51 const bool requantize =
src->quantization_info() !=
dst->quantization_info();
53 switch(
src->data_type())
58 create_arm_pooling_requant<uint8_t, uint8_t>(
src,
dst,
info, cpu_info);
62 create_arm_pooling<uint8_t, uint8_t>(
src,
dst,
info, cpu_info);
68 create_arm_pooling_requant<int8_t, int8_t>(
src,
dst,
info, cpu_info);
72 create_arm_pooling<int8_t, int8_t>(
src,
dst,
info, cpu_info);
75 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC 77 create_arm_pooling<float16_t, float16_t>(
src,
dst,
info, cpu_info);
81 create_arm_pooling<float, float>(
src,
dst,
info, cpu_info);
88 INEKernel::configure(win);
102 "Only AVG and MAX pooling are supported by assembly kernels");
104 if(
dst->total_size() > 0)
108 const auto src_qinfo =
src->quantization_info().uniform();
109 const auto dst_qinfo =
dst->quantization_info().uniform();
111 if(src_qinfo != dst_qinfo)
113 const float multiplier = src_qinfo.scale / dst_qinfo.scale;
114 int32_t dst_multiplier{};
122 const bool has_padding =
info.pad_stride_info.has_padding();
132 const bool has_padding =
info.pad_stride_info.has_padding();
152 const auto in_ptr =
src->buffer() +
src->info()->offset_first_element_in_bytes();
153 auto out_ptr =
dst->buffer() +
dst->info()->offset_first_element_in_bytes();
156 const auto src_shape =
src->info()->tensor_shape();
158 const auto src_padding =
src->info()->padding();
159 const auto dst_padding =
dst->info()->padding();
161 const size_t ld_src_col = src_shape[0] + src_padding.left + src_padding.right;
162 const size_t ld_src_row = ld_src_col * (src_shape[1] + src_padding.top + src_padding.bottom);
163 const size_t ld_src_batch = ld_src_row * src_shape[2];
164 const size_t ld_dst_col =
dst_shape[0] + dst_padding.left + dst_padding.right;
165 const size_t ld_dst_row = ld_dst_col * (
dst_shape[1] + dst_padding.top + dst_padding.bottom);
166 const size_t ld_dst_batch = ld_dst_row *
dst_shape[2];
168 _kernel_asm->execute(in_ptr, ld_src_col, ld_src_row, ld_src_batch,
169 out_ptr, ld_dst_col, ld_dst_row, ld_dst_batch,
170 working_space,
info.thread_id,
info.num_threads);
175 return _kernel_asm->get_working_size(num_threads);
180 return _kernel_asm !=
nullptr;
183 template <
typename Typesrc,
typename Typedst>
188 arm_conv::pooling::PoolingWindow window{};
189 window.cols = static_cast<unsigned int>(
info.pool_size.x());
190 window.rows = static_cast<unsigned int>(
info.pool_size.y());
192 arm_conv::pooling::PoolingStride stride{};
193 std::tie(stride.cols, stride.rows) =
info.pad_stride_info.stride();
195 const arm_conv::pooling::PaddingValues padding{
info.pad_stride_info.pad_left(),
info.pad_stride_info.pad_top(),
info.pad_stride_info.pad_right(),
info.pad_stride_info.pad_bottom() };
199 constexpr
unsigned int idx_channels = 0;
200 constexpr
unsigned int idx_batches = 3;
202 const unsigned int n_batches =
src->dimension(idx_batches);
204 const unsigned int src_cols =
src->dimension(
idx_width);
205 const unsigned int n_channels =
src->dimension(idx_channels);
207 const unsigned int dst_cols =
dst->dimension(
idx_width);
209 arm_conv::pooling::PoolingArgs
args(&cpu_info, pool_type, window, stride,
info.exclude_padding, n_batches, src_rows, src_cols, n_channels, dst_rows, dst_cols, padding,
nullptr);
212 auto pooling_kernel_asm = arm_conv::pooling::pooling<Typesrc, Typedst>(
args);
213 if(pooling_kernel_asm ==
nullptr)
219 _kernel_asm = std::move(pooling_kernel_asm);
222 template <
typename Typesrc,
typename Typedst>
223 void CpuPoolingAssemblyWrapperKernel::create_arm_pooling_requant(
const ITensorInfo *
src, ITensorInfo *
dst,
const PoolingLayerInfo &
info,
const CPUInfo &cpu_info)
227 arm_conv::pooling::PoolingWindow window{};
228 window.cols = static_cast<unsigned int>(
info.pool_size.x());
229 window.rows = static_cast<unsigned int>(
info.pool_size.y());
231 arm_conv::pooling::PoolingStride stride{};
232 std::tie(stride.cols, stride.rows) =
info.pad_stride_info.stride();
234 const arm_conv::pooling::PaddingValues padding{
info.pad_stride_info.pad_left(),
info.pad_stride_info.pad_top(),
info.pad_stride_info.pad_right(),
