forked from ccf-ai-infra/GPUCodeForces
133 lines
4.3 KiB
Python
133 lines
4.3 KiB
Python
# upsample_cuda.py
|
|
import torch
|
|
from torch.utils.cpp_extension import load_inline
|
|
|
|
from upsample_torch import BATCH_SIZE, CHANNELS, H_IN, W_IN, SCALE_FACTOR # 导入维度常量
|
|
|
|
|
|
W_OUT = W_IN * SCALE_FACTOR
|
|
assert W_OUT % 4 == 0, "Output width (W_in * scale) must be a multiple of 4 for float4 vectorization"
|
|
VEC_SIZE = 4
|
|
|
|
class ModelNew(torch.nn.Module):
|
|
|
|
def __init__(self):
|
|
super().__init__()
|
|
self._compile_cuda_kernel()
|
|
|
|
def _compile_cuda_kernel(self):
|
|
cpp_source = """
|
|
#include <torch/extension.h>
|
|
torch::Tensor upsample_forward_cuda(torch::Tensor input);
|
|
"""
|
|
|
|
cuda_source = """
|
|
#include <cuda_runtime.h>
|
|
#include <cmath>
|
|
|
|
#define BLOCK_SIZE 256
|
|
#define SCALE_VAL {scale_factor}
|
|
#define VEC_SIZE {vec_size}
|
|
|
|
__global__ void upsample_fused_vectorized_kernel(
|
|
const float* __restrict__ x,
|
|
float* __restrict__ y,
|
|
int N, int C, int H_in, int W_in
|
|
) {{
|
|
|
|
int H_out = H_in * SCALE_VAL;
|
|
int W_out = W_in * SCALE_VAL;
|
|
int W_out_vec = W_out / VEC_SIZE;
|
|
int CHW_out_vec = C * H_out * W_out_vec;
|
|
int C_HW_in = C * H_in * W_in;
|
|
|
|
int output_elements_vec = N * CHW_out_vec;
|
|
|
|
int grid_stride = gridDim.x * blockDim.x;
|
|
|
|
for (int idx_vec = blockIdx.x * blockDim.x + threadIdx.x;
|
|
idx_vec < output_elements_vec;
|
|
idx_vec += grid_stride)
|
|
{{
|
|
|
|
int n = idx_vec / CHW_out_vec;
|
|
int rem_n = idx_vec % CHW_out_vec;
|
|
|
|
int c = rem_n / (H_out * W_out_vec);
|
|
int rem_c = rem_n % (H_out * W_out_vec);
|
|
|
|
int h_out = rem_c / W_out_vec;
|
|
int w_out_vec = rem_c % W_out_vec;
|
|
|
|
|
|
int h_in = h_out / SCALE_VAL;
|
|
|
|
long base_input_idx = (long)n * C_HW_in +
|
|
(long)c * (H_in * W_in) +
|
|
(long)h_in * W_in;
|
|
|
|
|
|
float4* y4_ptr = reinterpret_cast<float4*>(y);
|
|
|
|
int w_out_start = w_out_vec * VEC_SIZE;
|
|
|
|
int w_in_start = w_out_start / SCALE_VAL;
|
|
|
|
|
|
float val_in_0 = x[base_input_idx + w_in_start];
|
|
float val_in_1 = x[base_input_idx + w_in_start + 1];
|
|
|
|
float4 output_vec;
|
|
|
|
output_vec.x = val_in_0;
|
|
output_vec.y = val_in_0;
|
|
output_vec.z = val_in_1;
|
|
output_vec.w = val_in_1;
|
|
|
|
y4_ptr[idx_vec] = output_vec;
|
|
}}
|
|
}}
|
|
|
|
torch::Tensor upsample_forward_cuda(torch::Tensor input) {{
|
|
TORCH_CHECK(input.is_cuda(), "Input must be a CUDA tensor");
|
|
input = input.contiguous();
|
|
|
|
int N = input.size(0);
|
|
int C = input.size(1);
|
|
int H_in = input.size(2);
|
|
int W_in = input.size(3);
|
|
int up_factor = SCALE_VAL;
|
|
|
|
int H_out = H_in * up_factor;
|
|
int W_out = W_in * up_factor;
|
|
|
|
TORCH_CHECK(W_out % VEC_SIZE == 0, "Output width must be divisible by 4 for float4 vectorization");
|
|
|
|
auto output = torch::empty({{N, C, H_out, W_out}}, input.options());
|
|
|
|
int n_elements_vec = output.numel() / VEC_SIZE;
|
|
|
|
const int block_size = 256;
|
|
const int grid_size = (n_elements_vec + block_size - 1) / block_size;
|
|
|
|
upsample_fused_vectorized_kernel<<<grid_size, block_size>>>(
|
|
input.data_ptr<float>(),
|
|
output.data_ptr<float>(),
|
|
N, C, H_in, W_in
|
|
);
|
|
|
|
return output;
|
|
}}
|
|
""".format(scale_factor=SCALE_FACTOR, vec_size=VEC_SIZE)
|
|
|
|
self.ps_op = load_inline(
|
|
name="upsample_fused_vectorized_op_final",
|
|
cpp_sources=cpp_source,
|
|
cuda_sources=cuda_source,
|
|
functions=["upsample_forward_cuda"],
|
|
extra_cuda_cflags=["-O3", "--use_fast_math"],
|
|
verbose=True
|
|
)
|
|
|
|
def forward(self, input: torch.Tensor) -> torch.Tensor:
|
|
return self.ps_op.upsample_forward_cuda(input) |