49 #if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) 69 using tag_type =
typename vtype::tag_type;
92 const int conv_stride_w = std::get<0>(conv_info.
stride());
93 const int conv_stride_h = std::get<1>(conv_info.
stride());
96 Window window_out = window;
108 constexpr
int num_elems_read_per_iteration = 16 /
sizeof(T);
113 const int in_w_start_t =
static_cast<int>(
id.x()) * conv_stride_w - conv_pad_left;
114 const int in_h_start_t =
static_cast<int>(
id.y()) * conv_stride_h - conv_pad_top;
115 const int in_w_end_t = in_w_start_t + kernel_dim_w;
116 const int in_h_end_t = in_h_start_t + kernel_dim_h;
119 const int in_w_start = std::max(in_w_start_t, 0);
120 const int in_h_start = std::max(in_h_start_t, 0);
121 const int in_w_end = std::min(in_w_end_t, input_dim_w);
122 const int in_h_end = std::min(in_h_end_t, input_dim_h);
125 const int wei_w_start = in_w_start - in_w_start_t;
126 const int wei_h_start = in_h_start - in_h_start_t;
127 const int wei_h_end = kernel_dim_h - (in_h_end_t - in_h_end);
133 const T *
const weights_ptr_start =
reinterpret_cast<const T *
>(wei.
ptr());
134 uint8_t *out_ptr = out.
ptr() + id_w[3] * output_stride_c;
135 T out_temp =
static_cast<T
>(0);
137 for(
int index_wei_c = 0, index_in_c = 0; index_wei_c < index_c_end; ++index_wei_c, ++index_in_c)
139 const T *
const in_ptr_row_0 = in_ptr_start + index_in_c * input_stride_c;
140 const T *
const weights_ptr_row_0 = weights_ptr_start + index_wei_c * kernel_stride_c;
141 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)
143 const T *in_ptr_row = in_ptr_row_0 + index_in_h * input_stride_h;
144 const T *weights_ptr_row = weights_ptr_row_0 + index_wei_h * kernel_stride_h;
145 int index_w = in_w_start;
146 int index_wei_w = wei_w_start;
147 vector_type out_temp_vec =
wrapper::vdup_n(static_cast<T>(0), tag_type());
148 for(; index_w <= ((in_w_end - num_elems_read_per_iteration)); index_w += num_elems_read_per_iteration, index_wei_w += num_elems_read_per_iteration)
150 const auto src_vec =
wrapper::vloadq(in_ptr_row + index_w * input_stride_w);
151 const auto w_vec =
wrapper::vloadq(weights_ptr_row + index_wei_w * kernel_stride_w);
154 out_temp +=
vreduce(out_temp_vec);
155 for(; index_w < in_w_end; ++index_w, ++index_wei_w)
157 const auto src_val = *(in_ptr_row + index_w * input_stride_w);
158 const auto w_val = *(weights_ptr_row + index_wei_w * kernel_stride_w);
159 out_temp += src_val * w_val;
163 *(
reinterpret_cast<T *
>(out_ptr)) = out_temp;
171 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC Window calculate_max_window(const ValidRegion &valid_region, const Steps &steps, bool skip_border, BorderSize border_size)
virtual size_t dimension(size_t index) const =0
Return the size of the requested dimension.
void convolve_nchw(const Window &window, const ITensor *src, const ITensor *weights, ITensor *dst, const PadStrideInfo &conv_info)
void neon_fp16_nchw_directconv2d(const Window &window, const ITensor *src, const ITensor *weights, ITensor *dst, const PadStrideInfo &conv_info)
uint8x16_t vloadq(const uint8_t *ptr)
Describe one of the image's dimensions with a start, end and step.
unsigned int pad_top() const
Get the top padding.
const size_t conv_pad_top
decltype(strategy::transforms) typedef type
Interface for CPU tensor.
SimpleTensor< float > src
Copyright (c) 2017-2022 Arm Limited.
template void convolve_nchw< float >(const Window &window, const ITensor *src, const ITensor *weights, ITensor *dst, const PadStrideInfo &conv_info)
const size_t conv_pad_left
T x() const
Alias to access the size of the first dimension.
static constexpr size_t DimX
Alias for dimension 0 also known as X dimension.
#define ARM_COMPUTE_UNUSED(...)
To avoid unused variables warnings.
void neon_fp32_nchw_directconv2d(const Window &window, const ITensor *src, const ITensor *weights, ITensor *dst, const PadStrideInfo &conv_info)
Class to describe a number of elements in each dimension.
T z() const
Alias to access the size of the third dimension.
virtual uint8_t * buffer() const =0
Interface to be implemented by the child class to return a pointer to CPU memory. ...
std::pair< unsigned int, unsigned int > stride() const
Get the stride.
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.
constexpr uint8_t * ptr() const
Return a pointer to the current pixel.
Padding and stride information class.
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.
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.
float vreduce(const float32x4_t &v)
Reduce a vector to be a scalar by accumulating all lanes in the vector.
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.
unsigned int pad_left() const
Get the left padding.
Describe a multidimensional execution window.