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>
// Vectorized and coalesced HardSigmoid kernel
// Computes y = clamp((x + 3) / 6, 0, 1) using vectorized global memory accesses
template <typename scalar_t>
__global__ void vectorized_coalesced_hardsigmoid_kernel(const scalar_t* __restrict__ input,
scalar_t* __restrict__ output,
size_t numel) {
// Calculate global thread index and stride
size_t idx = blockIdx.x * blockDim.x + threadIdx.x;
size_t stride = blockDim.x * gridDim.x;
// Use vectorized loads/stores for memory coalescing based on precision
if constexpr (std::is_same<scalar_t, float>::value) {
// For float, use float4 (4 floats at a time)
constexpr int vecSize = 4;
using vec_t = float4;
size_t num_vec = numel / vecSize; // number of vectorized elements
// Process main vectorized portion
for (size_t i = idx; i < num_vec; i += stride) {
vec_t in_vec = reinterpret_cast<const vec_t*>(input)[i];
vec_t out_vec;
out_vec.x = fminf(fmaxf((in_vec.x + 3.0f) / 6.0f, 0.0f), 1.0f);
out_vec.y = fminf(fmaxf((in_vec.y + 3.0f) / 6.0f, 0.0f), 1.0f);
out_vec.z = fminf(fmaxf((in_vec.z + 3.0f) / 6.0f, 0.0f), 1.0f);
out_vec.w = fminf(fmaxf((in_vec.w + 3.0f) / 6.0f, 0.0f), 1.0f);
reinterpret_cast<vec_t*>(output)[i] = out_vec;
}
// Process any remaining elements
size_t tail_start = num_vec * vecSize;
for (size_t i = idx; i < (numel - tail_start); i += stride) {
size_t index = tail_start + i;
float x = input[index];
float y = (x + 3.0f) / 6.0f;
y = fminf(fmaxf(y, 0.0f), 1.0f);
output[index] = y;
}
} else {
// For double, use double2 (2 doubles at a time)
constexpr int vecSize = 2;
using vec_t = double2;
size_t num_vec = numel / vecSize;
for (size_t i = idx; i < num_vec; i += stride) {
vec_t in_vec = reinterpret_cast<const vec_t*>(input)[i];
vec_t out_vec;
out_vec.x = fmin(fmax((in_vec.x + 3.0) / 6.0, 0.0), 1.0);
out_vec.y = fmin(fmax((in_vec.y + 3.0) / 6.0, 0.0), 1.0);
reinterpret_cast<vec_t*>(output)[i] = out_vec;
}
// Handle tail elements for double
size_t tail_start = num_vec * vecSize;
for (size_t i = idx; i < (numel - tail_start); i += stride) {
size_t index = tail_start + i;
double x = input[index];
double y = (x + 3.0) / 6.0;
y = fmin(fmax(y, 0.0), 1.0);
output[index] = y;
}
}
}
// Host function to launch the vectorized 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();
const int threads = 1024;
const int blocks = (numel + threads - 1) / threads;
AT_DISPATCH_FLOATING_TYPES(input.scalar_type(), "vectorized_coalesced_hardsigmoid_cuda", ([&] {
vectorized_coalesced_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, "Vectorized and coalesced HardSigmoid activation (CUDA)");
}
| Metric | Value | Unit | Variance | Samples |
|---|---|---|---|---|
| Executed Ipc Active | 1.086 | inst/cycle | 0.000 | 5 |
| Executed Ipc Elapsed | 0.364 | inst/cycle | 0.000 | 5 |
| Issue Slots Busy | 32.194 | % | 0.278 | 5 |
| Issued Ipc Active | 1.286 | inst/cycle | 0.000 | 5 |
| SM Busy | 32.194 | % | 0.278 | 5 |
| Memory Throughput | 280053929605.488 | byte/second | 15184891815191379968.000 | 5 |
| Mem Busy | 13.298 | % | 0.056 | 5 |
| Max Bandwidth | 12.184 | % | 0.048 | 5 |
| L1/TEX Hit Rate | 0.000 | % | 0.000 | 5 |
| L2 Hit Rate | 67.164 | % | 0.003 | 5 |
| Mem Pipes Busy | 5.292 | % | 0.008 | 5 |
| Warp Cycles Per Issued Instruction | 37.946 | cycle | 8.649 | 5 |
| Warp Cycles Per Executed Instruction | 44.994 | cycle | 12.198 | 5 |
| Avg. Active Threads Per Warp | 32.000 | 0.000 | 5 | |
| Avg. Not Predicated Off Threads Per Warp | 31.540 | 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 | 74.236 | % | 0.749 | 5 |
| Achieved Active Warps Per SM | 47.512 | warp | 0.310 | 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 (73.5%) 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 | 433959.03 | μs |
| Device Time | 40.10 | μs |
| Self CPU Time | 31.35 | 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 | 433927.68 | μs |
| Device Time | 40.10 | μs |
| Self CPU Time | 83.07 | 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::empty_strided | ||
| CPU Time | 453543.00 | μs |
| Device Time | 0.00 | μs |
| Self CPU Time | 20071.38 | 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 | 433290.78 | μs |
| Device Time | 0.00 | μs |
| Self CPU Time | 433290.78 | 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 | 502948.35 | μs |
| Device Time | 627.16 | μs |
| Self CPU Time | 502948.35 | μs |
| Self Device Time | 627.16 | 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_coalesced_hardsigmoid_kernel<float>(float const*, float*, unsigned long) | ||
| CPU Time | 0.00 | μs |
| Device Time | 26209.38 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 26209.38 | 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 | 22881.47 | μs |
| Device Time | 42460.51 | μs |
| Self CPU Time | 22881.47 | μs |
| Self Device Time | 42460.51 | 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 | 69153.78 | μs |
| Device Time | 648904.18 | μs |
| Self CPU Time | 12517.36 | 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 | 56638.10 | μs |
| Device Time | 648904.18 | μs |
| Self CPU Time | 16024.52 | μs |
| Self Device Time | 648904.18 | 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 | 648981.65 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 648981.65 | 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.