info.pad_stride_info.pad_bottom() };
238 constexpr
unsigned int idx_channels = 0;
239 constexpr
unsigned int idx_batches = 3;
241 const unsigned int n_batches =
src->dimension(idx_batches);
243 const unsigned int src_cols =
src->dimension(
idx_width);
244 const unsigned int n_channels =
src->dimension(idx_channels);
246 const unsigned int dst_cols =
dst->dimension(
idx_width);
248 arm_conv::pooling::PoolingArgs
args(&cpu_info, pool_type, window, stride,
info.exclude_padding, n_batches, src_rows, src_cols, n_channels, dst_rows, dst_cols, padding,
nullptr);
250 const auto src_qinfo =
src->quantization_info().uniform();
251 const auto dst_qinfo =
dst->quantization_info().uniform();
253 const float multiplier = src_qinfo.scale / dst_qinfo.scale;
254 int32_t dst_multiplier{};
258 const arm_conv::pooling::Requantize32 requant_args(src_qinfo.offset,
265 auto pooling_kernel_asm = arm_conv::pooling::pooling<Typesrc, Typedst, arm_conv::pooling::Requantize32>(
args, requant_args);
266 if(pooling_kernel_asm ==
nullptr)
272 _kernel_asm = std::move(pooling_kernel_asm);
Window calculate_max_window(const ValidRegion &valid_region, const Steps &steps, bool skip_border, BorderSize border_size)
#define ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(tensor)
bool empty() const
Checks if pack is empty.
1 channel, 1 F32 per channel
#define ARM_COMPUTE_ERROR_ON(cond)
If the condition is true then an error message is printed and an exception thrown.
Store the tensor's metadata.
Status calculate_quantized_multiplier(float multiplier, int32_t *quant_multiplier, int32_t *shift, bool ignore_epsilon=false)
Calculate quantized representation of multiplier.
#define ARM_COMPUTE_RETURN_ERROR_ON(cond)
If the condition is true, an error is returned.
Interface for CPU tensor.
SimpleTensor< float > src
Copyright (c) 2017-2021 Arm Limited.
1 channel, 1 F16 per channel
static Status validate(const ITensorInfo *src, const ITensorInfo *dst, const PoolingLayerInfo &info)
Indicates whether or not this function can be used to process the given parameters.
#define ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(...)
TensorShape compute_pool_shape(const ITensorInfo &input, PoolingLayerInfo pool_info)
Calculate the output pool shape of a tensor.
bool is_configured() const
Was the asm kernel successfully configured?
const ITensor * get_const_tensor(int id) const
Get constant tensor of a given id.
#define ARM_COMPUTE_UNUSED(...)
To avoid unused variables warnings.
void run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) override
Execute the kernel on the passed window.
quantized, asymmetric fixed-point 8-bit number unsigned
Class to describe a number of elements in each dimension.
virtual uint8_t * buffer() const =0
Interface to be implemented by the child class to return a pointer to CPU memory.
Pooling Layer Information struct.
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 ITensorInfo * info() const =0
Interface to be implemented by the child class to return the tensor's metadata.
size_t get_working_size(unsigned int num_threads) const
Get size of the workspace needed by the assembly kernel.
#define ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(k)
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.
PoolingType
Available pooling types.
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.
#define ARM_COMPUTE_RETURN_ERROR_MSG(...)
An error is returned with the given description.
Information about executing thread and CPU.
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(...)
Num samples, height, width, channels.
#define ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(t, c,...)
#define ARM_COMPUTE_RETURN_ERROR_ON_MSG(cond, msg)
If the condition is true, an error is returned.
#define ARM_COMPUTE_ERROR_ON_NULLPTR(...)
quantized, asymmetric fixed-point 8-bit number signed
DataLayout data_layout
Data layout to use.
Describe a multidimensional execution window.
void configure(const ITensorInfo *src, ITensorInfo *dst, const PoolingLayerInfo &info, const CPUInfo &cpu_info)
Initialise the kernel's src and dst.