GPUCodeForces/S1/uucoco_#29/DecayingSineUnit_cuda.py

73 lines
2.0 KiB
Python

import torch
import torch.nn as nn
from torch.utils.cpp_extension import load_inline
class ModelNew(nn.Module):
def __init__(self):
super().__init__()
self._compile_cuda_kernel()
def _compile_cuda_kernel(self):
cpp_source = """
torch::Tensor dsu_cuda(torch::Tensor x);
"""
cuda_source = """
#include <torch/extension.h>
#include <cuda_runtime.h>
#include <math.h>
__device__ __forceinline__ float dsu_op(float x) {
const float PI = 3.14159265358979323846f;
const float y = x - PI;
if (fabsf(y) < 1e-6f) {
return PI;
} else {
return sinf(PI * y) / y;
}
}
__global__ void dsu_kernel(
const float* __restrict__ x,
float* __restrict__ output,
const int n_elements)
{
const int tid = blockIdx.x * blockDim.x + threadIdx.x;
const int stride = blockDim.x * gridDim.x;
for (int i = tid; i < n_elements; i += stride) {
output[i] = dsu_op(x[i]);
}
}
torch::Tensor dsu_cuda(torch::Tensor x) {
auto x_c = x.contiguous();
const int n_elements = x_c.numel();
auto output = torch::empty_like(x_c);
const int threads = 256;
const int blocks = std::min((n_elements + threads - 1) / threads, 65535);
dsu_kernel<<<blocks, threads>>>(
x_c.data_ptr<float>(),
output.data_ptr<float>(),
n_elements
);
return output;
}
"""
self.op = load_inline(
name="dsu_v1",
cpp_sources=cpp_source,
cuda_sources=cuda_source,
functions=["dsu_cuda"],
extra_cuda_cflags=["-O3"],
verbose=False
)
def forward(self, x):
return self.op.dsu_cuda(x)