47 const int window_step_x = 16;
48 const auto window_start_x =
static_cast<int>(window.
x().
start());
49 const auto window_end_x =
static_cast<int>(window.
x().
end());
56 const float32x4_t invvscaleo = vdupq_n_f32(1.f / oq_info.
scale);
57 const float32x4_t voffseto = vdupq_n_f32(oq_info.
offset);
59 if(is_broadcast_across_x)
61 const bool is_broadcast_input_2 = input2_win.
x().
step() == 0;
62 Window broadcast_win = is_broadcast_input_2 ? input2_win : input1_win;
63 Window non_broadcast_win = !is_broadcast_input_2 ? input2_win : input1_win;
64 const ITensor *broadcast_tensor = is_broadcast_input_2 ? src1 : src0;
65 const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? src1 : src0;
69 const float32x4_t vscale1 = is_broadcast_input_2 ? vdupq_n_f32(iq1_info.
scale) : vdupq_n_f32(iq2_info.
scale);
70 const float32x4_t vscale2 = is_broadcast_input_2 ? vdupq_n_f32(iq2_info.
scale) : vdupq_n_f32(iq1_info.
scale);
71 const int32x4_t voffset1 = is_broadcast_input_2 ? vdupq_n_s32(iq1_info.
offset) : vdupq_n_s32(iq2_info.
offset);
72 const int32x4_t voffset2 = is_broadcast_input_2 ? vdupq_n_s32(iq2_info.
offset) : vdupq_n_s32(iq1_info.
offset);
77 Iterator broadcast_input(broadcast_tensor, broadcast_win);
78 Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win);
83 const auto non_broadcast_input_ptr =
reinterpret_cast<const uint8_t *
>(non_broadcast_input.
ptr());
84 const auto output_ptr =
reinterpret_cast<uint8_t *
>(output.
ptr());
86 const uint8_t broadcast_value = *
reinterpret_cast<const uint8_t *
>(broadcast_input.
ptr());
87 const uint8x16_t broadcast_value_vec = vdupq_n_u8(broadcast_value);
89 const auto bf_0 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_low_u8(broadcast_value_vec))))), voffset2)), vscale2);
90 const auto bf_1 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_low_u8(broadcast_value_vec))))), voffset2)), vscale2);
91 const auto bf_2 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_high_u8(broadcast_value_vec))))), voffset2)), vscale2);
92 const auto bf_3 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_high_u8(broadcast_value_vec))))), voffset2)), vscale2);
94 const float bfs =
static_cast<int32_t
>(broadcast_value - broadcast_qinfo.
offset) * broadcast_qinfo.
scale;
97 int x = window_start_x;
98 for(; x <= (window_end_x - window_step_x); x += window_step_x)
100 const uint8x16_t a = vld1q_u8(non_broadcast_input_ptr + x);
101 const auto af_0 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_low_u8(a))))), voffset1)), vscale1);
102 const auto af_1 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_low_u8(a))))), voffset1)), vscale1);
103 const auto af_2 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_high_u8(a))))), voffset1)), vscale1);
104 const auto af_3 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_high_u8(a))))), voffset1)), vscale1);
112 rf_0 = vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_0, bf_0), invvscaleo));
113 rf_1 = vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_1, bf_1), invvscaleo));
114 rf_2 = vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_2, bf_2), invvscaleo));
115 rf_3 = vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_3, bf_3), invvscaleo));
117 rf_0 = vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_0, bf_0), invvscaleo));
118 rf_1 = vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_1, bf_1), invvscaleo));
119 rf_2 = vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_2, bf_2), invvscaleo));
120 rf_3 = vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_3, bf_3), invvscaleo));
123 const uint8x8_t pa = vqmovun_s16(vcombine_s16(vqmovn_s32(rf_0), vqmovn_s32(rf_1)));
124 const uint8x8_t pb = vqmovun_s16(vcombine_s16(vqmovn_s32(rf_2), vqmovn_s32(rf_3)));
125 vst1q_u8(output_ptr + x, vcombine_u8(pa, pb));
129 for(; x < window_end_x; ++x)
131 const float afs =
static_cast<int32_t
>(*(non_broadcast_input_ptr + x) - non_broadcast_qinfo.
offset) * non_broadcast_qinfo.
