COMPMID-791: Generic Depthwise Convolution Layer NEON QASYMM8

Change-Id: I33cf54e68f6c097ac58b6f16c3f9a720978f09cd
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/117289
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Michalis Spyrou <michalis.spyrou@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

1 files changed, 155 insertions, 36 deletions

diff --git a/src/core/NEON/kernels/NEGEMMMatrixVectorMultiplyKernel.cpp b/src/core/NEON/kernels/NEGEMMMatrixVectorMultiplyKernel.cpp
index fe79df2528..c1e975e77e 100644
--- a/src/core/NEON/kernels/NEGEMMMatrixVectorMultiplyKernel.cpp
+++ b/src/core/NEON/kernels/NEGEMMMatrixVectorMultiplyKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016, 2017 ARM Limited.
+ * Copyright (c) 2016-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -39,24 +39,170 @@
 
 using namespace arm_compute;
 
+template <typename I0, typename I1, typename O>
+void NEGEMMMatrixVectorMultiplyKernel::matrix_vector_multiply(const Window &window_in, const Window &window_w, const Window &window_out)
+{
+    ARM_COMPUTE_ERROR("Unsupported data types");
+    ARM_COMPUTE_UNUSED(window_in);
+    ARM_COMPUTE_UNUSED(window_w);
+    ARM_COMPUTE_UNUSED(window_out);
+}
+
+namespace arm_compute
+{
+template <>
+void NEGEMMMatrixVectorMultiplyKernel::matrix_vector_multiply<float, float, float>(const Window &window_in,
+                                                                                   const Window &window_w,
+                                                                                   const Window &window_out)
+{
+    Iterator in(_input0, window_in);
+    Iterator in2(_input1, window_w);
+    Iterator out(_output, window_out);
+
+    const int input_w          = _input0->info()->dimension(0);
+    const int input_h          = _input0->info()->dimension(1);
+    const int input_stride_x   = _input0->info()->strides_in_bytes().x();
+    const int weights_stride_x = _input1->info()->strides_in_bytes().x();
+    const int weights_stride_y = _input1->info()->strides_in_bytes().y();
+    const int output_stride_x  = _output->info()->strides_in_bytes().x();
+
+    execute_window_loop(window_in, [&](const Coordinates & id)
+    {
+        // Get pointers
+        const uint8_t *const input_ptr   = in.ptr();
+        const uint8_t *const weights_ptr = in2.ptr() + id.z() * weights_stride_y;
+        auto                 output_ptr  = reinterpret_cast<float *>(out.ptr() + (id.y() + id.z() * input_h) * output_stride_x);
+
+        float32x4_t row_dot = vdupq_n_f32(0.f);
+        for(int i = 0; i < input_w; i += 4)
+        {
+            const auto input   = vld1q_f32(reinterpret_cast<const float *>(input_ptr + i * input_stride_x));
+            const auto weights = vld1q_f32(reinterpret_cast<const float *>(weights_ptr + i * weights_stride_x));
+            row_dot            = vaddq_f32(row_dot, vmulq_f32(input, weights));
+        }
+
+        auto temp = vadd_f32(vget_high_f32(row_dot), vget_low_f32(row_dot));
+        temp      = vpadd_f32(temp, temp);
+
+        *output_ptr = vget_lane_f32(temp, 0);
+    },
+    in, in2, out);
+}
+
+template <>
+void NEGEMMMatrixVectorMultiplyKernel::matrix_vector_multiply<uint8_t, uint8_t, int32_t>(const Window &window_in,
+                                                                                         const Window &window_w,
+                                                                                         const Window &window_out)
+{
+    Iterator in(_input0, window_in);
+    Iterator in2(_input1, window_w);
+    Iterator out(_output, window_out);
+
+    const int input_offset   = -_input0->info()->quantization_info().offset;
+    const int weights_offset = -_input1->info()->quantization_info().offset;
+
+    const int input_w          = _input0->info()->dimension(0);
+    const int input_h          = _input0->info()->dimension(1);
+    const int input_stride_x   = _input0->info()->strides_in_bytes().x();
