GPUCodeForces/S1/uucoco_#42/PSMish_cuda.py

79 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, alpha: float = 1.0, beta: float = 1.0):
super().__init__()
self.alpha = alpha
self.beta = beta
self._compile_cuda_kernel()
def _compile_cuda_kernel(self):
cpp_source = """
torch::Tensor psmish_cuda(torch::Tensor x, float alpha, float beta);
"""
cuda_source = """
#include <torch/extension.h>
#include <cuda_runtime.h>
#include <math.h>
__device__ __forceinline__ double psmish_op_double(double x, double alpha, double beta) {
double bx = beta * x;
double gate = tanh(log(1.0 + exp(bx)));
return alpha * x * gate;
}
__global__ void psmish_kernel_double_intermediate(
const float* __restrict__ x,
float* __restrict__ output,
const int n_elements,
const float alpha,
const float beta)
{
const int block_start = blockIdx.x * blockDim.x;
const int block_end = min(block_start + blockDim.x, n_elements);
double alpha_d = (double)alpha;
double beta_d = (double)beta;
for (int i = block_start + threadIdx.x; i < block_end; i += blockDim.x) {
double val_d = (double)x[i];
double result_d = psmish_op_double(val_d, alpha_d, beta_d);
output[i] = (float)result_d;
}
}
torch::Tensor psmish_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 - 1) / threads, max_blocks);
psmish_kernel_double_intermediate<<<blocks, threads>>>(
x_c.data_ptr<float>(),
output.data_ptr<float>(),
n_elements,
alpha,
beta
);
return output;
}
"""
self.op = load_inline(
name="psmish_v3",
cpp_sources=cpp_source,
cuda_sources=cuda_source,
functions=["psmish_cuda"],
extra_cuda_cflags=["-O3"],
verbose=False
)
def forward(self, x):
return self.op.psmish_cuda(x, self.alpha, self.beta)