COMPMID-675 - Reworked NEGEMMLowp interface/function

The new interface makes NEGEMMLowp able to work with ASYMM8 data types.

Implemented 2 new functions:
- NEGEMMLowpMatrixMultiplyCore
- NEGEMMLowpOutputStage

These functions should make the integration in android NN doable

For more information about GEMMLowp:
https://github.com/google/gemmlowp/blob/master/doc/low-precision.md

Change-Id: Ie2c775f45234f68ca53dba644b3a912b997fd890
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/95504
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Pablo Tello <pablo.tello@arm.com>

1 files changed, 79 insertions, 0 deletions

diff --git a/arm_compute/core/NEON/kernels/NEGEMMLowpOffsetContributionKernel.h b/arm_compute/core/NEON/kernels/NEGEMMLowpOffsetContributionKernel.h
new file mode 100644
index 0000000000..04b84339b0
--- /dev/null
+++ b/arm_compute/core/NEON/kernels/NEGEMMLowpOffsetContributionKernel.h
@@ -0,0 +1,79 @@
+/*
+ * Copyright (c) 2017 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_NEGEMMLOWPOFFSETCONTRIBUTIONKERNEL_H__
+#define __ARM_COMPUTE_NEGEMMLOWPOFFSETCONTRIBUTIONKERNEL_H__
+
+#include "arm_compute/core/NEON/INEKernel.h"
+
+namespace arm_compute
+{
+class ITensor;
+
+/* NEON kernel used to add the offset contribution after @ref NEGEMMLowpMatrixMultiplyKernel. The computation is performed in-place
+ *
+ * This kernel takes a final int32 accumulator value (the output of @NEGEMMLowpMatrixMultiplyKernel),
+ * and adds to it the offset contribution of matrix A and matrix B in-place.
+ *
+ */
+class NEGEMMLowpOffsetContributionKernel : public INEKernel
+{
+public:
+    /** Constructor */
+    NEGEMMLowpOffsetContributionKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers)*/
+    NEGEMMLowpOffsetContributionKernel(const NEGEMMLowpOffsetContributionKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers)*/
+    NEGEMMLowpOffsetContributionKernel &operator=(const NEGEMMLowpOffsetContributionKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    NEGEMMLowpOffsetContributionKernel(NEGEMMLowpOffsetContributionKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    NEGEMMLowpOffsetContributionKernel &operator=(NEGEMMLowpOffsetContributionKernel &&) = default;
+    /** Initialise the kernel's input and output.
+     *
+     * @param[in, out] 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.
+     *                                Note: vector_sum_row can be a nullptr in case b_offset = 0. Data type supported: same as @p mm_result
+     * @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.
+     */
+    void configure(ITensor *mm_result, const ITensor *vector_sum_col, const ITensor *vector_sum_row, int32_t k, int32_t a_offset, int32_t b_offset);
+
+    // Inherited methods overridden:
+    void run(const Window &window, const ThreadInfo &info) override;
+
+private:
+    const ITensor *_vector_sum_col;
+    const ITensor *_vector_sum_row;
+    ITensor       *_mm_result;
+    int32_t        _a_offset;
+    int32_t        _b_offset;
+    int32_t        _k_offset;
+    bool           _slide_vector_sum_col;
+};
+} // namespace arm_compute
+
+#endif /* __ARM_COMPUTE_NEGEMMLOWPOFFSETCONTRIBUTIONKERNEL_H__ */