Blame - src/cpu/kernels/conv3d/neon/list.h - ml/ComputeLibrary

void directconv3d_float_neon_ndhwc(const ITensor *src0, const ITensor *src1, const ITensor *src2, ITensor *dst, const Conv3dInfo &conv_info, const Window &window)

40

{

41

const ITensor *src = src0;

42

const ITensor *weights = src1;

43

const ITensor *biases = src2;

44

45

using vtype = wrapper::traits::neon_bitvector<T, wrapper::traits::BitWidth::W128>;

46

using vector_type = typename vtype::type;

47

using tag_type = typename vtype::tag_type;

48

constexpr int num_elems_read_per_iteration = 16 / sizeof(T);

49

50

// Scalar quantities (N D H W Cin)

51

const int element_size = src->info()->element_size();

52

const int input_stride_w = src->info()->strides_in_bytes().y() / element_size;

53

const int input_stride_h = src->info()->strides_in_bytes().z() / element_size;

54

const int input_stride_d = src->info()->strides_in_bytes()[3] / element_size;

55

const int input_stride_n = src->info()->strides_in_bytes()[4] / element_size;

56

const int input_dim_w = src->info()->dimension(1);

57

const int input_dim_h = src->info()->dimension(2);

58

const int input_dim_d = src->info()->dimension(3);

59

60

// Kernel info (D H W Cin Cout)

61

const unsigned int kernel_stride_w = weights->info()->strides_in_bytes()[2] / element_size;

62

const unsigned int kernel_stride_h = weights->info()->strides_in_bytes()[3] / element_size;

63

const unsigned int kernel_stride_d = weights->info()->strides_in_bytes()[4] / element_size;

64

const int kernel_dim_w = weights->info()->dimension(2);

65

const int kernel_dim_h = weights->info()->dimension(3);

66

const int kernel_dim_d = weights->info()->dimension(4);

67

68

// Convolution padding and stride

69

const int conv_pad_top = conv_info.padding.top;

70

const int conv_pad_left = conv_info.padding.left;

71

const int conv_pad_front = conv_info.padding.front;

72

const int conv_stride_w = conv_info.stride.width;

73

const int conv_stride_h = conv_info.stride.height;

74

const int conv_stride_d = conv_info.stride.depth;

75

76

// Setup input window for the output iterator

77

Window window_out = window;

78

window_out.set(Window::DimX, Window::Dimension(0, 1, 1));

79

80

// Setup input window for the weights iterator

81

Window window_w = calculate_max_window(*weights->info(), Steps());

82

window_w.set(Window::DimY, Window::Dimension(0, 1, 1));

83

window_w.set(Window::DimZ, Window::Dimension(0, 1, 1));

84

window_w.set(Window::DimW, Window::Dimension(0, 1, 1));

85

window_w.set(4, Window::Dimension(0, 1, 1));

86

87

Iterator out(dst, window_out);

88

Iterator wei(weights, window_w);

89

90

const T *biases_ptr = nullptr;

91

if(biases != nullptr)

92

{

93

biases_ptr = reinterpret_cast<T *>(biases->buffer() + biases->info()->offset_first_element_in_bytes());

94

}

95

execute_window_loop(window_out, [&](const Coordinates & id)

96

{

97

// We are computing the theoretical input starting points

98

const int in_w_start_t = static_cast<int>(id.y()) * conv_stride_w - conv_pad_left;

99

const int in_h_start_t = static_cast<int>(id.z()) * conv_stride_h - conv_pad_top;

100

const int in_d_start_t = static_cast<int>(id[3]) * conv_stride_d - conv_pad_front;

101

const int in_w_end_t = in_w_start_t + kernel_dim_w;

102

const int in_h_end_t = in_h_start_t + kernel_dim_h;

103

const int in_d_end_t = in_d_start_t + kernel_dim_d;

104

105

// We are computing the valid initial and ending input points by checking the borders

106

const int in_w_start = std::max(in_w_start_t, 0);

107

const int in_h_start = std::max(in_h_start_t, 0);

108

const int in_d_start = std::max(in_d_start_t, 0);