+    const int weights_stride_x = _input1->info()->strides_in_bytes().x();
+    const int weights_stride_y = _input1->info()->strides_in_bytes().y();
+    const int output_stride_x  = _output->info()->strides_in_bytes().x();
+    const int read_step        = 16 / _input0->info()->element_size();
+
+    const int32x4_t v_input_offset   = vdupq_n_s32(input_offset);
+    const int32x4_t v_weights_offset = vdupq_n_s32(weights_offset);
+
+    execute_window_loop(window_in, [&](const Coordinates & id)
+    {
+        // Get pointers
+        const uint8_t *const input_ptr   = in.ptr();
+        const uint8_t *const weights_ptr = in2.ptr() + id.z() * weights_stride_y;
+        auto                 output_ptr  = reinterpret_cast<int32_t *>(out.ptr() + (id.y() + id.z() * input_h) * output_stride_x);
+
+        int32x4_t row_dot = vdupq_n_s32(0);
+        for(int i = 0; i < input_w; i += read_step)
+        {
+            // Read values
+            const auto input   = vld1q_u8(reinterpret_cast<const uint8_t *>(input_ptr + i * input_stride_x));
+            const auto weights = vld1q_u8(reinterpret_cast<const uint8_t *>(weights_ptr + i * weights_stride_x));
+
+            // Add offsets
+            const int32x4x4_t input_s32 =
+            {
+                {
+                    vaddw_s16(v_input_offset, vreinterpret_s16_u16(vget_low_u16(vmovl_u8(vget_low_u8(input))))),
+                    vaddw_s16(v_input_offset, vreinterpret_s16_u16(vget_high_u16(vmovl_u8(vget_low_u8(input))))),
+                    vaddw_s16(v_input_offset, vreinterpret_s16_u16(vget_low_u16(vmovl_u8(vget_high_u8(input))))),
+                    vaddw_s16(v_input_offset, vreinterpret_s16_u16(vget_high_u16(vmovl_u8(vget_high_u8(input)))))
+                }
+            };
+            const int32x4x4_t weights_s32 =
+            {
+                {
+                    vaddw_s16(v_weights_offset, vreinterpret_s16_u16(vget_low_u16(vmovl_u8(vget_low_u8(weights))))),
+                    vaddw_s16(v_weights_offset, vreinterpret_s16_u16(vget_high_u16(vmovl_u8(vget_low_u8(weights))))),
+                    vaddw_s16(v_weights_offset, vreinterpret_s16_u16(vget_low_u16(vmovl_u8(vget_high_u8(weights))))),
+                    vaddw_s16(v_weights_offset, vreinterpret_s16_u16(vget_high_u16(vmovl_u8(vget_high_u8(weights)))))
+                }
+            };
+
+            // Dot
+            row_dot = vaddq_s32(row_dot, vmulq_s32(input_s32.val[0], weights_s32.val[0]));
+            row_dot = vaddq_s32(row_dot, vmulq_s32(input_s32.val[1], weights_s32.val[1]));
+            row_dot = vaddq_s32(row_dot, vmulq_s32(input_s32.val[2], weights_s32.val[2]));
+            row_dot = vaddq_s32(row_dot, vmulq_s32(input_s32.val[3], weights_s32.val[3]));
+        }
+
+        // Reduction
+        auto temp = vadd_s32(vget_high_s32(row_dot), vget_low_s32(row_dot));
+        temp      = vpadd_s32(temp, temp);
+
+        *output_ptr = vget_lane_s32(temp, 0);
+    },
+    in, in2, out);
+}
+} //namespace arm_compute
+
 NEGEMMMatrixVectorMultiplyKernel::NEGEMMMatrixVectorMultiplyKernel()
-    : _input0(nullptr), _input1(nullptr), _output(nullptr)
+    : _func(nullptr), _input0(nullptr), _input1(nullptr), _output(nullptr), _border_size(0)
+{
+}
+
+BorderSize NEGEMMMatrixVectorMultiplyKernel::border_size() const
 {
+    return _border_size;
 }
 
 void NEGEMMMatrixVectorMultiplyKernel::configure(const ITensor *input0, const ITensor *input1, ITensor *output)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::F16, DataType::F32);
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1, output);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::QASYMM8, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input0, input1, output);
+    ARM_COMPUTE_ERROR_ON(is_data_type_quantized_asymmetric(input0->info()->data_type()) && (output->info()->data_type() != DataType::S32));
     ARM_COMPUTE_ERROR_ON(input0->info()->dimension(2) != input1->info()->dimension(1));
 
