import torch
import torch.nn as nn
import torch.nn.functional as F
def module_fn(x: torch.Tensor) -> torch.Tensor:
"""
Applies Softplus activation to the input tensor.
Args:
x (torch.Tensor): Input tensor of any shape.
Returns:
torch.Tensor: Output tensor with Softplus applied, same shape as input.
"""
return F.softplus(x)
class Model(nn.Module):
"""
Simple model that performs a Softplus 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 Softplus activation.
"""
def __init__(self):
super(Model, self).__init__()
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""
Applies Softplus activation to the input tensor.
Args:
x (torch.Tensor): Input tensor of any shape.
Returns:
torch.Tensor: Output tensor with Softplus applied, same shape as input.
"""
return torch.nn.functional.softplus(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 <cmath>
// Branchless formulation of softplus: f(x) = max(x, 0) + log1p(exp(-|x|))
// This formulation avoids conditional branches and hence reduces warp divergence.
__device__ __forceinline__ float branchless_softplus(float x) {
return fmaxf(x, 0.0f) + log1pf(expf(-fabsf(x)));
}
__device__ __forceinline__ double branchless_softplus(double x) {
return fmax(x, 0.0) + log1p(exp(-fabs(x)));
}
// CUDA kernel using block-stride loop with branchless softplus
template <typename scalar_t>
__global__ void softplus_kernel_branchless(
const scalar_t* __restrict__ input,
scalar_t* __restrict__ output,
const int size) {
int idx = blockIdx.x * blockDim.x + threadIdx.x;
int stride = blockDim.x * gridDim.x;
for (; idx < size; idx += stride) {
scalar_t x = input[idx];
// Use the branchless softplus formulation for uniform control flow
output[idx] = branchless_softplus(x);
}
}
// CUDA forward function
torch::Tensor softplus_cuda_forward(torch::Tensor input) {
auto output = torch::empty_like(input);
const int size = input.numel();
const int threads = 256;
const int blocks = (size + threads - 1) / threads;
AT_DISPATCH_FLOATING_TYPES(input.scalar_type(), "softplus_forward_cuda", ([&] {
softplus_kernel_branchless<scalar_t><<<blocks, threads>>>(
input.data_ptr<scalar_t>(),
output.data_ptr<scalar_t>(),
size);
}));
return output;
}
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
m.def("forward", &softplus_cuda_forward, "Softplus forward (CUDA)");
}
| Metric | Value | Unit | Variance | Samples |
|---|---|---|---|---|
| Executed Ipc Active | 1.800 | inst/cycle | 0.003 | 5 |
| Executed Ipc Elapsed | 0.818 | inst/cycle | 0.000 | 5 |
| Issue Slots Busy | 47.314 | % | 1.577 | 5 |
| Issued Ipc Active | 1.892 | inst/cycle | 0.003 | 5 |
| SM Busy | 47.314 | % | 1.577 | 5 |
| Memory Throughput | 259609843132.966 | byte/second | 17805477762727696384.000 | 5 |
| Mem Busy | 12.328 | % | 0.063 | 5 |
| Max Bandwidth | 11.456 | % | 0.044 | 5 |
| L1/TEX Hit Rate | 0.000 | % | 0.000 | 5 |
| L2 Hit Rate | 66.480 | % | 0.070 | 5 |
| Mem Pipes Busy | 12.972 | % | 0.049 | 5 |
| Warp Cycles Per Issued Instruction | 26.780 | cycle | 0.015 | 5 |
| Warp Cycles Per Executed Instruction | 28.174 | cycle | 0.016 | 5 |
| Avg. Active Threads Per Warp | 32.000 | 0.000 | 5 | |
| Avg. Not Predicated Off Threads Per Warp | 29.330 | 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 | 80.404 | % | 0.116 | 5 |
| Achieved Active Warps Per SM | 51.462 | warp | 0.047 | 5 |
| Rule | Description |
|---|---|
| INF HighPipeUtilization | ALU is the highest-utilized pipeline (30.4%) 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 (79.8%) 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. |
| 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. |
| Operation / Metric | Value | Unit |
|---|---|---|
| aten::to | ||
| CPU Time | 639823.87 | μs |
| Device Time | 40.19 | μs |
| Self CPU Time | 47.57 | 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 | 639776.31 | μs |
| Device Time | 40.19 | μs |
| Self CPU Time | 93.70 | 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 | 655389.49 | μs |
| Device Time | 0.00 | μs |
| Self CPU Time | 16067.28 | 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 | 639092.94 | μs |
| Device Time | 0.00 | μs |
| Self CPU Time | 639092.94 | 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 | 396161.30 | μs |
| Device Time | 17771.32 | μs |
| Self CPU Time | 396161.30 | μs |
| Self Device Time | 17771.32 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
| void softplus_kernel_branchless<float>(float const*, float*, int) | ||
| CPU Time | 0.00 | μs |
| Device Time | 27559.53 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 27559.53 | 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 | 17822.34 | μs |
| Device Time | 34230.12 | μs |
| Self CPU Time | 17822.34 | μs |
| Self Device Time | 34230.12 | 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 | 56682.85 | μs |
| Device Time | 509225.31 | μs |
| Self CPU Time | 10018.80 | 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 | 46665.69 | μs |
| Device Time | 509225.31 | μs |
| Self CPU Time | 12560.16 | μs |
| Self Device Time | 509225.31 | 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 | 509304.18 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 509304.18 | 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.