COMPMID-1694: Fuse offset contribution with the output stage when we use NEGEMMLowpMatrixMultiplyCore
Change-Id: Ic1a681e4cc03e1eba3bf8485d9cdb17b3e926047
Signed-off-by: giuros01 <giuseppe.rossini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/561
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
diff --git a/arm_compute/core/CL/kernels/CLGEMMLowpOffsetContributionKernel.h b/arm_compute/core/CL/kernels/CLGEMMLowpOffsetContributionKernel.h
index e6b7917..c66a470 100644
--- a/arm_compute/core/CL/kernels/CLGEMMLowpOffsetContributionKernel.h
+++ b/arm_compute/core/CL/kernels/CLGEMMLowpOffsetContributionKernel.h
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -58,7 +58,7 @@
CLGEMMLowpOffsetContributionKernel &operator=(CLGEMMLowpOffsetContributionKernel &&) = default;
/** Initialise the kernel's input and output.
*
- * @param[in, out] mm_result Input tensor containing the result of @ref CLGEMMLowpMatrixMultiplyKernel
+ * @param[in, out] mm_result Input tensor containing the result of @ref CLGEMMLowpMatrixMultiplyKernel. 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.
diff --git a/arm_compute/core/NEON/NEAsymm.h b/arm_compute/core/NEON/NEAsymm.h
index c7f59e9..997f28f 100644
--- a/arm_compute/core/NEON/NEAsymm.h
+++ b/arm_compute/core/NEON/NEAsymm.h
@@ -45,6 +45,17 @@
*/
int32x4_t rounding_divide_by_pow2(int32x4_t x, int exponent);
+/** Round to the nearest division by a power-of-two using exponent
+ *
+ * @note This function calculates the following expression: (x + 2^n -1 ) / 2^n where n = exponent
+ *
+ * @param[in] x Element to divide.
+ * @param[in] exponent Integer value used to round to nearest division by a power-of-two
+ *
+ * @return the nearest division by a power-of-two using exponent
+ */
+int32_t rounding_divide_by_pow2(int32_t x, int exponent);
+
/** Perform a multiply-accumulate on all 16 components of a QASYMM8 vector
*
* vd*vs + vo
@@ -125,6 +136,45 @@
return out_u8;
}
+/** Performs final quantization step on single 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] result_offset_after_shift_s32 Result offset parameter
+ * @param[in] min_u8 Relu lower bound
+ * @param[in] max_u8 Relu upper bound
+ *
+ * @return Quantized value
+ */
+template <bool is_bounded_relu>
+inline uint8_t finalize_quantization(int32_t in_value, int result_fixedpoint_multiplier,
+ int32_t result_shift, int32_t result_offset_after_shift_s32,
+ uint8_t min_u8, uint8_t max_u8)
+{
+ 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);
+
+ // Add the offset term
+ in_value += result_offset_after_shift_s32;
+
+ // Bound the result
+ uint8_t out_u8 = static_cast<uint8_t>(std::max(0, std::min(255, in_value)));
+ if(is_bounded_relu)
+ {
+ out_u8 = static_cast<uint8_t>(std::max(min_u8, std::min(max_u8, out_u8)));
+ }
+
+ return out_u8;
+}
+
/** Dequantize a neon vector holding 16 quantized values.
*
* @param qv Input values to be dequantized.
diff --git a/arm_compute/core/NEON/NEAsymm.inl b/arm_compute/core/NEON/NEAsymm.inl
index ce999a5..209785d 100644
--- a/arm_compute/core/NEON/NEAsymm.inl
+++ b/arm_compute/core/NEON/NEAsymm.inl
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -31,6 +31,13 @@
return vrshlq_s32(fixed_up_x, shift_vec);
}
+inline int32_t rounding_divide_by_pow2(int32_t x, int exponent)
+{
+ const int32_t mask = (1 << exponent) - 1;
+ const int32_t threshold = (mask >> 1) + (x < 0 ? 1 : 0);
+ return (x >> exponent) + ((x & mask) > threshold ? 1 : 0);
+}
+
inline qasymm8x16_t vmlaq_qasymm8(qasymm8x16_t vd, float32x4_t vs, float32x4_t vo)
{
// Convert uint8 vectors to uint16 vectors
diff --git a/arm_compute/core/NEON/NEKernels.h b/arm_compute/core/NEON/NEKernels.h
index f1d94c8..5b1b701 100644
--- a/arm_compute/core/NEON/NEKernels.h
+++ b/arm_compute/core/NEON/NEKernels.h
@@ -73,6 +73,7 @@
#include "arm_compute/core/NEON/kernels/NEGEMMInterleave4x4Kernel.h"
#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/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
new file mode 100644
index 0000000..c284ca5
--- /dev/null
+++ b/arm_compute/core/NEON/kernels/NEGEMMLowpOffsetContributionOutputStageKernel.h
@@ -0,0 +1,136 @@
+/*
+ * 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_NEGEMMLOWPOFFSETCONTRIBUTIONOUTPUTSTAGEKERNEL_H__
+#define __ARM_COMPUTE_NEGEMMLOWPOFFSETCONTRIBUTIONOUTPUTSTAGEKERNEL_H__
+
+#include "arm_compute/core/NEON/INEKernel.h"
+
+namespace arm_compute
+{
+class ITensor;
+
+/** NEON kernel used to add the offset contribution and perform the output stage after @ref NEGEMMLowpMatrixMultiplyKernel.
