forked from ccf-ai-infra/GPUCodeForces
148 lines
4.2 KiB
Python
148 lines
4.2 KiB
Python
import torch
|
|
import torch.nn as nn
|
|
from torch.utils.cpp_extension import load_inline
|
|
|
|
N, C, H, W = 32, 64, 56, 56
|
|
EPS = 1e-6
|
|
|
|
assert (C * H * W) % 4 == 0, "Instance size (C*H*W) must be a multiple of 4"
|
|
|
|
|
|
class ModelNew(nn.Module):
|
|
def __init__(self, p=3.0):
|
|
super().__init__()
|
|
self.p = float(p)
|
|
self.eps = EPS
|
|
self.blocks_per_instance = 16
|
|
self._compile_cuda_kernel()
|
|
|
|
def _compile_cuda_kernel(self):
|
|
cpp_source = """
|
|
#include <torch/extension.h>
|
|
void minkowski_sum_cuda(
|
|
torch::Tensor x,
|
|
torch::Tensor y,
|
|
torch::Tensor out,
|
|
float p,
|
|
int N,
|
|
int D,
|
|
int blocks_per_instance);
|
|
"""
|
|
|
|
cuda_source = """
|
|
#include <cuda_runtime.h>
|
|
#include <cmath>
|
|
#define WARP_SIZE 32
|
|
|
|
__inline__ __device__ float warp_reduce_sum(float val) {
|
|
for (int offset = WARP_SIZE / 2; offset > 0; offset /= 2) {
|
|
val += __shfl_down_sync(0xffffffff, val, offset);
|
|
}
|
|
return val;
|
|
}
|
|
|
|
__inline__ __device__ float block_reduce_sum(float val) {
|
|
__shared__ float shared[32];
|
|
int lane = threadIdx.x % WARP_SIZE;
|
|
int wid = threadIdx.x / WARP_SIZE;
|
|
|
|
val = warp_reduce_sum(val);
|
|
if (lane == 0) shared[wid] = val;
|
|
__syncthreads();
|
|
|
|
val = (threadIdx.x < blockDim.x / WARP_SIZE) ? shared[lane] : 0.0f;
|
|
if (wid == 0) val = warp_reduce_sum(val);
|
|
return val;
|
|
}
|
|
|
|
__global__ void minkowski_sum_kernel(
|
|
const float* __restrict__ x,
|
|
const float* __restrict__ y,
|
|
float* __restrict__ out,
|
|
float p,
|
|
int N,
|
|
int D_vec
|
|
) {
|
|
int n_idx = blockIdx.x;
|
|
int stride = blockDim.x * gridDim.y;
|
|
int d_start = blockIdx.y * blockDim.x + threadIdx.x;
|
|
int offset_base = n_idx * D_vec * 4;
|
|
|
|
const float4* x4_ptr = reinterpret_cast<const float4*>(x + offset_base);
|
|
const float4* y4_ptr = reinterpret_cast<const float4*>(y + offset_base);
|
|
|
|
float local_sum = 0.0f;
|
|
|
|
for (int i = d_start; i < D_vec; i += stride) {
|
|
float4 vx = x4_ptr[i];
|
|
float4 vy = y4_ptr[i];
|
|
|
|
local_sum += powf(fabsf(vx.x - vy.x), p);
|
|
local_sum += powf(fabsf(vx.y - vy.y), p);
|
|
local_sum += powf(fabsf(vx.z - vy.z), p);
|
|
local_sum += powf(fabsf(vx.w - vy.w), p);
|
|
}
|
|
|
|
float block_sum = block_reduce_sum(local_sum);
|
|
|
|
if (threadIdx.x == 0) {
|
|
atomicAdd(&out[n_idx], block_sum);
|
|
}
|
|
}
|
|
|
|
void minkowski_sum_cuda(
|
|
torch::Tensor x,
|
|
torch::Tensor y,
|
|
torch::Tensor out,
|
|
float p,
|
|
int N,
|
|
int D,
|
|
int blocks_per_instance)
|
|
{
|
|
dim3 blocks(N, blocks_per_instance);
|
|
dim3 threads(256);
|
|
int D_vec = D / 4;
|
|
|
|
minkowski_sum_kernel<<<blocks, threads>>>(
|
|
x.data_ptr<float>(),
|
|
y.data_ptr<float>(),
|
|
out.data_ptr<float>(),
|
|
p,
|
|
N,
|
|
D_vec
|
|
);
|
|
}
|
|
"""
|
|
|
|
self.op = load_inline(
|
|
name='minkowski_cuda_opt_v2',
|
|
cpp_sources=cpp_source,
|
|
cuda_sources=cuda_source,
|
|
functions=['minkowski_sum_cuda'],
|
|
extra_cuda_cflags=['-O3', '--use_fast_math'],
|
|
verbose=False
|
|
)
|
|
|
|
def forward(self, x: torch.Tensor, y: torch.Tensor) -> torch.Tensor:
|
|
if not x.is_cuda: x = x.cuda()
|
|
if not y.is_cuda: y = y.cuda()
|
|
|
|
x = x.contiguous()
|
|
y = y.contiguous()
|
|
|
|
N, C, H, W = x.size()
|
|
D = C * H * W
|
|
|
|
out = torch.zeros(N, device=x.device, dtype=x.dtype)
|
|
|
|
self.op.minkowski_sum_cuda(
|
|
x,
|
|
y,
|
|
out,
|
|
self.p,
|
|
N,
|
|
D,
|
|
self.blocks_per_instance
|
|
)
|
|
|
|
return torch.pow(out + self.eps, 1.0 / self.p) |