SimaBEVPoolOptions Struct

Construction options for a SimaBEVPool Node. More...

Declaration

struct simaai::neat::SimaBEVPoolOptions { ... }

Included Headers

#include <SimaBEVPool.h>

Public Member Attributes Index

std::array< int, 3 >	cells {}
	BEV cell counts nX, nY, nZ; must match the plan. More...
int	depth_zero_point = 0
	INT8 zero point of depth; 0 for BF16. More...
int	context_zero_point = 0
	INT8 zero point of context; 0 for BF16. More...
SimaBEVPoolPlanFormat	plan_format = SimaBEVPoolPlanFormat::V1
	Layout of the bound plan: the producing SimaBEVPoolPlan's format. More...
SimaKernelQuantization	quantization
	Applied to each cell's binary32 sum for an INT8 bev. More...
SimaBEVPoolStorage	input_storage = SimaBEVPoolStorage::ChannelsLast
	depth/context order. More...
SimaBEVPoolStorage	output_storage = SimaBEVPoolStorage::ChannelsLast
	bev byte order. More...
TensorDType	output_dtype = TensorDType::Int8
	bev element type. More...
SimaKernelTarget	target = SimaKernelTarget::A65
	Execution target; no fallback. More...

Description

Construction options for a SimaBEVPool Node.

The input Sample holds three tensors, in order: depth, context (both INT8 or both BF16, layout per input_storage) and the UINT8 plan [P] produced by SimaBEVPoolPlan for the same frustum and grid. The output Sample holds one bev tensor. Per-channel quantization metadata is unsupported and rejected rather than discarded.

BF16 inputs need zero points of 0. Their sums are binary32, in the plan's point order. A BF16 bev stores each sum rounded to nearest-even (NaN as 32704) and needs the identity quantization; an INT8 bev from BF16 inputs quantizes the sum with multiplier = 1 / output_scale. INT8 inputs cannot produce a BF16 bev.

Definition at line 59 of file SimaBEVPool.h.

Public Member Attributes

cells

std::array<int, 3> simaai::neat::SimaBEVPoolOptions::cells {}

BEV cell counts nX, nY, nZ; must match the plan.

Definition at line 60 of file SimaBEVPool.h.

context_zero_point

int simaai::neat::SimaBEVPoolOptions::context_zero_point = 0

INT8 zero point of context; 0 for BF16.

Definition at line 62 of file SimaBEVPool.h.

depth_zero_point

int simaai::neat::SimaBEVPoolOptions::depth_zero_point = 0

INT8 zero point of depth; 0 for BF16.

Definition at line 61 of file SimaBEVPool.h.

input_storage

SimaBEVPoolStorage simaai::neat::SimaBEVPoolOptions::input_storage = SimaBEVPoolStorage::ChannelsLast

depth/context order.

Definition at line 69 of file SimaBEVPool.h.

output_dtype

TensorDType simaai::neat::SimaBEVPoolOptions::output_dtype = TensorDType::Int8

bev element type.

Definition at line 71 of file SimaBEVPool.h.

output_storage

SimaBEVPoolStorage simaai::neat::SimaBEVPoolOptions::output_storage = SimaBEVPoolStorage::ChannelsLast

bev byte order.

Definition at line 70 of file SimaBEVPool.h.

plan_format

SimaBEVPoolPlanFormat simaai::neat::SimaBEVPoolOptions::plan_format = SimaBEVPoolPlanFormat::V1

Layout of the bound plan: the producing SimaBEVPoolPlan's format.

Definition at line 64 of file SimaBEVPool.h.

quantization

SimaKernelQuantization simaai::neat::SimaBEVPoolOptions::quantization

Applied to each cell's binary32 sum for an INT8 bev.

For INT8 inputs set multiplier = depth_scale * context_scale / output_scale; for BF16 inputs, 1 / output_scale. A BF16 bev needs the identity.

Definition at line 68 of file SimaBEVPool.h.

target

SimaKernelTarget simaai::neat::SimaBEVPoolOptions::target = SimaKernelTarget::A65

Execution target; no fallback.

Definition at line 73 of file SimaBEVPool.h.

---

The documentation for this struct was generated from the following file:

• SimaBEVPool.h

---

Generated via doxygen2docusaurus 2.0.0 by Doxygen 1.15.0.