24 #ifndef SRC_CORE_NEON_KERNELS_CONV3D_QUANTIZED_H 25 #define SRC_CORE_NEON_KERNELS_CONV3D_QUANTIZED_H 48 using tag_type =
typename vtype::tag_type;
49 constexpr
int num_elems_read_per_iteration = 16 /
sizeof(T);
61 int32_t output_multiplier = 0;
62 int32_t output_shift = 0;
63 const float multiplier = input_scale * weights_scale / output_scale;
93 Window window_out = window;
106 const int32_t *biases_ptr =
nullptr;
107 if(biases !=
nullptr)
114 const int in_w_start_t =
static_cast<int>(
id.y()) * conv_stride_w - conv_pad_left;
115 const int in_h_start_t =
static_cast<int>(
id.z()) * conv_stride_h - conv_pad_top;
116 const int in_d_start_t =
static_cast<int>(
id[3]) * conv_stride_d - conv_pad_front;
117 const int in_w_end_t = in_w_start_t + kernel_dim_w;
118 const int in_h_end_t = in_h_start_t + kernel_dim_h;
119 const int in_d_end_t = in_d_start_t + kernel_dim_d;
122 const int in_w_start = std::max(in_w_start_t, 0);
123 const int in_h_start = std::max(in_h_start_t, 0);
124 const int in_d_start = std::max(in_d_start_t, 0);
125 const int in_w_end = std::min(in_w_end_t, input_dim_w);
126 const int in_h_end = std::min(in_h_end_t, input_dim_h);
127 const int in_d_end = std::min(in_d_end_t, input_dim_d);
130 const int wei_w_start = in_w_start - in_w_start_t;
131 const int wei_h_start = in_h_start - in_h_start_t;
132 const int wei_d_start = in_d_start - in_d_start_t;
133 const int wei_w_end = kernel_dim_w - (in_w_end_t - in_w_end);
134 const int wei_h_end = kernel_dim_h - (in_h_end_t - in_h_end);
135 const int wei_d_end = kernel_dim_d - (in_d_end_t - in_d_end);
146 const auto weights_ptr_start =
reinterpret_cast<const T *
>(wei.
ptr());
147 int32_t acc =
static_cast<int32_t
>(0);
148 T *out_ptr =
reinterpret_cast<T *
>(out.
ptr());
149 for(
int index_wei_d = wei_d_start, index_in_d = in_d_start; index_wei_d < wei_d_end; ++index_wei_d, ++index_in_d)
151 const auto in_ptr_d = in_ptr_start + index_in_d * input_stride_d;
152 const auto weights_ptr_d = weights_ptr_start + index_wei_d * kernel_stride_d;
153 for(
int index_wei_h = wei_h_start, index_in_h = in_h_start; index_wei_h < wei_h_end; ++index_wei_h, ++index_in_h)
155 const T *
const in_ptr_row = in_ptr_d + index_in_h * input_stride_h;
156 const T *
const weights_ptr_row = weights_ptr_d + index_wei_h * kernel_stride_h;
157 for(
int index_wei_w = wei_w_start, index_in_w = in_w_start; index_wei_w < wei_w_end; ++index_wei_w, ++index_in_w)
159 const T *in_ptr_mover = in_ptr_row + index_in_w * input_stride_w;
160 const T *weights_ptr_mover = weights_ptr_row + index_wei_w * kernel_stride_w;
164 q32x4_t acc_q32_0 =
wrapper::vdup_n(static_cast<q32_t>(0), tag_type());
165 q32x4_t acc_q32_1 =
wrapper::vdup_n(static_cast<q32_t>(0), tag_type());
166 q32x4_t acc_q32_2 =
wrapper::vdup_n(static_cast<q32_t>(0), tag_type());
167 q32x4_t acc_q32_3 =
wrapper::vdup_n(static_cast<q32_t>(0), tag_type());
169 for(; index_c_in <= index_c_in_end - num_elems_read_per_iteration;
170 index_c_in += num_elems_read_per_iteration, in_ptr_mover += num_elems_read_per_iteration)
174 for(
int k = 0;
k < num_elems_read_per_iteration; ++
k, weights_ptr_mover += index_c_out_end)
178 q32x4_t src_q32_0 =
wrapper::vdup_n(static_cast<q32_t>(input_offset), tag_type());
179 q32x4_t src_q32_1 =
wrapper::vdup_n(static_cast<q32_t>(input_offset), tag_type());
180 q32x4_t src_q32_2 =
wrapper::vdup_n(static_cast<q32_t>(input_offset), tag_type());
181 q32x4_t src_q32_3 =
wrapper::vdup_n(static_cast<q32_t>(input_offset), tag_type());
183 q32x4_t wei_q32_0 =
wrapper::vdup_n(static_cast<q32_t>(weights_offset), tag_type());
184 q32x4_t wei_q32_1 =
wrapper::vdup_n(static_cast<q32_t>(weights_offset), tag_type());
185 q32x4_t wei_q32_2 =
wrapper::vdup_n(static_cast<q32_t>(weights_offset), tag_type());
186 q32x4_t wei_q32_3 =
wrapper::vdup_n(static_cast<q32_t>(weights_offset), tag_type());
208 #if defined(__aarch64__) 209 acc += wrapper::vaddv(acc_q32_0);
210 acc += wrapper::vaddv(acc_q32_1);
211 acc += wrapper::vaddv(acc_q32_2);
212 acc += wrapper::vaddv(acc_q32_3);
230 #endif // __aarch64__ 232 for(; index_c_in < index_c_in_end; ++index_c_in, ++in_ptr_mover, weights_ptr_mover += index_c_out_end)
234 const auto src_val = *(in_ptr_mover) + input_offset;
235 const auto w_val = *(weights_ptr_mover) + weights_offset;
236 acc += src_val * w_val;
244 acc += *
reinterpret_cast<const int32_t *
>(biases_ptr + id_w[0]);
247 T out_val =
finalize_quantization(acc, output_multiplier, output_shift, output_offset, T(0), T(0),
false);
248 *(
reinterpret_cast<T *
>(out_ptr + id_w[0])) = out_val;
256 #endif // SRC_CORE_NEON_KERNELS_CONV3D_QUANTIZED_H Window calculate_max_window(const ValidRegion &valid_region, const Steps &steps, bool skip_border, BorderSize border_size)
Descriptor used by the 3d Convolution function.
