forked from ccf-ai-infra/GPUCodeForces
112 lines
3.4 KiB
Python
112 lines
3.4 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()
|
|
self.epsilon = 1e-5
|
|
|
|
def _compile_cuda_kernel(self):
|
|
cpp_source = """
|
|
torch::Tensor rsg_apply_cuda(torch::Tensor x, torch::Tensor x_flat_sorted, float epsilon);
|
|
"""
|
|
|
|
cuda_source = """
|
|
#include <torch/extension.h>
|
|
#include <cuda_runtime.h>
|
|
#include <math.h>
|
|
|
|
__device__ __forceinline__ float sigmoid_op(float x) {
|
|
if (x >= 0.0f) {
|
|
return 1.0f / (1.0f + expf(-x));
|
|
} else {
|
|
float z = expf(x);
|
|
return z / (1.0f + z);
|
|
}
|
|
}
|
|
|
|
__global__ void rsg_apply_kernel(
|
|
const float* __restrict__ x,
|
|
float* __restrict__ output,
|
|
const int n_elements,
|
|
const float median_val,
|
|
const float iqr_recip)
|
|
{
|
|
const int tid = blockIdx.x * blockDim.x + threadIdx.x;
|
|
const int stride = blockDim.x * gridDim.x;
|
|
|
|
for (int i = tid; i < n_elements; i += stride) {
|
|
float val = x[i];
|
|
float z_robust = (val - median_val) * iqr_recip;
|
|
float gate = sigmoid_op(val);
|
|
output[i] = z_robust * gate;
|
|
}
|
|
}
|
|
|
|
torch::Tensor rsg_apply_cuda(torch::Tensor x, torch::Tensor x_flat_sorted, float epsilon) {
|
|
auto x_c = x.contiguous();
|
|
const int n_elements = x_c.numel();
|
|
|
|
int n = x_flat_sorted.size(0);
|
|
|
|
float median_val, q1_val, q3_val;
|
|
|
|
int median_idx = n / 2;
|
|
int q1_idx = n / 4;
|
|
int q3_idx = (3 * n) / 4;
|
|
|
|
if (n % 2 == 0) {
|
|
median_val = (x_flat_sorted[median_idx - 1].item<float>() + x_flat_sorted[median_idx].item<float>()) * 0.5f;
|
|
} else {
|
|
median_val = x_flat_sorted[median_idx].item<float>();
|
|
}
|
|
|
|
if (n % 4 == 0) {
|
|
q1_val = (x_flat_sorted[q1_idx - 1].item<float>() + x_flat_sorted[q1_idx].item<float>()) * 0.5f;
|
|
} else {
|
|
q1_val = x_flat_sorted[q1_idx].item<float>();
|
|
}
|
|
|
|
if ((3 * n) % 4 == 0) {
|
|
q3_val = (x_flat_sorted[q3_idx - 1].item<float>() + x_flat_sorted[q3_idx].item<float>()) * 0.5f;
|
|
} else {
|
|
q3_val = x_flat_sorted[q3_idx].item<float>();
|
|
}
|
|
|
|
float iqr = q3_val - q1_val;
|
|
float iqr_recip = 1.0f / (iqr + epsilon);
|
|
|
|
auto output = torch::empty_like(x_c);
|
|
|
|
const int threads = 256;
|
|
const int blocks = min((n_elements + threads - 1) / threads, 65535);
|
|
|
|
rsg_apply_kernel<<<blocks, threads>>>(
|
|
x_c.data_ptr<float>(),
|
|
output.data_ptr<float>(),
|
|
n_elements,
|
|
median_val,
|
|
iqr_recip
|
|
);
|
|
|
|
return output;
|
|
}
|
|
"""
|
|
|
|
self.op = load_inline(
|
|
name="rsg_v4",
|
|
cpp_sources=cpp_source,
|
|
cuda_sources=cuda_source,
|
|
functions=["rsg_apply_cuda"],
|
|
extra_cuda_cflags=["-O3"],
|
|
verbose=False
|
|
)
|
|
|
|
def forward(self, x):
|
|
x_flat = x.flatten()
|
|
x_flat_sorted = x_flat.sort()[0]
|
|
|
|
return self.op.rsg_apply_cuda(x, x_flat_sorted, self.epsilon) |