GPUCodeForces/S1/uucoco_#40/PReLU_cuda.py

128 lines
4.0 KiB
Python

import torch
import torch.nn as nn
from torch.utils.cpp_extension import load_inline
class ModelNew(nn.Module):
def __init__(self, num_features=64, init=0.25):
super().__init__()
self.num_features = num_features
self.weight = nn.Parameter(torch.full((num_features,), init))
self._compile_cuda_kernel()
def _compile_cuda_kernel(self):
cpp_source = """
#include <torch/extension.h>
torch::Tensor prelu_cuda(torch::Tensor x, torch::Tensor w);
"""
cuda_source = """
#include <cuda_runtime.h>
__global__ void prelu_channel_vec4(
const float* __restrict__ x,
float* __restrict__ y,
const float* __restrict__ w,
int total_vecs,
int spatial_vecs, // spatial_size / 4
int channels)
{
int idx = blockIdx.x * blockDim.x + threadIdx.x;
int stride = gridDim.x * blockDim.x;
const float4* x_vec = reinterpret_cast<const float4*>(x);
float4* y_vec = reinterpret_cast<float4*>(y);
for (int i = idx; i < total_vecs; i += stride) {
float4 v = x_vec[i];
float4 out;
int c = (i / spatial_vecs) % channels;
float slope = w[c];
out.x = (v.x < 0.0f) ? (v.x * slope) : v.x;
out.y = (v.y < 0.0f) ? (v.y * slope) : v.y;
out.z = (v.z < 0.0f) ? (v.z * slope) : v.z;
out.w = (v.w < 0.0f) ? (v.w * slope) : v.w;
y_vec[i] = out;
}
}
__global__ void prelu_scalar_kernel(
const float* __restrict__ x,
float* __restrict__ y,
const float* __restrict__ w,
int total_elements,
int spatial_size,
int channels)
{
int idx = blockIdx.x * blockDim.x + threadIdx.x;
int stride = gridDim.x * blockDim.x;
for (int i = idx; i < total_elements; i += stride) {
int c = (i / spatial_size) % channels;
float val = x[i];
float slope = w[c];
y[i] = (val < 0.0f) ? (val * slope) : val;
}
}
torch::Tensor prelu_cuda(torch::Tensor x, torch::Tensor w) {
auto x_c = x.contiguous();
auto w_c = w.contiguous();
auto output = torch::empty_like(x_c);
int N = x_c.size(0);
int C = x_c.size(1);
int total_elements = x_c.numel();
// Spatial size (H*W or L) = Total / (N*C)
int S = total_elements / (N * C);
if (S % 4 == 0) {
int total_vecs = total_elements / 4;
int spatial_vecs = S / 4;
int threads = 256;
int blocks = (total_vecs + threads - 1) / threads;
if (blocks > 65535) blocks = 65535;
prelu_channel_vec4<<<blocks, threads>>>(
x_c.data_ptr<float>(),
output.data_ptr<float>(),
w_c.data_ptr<float>(),
total_vecs,
spatial_vecs,
C
);
} else {
int threads = 256;
int blocks = (total_elements + threads - 1) / threads;
if (blocks > 65535) blocks = 65535;
prelu_scalar_kernel<<<blocks, threads>>>(
x_c.data_ptr<float>(),
output.data_ptr<float>(),
w_c.data_ptr<float>(),
total_elements,
S,
C
);
}
return output;
}
"""
self.op = load_inline(
name="prelu_opt_v3_fix",
cpp_sources=cpp_source,
cuda_sources=cuda_source,
functions=["prelu_cuda"],
extra_cuda_cflags=["-O3", "--use_fast_math"],
verbose=False
)
def forward(self, x):
return self.op.prelu_cuda(x, self.weight)