COMPMID-706 - Add GEMMLowp output stage for scaling by a fixed point number

DoD:
- Implement NEON kernel for quantizing down the gemmlowp result. The
  result should be scaled by a fixedpoint number
- Implement OpenCL kernel for quantizing down the gemmlowp result. The
  result should be scaled by a fixedpoint number
- Add test for validating the result

Required for:
- Integration of GEMMLowp in Android NN
- Convolution quantized
- Fully connected quantized

Change-Id: Ia963d25d695471e963961fb49a5600e78374ac4f
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/110981
Tested-by: BSG Visual Compute Jenkins server to access repositories on http://mpd-gerrit.cambridge.arm.com <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

1 files changed, 274 insertions, 0 deletions

diff --git a/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp b/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp
new file mode 100644
index 0000000000..102d08c7ba
--- /dev/null
+++ b/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp
@@ -0,0 +1,274 @@
+/*
+ * 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.
+ */
+#include "arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.h"
+
+#include "arm_compute/core/AccessWindowStatic.h"
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/NEON/NEAsymm.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Utils.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/Window.h"
+
+#include <arm_neon.h>
+#include <cstddef>
+#include <cstdint>
+
+using namespace arm_compute;
+
+namespace
+{
+Error validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min, int max)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::S32);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QASYMM8);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
+    ARM_COMPUTE_RETURN_ERROR_ON(max > 255);
+    ARM_COMPUTE_RETURN_ERROR_ON(min < 0 || min > max);
+
+    // Check biases if exist
+    if(bias != nullptr)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, bias);
+        ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1);
+        ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) != bias->dimension(0));
+    }
+    return Error{};
+}
+
+std::pair<Error, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *bias, ITensorInfo *output)
+{
+    constexpr unsigned int num_elems_processed_per_iteration = 16;
+
+    // Configure kernel window
+    Window win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration));
+
+    AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration);
+    AccessWindowHorizontal output_result_access(output, 0, num_elems_processed_per_iteration);
+
+    bool window_changed = update_window_and_padding(win,
+                                                    input_access,
+                                                    output_result_access);
+
+    if(bias != nullptr)
+    {
+        AccessWindowStatic bias_access(bias, 0, 0, ceil_to_multiple(bias->dimension(0), num_elems_processed_per_iteration), bias->tensor_shape()[1]);
+        window_changed = window_changed || update_window_and_padding(win, bias_access);
+    }
+
+    output_result_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape()));
+
+    Error err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Error{};
+    return std::make_pair(err, win);
+}
+
+template <bool    is_bounded_relu>
+inline uint8x16_t finalize_quantization(int32x4x4_t &in_s32, int result_fixedpoint_multiplier, int32_t result_shift, int32x4_t result_offset_after_shift_s32, uint8x16_t min_u8,
+                                        uint8x16_t max_u8)
+{
+    const static int32x4_t zero_s32 = vdupq_n_s32(0);
+
+    // 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);
+
+    // Saturate negative values
+    in_s32.val[0] = vmaxq_s32(in_s32.val[0], zero_s32);
+    in_s32.val[1] = vmaxq_s32(in_s32.val[1], zero_s32);
+    in_s32.val[2] = vmaxq_s32(in_s32.val[2], zero_s32);
+    in_s32.val[3] = vmaxq_s32(in_s32.val[3], zero_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 U8
+    uint8x16_t out_u8 = vcombine_u8(vqmovun_s16(in_s16.val[0]), vqmovun_s16(in_s16.val[1]));
+
+    if(is_bounded_relu)
+    {
+        out_u8 = vmaxq_u8(out_u8, min_u8);
+        out_u8 = vminq_u8(out_u8, max_u8);
+    }
+
+    return out_u8;
+}
+} // namespace
+
+namespace arm_compute
+{
+class Coordinates;
+} // namespace arm_compute
+
+template <bool is_bounded_relu>
+void NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::run(const Window &window)
+{
+    const int32x4_t  result_offset_after_shift_s32 = vdupq_n_s32(_result_offset_after_shift);
+    const uint8x16_t min_u8                        = vdupq_n_u8(static_cast<uint8_t>(_min));
+    const uint8x16_t max_u8                        = vdupq_n_u8(static_cast<uint8_t>(_max));
+
+    ARM_COMPUTE_UNUSED(min_u8);
+    ARM_COMPUTE_UNUSED(max_u8);
+
+    Iterator in(_input, window);
+    Iterator out(_output, window);
+
+    if(_bias != nullptr)
+    {
+        Window win_biases;
+        win_biases.set(Window::DimX, Window::Dimension(window.x().start(), window.x().end(), window.x().step()));
+        win_biases.set(Window::DimY, Window::Dimension(0, 1, 1));
