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>
// Device function to compute HardSigmoid: y = clamp((x + 3) / 6, 0, 1)
template <typename scalar_t>
__device__ inline scalar_t hardsigmoid_func(scalar_t x) {
scalar_t y = (x + static_cast<scalar_t>(3)) / static_cast<scalar_t>(6);
return (y < static_cast<scalar_t>(0)) ? static_cast<scalar_t>(0) :
(y > static_cast<scalar_t>(1)) ? static_cast<scalar_t>(1) : y;
}
// Vectorized kernel using pack types to ensure memory coalescing
// For float, we use float4 (pack_size = 4); for double, we use double2 (pack_size = 2).
template <typename scalar_t, typename pack_t, int pack_size>
__global__ void vectorized_hardsigmoid_kernel(const scalar_t* __restrict__ input,
scalar_t* __restrict__ output,
size_t numel) {
// Number of complete packs
size_t num_pack = numel / pack_size;
size_t idx = blockIdx.x * blockDim.x + threadIdx.x;
size_t stride = blockDim.x * gridDim.x;
// Reinterpret memory as vectorized packs
const pack_t* input_pack = reinterpret_cast<const pack_t*>(input);
pack_t* output_pack = reinterpret_cast<pack_t*>(output);
// Process vectorized elements
for (size_t i = idx; i < num_pack; i += stride) {
pack_t in_pack = input_pack[i];
pack_t out_pack;
// Process each element in the pack
scalar_t* in_vals = reinterpret_cast<scalar_t*>(&in_pack);
scalar_t* out_vals = reinterpret_cast<scalar_t*>(&out_pack);
#pragma unroll
for (int j = 0; j < pack_size; j++) {
out_vals[j] = hardsigmoid_func<scalar_t>(in_vals[j]);
}
output_pack[i] = out_pack;
}
// Handle leftover elements that don't fit in a complete pack
size_t remainder_start = num_pack * pack_size;
for (size_t i = remainder_start + idx; i < numel; i += stride) {
scalar_t x = input[i];
output[i] = hardsigmoid_func<scalar_t>(x);
}
}
// Forward function dispatching the appropriate kernel
torch::Tensor forward(torch::Tensor input) {
TORCH_CHECK(input.is_cuda(), "Input tensor must be on CUDA");
auto output = torch::empty_like(input);
size_t numel = input.numel();
int threads = 1024;
int blocks = 0;
AT_DISPATCH_FLOATING_TYPES(input.scalar_type(), "vectorized_coalesced_hardsigmoid_cuda", ([&] {
if (std::is_same<scalar_t, float>::value) {
constexpr int pack_size = 4;
using pack_t = float4;
// Adjust blocks for the vectorized loop
blocks = ((numel / pack_size) + threads - 1) / threads;
vectorized_hardsigmoid_kernel<scalar_t, pack_t, pack_size><<<blocks, threads>>>(
input.data_ptr<scalar_t>(),
output.data_ptr<scalar_t>(),
numel);
} else if (std::is_same<scalar_t, double>::value) {
constexpr int pack_size = 2;
using pack_t = double2;
blocks = ((numel / pack_size) + threads - 1) / threads;
vectorized_hardsigmoid_kernel<scalar_t, pack_t, pack_size><<<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, "Vectorized, coalesced HardSigmoid activation forward (CUDA)");
}
| Metric | Value | Unit | Variance | Samples |
|---|---|---|---|---|
| Executed Ipc Active | 0.872 | inst/cycle | 0.002 | 5 |
| Executed Ipc Elapsed | 0.174 | inst/cycle | 0.000 | 5 |
| Issue Slots Busy | 24.384 | % | 1.337 | 5 |
| Issued Ipc Active | 0.976 | inst/cycle | 0.002 | 5 |
| SM Busy | 24.384 | % | 1.337 | 5 |
| Memory Throughput | 281358297070.950 | byte/second | 87093134701075513344.000 | 5 |
| Mem Busy | 13.396 | % | 0.214 | 5 |
| Max Bandwidth | 12.314 | % | 0.173 | 5 |
| L1/TEX Hit Rate | 0.000 | % | 0.000 | 5 |
| L2 Hit Rate | 67.046 | % | 0.023 | 5 |
| Mem Pipes Busy | 1.586 | % | 0.003 | 5 |
| Warp Cycles Per Issued Instruction | 27.836 | cycle | 1.446 | 5 |
| Warp Cycles Per Executed Instruction | 31.104 | cycle | 1.807 | 5 |
| Avg. Active Threads Per Warp | 32.000 | 0.000 | 5 | |
| Avg. Not Predicated Off Threads Per Warp | 31.030 | 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 | 43.114 | % | 0.173 | 5 |
| Achieved Active Warps Per SM | 27.592 | warp | 0.071 | 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 (42.8%) 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 | 377524.25 | μs |
| Device Time | 40.13 | μs |
| Self CPU Time | 36.07 | 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 | 377488.17 | μs |
| Device Time | 40.13 | μs |
| Self CPU Time | 87.91 | 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 | 391700.31 | μs |
| Device Time | 0.00 | μs |
| Self CPU Time | 14665.57 | 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 |
| cudaDeviceGetStreamPriorityRange | ||
| CPU Time | 376850.62 | μs |
| Device Time | 0.00 | μs |
| Self CPU Time | 376850.62 | 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 | 353781.99 | μs |
| Device Time | 549.88 | μs |
| Self CPU Time | 353781.99 | μs |
| Self Device Time | 549.88 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
| void vectorized_hardsigmoid_kernel<float, float4, 4>(float const*, float*, unsigned long) | ||
| CPU Time | 0.00 | μs |
| Device Time | 17888.21 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 17888.21 | 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 | 16163.45 | μs |
| Device Time | 30204.45 | μs |
| Self CPU Time | 16163.45 | μs |
| Self Device Time | 30204.45 | 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 | 53383.74 | μs |
| Device Time | 466214.72 | μs |
| Self CPU Time | 9437.94 | 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 |
| aten::fill_ | ||
| CPU Time | 43947.59 | μs |
| Device Time | 466214.72 | μs |
| Self CPU Time | 13382.71 | μs |
| Self Device Time | 466214.72 | 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 | 466214.72 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 466214.72 | 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.