forked from ccf-ai-infra/GPUCodeForces
114 lines
3.5 KiB
Python
114 lines
3.5 KiB
Python
import torch
|
|
import torch.nn as nn
|
|
from torch.utils.cpp_extension import load_inline
|
|
|
|
|
|
class ModelNew(nn.Module):
|
|
def __init__(self, value_coef):
|
|
super().__init__()
|
|
self.value_coef = value_coef.item() if isinstance(value_coef, torch.Tensor) else value_coef
|
|
self._compile_cuda_kernel()
|
|
|
|
def _compile_cuda_kernel(self):
|
|
cpp_source = """
|
|
torch::Tensor ac_loss_cuda(
|
|
torch::Tensor log_probs,
|
|
torch::Tensor values,
|
|
torch::Tensor returns,
|
|
torch::Tensor advantages,
|
|
float value_coef);
|
|
"""
|
|
|
|
cuda_source = """
|
|
#include <torch/extension.h>
|
|
#include <cuda_runtime.h>
|
|
|
|
__global__ void ac_loss_kernel(
|
|
const float* __restrict__ log_probs,
|
|
const float* __restrict__ values,
|
|
const float* __restrict__ returns,
|
|
const float* __restrict__ advantages,
|
|
float* __restrict__ output,
|
|
float value_coef,
|
|
int n)
|
|
{
|
|
extern __shared__ float sdata[];
|
|
unsigned int tid = threadIdx.x;
|
|
unsigned int i = blockIdx.x * blockDim.x + threadIdx.x;
|
|
unsigned int gridSize = blockDim.x * gridDim.x;
|
|
|
|
float local_sum = 0.0f;
|
|
|
|
while (i < n) {
|
|
float lp = log_probs[i];
|
|
float adv = advantages[i];
|
|
float val = values[i];
|
|
float ret = returns[i];
|
|
|
|
float actor_term = -lp * adv;
|
|
float diff = val - ret;
|
|
float critic_term = diff * diff;
|
|
|
|
local_sum += actor_term + value_coef * critic_term;
|
|
i += gridSize;
|
|
}
|
|
|
|
sdata[tid] = local_sum;
|
|
__syncthreads();
|
|
|
|
for (unsigned int s = blockDim.x / 2; s > 0; s >>= 1) {
|
|
if (tid < s) {
|
|
sdata[tid] += sdata[tid + s];
|
|
}
|
|
__syncthreads();
|
|
}
|
|
|
|
if (tid == 0) {
|
|
atomicAdd(output, sdata[0] / n);
|
|
}
|
|
}
|
|
|
|
torch::Tensor ac_loss_cuda(
|
|
torch::Tensor log_probs,
|
|
torch::Tensor values,
|
|
torch::Tensor returns,
|
|
torch::Tensor advantages,
|
|
float value_coef)
|
|
{
|
|
auto log_probs_c = log_probs.contiguous();
|
|
auto values_c = values.contiguous();
|
|
auto returns_c = returns.contiguous();
|
|
auto advantages_c = advantages.contiguous();
|
|
|
|
int n = log_probs_c.numel();
|
|
auto output = torch::zeros({1}, log_probs.options());
|
|
|
|
const int threads = 256;
|
|
const int blocks = min((n + threads - 1) / threads, 1024);
|
|
const int shared_mem = threads * sizeof(float);
|
|
|
|
ac_loss_kernel<<<blocks, threads, shared_mem>>>(
|
|
log_probs_c.data_ptr<float>(),
|
|
values_c.data_ptr<float>(),
|
|
returns_c.data_ptr<float>(),
|
|
advantages_c.data_ptr<float>(),
|
|
output.data_ptr<float>(),
|
|
value_coef,
|
|
n
|
|
);
|
|
|
|
return output[0];
|
|
}
|
|
"""
|
|
|
|
self.op = load_inline(
|
|
name="ac_loss_op",
|
|
cpp_sources=cpp_source,
|
|
cuda_sources=cuda_source,
|
|
functions=["ac_loss_cuda"],
|
|
extra_cuda_cflags=["-O3"],
|
|
verbose=False
|
|
)
|
|
|
|
def forward(self, log_probs, values, returns, advantages):
|
|
return self.op.ac_loss_cuda(log_probs, values, returns, advantages, self.value_coef) |