forked from ccf-ai-infra/GPUCodeForces
121 lines
3.4 KiB
Python
121 lines
3.4 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
|
|
GROUPS = 4
|
|
|
|
assert (H * W) % 4 == 0, "Spatial size (H*W) must be a multiple of 4"
|
|
|
|
|
|
class ModelNew(nn.Module):
|
|
|
|
def __init__(self, groups=GROUPS):
|
|
super().__init__()
|
|
self.groups = groups
|
|
self.block_size = 256
|
|
self._compile_cuda_kernel()
|
|
|
|
def _compile_cuda_kernel(self):
|
|
cpp_source = """
|
|
#include <torch/extension.h>
|
|
|
|
torch::Tensor channel_shuffle_cuda(torch::Tensor input, int groups);
|
|
"""
|
|
|
|
cuda_source = f"""
|
|
#include <cuda_runtime.h>
|
|
|
|
#define BLOCK_SIZE {self.block_size}
|
|
|
|
|
|
__global__ void channel_shuffle_kernel(
|
|
const float* __restrict__ input,
|
|
float* __restrict__ output,
|
|
int N,
|
|
int C,
|
|
int S_vec, // H * W / 4
|
|
int groups,
|
|
int channels_per_group
|
|
) {{
|
|
|
|
int total_threads = N * C * S_vec;
|
|
int idx = blockIdx.x * blockDim.x + threadIdx.x;
|
|
|
|
for (; idx < total_threads; idx += gridDim.x * blockDim.x) {{
|
|
|
|
|
|
int s = idx % S_vec;
|
|
int tmp = idx / S_vec;
|
|
int c_in = tmp % C;
|
|
int n = tmp / C;
|
|
|
|
|
|
|
|
int g_idx = c_in / channels_per_group;
|
|
int c_idx = c_in % channels_per_group;
|
|
int c_out = c_idx * groups + g_idx;
|
|
|
|
|
|
int out_global_idx = (n * C + c_out) * S_vec + s;
|
|
|
|
|
|
const float4* in_ptr = reinterpret_cast<const float4*>(input);
|
|
float4* out_ptr = reinterpret_cast<float4*>(output);
|
|
|
|
out_ptr[out_global_idx] = in_ptr[idx];
|
|
}}
|
|
}}
|
|
|
|
|
|
torch::Tensor channel_shuffle_cuda(torch::Tensor input, int groups) {{
|
|
|
|
TORCH_CHECK(input.is_cuda(), "Input must be a CUDA tensor");
|
|
TORCH_CHECK(input.dim() == 4, "Input must be 4D (N, C, H, W)");
|
|
|
|
int N = input.size(0);
|
|
int C = input.size(1);
|
|
int H = input.size(2);
|
|
int W = input.size(3);
|
|
|
|
TORCH_CHECK(C % groups == 0, "Channels must be divisible by groups");
|
|
TORCH_CHECK((H * W) % 4 == 0, "Spatial size must be divisible by 4 for float4 optimization");
|
|
|
|
|
|
input = input.contiguous();
|
|
|
|
auto output = torch::empty_like(input);
|
|
|
|
int S = H * W;
|
|
int S_vec = S / 4;
|
|
int channels_per_group = C / groups;
|
|
|
|
int total_threads = N * C * S_vec;
|
|
int blocks = std::min((total_threads + BLOCK_SIZE - 1) / BLOCK_SIZE, 1024);
|
|
|
|
channel_shuffle_kernel<<<blocks, BLOCK_SIZE>>>(
|
|
input.data_ptr<float>(),
|
|
output.data_ptr<float>(),
|
|
N,
|
|
C,
|
|
S_vec,
|
|
groups,
|
|
channels_per_group
|
|
);
|
|
|
|
return output;
|
|
}}
|
|
"""
|
|
|
|
self.op = load_inline(
|
|
name='channel_shuffle_cuda_opt_v1',
|
|
cpp_sources=cpp_source,
|
|
cuda_sources=cuda_source,
|
|
functions=['channel_shuffle_cuda'],
|
|
extra_cuda_cflags=['-O3'],
|
|
verbose=False
|
|
)
|
|
|
|
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
|
if not x.is_cuda: x = x.cuda()
|
|
return self.op.channel_shuffle_cuda(x, self.groups) |