+ *
+ * The computation is performed in-place
+ *
+ * This kernel takes a final int32 accumulator value (the output of @ref NEGEMMLowpMatrixMultiplyKernel),
+ * and adds to it the offset contribution of matrix A and matrix B in-place.
+ *
+ * The output stage can perform either QuantizeDownInt32ToUint8Scale or QuantizeDownInt32ToUint8ScaleByFixedPoint.
+ *
+ * For QuantizeDownInt32ToUint8Scale the final result is:
+ *
+ * ((mm_result'[i][k] + result_offset) * result_mult_int) >> result_shift
+ *
+ * For QuantizeDownInt32ToUint8ScaleByFixedPoint the final result is:
+ *
+ * (FixedPointMul(mm_result'[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
+ *
+ * and mm_result'[i][k] = mm_result[i][k] +
+ * (vector_sum_col[k] * a_offset) +
+ * (vector_sum_row[i] * b_offset) +
+ * (a_offset * b_offset * k)
+ */
+
+class NEGEMMLowpOffsetContributionOutputStageKernel : public INEKernel
+{
+public:
+ const char *name() const override
+ {
+ return "NEGEMMLowpOffsetContributionOutputStageKernel";
+ }
+ /** Constructor */
+ NEGEMMLowpOffsetContributionOutputStageKernel();
+ /** Prevent instances of this class from being copied (As this class contains pointers)*/
+ NEGEMMLowpOffsetContributionOutputStageKernel(const NEGEMMLowpOffsetContributionOutputStageKernel &) = delete;
+ /** Prevent instances of this class from being copied (As this class contains pointers)*/
+ NEGEMMLowpOffsetContributionOutputStageKernel &operator=(const NEGEMMLowpOffsetContributionOutputStageKernel &) = delete;
+ /** Allow instances of this class to be moved */
+ NEGEMMLowpOffsetContributionOutputStageKernel(NEGEMMLowpOffsetContributionOutputStageKernel &&) = default;
+ /** Allow instances of this class to be moved */
+ NEGEMMLowpOffsetContributionOutputStageKernel &operator=(NEGEMMLowpOffsetContributionOutputStageKernel &&) = default;
+ /** Initialise the kernel's input and output.
+ *
+ * @param[in] mm_result Input tensor containing the result of @ref NEGEMMLowpMatrixMultiplyKernel. 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.
+ * @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[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.
+ * @param[in] output_stage GEMMLowp output stage info, providing the type of quantization and the necessary parameters.
+ */
+ void configure(const ITensor *mm_result, const ITensor *vector_sum_col, const ITensor *vector_sum_row, const ITensor *bias, ITensor *output, int32_t k, int32_t a_offset, int32_t b_offset,
+ GEMMLowpOutputStageInfo output_stage);
+ /** Static function to check if given info will lead to a valid configuration of @ref NEGEMMLowpOffsetContributionOutputStageKernel
+ *
+ * @param[in] mm_result Input tensor info containing the result of @ref NEGEMMLowpMatrixMultiplyKernel. Data type supported: S32
+ * @param[in] vector_sum_col Tensor info for the 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 Tensor info for the 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 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] 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.
+ *
+ * @return a status
+ */
+ static Status validate(const ITensorInfo *mm_result, const ITensorInfo *vector_sum_col, const ITensorInfo *vector_sum_row, const ITensorInfo *bias, const ITensorInfo *output, int32_t a_offset,
+ int32_t b_offset,
+ GEMMLowpOutputStageInfo output_stage);
+
+ // Inherited methods overridden:
+ void run(const Window &window, const ThreadInfo &info) override;
+
+ using NEGEMMLowpOffsetContributionOutputStageFunction = std::function<void(const Window, const ITensor *, const ITensor *, const ITensor *, const ITensor *,
+ ITensor *, int32_t, int32_t, int32_t, bool, GEMMLowpOutputStageInfo)>;
+
+private:
+ /** Function to use for the particular tensors passed to configure() */
+ NEGEMMLowpOffsetContributionOutputStageFunction _function;
+ const ITensor *_vector_sum_col;
+ const ITensor *_vector_sum_row;
+ const ITensor *_bias;
+ const ITensor *_mm_result;
+ ITensor *_output;
+ int32_t _a_offset;
+ int32_t _b_offset;
+ int32_t _k_offset;
+ bool _slide_vector_sum_col;
+ GEMMLowpOutputStageInfo _output_stage;
+};
+} // namespace arm_compute
+
+#endif /* __ARM_COMPUTE_NEGEMMLOWPOFFSETCONTRIBUTIONOUTPUTSTAGEKERNEL_H__ */