forked from ccf-ai-infra/GPUCodeForces
115 lines
3.5 KiB
Python
115 lines
3.5 KiB
Python
# bcewithlogitsloss_cuda.py
|
|
import torch
|
|
import torch.nn as nn
|
|
from torch.utils.cpp_extension import load_inline
|
|
|
|
from bcewithlogitsloss_torch import BATCH_SIZE, FEATURE_DIM # 导入维度常量
|
|
|
|
N_ELEMENTS = BATCH_SIZE * FEATURE_DIM
|
|
|
|
class ModelNew(nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
self._compile_cuda_kernel()
|
|
|
|
def _compile_cuda_kernel(self):
|
|
cpp_source = """
|
|
#include <torch/extension.h>
|
|
torch::Tensor bce_forward_cuda(torch::Tensor logits, torch::Tensor target);
|
|
"""
|
|
|
|
cuda_source = """
|
|
#include <cuda_runtime.h>
|
|
#include <cmath>
|
|
#include <float.h> // For FLT_MAX
|
|
|
|
#define BLOCK_SIZE 256
|
|
#define WARP_SIZE 32
|
|
|
|
__device__ __forceinline__ 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;
|
|
}
|
|
|
|
__global__ void bce_fused_kernel(
|
|
const float* __restrict__ logits,
|
|
const float* __restrict__ target,
|
|
float* __restrict__ loss_sum_out,
|
|
int n_elements
|
|
) {
|
|
__shared__ float s_data[BLOCK_SIZE];
|
|
|
|
|
|
float thread_loss = 0.0f;
|
|
int grid_stride = gridDim.x * blockDim.x;
|
|
|
|
for (int idx = blockIdx.x * blockDim.x + threadIdx.x;
|
|
idx < n_elements;
|
|
idx += grid_stride)
|
|
{
|
|
float x = logits[idx];
|
|
float y = target[idx];
|
|
|
|
float max_val = fmaxf(0.0f, x);
|
|
|
|
|
|
if (x >= 0) {
|
|
thread_loss += (x - x * y) + logf(1.0f + expf(-x));
|
|
} else {
|
|
thread_loss += (-x * y) + logf(1.0f + expf(x));
|
|
}
|
|
}
|
|
|
|
|
|
s_data[threadIdx.x] = thread_loss;
|
|
__syncthreads();
|
|
|
|
|
|
for (int offset = BLOCK_SIZE / 2; offset > 0; offset >>= 1) {
|
|
if (threadIdx.x < offset) {
|
|
s_data[threadIdx.x] += s_data[threadIdx.x + offset];
|
|
}
|
|
__syncthreads();
|
|
}
|
|
|
|
|
|
if (threadIdx.x == 0) {
|
|
loss_sum_out[blockIdx.x] = s_data[0];
|
|
}
|
|
}
|
|
|
|
torch::Tensor bce_forward_cuda(torch::Tensor logits, torch::Tensor target) {
|
|
TORCH_CHECK(logits.is_cuda(), "Input must be a CUDA tensor");
|
|
logits = logits.contiguous();
|
|
target = target.contiguous();
|
|
|
|
int n_elements = logits.numel();
|
|
|
|
const int block_size = BLOCK_SIZE;
|
|
const int grid_size = (n_elements + block_size - 1) / block_size;
|
|
|
|
auto block_loss_sums = torch::empty({grid_size}, logits.options());
|
|
|
|
bce_fused_kernel<<<grid_size, block_size>>>(
|
|
logits.data_ptr<float>(),
|
|
target.data_ptr<float>(),
|
|
block_loss_sums.data_ptr<float>(),
|
|
n_elements
|
|
);
|
|
|
|
return block_loss_sums.sum() / n_elements;
|
|
}
|
|
"""
|
|
|
|
self.bce_op = load_inline(
|
|
name="bce_fused_op",
|
|
cpp_sources=cpp_source,
|
|
cuda_sources=cuda_source,
|
|
functions=["bce_forward_cuda"],
|
|
extra_cuda_cflags=["-O3", "--use_fast_math"],
|
|
verbose=True
|
|
)
|
|
|
|
def forward(self, logits: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
|
|
return self.bce_op.bce_forward_cuda(logits, target) |