GPUCodeForces/S1/gsd123_#42/CosLU_cuda.py

95 lines
3.1 KiB
Python

import torch
import torch.nn as nn
from torch.utils.cpp_extension import load_inline
class ModelNew(nn.Module):
def __init__(self, alpha=1.0, beta=1.0):
super().__init__()
self.alpha = alpha
self.beta = beta
self._compile_cuda_kernel()
def _compile_cuda_kernel(self):
cpp_source = """
torch::Tensor coslu_cuda(torch::Tensor x, float alpha, float beta);
"""
cuda_source = """
#include <torch/extension.h>
#include <cuda_runtime.h>
#include <math.h>
__device__ __forceinline__ float sigmoid_f(float x) {
return 1.0f / (1.0f + expf(-x));
}
__device__ __forceinline__ float coslu_op(float x, float alpha, float beta) {
// CosLU(x) = (x + alpha * cos(beta * x)) * sigmoid(x)
float term_cos = alpha * cosf(beta * x);
return (x + term_cos) * sigmoid_f(x);
}
__global__ void coslu_kernel(
const float* __restrict__ x,
float* __restrict__ output,
const int n_elements,
const float alpha,
const float beta)
{
const int tid = blockIdx.x * blockDim.x + threadIdx.x;
const int stride = blockDim.x * gridDim.x;
const int vec_loops = n_elements >> 2;
const float4* x_vec = reinterpret_cast<const float4*>(x);
float4* out_vec = reinterpret_cast<float4*>(output);
for (int i = tid; i < vec_loops; i += stride) {
float4 v = __ldg(&x_vec[i]);
float4 r;
r.x = coslu_op(v.x, alpha, beta);
r.y = coslu_op(v.y, alpha, beta);
r.z = coslu_op(v.z, alpha, beta);
r.w = coslu_op(v.w, alpha, beta);
out_vec[i] = r;
}
const int tail_start = vec_loops << 2;
for (int i = tail_start + tid; i < n_elements; i += stride) {
output[i] = coslu_op(x[i], alpha, beta);
}
}
torch::Tensor coslu_cuda(torch::Tensor x, float alpha, float beta) {
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 max_blocks = 65535;
const int blocks = std::min((n_elements + threads * 4 - 1) / (threads * 4), max_blocks);
coslu_kernel<<<blocks, threads>>>(
x_c.data_ptr<float>(),
output.data_ptr<float>(),
n_elements,
alpha,
beta
);
return output;
}
"""
self.op = load_inline(
name="coslu_v1",
cpp_sources=cpp_source,
cuda_sources=cuda_source,
functions=["coslu_cuda"],
extra_cuda_cflags=["-O3", "--use_fast_math"],
verbose=False
)
def forward(self, x):
return self.op.coslu_cuda(x, self.alpha, self.beta)