import torch
import torch.nn as nn
import torch.nn.functional as F
def module_fn(x: torch.Tensor) -> torch.Tensor:
"""
Applies HardSigmoid activation to the input tensor.
Args:
x (torch.Tensor): Input tensor of any shape.
Returns:
torch.Tensor: Output tensor with HardSigmoid applied, same shape as input.
"""
return F.hardsigmoid(x)
class Model(nn.Module):
"""
Simple model that performs a HardSigmoid activation.
"""
def __init__(self):
super(Model, self).__init__()
def forward(self, x: torch.Tensor, fn=module_fn) -> torch.Tensor:
"""
Applies HardSigmoid activation to the input tensor.
Args:
x (torch.Tensor): Input tensor of any shape.
Returns:
torch.Tensor: Output tensor with HardSigmoid applied, same shape as input.
"""
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 HardSigmoid activation.
"""
def __init__(self):
super(Model, self).__init__()
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""
Applies HardSigmoid activation to the input tensor.
Args:
x (torch.Tensor): Input tensor of any shape.
Returns:
torch.Tensor: Output tensor with HardSigmoid applied, same shape as input.
"""
return torch.nn.functional.hardsigmoid(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 <type_traits>
// Branchless clamp function using CUDA intrinsics to minimize warp divergence
template <typename scalar_t>
__device__ inline scalar_t clamp_val(scalar_t x) {
if constexpr (std::is_same<scalar_t, float>::value) {
return fminf(fmaxf(x, 0.f), 1.f);
} else {
return fmin(fmax(x, static_cast<scalar_t>(0)), static_cast<scalar_t>(1));
}
}
// CUDA kernel: computes HardSigmoid activation: y = clamp((x + 3) / 6, 0, 1)
// using branchless intrinsics to reduce warp divergence
template <typename scalar_t>
__global__ void warp_optimized_hardsigmoid_kernel(const scalar_t* __restrict__ input,
scalar_t* __restrict__ output,
size_t numel) {
const size_t idx = blockIdx.x * blockDim.x + threadIdx.x;
const size_t stride = blockDim.x * gridDim.x;
for (size_t i = idx; i < numel; i += stride) {
const scalar_t x = input[i];
scalar_t y = (x + static_cast<scalar_t>(3)) / static_cast<scalar_t>(6);
// Apply branchless clamp to maintain uniform control flow
y = clamp_val(y);
output[i] = y;
}
}
// Host function
torch::Tensor forward(torch::Tensor input) {
TORCH_CHECK(input.is_cuda(), "Input tensor must be on CUDA");
auto output = torch::empty_like(input);
const size_t numel = input.numel();
const int threads = 1024;
const int blocks = (numel + threads - 1) / threads;
AT_DISPATCH_FLOATING_TYPES(input.scalar_type(), "warp_optimized_hardsigmoid_cuda", ([&] {
warp_optimized_hardsigmoid_kernel<scalar_t><<<blocks, threads>>>(
input.data_ptr<scalar_t>(),
output.data_ptr<scalar_t>(),
numel);
}));
cudaError_t err = cudaGetLastError();
TORCH_CHECK(err == cudaSuccess, "CUDA kernel failed: ", cudaGetErrorString(err));
return output;
}
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
m.def("forward", &forward, "HardSigmoid activation forward (CUDA) with warp optimization");
}
| Metric | Value | Unit | Variance | Samples |
|---|---|---|---|---|
| Executed Ipc Active | 0.900 | inst/cycle | 0.000 | 5 |
| Executed Ipc Elapsed | 0.336 | inst/cycle | 0.000 | 5 |
| Issue Slots Busy | 25.082 | % | 0.046 | 5 |
| Issued Ipc Active | 1.004 | inst/cycle | 0.000 | 5 |
| SM Busy | 25.082 | % | 0.046 | 5 |
| Memory Throughput | 279517471770.026 | byte/second | 31936616788688674816.000 | 5 |
| Mem Busy | 13.270 | % | 0.087 | 5 |
| Max Bandwidth | 12.342 | % | 0.071 | 5 |
| L1/TEX Hit Rate | 0.000 | % | 0.000 | 5 |
| L2 Hit Rate | 67.074 | % | 0.080 | 5 |
| Mem Pipes Busy | 5.288 | % | 0.014 | 5 |
| Warp Cycles Per Issued Instruction | 51.238 | cycle | 0.106 | 5 |
| Warp Cycles Per Executed Instruction | 57.112 | cycle | 0.132 | 5 |
| Avg. Active Threads Per Warp | 32.000 | 0.000 | 5 | |
| Avg. Not Predicated Off Threads Per Warp | 30.000 | 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 | 4.000 | 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 | 81.346 | % | 0.043 | 5 |
| Achieved Active Warps Per SM | 52.062 | warp | 0.018 | 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 (81.3%) 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 | 635542.45 | μs |
| Device Time | 40.29 | μs |
| Self CPU Time | 43.24 | 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 | 635499.21 | μs |
| Device Time | 40.29 | μs |
| Self CPU Time | 86.23 | 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 | 655835.78 | μs |
| Device Time | 0.00 | μs |
| Self CPU Time | 20764.58 | 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 | 634358.66 | μs |
| Device Time | 0.00 | μs |
| Self CPU Time | 634358.66 | 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 | 483834.10 | μs |
| Device Time | 626.94 | μs |
| Self CPU Time | 483834.10 | μs |
| Self Device Time | 626.94 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
| void warp_optimized_hardsigmoid_kernel<float>(float const*, float*, unsigned long) | ||
| CPU Time | 0.00 | μs |
| Device Time | 26992.62 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 26992.62 | 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 | 22649.58 | μs |
| Device Time | 41689.60 | μs |
| Self CPU Time | 22649.58 | μs |
| Self Device Time | 41689.60 | 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 | 68950.46 | μs |
| Device Time | 638280.81 | μs |
| Self CPU Time | 12944.88 | 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 | 56006.58 | μs |
| Device Time | 638280.81 | μs |
| Self CPU Time | 16532.61 | μs |
| Self Device Time | 638280.81 | 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 | 638359.40 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 638359.40 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
45279 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.