import torch
import torch.nn as nn
import torch.nn.functional as F
def module_fn(predictions: torch.Tensor, targets: torch.Tensor) -> torch.Tensor:
"""
Computes the Cross Entropy Loss for multi-class classification tasks.
Args:
predictions (torch.Tensor): Predicted values.
targets (torch.Tensor): Target values.
Returns:
torch.Tensor: Cross Entropy Loss.
"""
return F.cross_entropy(predictions, targets)
class Model(nn.Module):
"""
A model that computes Cross Entropy Loss for multi-class classification tasks.
Parameters:
None
"""
def __init__(self):
super(Model, self).__init__()
def forward(self, predictions, targets, fn=module_fn):
return fn(predictions, targets)
batch_size = 4096
num_classes = 10
input_shape = (num_classes,) # Output for each class
dim = 1
def get_inputs():
return [
torch.randn(batch_size, *input_shape),
torch.randint(0, num_classes, (batch_size,)),
]
def get_init_inputs():
return []
import torch
import torch.nn as nn
class Model(nn.Module):
"""
A model that computes Cross Entropy Loss for multi-class classification tasks.
Parameters:
None
"""
def __init__(self):
super(Model, self).__init__()
def forward(self, predictions, targets):
return torch.nn.functional.cross_entropy(predictions, targets)
batch_size = 4096
num_classes = 10
input_shape = (num_classes, ) # Output for each class
dim = 1
def get_inputs():
return [torch.randn(batch_size, *input_shape), torch.randint(0, num_classes, (batch_size,))]
def get_init_inputs():
return []
#include <torch/extension.h>
__global__ void cross_entropy_loss_kernel_stride(
const float* __restrict__ logits,
const int64_t* __restrict__ targets,
float* __restrict__ losses,
int batch_size,
int num_classes
)
{
const int total_threads = blockDim.x * gridDim.x;
int tid = blockIdx.x * blockDim.x + threadIdx.x;
// Stride loop for workload distribution
for (int i = tid; i < batch_size; i += total_threads) {
const float* logits_i = logits + i * num_classes;
int64_t target = targets[i];
// Branchless max reduction
float max_logit = logits_i[0];
for (int j = 1; j < num_classes; ++j) {
max_logit = fmaxf(max_logit, logits_i[j]);
}
// Vectorized exp sum
float sum_exp = 0.0f;
for (int j = 0; j < num_classes; ++j) {
sum_exp += expf(logits_i[j] - max_logit);
}
losses[i] = -(logits_i[target] - max_logit - logf(sum_exp));
}
}
torch::Tensor forward(torch::Tensor predictions, torch::Tensor targets) {
TORCH_CHECK(predictions.is_cuda(), "predictions must be a CUDA tensor");
TORCH_CHECK(targets.is_cuda(), "targets must be a CUDA tensor");
TORCH_CHECK(predictions.dim() == 2, "predictions must be a 2D tensor");
TORCH_CHECK(targets.dim() == 1, "targets must be a 1D tensor");
const int batch_size = predictions.size(0);
const int num_classes = predictions.size(1);
auto losses = torch::empty({batch_size}, predictions.options());
// Optimized launch configuration
const int threads = 256;
const int blocks = (batch_size + threads - 1) / threads;
cross_entropy_loss_kernel_stride<<<blocks, threads>>>(
predictions.data_ptr<float>(),
targets.data_ptr<int64_t>(),
losses.data_ptr<float>(),
batch_size,
num_classes
);
cudaError_t err = cudaGetLastError();
TORCH_CHECK(err == cudaSuccess, "CUDA error: ", cudaGetErrorString(err));
return losses.mean();
}
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
m.def("forward", &forward, "Optimized CrossEntropyLoss forward (CUDA)");
}| Metric | Value | Unit | Variance | Samples |
|---|---|---|---|---|
| Executed Ipc Active | 0.514 | inst/cycle | 0.000 | 5 |
| Executed Ipc Elapsed | 0.030 | inst/cycle | 0.000 | 5 |
| Issue Slots Busy | 13.278 | % | 0.105 | 5 |
| Issued Ipc Active | 0.530 | inst/cycle | 0.000 | 5 |
| SM Busy | 13.278 | % | 0.105 | 5 |
| Memory Throughput | 52731781883.362 | byte/second | 331259100094932352.000 | 5 |
| Mem Busy | 8.658 | % | 0.019 | 5 |
| Max Bandwidth | 4.742 | % | 0.001 | 5 |
| L1/TEX Hit Rate | 92.360 | % | 0.000 | 5 |
| L2 Hit Rate | 90.042 | % | 0.027 | 5 |
| Mem Pipes Busy | 0.386 | % | 0.000 | 5 |
| Warp Cycles Per Issued Instruction | 14.540 | cycle | 0.416 | 5 |
| Warp Cycles Per Executed Instruction | 15.026 | cycle | 0.442 | 5 |
| Avg. Active Threads Per Warp | 32.000 | 0.000 | 5 | |
| Avg. Not Predicated Off Threads Per Warp | 28.980 | 0.000 | 5 | |