109

const int in_w_end = std::min(in_w_end_t, input_dim_w);

110

const int in_h_end = std::min(in_h_end_t, input_dim_h);

111

const int in_d_end = std::min(in_d_end_t, input_dim_d);

112

113

// We use the input points to select the valid weight points to use

114

const int wei_w_start = in_w_start - in_w_start_t;

115

const int wei_h_start = in_h_start - in_h_start_t;

116

const int wei_d_start = in_d_start - in_d_start_t;

117

const int wei_w_end = kernel_dim_w - (in_w_end_t - in_w_end);

118

const int wei_h_end = kernel_dim_h - (in_h_end_t - in_h_end);

119

const int wei_d_end = kernel_dim_d - (in_d_end_t - in_d_end);

120

121

const int index_c_out_end = weights->info()->dimension(0);

122

const int index_c_in_end = weights->info()->dimension(1);

123

const T *const in_ptr_start = reinterpret_cast<const T *>(src->buffer() + src->info()->offset_first_element_in_bytes()) + id[4] * input_stride_n;

124

125

execute_window_loop(window_w, [&](const Coordinates & id_w)

126

{

127

/*

128

* This is the loop in the weights, and it goes along OFM (output feature map)

129

*/

130

const auto weights_ptr_start = reinterpret_cast<const T *>(wei.ptr());

131

T out_temp = static_cast<T>(0);

132

T *out_ptr = reinterpret_cast<T *>(out.ptr());

133

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)

134

{

135

const auto in_ptr_d = in_ptr_start + index_in_d * input_stride_d;

136

const auto weights_ptr_d = weights_ptr_start + index_wei_d * kernel_stride_d;

137

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)

138

{

139

const T *const in_ptr_row = in_ptr_d + index_in_h * input_stride_h;

140

const T *const weights_ptr_row = weights_ptr_d + index_wei_h * kernel_stride_h;

141

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)

142

{

143

const T *in_ptr_mover = in_ptr_row + index_in_w * input_stride_w;

144

const T *weights_ptr_mover = weights_ptr_row + index_wei_w * kernel_stride_w;

145

int index_c_in = 0;

146

vector_type out_temp_vec = wrapper::vdup_n(static_cast<T>(0), tag_type());

147

vector_type w_vec = wrapper::vdup_n(static_cast<T>(0), tag_type());

148

for(; index_c_in <= index_c_in_end - num_elems_read_per_iteration;

149

index_c_in += num_elems_read_per_iteration, in_ptr_mover += num_elems_read_per_iteration)

150

{

151

const auto src_vec = wrapper::vloadq(in_ptr_mover);

152

//Load Cin weights

Freddie Liardet

ebefe52

2021-11-25 16:19:28 +0000

[diff] [blame]

153

for(int k = 0; k < num_elems_read_per_iteration; ++k, weights_ptr_mover += index_c_out_end)

Sheri Zhang

5dda217

2021-10-15 19:54:17 +0100

[diff] [blame]

154

{

155

w_vec = wrapper::vsetlane(*weights_ptr_mover, w_vec, k);

156

}

157

out_temp_vec = wrapper::vmla(out_temp_vec, w_vec, src_vec);

158

}

159

out_temp += vreduce(out_temp_vec);

160

for(; index_c_in < index_c_in_end; ++index_c_in, ++in_ptr_mover, weights_ptr_mover += index_c_out_end)

161

{

162

const auto src_val = *(in_ptr_mover);

163

const auto w_val = *(weights_ptr_mover);

164

out_temp += src_val * w_val;

}

}

}

}

*(reinterpret_cast<T *>(out_ptr + id_w[0])) = (biases_ptr != nullptr) ? out_temp + biases_ptr[id_w[0]] : out_temp;

},

wei);

},

out);

}

Freddie Liardet

f727ef4

2021-10-18 13:28:57 +0100

[diff] [blame]

175

Sheri Zhang

5dda217

2021-10-15 19:54:17 +0100

[diff] [blame]

176

} // namespace cpu

177

} // namespace arm_compute

178

#endif // SRC_CORE_NEON_KERNELS_CONV3D_LIST_H