     _input0 = input0;
     _input1 = input1;
     _output = output;
 
+    // Set appropriate function to run
+    switch(input0->info()->data_type())
+    {
+        case DataType::QASYMM8:
+            _func = &NEGEMMMatrixVectorMultiplyKernel::matrix_vector_multiply<uint8_t, uint8_t, int32_t>;
+            break;
+        case DataType::F32:
+            _func = &NEGEMMMatrixVectorMultiplyKernel::matrix_vector_multiply<float, float, float>;
+            break;
+        default:
+            ARM_COMPUTE_ERROR("Unsupported data type");
+    }
+
     // Configure kernel window
-    const unsigned int num_elems_read_per_iteration = 4;
+    const unsigned int num_elems_read_per_iteration = 16 / _input0->info()->element_size();
+
+    const unsigned int border_x = ceil_to_multiple(input0->info()->dimension(0), num_elems_read_per_iteration) - input0->info()->dimension(0);
+    _border_size                = BorderSize(0, border_x);
 
     Window win = calculate_max_window(*input0->info(), Steps(num_elems_read_per_iteration));
 
@@ -75,6 +221,7 @@ void NEGEMMMatrixVectorMultiplyKernel::run(const Window &window, const ThreadInf
 {
     ARM_COMPUTE_UNUSED(info);
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
 
     Window window_slice = window.first_slice_window_3D();
 
@@ -96,36 +243,8 @@ void NEGEMMMatrixVectorMultiplyKernel::run(const Window &window, const ThreadInf
     window_out.set(Window::DimY, Window::Dimension(0, 0, 0));
     window_out.set(Window::DimZ, Window::Dimension(0, 0, 0));
 
-    Iterator in(_input0, window_in);
-    Iterator in2(_input1, window_weights);
-    Iterator out(_output, window_out);
-
-    const int input_w          = _input0->info()->dimension(0);
-    const int input_h          = _input0->info()->dimension(1);
-    const int input_stride_x   = _input0->info()->strides_in_bytes().x();
-    const int weights_stride_x = _input1->info()->strides_in_bytes().x();
-    const int weights_stride_y = _input1->info()->strides_in_bytes().y();
-    const int output_stride_x  = _output->info()->strides_in_bytes().x();
-
-    execute_window_loop(window_in, [&](const Coordinates & id)
+    if(_func != nullptr)
     {
-        // Get pointers
-        const uint8_t *const input_ptr   = in.ptr();
-        const uint8_t *const weights_ptr = in2.ptr() + id.z() * weights_stride_y;
-        auto                 output_ptr  = reinterpret_cast<float *>(out.ptr() + (id.y() + id.z() * input_h) * output_stride_x);
-
-        float32x4_t row_dot = vdupq_n_f32(0.f);
-        for(int i = 0; i < input_w; i += 4)
-        {
-            const auto input   = vld1q_f32(reinterpret_cast<const float *>(input_ptr + i * input_stride_x));
-            const auto weights = vld1q_f32(reinterpret_cast<const float *>(weights_ptr + i * weights_stride_x));
-            row_dot            = vaddq_f32(row_dot, vmulq_f32(input, weights));
-        }
-
-        auto temp = vadd_f32(vget_high_f32(row_dot), vget_low_f32(row_dot));
-        temp      = vpadd_f32(temp, temp);
-
-        *output_ptr = vget_lane_f32(temp, 0);
-    },
-    in, in2, out);
+        (this->*_func)(window_in, window_weights, window_out);
+    }
 }