49 Status validate_arguments(
const ITensorInfo *
src,
const ITensorInfo *bias,
const ITensorInfo *
dst,
int min,
int max)
63 if(dst->total_size() != 0)
73 template <
bool is_bounded_relu>
74 void CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::run_internal(
const ITensor *src,
const ITensor *bias, ITensor *dst,
const Window &window)
76 const int16x8_t min_s16 = vdupq_n_s16(static_cast<int16_t>(_min));
77 const int16x8_t max_s16 = vdupq_n_s16(static_cast<int16_t>(_max));
82 const int window_step_x = 8;
83 const auto window_start_x =
static_cast<int>(window.x().start());
84 const auto window_end_x =
static_cast<int>(window.x().end());
86 Window win_collapsed = window.collapse_if_possible(window,
Window::DimZ);
87 win_collapsed.set(
Window::DimX, Window::Dimension(0, 1, 1));
89 Iterator in(src, win_collapsed);
90 Iterator out(dst, win_collapsed);
94 win_biases.set(
Window::DimX, Window::Dimension(0, 1, 1));
95 win_biases.set(
Window::DimY, Window::Dimension(0, 1, 1));
97 Iterator bias_i(bias, win_biases);
101 int x = window_start_x;
102 for(; x <= (window_end_x - window_step_x); x += window_step_x)
107 vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 0),
108 vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 4)
112 const int32x4x2_t bias_s32 =
115 vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 0),
116 vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 4)
121 in_s32.val[0] = vaddq_s32(in_s32.val[0], bias_s32.val[0]);
122 in_s32.val[1] = vaddq_s32(in_s32.val[1], bias_s32.val[1]);
124 vst1q_s16(reinterpret_cast<int16_t *>(out.ptr()) + x, finalize_quantization_int16<is_bounded_relu>(in_s32, _result_fixedpoint_multiplier, _result_shift, min_s16, max_s16));
128 for(; x < window_end_x; ++x)
130 const int32_t bias_value = *(
reinterpret_cast<const int32_t *
>(bias_i.ptr()) + x);
131 int32_t in_value = *(
reinterpret_cast<const int32_t *
>(in.ptr()) + x);
134 in_value += bias_value;
136 *(
reinterpret_cast<int16_t *
>(out.ptr()) + x) = finalize_quantization_int16<is_bounded_relu>(in_value, _result_fixedpoint_multiplier, _result_shift,
static_cast<int16_t
>(_min),
137 static_cast<int16_t>(_max));
147 int x = window_start_x;
148 for(; x <= (window_end_x - window_step_x); x += window_step_x)
153 vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 0),
154 vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 4)
158 vst1q_s16(reinterpret_cast<int16_t *>(out.ptr()) + x, finalize_quantization_int16<is_bounded_relu>(in_s32, _result_fixedpoint_multiplier, _result_shift, min_s16, max_s16));
162 for(; x < window_end_x; ++x)
164 const int32_t in_value = *(
reinterpret_cast<const int32_t *
>(in.ptr()) + x);
167 *(
reinterpret_cast<int16_t *
>(out.ptr()) + x) = finalize_quantization_int16<is_bounded_relu>(in_value, _result_fixedpoint_multiplier, _result_shift,
static_cast<int16_t
>(_min),
168 static_cast<int16_t>(_max));
183 _result_fixedpoint_multiplier = result_fixedpoint_multiplier;
184 _result_shift = result_shift;
192 ICpuKernel::configure(win_config);
195 const bool is_bounded_relu = !(min <= -32768 && max >= 32767);
196 _func = is_bounded_relu ? &CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::run_internal<true> :
197 &CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::run_internal<false>;
218 (this->*_func)(src, bias, dst, window);
223 return "CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel";
Window calculate_max_window(const ValidRegion &valid_region, const Steps &steps, bool skip_border, BorderSize border_size)
const Window & window() const
The maximum window the kernel can be executed on.
quantized, symmetric fixed-point 16-bit number
void configure(ITensorInfo *src, ITensorInfo *bias, ITensorInfo *dst, int result_fixedpoint_multiplier, int result_shift, int min=0, int max=0)
Initialise the kernel's input and output.
void run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) override
Execute the kernel on the passed window.
bool empty() const
Checks if pack is empty.
#define ARM_COMPUTE_RETURN_ON_ERROR(status)
Checks if a status contains an error and returns it.
Store the tensor's metadata.
#define ARM_COMPUTE_ERROR_THROW_ON(status)
#define ARM_COMPUTE_RETURN_ERROR_ON(cond)
If the condition is true, an error is returned.
SimpleTensor< float > src
Copyright (c) 2017-2021 Arm Limited.
1 channel, 1 S32 per channel
const ITensor * get_const_tensor(int id) const
Get constant tensor of a given id.
static constexpr size_t DimX
Alias for dimension 0 also known as X dimension.
#define ARM_COMPUTE_UNUSED(...)
To avoid unused variables warnings.
Class to describe a number of elements in each dimension.
#define ARM_COMPUTE_ERROR_ON_MSG(cond, msg)
bool auto_init_if_empty(ITensorInfo &info, const TensorShape &shape, int num_channels, DataType data_type, QuantizationInfo quantization_info=QuantizationInfo())
Auto initialize the tensor info (shape, number of channels and data type) if the current assignment i...
virtual std::unique_ptr< T > clone() const =0
Provide a clone of the current object of class T.
const char * name() const override
Name of the kernel.
#define ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(k)
static constexpr size_t DimY
Alias for dimension 1 also known as Y dimension.
ScaleKernelInfo info(interpolation_policy, default_border_mode, PixelValue(), sampling_policy, false)
ITensor * get_tensor(int id)
Get tensor of a given id from the pac.
Information about executing thread and CPU.
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(...)
static constexpr size_t DimZ
Alias for dimension 2 also known as Z dimension.
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(...)
#define ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(t, c,...)
static Status validate(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, int min=0, int max=0)
Static function to check if given info will lead to a valid configuration.
#define ARM_COMPUTE_ERROR_ON_NULLPTR(...)
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...
Describe a multidimensional execution window.
#define ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(f, s)