+
+        Iterator bias(_bias, win_biases);
+        execute_window_loop(window, [&](const Coordinates & id)
+        {
+            int32x4x4_t in_s32 =
+            {
+                {
+                    vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + 0),
+                    vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + 4),
+                    vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + 8),
+                    vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + 12)
+                }
+            };
+
+            const int32x4x4_t bias_s32 =
+            {
+                {
+                    vld1q_s32(reinterpret_cast<const int32_t *>(bias.ptr()) + 0),
+                    vld1q_s32(reinterpret_cast<const int32_t *>(bias.ptr()) + 4),
+                    vld1q_s32(reinterpret_cast<const int32_t *>(bias.ptr()) + 8),
+                    vld1q_s32(reinterpret_cast<const int32_t *>(bias.ptr()) + 12)
+                }
+            };
+
+            // Add the bias to GEMM's result
+            in_s32.val[0] = vaddq_s32(in_s32.val[0], bias_s32.val[0]);
+            in_s32.val[1] = vaddq_s32(in_s32.val[1], bias_s32.val[1]);
+            in_s32.val[2] = vaddq_s32(in_s32.val[2], bias_s32.val[2]);
+            in_s32.val[3] = vaddq_s32(in_s32.val[3], bias_s32.val[3]);
+
+            vst1q_u8(out.ptr(), finalize_quantization<is_bounded_relu>(in_s32, _result_fixedpoint_multiplier, _result_shift, result_offset_after_shift_s32, min_u8, max_u8));
+        },
+        in, bias, out);
+    }
+    else
+    {
+        execute_window_loop(window, [&](const Coordinates & id)
+        {
+            int32x4x4_t in_s32 =
+            {
+                {
+                    vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + 0),
+                    vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + 4),
+                    vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + 8),
+                    vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + 12)
+                }
+            };
+
+            vst1q_u8(out.ptr(), finalize_quantization<is_bounded_relu>(in_s32, _result_fixedpoint_multiplier, _result_shift, result_offset_after_shift_s32, min_u8, max_u8));
+        },
+        in, out);
+    }
+}
+
+NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel()
+    : _func(nullptr), _input(nullptr), _bias(nullptr), _output(nullptr), _result_fixedpoint_multiplier(0), _result_shift(0), _result_offset_after_shift(0), _min(0), _max(0)
+{
+}
+
+void NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::configure(const ITensor *input, const ITensor *bias, ITensor *output, int result_fixedpoint_multiplier, int result_shift,
+                                                                          int result_offset_after_shift, int min, int max)
+{
+    // Perform validate step
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+
+    // Output auto inizialitation if not yet initialized
+    auto_init_if_empty(*output->info(), input->info()->clone()->set_data_type(DataType::QASYMM8));
+
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(),
+                                                  (bias != nullptr) ? bias->info() : nullptr,
+                                                  output->info(),
+                                                  min,
+                                                  max));
+
+    _input                        = input;
+    _bias                         = bias;
+    _output                       = output;
+    _result_fixedpoint_multiplier = result_fixedpoint_multiplier;
+    _result_shift                 = result_shift;
+    _result_offset_after_shift    = result_offset_after_shift;
+    _min                          = min;
+    _max                          = max;
+
+    // Configure kernel window
+    auto win_config = validate_and_configure_window(input->info(), (bias != nullptr) ? bias->info() : nullptr, output->info());
+    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+    INEKernel::configure(win_config.second);
+
+    // Check if we need to clamp the result using min and max
+    const bool is_bounded_relu = ((min != max) && !(min == 0 && max == 255));
+    _func                      = is_bounded_relu ? &NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::run<true> : &NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::run<false>;
+}
+
+Error NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min, int max)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, bias, output, min, max));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(),
+                                                              (bias != nullptr) ? bias->clone().get() : nullptr,
+                                                              output->clone().get())
+                                .first);
+
+    return Error{};
+}
+
+void NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::run(const Window &window, const ThreadInfo &info)
+{
+    ARM_COMPUTE_UNUSED(info);
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
+
+    (this->*_func)(window);
+}
\ No newline at end of file