GPUCodeForces/S1/uucoco_#23/LogSigmoid_cuda.py

88 lines
2.5 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 = """
#include <torch/extension.h>
torch::Tensor logsigmoid_cuda(torch::Tensor x);
"""
cuda_source = """
#include <cuda_runtime.h>
__device__ __forceinline__ float logsigmoid_op(float x) {
if (x > 0.0f) {
return -log1pf(expf(-x));
} else {
return x - log1pf(expf(x));
}
}
__global__ void logsigmoid_kernel_vec4(
const float* __restrict__ x,
float* __restrict__ y,
int n)
{
int idx = blockIdx.x * blockDim.x + threadIdx.x;
int stride = blockDim.x * gridDim.x;
int vec_n = n / 4;
const float4* x_vec = reinterpret_cast<const float4*>(x);
float4* y_vec = reinterpret_cast<float4*>(y);
for (int i = idx; i < vec_n; i += stride) {
float4 v = x_vec[i];
float4 out;
out.x = logsigmoid_op(v.x);
out.y = logsigmoid_op(v.y);
out.z = logsigmoid_op(v.z);
out.w = logsigmoid_op(v.w);
y_vec[i] = out;
}
int tail = vec_n * 4;
for (int i = tail + idx; i < n; i += stride) {
y[i] = logsigmoid_op(x[i]);
}
}
torch::Tensor logsigmoid_cuda(torch::Tensor x) {
auto x_c = x.contiguous();
auto output = torch::empty_like(x_c);
int n = x_c.numel();
int threads = 256;
int blocks = (n / 4 + threads - 1) / threads;
if (blocks > 65535) blocks = 65535;
if (blocks == 0) blocks = 1;
logsigmoid_kernel_vec4<<<blocks, threads>>>(
x_c.data_ptr<float>(),
output.data_ptr<float>(),
n
);
return output;
}
"""
self.op = load_inline(
name="logsigmoid_opt_vec4",
cpp_sources=cpp_source,
cuda_sources=cuda_source,
functions=["logsigmoid_cuda"],
extra_cuda_cflags=["-O3", "--use_fast_math"],
verbose=False
)
def forward(self, x):
return self.op.logsigmoid_cuda(x)