GPUCodeForces/S1/gsd123_#45/FReLU_cuda.py

87 lines
2.7 KiB
Python

import torch
import torch.nn as nn
from torch.utils.cpp_extension import load_inline
class ModelNew(nn.Module):
def __init__(self, b=0.0):
super().__init__()
self.b = b
self._compile_cuda_kernel()
def _compile_cuda_kernel(self):
cpp_source = """
torch::Tensor frelu_cuda(torch::Tensor x, float b);
"""
cuda_source = """
#include <torch/extension.h>
#include <cuda_runtime.h>
#include <math.h>
__device__ __forceinline__ float frelu_op(float x, float b) {
// FReLU(x) = x + b if x > 0 else b
return (x > 0.0f) ? (x + b) : b;
}
__global__ void frelu_kernel(
const float* __restrict__ x,
float* __restrict__ output,
const int n_elements,
const float b)
{
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 = frelu_op(v.x, b);
r.y = frelu_op(v.y, b);
r.z = frelu_op(v.z, b);
r.w = frelu_op(v.w, b);
out_vec[i] = r;
}
const int tail_start = vec_loops << 2;
for (int i = tail_start + tid; i < n_elements; i += stride) {
output[i] = frelu_op(x[i], b);
}
}
torch::Tensor frelu_cuda(torch::Tensor x, float b) {
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);
frelu_kernel<<<blocks, threads>>>(
x_c.data_ptr<float>(),
output.data_ptr<float>(),
n_elements,
b
);
return output;
}
"""
self.op = load_inline(
name="frelu_v1",
cpp_sources=cpp_source,
cuda_sources=cuda_source,
functions=["frelu_cuda"],
extra_cuda_cflags=["-O3", "--use_fast_math"],
verbose=False
)
def forward(self, x):
return self.op.frelu_cuda(x, self.b)