scale;
135 broadcast_input, non_broadcast_input, output);
147 const float32x4_t vscale1 = vdupq_n_f32(iq1_info.
scale);
148 const float32x4_t vscale2 = vdupq_n_f32(iq2_info.
scale);
149 const int32x4_t voffset1 = vdupq_n_s32(iq1_info.
offset);
150 const int32x4_t voffset2 = vdupq_n_s32(iq2_info.
offset);
154 const auto input1_ptr =
reinterpret_cast<const uint8_t *
>(input1.
ptr());
155 const auto input2_ptr =
reinterpret_cast<const uint8_t *
>(input2.
ptr());
156 const auto output_ptr =
reinterpret_cast<uint8_t *
>(output.
ptr());
159 int x = window_start_x;
160 for(; x <= (window_end_x - window_step_x); x += window_step_x)
162 const uint8x16_t a = vld1q_u8(input1_ptr + x);
163 const uint8x16_t
b = vld1q_u8(input2_ptr + x);
165 const auto af_0 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_low_u8(a))))), voffset1)), vscale1);
166 const auto af_1 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_low_u8(a))))), voffset1)), vscale1);
167 const auto af_2 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_high_u8(a))))), voffset1)), vscale1);
168 const auto af_3 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_high_u8(a))))), voffset1)), vscale1);
170 const auto bf_0 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_low_u8(b))))), voffset2)), vscale2);
171 const auto bf_1 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_low_u8(b))))), voffset2)), vscale2);
172 const auto bf_2 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_high_u8(b))))), voffset2)), vscale2);
173 const auto bf_3 = vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_high_u8(b))))), voffset2)), vscale2);
181 rf_0 = vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_0, bf_0), invvscaleo));
182 rf_1 = vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_1, bf_1), invvscaleo));
183 rf_2 = vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_2, bf_2), invvscaleo));
184 rf_3 = vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_3, bf_3), invvscaleo));
186 rf_0 = vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_0, bf_0), invvscaleo));
187 rf_1 = vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_1, bf_1), invvscaleo));
188 rf_2 = vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_2, bf_2), invvscaleo));
189 rf_3 = vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af_3, bf_3), invvscaleo));
192 const uint8x8_t pa = vqmovun_s16(vcombine_s16(vqmovn_s32(rf_0), vqmovn_s32(rf_1)));
193 const uint8x8_t pb = vqmovun_s16(vcombine_s16(vqmovn_s32(rf_2), vqmovn_s32(rf_3)));
194 vst1q_u8(output_ptr + x, vcombine_u8(pa, pb));
198 for(; x < window_end_x; ++x)
200 const float afs =
static_cast<int32_t
>((*(input1_ptr + x)) - iq1_info.
offset) * iq1_info.
scale;
201 const float bfs = static_cast<int32_t>((*(input2_ptr + x)) - iq2_info.
offset) * iq2_info.
scale;
205 input1, input2, output);
uint8_t quantize_qasymm8(float value, const INFO_TYPE &qinfo, RoundingPolicy rounding_policy=RoundingPolicy::TO_NEAREST_UP)
Quantize a value given an unsigned 8-bit asymmetric quantization scheme.
constexpr int step() const
Return the step of the dimension.
Describe one of the image's dimensions with a start, end and step.
Interface for CPU tensor.
Copyright (c) 2017-2021 Arm Limited.
void add_qasymm8_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
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.
virtual const TensorShape & tensor_shape() const =0
Size for each dimension of the tensor.
UniformQuantizationInfo uniform() const
Return per layer quantization info.
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.
void set(size_t dimension, const Dimension &dim)
Set the values of a given dimension.
Window broadcast_if_dimension_le_one(const TensorShape &shape) const
Don't advance in the dimension where shape is less equal to 1.
virtual QuantizationInfo quantization_info() const =0
Get the quantization settings (scale and offset) of the tensor.
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...
std::vector< NodeID > bfs(Graph &g)
Breadth first search traversal.
constexpr int end() const
Return the end of the dimension.
Iterator updated by execute_window_loop for each window element.
constexpr int start() const
Return the start of the dimension.
Describe a multidimensional execution window.
ConvertPolicy
Policy to handle overflow.
constexpr const Dimension & x() const
Alias to access the first dimension of the window.