define __ARM_COMPUTE_NEGEMMLOWPQUANTIZEDOWNINT32TOINT16SCALEBYFIXEDPOINTKERNEL_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 QSYMM16
+ *
+ * This kernel takes a final int32 accumulator value (the output of @ref NEGEMMLowpMatrixMultiplyKernel), and processes it to obtain the final QSYMM16 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
+ *  -# Clamp the value between the specified min and max bounds
+ *  -# Clamp the resulting int32 values to the [-32768, 32767] range and cast to QSYMM16.
+ *
+ */
+class NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel : public INEKernel
+{
+public:
+    const char *name() const override
+    {
+        return "NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel";
+    }
+    /** Constructor */
+    NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers)*/
+    NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel(const NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers)*/
+    NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel &operator=(const NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel(NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel &operator=(NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel &&) = 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: 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                 Integer value used to round to nearest division by a power-of-two 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 0.
+     * @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 0.
+     */
+    void configure(const ITensor *input, const ITensor *bias, ITensor *output, int result_fixedpoint_multiplier, int result_shift, int min = 0, int max = 0);
+    /** Static function to check if given info will lead to a valid configuration of @ref NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel
+     *
+     * @param[in] input  Input tensor info. Data type supported: S32
+     * @param[in] bias   Biases tensor info. Only shared biases supported and it can be a nullptr if the biases addition is not required.
+     *                   Biases are 1D tensor info 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] min    (Optional) Min value used to saturate down the output result before converting back to QSYMM16. Defaults to 0.
+     * @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 0.
+     *
+     * @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 NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel
+     *
+     * @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 NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel functions
+     *
+     * @param[in] window Region on which to execute the kernel.
+     */
+    using QuantizeDownFunctionPtr = void (NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::*)(const Window &window);
+
+    QuantizeDownFunctionPtr _func;
+    const ITensor          *_input;
+    const ITensor          *_bias;
+    ITensor                *_output;
+    int                     _result_fixedpoint_multiplier;
+    int                     _result_shift;
+    int                     _min;
+    int                     _max;
+};
+} // namespace arm_compute
+#endif /* __ARM_COMPUTE_NEGEMMLOWPQUANTIZEDOWNINT32TOINT16SCALEBYFIXEDPOINTKERNEL_H__ */
diff --git a/arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h b/arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h
index 77bfb98ba6..5ece753660 100644
--- a/arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h
+++ b/arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -147,5 +147,63 @@ public:
      */
     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:
+ *
+ *  result_fixedpoint_multiplier, result_shift
+ *
+ * The final result is:
+ *
+ * (FixedPointMul(input[i][k], result_fixedpoint_multiplier) >> result_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 NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel
+ *
+ * @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 NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint : 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: 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 0.
+     * @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 0.
+     */
+    void configure(const ITensor *input, const ITensor *bias, ITensor *output, int result_fixedpoint_multiplier, int result_shift, int min = 0, int max = 0);
+    /** Static function to check if given info will lead to a valid configuration of @ref NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint
+     *
+     * @param[in] input  Input tensor info. It is the output of @ref NEGEMMLowpMatrixMultiplyCore 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] min    (Optional) Min value used to saturate down the output result before converting back to QSYMM16. Defaults to 0.
+     * @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 0.
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min = 0, int max = 0);
+};
 } // namespace arm_compute
-#endif /*__ARM_COMPUTE_NEGEMMLOWPOUTPUTSTAGE_H__ */
\ No newline at end of file
+#endif /*__ARM_COMPUTE_NEGEMMLOWPOUTPUTSTAGE_H__ */