forked from ccf-ai-infra/GPUCodeForces
67 lines
1.9 KiB
Python
67 lines
1.9 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 serf_cuda(torch::Tensor x);
|
|
"""
|
|
|
|
cuda_source = """
|
|
#include <torch/extension.h>
|
|
#include <cuda_runtime.h>
|
|
#include <math.h>
|
|
|
|
__device__ __forceinline__ float serf_op(float x) {
|
|
return x * erff(logf(1.0f + expf(x)));
|
|
}
|
|
|
|
__global__ void serf_tiled_kernel(
|
|
const float* __restrict__ x,
|
|
float* __restrict__ output,
|
|
const int n_elements)
|
|
{
|
|
const int block_start = blockIdx.x * blockDim.x;
|
|
const int block_end = min(block_start + blockDim.x, n_elements);
|
|
|
|
for (int i = block_start + threadIdx.x; i < block_end; i += blockDim.x) {
|
|
output[i] = serf_op(x[i]);
|
|
}
|
|
}
|
|
|
|
torch::Tensor serf_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 max_blocks = 65535;
|
|
const int blocks = std::min((n_elements + threads - 1) / threads, max_blocks);
|
|
|
|
serf_tiled_kernel<<<blocks, threads>>>(
|
|
x_c.data_ptr<float>(),
|
|
output.data_ptr<float>(),
|
|
n_elements
|
|
);
|
|
|
|
return output;
|
|
}
|
|
"""
|
|
|
|
self.op = load_inline(
|
|
name="serf_v2",
|
|
cpp_sources=cpp_source,
|
|
cuda_sources=cuda_source,
|
|
functions=["serf_cuda"],
|
|
extra_cuda_cflags=["-O3"],
|
|
verbose=False
|
|
)
|
|
|
|
def forward(self, x):
|
|
return self.op.serf_cuda(x)
|