| Max Active Clusters | 0.000 | cluster | 0.000 | 5 |
| Max Cluster Size | 8.000 | block | 0.000 | 5 |
| Overall GPU Occupancy | 0.000 | % | 0.000 | 5 |
| Cluster Occupancy | 0.000 | % | 0.000 | 5 |
| Block Limit SM | 32.000 | block | 0.000 | 5 |
| Block Limit Registers | 8.000 | block | 0.000 | 5 |
| Block Limit Shared Mem | 32.000 | block | 0.000 | 5 |
| Block Limit Warps | 8.000 | block | 0.000 | 5 |
| Theoretical Active Warps per SM | 64.000 | warp | 0.000 | 5 |
| Theoretical Occupancy | 100.000 | % | 0.000 | 5 |
| Achieved Occupancy | 11.900 | % | 0.013 | 5 |
| Achieved Active Warps Per SM | 7.614 | warp | 0.005 | 5 |
| Rule | Description |
|---|---|
| WRN HighPipeUtilization | All compute pipelines are under-utilized. Either this kernel is very small or it doesn't issue enough warps per scheduler. Check the Launch Statistics and Scheduler Statistics sections for further details. |
| INF CPIStall | Check the Warp Stall Sampling (All Cycles) table for the top stall locations in your source based on sampling data. The Kernel Profiling Guide (https://docs.nvidia.com/nsight-compute/ProfilingGuide/index.html#metrics-reference) provides more details on each stall reason. |
| WRN Occupancy | This kernel's theoretical occupancy is not impacted by any block limit. The difference between calculated theoretical (100.0%) and measured achieved occupancy (11.8%) can be the result of warp scheduling overheads or workload imbalances during the kernel execution. Load imbalances can occur between warps within a block as well as across blocks of the same kernel. See the CUDA Best Practices Guide (https://docs.nvidia.com/cuda/cuda-c-best-practices-guide/index.html#occupancy) for more details on optimizing occupancy. |
| Operation / Metric | Value | Unit |
|---|---|---|
| aten::to | ||
| CPU Time | 743948.57 | μs |
| Device Time | 11.33 | μs |
| Self CPU Time | 55.31 | μs |
| Self Device Time | 0.00 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
| aten::_to_copy | ||
| CPU Time | 743893.26 | μs |
| Device Time | 11.33 | μs |
| Self CPU Time | 127.19 | μs |
| Self Device Time | 0.00 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
| aten::empty_strided | ||
| CPU Time | 743612.59 | μs |
| Device Time | 0.00 | μs |
| Self CPU Time | 118.60 | μs |
| Self Device Time | 0.00 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
| cudaDeviceGetStreamPriorityRange | ||
| CPU Time | 740445.71 | μs |
| Device Time | 0.00 | μs |
| Self CPU Time | 740445.71 | μs |
| Self Device Time | 0.00 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
| cudaLaunchKernel | ||
| CPU Time | 778315.46 | μs |
| Device Time | 8437.38 | μs |
| Self CPU Time | 778315.46 | μs |
| Self Device Time | 8437.38 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
| aten::mean | ||
| CPU Time | 222407.66 | μs |
| Device Time | 56045.11 | μs |
| Self CPU Time | 101369.75 | μs |
| Self Device Time | 56045.11 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
| void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4>) | ||
| CPU Time | 0.00 | μs |
| Device Time | 56045.11 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 56045.11 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
| aten::zero_ | ||
| CPU Time | 102200.49 | μs |
| Device Time | 1018201.95 | μs |
| Self CPU Time | 23969.52 | μs |
| Self Device Time | 0.00 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
| aten::fill_ | ||
| CPU Time | 78232.71 | μs |
| Device Time | 1018201.95 | μs |
| Self CPU Time | 25627.70 | μs |
| Self Device Time | 1018201.95 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
| void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, at::detail::Array<char*, 1> >(int, at::native::FillFunctor<int>, at::detail::Array<char*, 1>) | ||
| CPU Time | 0.00 | μs |
| Device Time | 1018280.57 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 1018280.57 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
45283 warnings generated when compiling for host. Suppressed 45322 warnings (45275 in non-user code, 47 NOLINT). Use -header-filter=.* to display errors from all non-system headers. Use -system-headers to display errors from system headers as well.