import torch
import torch.nn as nn
import torch.nn.functional as F
def module_fn(x: torch.Tensor) -> torch.Tensor:
"""
Applies SELU activation to the input tensor.
Args:
x (torch.Tensor): Input tensor of any shape.
Returns:
torch.Tensor: Output tensor with SELU applied, same shape as input.
"""
return F.selu(x)
class Model(nn.Module):
"""
Simple model that performs a SELU activation.
"""
def __init__(self):
super(Model, self).__init__()
def forward(self, x: torch.Tensor, fn=module_fn) -> torch.Tensor:
return fn(x)
batch_size = 16
dim = 16384
def get_inputs():
x = torch.randn(batch_size, dim)
return [x]
def get_init_inputs():
return [] # No special initialization inputs needed
import torch
import torch.nn as nn
class Model(nn.Module):
"""
Simple model that performs a SELU activation.
"""
def __init__(self):
super(Model, self).__init__()
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""
Applies SELU activation to the input tensor.
Args:
x (torch.Tensor): Input tensor of any shape.
Returns:
torch.Tensor: Output tensor with SELU applied, same shape as input.
"""
return torch.selu(x)
batch_size = 16
dim = 16384
def get_inputs():
x = torch.randn(batch_size, dim)
return [x]
def get_init_inputs():
return [] # No special initialization inputs needed#include <torch/extension.h>
#include <cuda.h>
#include <cuda_runtime.h>
#include <math.h>
// Device helper: define an inline exponential function for float and double.
template <typename scalar_t>
__device__ inline scalar_t my_exp(scalar_t x);
template <>
__device__ inline float my_exp<float>(float x) {
return expf(x);
}
template <>
__device__ inline double my_exp<double>(double x) {
return exp(x);
}
// CUDA kernel that evenly distributes the workload among threads.
// Each thread computes a contiguous segment of the data based on its global thread ID.
// This ensures balanced workload and minimizes underutilization or bottlenecks.
template <typename scalar_t>
__global__ void selu_kernel_even_load_balance(const scalar_t* __restrict__ input,
scalar_t* __restrict__ output,
size_t numel) {
// Compute a unique thread ID
const int tid = blockIdx.x * blockDim.x + threadIdx.x;
const int total_threads = gridDim.x * blockDim.x;
// Calculate the number of elements per thread and the residue
size_t base = numel / total_threads;
size_t residue = numel % total_threads;
// Each thread processes base elements, plus one extra if its ID is less than the residue
size_t start = tid * base + (tid < residue ? tid : residue);
size_t count = base + (tid < residue ? 1 : 0);
size_t end = start + count;
for (size_t i = start; i < end; i++) {
// Load input using __ldg for potential caching benefits
scalar_t x = __ldg(&input[i]);
scalar_t res = (x > static_cast<scalar_t>(0))
? x
: static_cast<scalar_t>(1.67326324235437728481) *
(my_exp(x) - static_cast<scalar_t>(1));
output[i] = static_cast<scalar_t>(1.05070098735548049342) * res;
}
}
// Host function launching the kernel
torch::Tensor selu_forward(torch::Tensor input) {
TORCH_CHECK(input.is_cuda(), "Input tensor must be a CUDA tensor");
auto output = torch::empty_like(input);
size_t numel = input.numel();
// Launch configuration: using 1024 threads per block.
const int threads = 1024;
int blocks = (numel + threads - 1) / threads; // Ensures enough threads are launched
AT_DISPATCH_FLOATING_TYPES(input.scalar_type(), "selu_forward_even_load_balance_cuda", ([&] {
const scalar_t* input_ptr = input.data_ptr<scalar_t>();
scalar_t* output_ptr = output.data_ptr<scalar_t>();
selu_kernel_even_load_balance<scalar_t><<<blocks, threads>>>(input_ptr, output_ptr, numel);
}));
return output;
}
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
m.def("forward", &selu_forward, "SELU Activation Forward with Even Load Balancing (CUDA)");
}
| Metric | Value | Unit | Variance | Samples |
|---|---|---|---|---|
| Executed Ipc Active | 1.660 | inst/cycle | 0.000 | 5 |
| Executed Ipc Elapsed | 0.778 | inst/cycle | 0.000 | 5 |
| Issue Slots Busy | 43.848 | % | 0.059 | 5 |
| Issued Ipc Active | 1.756 | inst/cycle | 0.000 | 5 |
| SM Busy | 43.848 | % | 0.059 | 5 |
| Memory Throughput | 236459105348.264 | byte/second | 17333745306692784128.000 | 5 |
| Mem Busy | 11.180 | % | 0.044 | 5 |
| Max Bandwidth | 10.430 | % | 0.037 | 5 |
| L1/TEX Hit Rate | 0.000 | % | 0.000 | 5 |
| L2 Hit Rate | 67.524 | % | 0.128 | 5 |
| Mem Pipes Busy | 6.628 | % | 0.017 | 5 |
| Warp Cycles Per Issued Instruction | 30.316 | cycle | 0.616 | 5 |
| Warp Cycles Per Executed Instruction | 31.986 | cycle | 0.682 | 5 |
| Avg. Active Threads Per Warp | 32.000 | 0.000 | 5 | |
| Avg. Not Predicated Off Threads Per Warp | 29.460 | 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 | 2.000 | block | 0.000 | 5 |
| Block Limit Shared Mem | 8.000 | block | 0.000 | 5 |
| Block Limit Warps | 2.000 | block | 0.000 | 5 |
| Theoretical Active Warps per SM | 64.000 | warp | 0.000 | 5 |
| Theoretical Occupancy | 100.000 | % | 0.000 | 5 |
| Achieved Occupancy | 82.626 | % | 0.135 | 5 |
| Achieved Active Warps Per SM | 52.878 | warp | 0.055 | 5 |
| Rule | Description |
|---|---|
| INF HighPipeUtilization | ALU is the highest-utilized pipeline (36.9%) based on active cycles, taking into account the rates of its different instructions. It executes integer and logic operations. It is well-utilized, but should not be a bottleneck. |
| 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 (82.4%) 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 | 342781.39 | μs |
| Device Time | 39.90 | μs |
| Self CPU Time | 30.74 | 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 | 342750.65 | μs |
| Device Time | 39.90 | μs |
| Self CPU Time | 79.71 | 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 |
| aten::empty_strided | ||
| CPU Time | 357416.67 | μs |
| Device Time | 0.00 | μs |
| Self CPU Time | 15053.69 | 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 | 342171.66 | μs |
| Device Time | 0.00 | μs |
| Self CPU Time | 342171.66 | 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 | 426440.94 | μs |
| Device Time | 18846.26 | μs |
| Self CPU Time | 426440.94 | μs |
| Self Device Time | 18846.26 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
| void selu_kernel_even_load_balance<float>(float const*, float*, unsigned long) | ||
| CPU Time | 0.00 | μs |
| Device Time | 24911.91 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 24911.91 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
| cudaEventRecord | ||
| CPU Time | 15041.88 | μs |
| Device Time | 36367.53 | μs |
| Self CPU Time | 15041.88 | μs |
| Self Device Time | 36367.53 | 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 | 55908.52 | μs |
| Device Time | 539443.27 | μs |
| Self CPU Time | 10225.34 | 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 | 45684.22 | μs |
| Device Time | 539443.27 | μs |
| Self CPU Time | 13412.20 | μs |
| Self Device Time | 539443.27 | 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 | 539443.27 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 539443.27 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
45280 warnings generated when compiling for host. Suppressed 45321 warnings (45274 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.