virtual size_t dimension(size_t index) const =0
Return the size of the requested dimension.
uint8x16_t vloadq(const uint8_t *ptr)
uint8x8_t vadd(const uint8x8_t &a, const uint8x8_t &b)
Describe one of the image's dimensions with a start, end and step.
Status calculate_quantized_multiplier(float multiplier, int32_t *quant_multiplier, int32_t *shift, bool ignore_epsilon=false)
Calculate quantized representation of multiplier.
const size_t conv_pad_top
decltype(strategy::transforms) typedef type
Interface for CPU tensor.
SimpleTensor< float > src
Copyright (c) 2017-2022 Arm Limited.
const size_t conv_pad_left
typename promote< T >::type promote_t
Get promoted type.
uint8x8_t vpadd(const uint8x8_t &a, const uint8x8_t &b)
uint8_t vgetlane(const uint8x8_t vector, const unsigned int lane)
static constexpr size_t DimX
Alias for dimension 0 also known as X dimension.
size_t front
Padding across the depth dimenstion on the front, in elements.
Create the appropriate SIMD vector given its type and size in terms of elements.
void directconv3d_quantized_neon_ndhwc(const ITensor *src0, const ITensor *src1, const ITensor *src2, ITensor *dst, const Conv3dInfo &conv_info, const Window &window)
Class to describe a number of elements in each dimension.
T z() const
Alias to access the size of the third dimension.
size_t height
Height of the 3D shape or object.
virtual uint8_t * buffer() const =0
Interface to be implemented by the child class to return a pointer to CPU memory. ...
UniformQuantizationInfo uniform() const
Return per layer quantization info.
Create the appropriate SIMD vector given its type and size in terms of bits.
virtual ITensorInfo * info() const =0
Interface to be implemented by the child class to return the tensor's metadata.
size_t top
Padding across the height dimenstion on the top, in elements.
constexpr uint8_t * ptr() const
Return a pointer to the current pixel.
uint8x8_t vgetlow(const uint8x16_t val)
size_t left
Padding across the width dimenstion on the left, in elements.
size_t width
Width of the 3D shape or object.
virtual size_t element_size() const =0
Element size in bytes calculated as data_size() * num_channels()
void set(size_t dimension, const Dimension &dim)
Set the values of a given dimension.
uint8x8_t vsetlane(const uint8_t value, const uint8x8_t vector, const unsigned int lane)
static constexpr size_t DimW
Alias for dimension 3 also known as W dimension.
virtual QuantizationInfo quantization_info() const =0
Get the quantization settings (scale and offset) of the tensor.
uint8x8_t vgethigh(const uint8x16_t val)
virtual size_t offset_first_element_in_bytes() const =0
The offset from the beginning of the memory allocation to the first element of the tensor...
static constexpr size_t DimY
Alias for dimension 1 also known as Y dimension.
size_t depth
Depth of the 3D shape or object.
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, bool is_bounded_relu)
Performs final quantization step on 16 elements.
static constexpr size_t DimZ
Alias for dimension 2 also known as Z dimension.
uint8x8_t vdup_n(uint8_t value, traits::vector_64_tag)
void execute_window_loop(const Window &w, L &&lambda_function, Ts &&... iterators)
Iterate through the passed window, automatically adjusting the iterators and calling the lambda_funct...
T y() const
Alias to access the size of the second dimension.
Includes all wrapper headers at once.
virtual const Strides & strides_in_bytes() const =0
The strides in bytes for accessing each dimension of the tensor.
uint8x8_t vmla(const uint8x8_t &a, const uint8x8_t &b, const uint8x8_t &c)
Iterator updated by execute_window_loop for each window element.
uint16x8_t vmovl(const uint8x8_t &a)
Describe a multidimensional execution window.