"""GPU elementwise addition kernel."""
from __future__ import annotations
from typing import Any, Mapping
import torch
from shmpipeline.config import KernelConfig, SharedMemoryConfig
from shmpipeline.kernels.gpu._common import (
validate_binary_same_shape_and_dtype,
)
from shmpipeline.kernels.gpu.base import GpuKernel, as_gpu_tensor
[docs]
class ElementwiseAddGpuKernel(GpuKernel):
"""Compute output = input + auxiliary elementwise."""
kind = "gpu.elementwise_add"
auxiliary_arity = 1
[docs]
@classmethod
def validate_config(
cls,
config: KernelConfig,
shared_memory: Mapping[str, SharedMemoryConfig],
) -> None:
super().validate_config(config, shared_memory)
validate_binary_same_shape_and_dtype(config, shared_memory)
[docs]
def compute_into(
self,
trigger_input: Any,
output: Any,
auxiliary_inputs: Mapping[str, Any],
) -> None:
alias = self.context.config.auxiliary_aliases[0]
rhs = as_gpu_tensor(auxiliary_inputs[alias], device=self.device)
torch.add(
as_gpu_tensor(trigger_input, device=self.device), rhs, out=output
)
torch.cuda.synchronize(output.device)