forked from ccf-ai-infra/GPUCodeForces
121 lines
3.8 KiB
Python
121 lines
3.8 KiB
Python
import torch
|
|
import torch.nn as nn
|
|
from torch.utils.cpp_extension import load_inline
|
|
|
|
|
|
class ModelNew(nn.Module):
|
|
def __init__(self, feature_dim):
|
|
super().__init__()
|
|
self.register_buffer("center", torch.zeros(feature_dim))
|
|
self._compile_cuda_kernel()
|
|
|
|
def _compile_cuda_kernel(self):
|
|
cpp_source = """
|
|
#include <torch/extension.h>
|
|
torch::Tensor chebyshev_cuda(torch::Tensor x, torch::Tensor center);
|
|
"""
|
|
|
|
cuda_source = """
|
|
#include <cuda_runtime.h>
|
|
#include <math.h>
|
|
|
|
__device__ __forceinline__ float warp_reduce_max(float val) {
|
|
#pragma unroll
|
|
for (int offset = 16; offset > 0; offset /= 2) {
|
|
float other = __shfl_down_sync(0xffffffff, val, offset);
|
|
val = fmaxf(val, other);
|
|
}
|
|
return val;
|
|
}
|
|
|
|
__device__ __forceinline__ float block_reduce_max(float val) {
|
|
static __shared__ float shared[32];
|
|
int lane = threadIdx.x % 32;
|
|
int wid = threadIdx.x / 32;
|
|
|
|
val = warp_reduce_max(val);
|
|
|
|
if (lane == 0) shared[wid] = val;
|
|
__syncthreads();
|
|
|
|
val = (threadIdx.x < blockDim.x / 32) ? shared[lane] : 0.0f;
|
|
|
|
if (wid == 0) val = warp_reduce_max(val);
|
|
|
|
return val;
|
|
}
|
|
|
|
__global__ void chebyshev_kernel_vec4(
|
|
const float* __restrict__ x,
|
|
const float* __restrict__ center,
|
|
float* __restrict__ output,
|
|
int batch_size,
|
|
int feature_dim)
|
|
{
|
|
int bid = blockIdx.x;
|
|
if (bid >= batch_size) return;
|
|
|
|
const float* row_x = x + bid * feature_dim;
|
|
|
|
float max_val = 0.0f;
|
|
int vec_loops = feature_dim / 4;
|
|
int vec_remainder = feature_dim % 4;
|
|
|
|
for (int i = threadIdx.x; i < vec_loops; i += blockDim.x) {
|
|
float4 vx = reinterpret_cast<const float4*>(row_x)[i];
|
|
float4 vc = reinterpret_cast<const float4*>(center)[i];
|
|
|
|
max_val = fmaxf(max_val, fabsf(vx.x - vc.x));
|
|
max_val = fmaxf(max_val, fabsf(vx.y - vc.y));
|
|
max_val = fmaxf(max_val, fabsf(vx.z - vc.z));
|
|
max_val = fmaxf(max_val, fabsf(vx.w - vc.w));
|
|
}
|
|
|
|
int tail_start = vec_loops * 4;
|
|
if (threadIdx.x < vec_remainder) {
|
|
int idx = tail_start + threadIdx.x;
|
|
max_val = fmaxf(max_val, fabsf(row_x[idx] - center[idx]));
|
|
}
|
|
|
|
max_val = block_reduce_max(max_val);
|
|
|
|
if (threadIdx.x == 0) {
|
|
output[bid] = max_val;
|
|
}
|
|
}
|
|
|
|
torch::Tensor chebyshev_cuda(torch::Tensor x, torch::Tensor center) {
|
|
auto x_c = x.contiguous();
|
|
auto c_c = center.contiguous();
|
|
|
|
int batch_size = x_c.size(0);
|
|
int feature_dim = x_c.size(1);
|
|
|
|
auto output = torch::empty({batch_size}, x.options());
|
|
|
|
int threads = 256;
|
|
int blocks = batch_size;
|
|
|
|
chebyshev_kernel_vec4<<<blocks, threads>>>(
|
|
x_c.data_ptr<float>(),
|
|
c_c.data_ptr<float>(),
|
|
output.data_ptr<float>(),
|
|
batch_size,
|
|
feature_dim
|
|
);
|
|
|
|
return output;
|
|
}
|
|
"""
|
|
|
|
self.op = load_inline(
|
|
name="chebyshev_opt_v1",
|
|
cpp_sources=cpp_source,
|
|
cuda_sources=cuda_source,
|
|
functions=["chebyshev_cuda"],
|
|
extra_cuda_cflags=["-O3"],
|
|
verbose=False
|
|
)
|
|
|
|
def forward(self, x):
|
|
return self.op.chebyshev_cuda(x, self.center) |