forked from ccf-ai-infra/GPUCodeForces
91 lines
2.3 KiB
Python
91 lines
2.3 KiB
Python
import torch
|
|
import torch.nn as nn
|
|
from torch.utils.cpp_extension import load_inline
|
|
|
|
cuda_source = """
|
|
#include <torch/extension.h>
|
|
#include <cuda_runtime.h>
|
|
|
|
__global__ void complex_conj_mul_div_kernel(
|
|
const float* __restrict__ a,
|
|
const float* __restrict__ b,
|
|
const float* __restrict__ c,
|
|
float* __restrict__ output,
|
|
int N, float eps
|
|
) {
|
|
int idx = blockIdx.x * blockDim.x + threadIdx.x;
|
|
if (idx < N) {
|
|
int offset = idx * 2;
|
|
|
|
// Input A = Xa + iYa
|
|
float Xa = a[offset];
|
|
float Ya = a[offset + 1];
|
|
|
|
// Input B = Xb + iYb
|
|
float Xb = b[offset];
|
|
float Yb = b[offset + 1];
|
|
|
|
// Input C = Xc + iYc
|
|
float Xc = c[offset];
|
|
float Yc = c[offset + 1];
|
|
|
|
// Step 1 & 2: P = conj(A) * B = Xp + iYp
|
|
// Xp = Xa*Xb + Ya*Yb
|
|
// Yp = Xa*Yb - Ya*Xb
|
|
float Xp = fmaf(Xa, Xb, Ya * Yb);
|
|
float Yp = fmaf(Xa, Yb, -Ya * Xb);
|
|
|
|
// Step 3: Division Out = P / C
|
|
// Denominator D = |C|^2 = Xc^2 + Yc^2
|
|
float D = fmaf(Xc, Xc, Yc * Yc);
|
|
float inv_D = 1.0f / (D + eps);
|
|
|
|
// Out_Re = (Xp*Xc + Yp*Yc) / D
|
|
// Out_Im = (Yp*Xc - Xp*Yc) / D
|
|
float Out_Re = fmaf(Xp, Xc, Yp * Yc) * inv_D;
|
|
float Out_Im = fmaf(Yp, Xc, -Xp * Yc) * inv_D;
|
|
|
|
output[offset] = Out_Re;
|
|
output[offset + 1] = Out_Im;
|
|
}
|
|
}
|
|
|
|
torch::Tensor complex_conj_mul_div_cuda(torch::Tensor a, torch::Tensor b, torch::Tensor c) {
|
|
auto output = torch::empty_like(a);
|
|
int N = a.size(0);
|
|
|
|
const int block_size = 256;
|
|
int num_blocks = (N + block_size - 1) / block_size;
|
|
|
|
complex_conj_mul_div_kernel<<<num_blocks, block_size>>>(
|
|
a.data_ptr<float>(),
|
|
b.data_ptr<float>(),
|
|
c.data_ptr<float>(),
|
|
output.data_ptr<float>(),
|
|
N, 1e-12f
|
|
);
|
|
|
|
return output;
|
|
}
|
|
"""
|
|
|
|
cpp_source = """
|
|
torch::Tensor complex_conj_mul_div_cuda(torch::Tensor a, torch::Tensor b, torch::Tensor c);
|
|
"""
|
|
|
|
module = load_inline(
|
|
name="complex_conj_mul_div",
|
|
cpp_sources=cpp_source,
|
|
cuda_sources=cuda_source,
|
|
functions=["complex_conj_mul_div_cuda"],
|
|
verbose=True
|
|
)
|
|
|
|
|
|
class ModelNew(nn.Module):
|
|
def __init__(self):
|
|
super(ModelNew, self).__init__()
|
|
self.module = module
|
|
|
|
def forward(self, a, b, c):
|
|
return self.module.complex_conj_mul_div_cuda(a, b, c) |