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>
// Device function to compute the Softplus activation
template <typename scalar_t>
__device__ __forceinline__ scalar_t compute_softplus(const scalar_t x) {
if (x > static_cast<scalar_t>(20.0)) {
return x;
} else if (x < static_cast<scalar_t>(-20.0)) {
return exp(x);
}
return log1p(exp(x));
}
// CUDA kernel with manual loop unrolling (factor of 4) using #pragma unroll
template <typename scalar_t>
__global__ void softplus_kernel_unrolled(
const scalar_t* __restrict__ input,
scalar_t* __restrict__ output,
const int size) {
const int tid = blockIdx.x * blockDim.x + threadIdx.x;
const int stride = blockDim.x * gridDim.x;
int i = tid;
// Unrolled loop: process 4 elements per iteration
for (; i + 3 * stride < size; i += 4 * stride) {
#pragma unroll
{
scalar_t in0 = input[i];
scalar_t in1 = input[i + stride];
scalar_t in2 = input[i + 2 * stride];
scalar_t in3 = input[i + 3 * stride];
output[i] = compute_softplus(in0);
output[i + stride] = compute_softplus(in1);
output[i + 2 * stride] = compute_softplus(in2);
output[i + 3 * stride] = compute_softplus(in3);
}
}
// Process any remaining elements
for (; i < size; i += stride) {
output[i] = compute_softplus(input[i]);
}
}
// 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_unrolled<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 | 1.906 | inst/cycle | 0.000 | 5 |
| Executed Ipc Elapsed | 0.884 | inst/cycle | 0.000 | 5 |
| Issue Slots Busy | 51.406 | % | 0.106 | 5 |
| Issued Ipc Active | 2.056 | inst/cycle | 0.000 | 5 |
| SM Busy | 51.406 | % | 0.106 | 5 |
| Memory Throughput | 254029531171.830 | byte/second | 5144178277050267648.000 | 5 |
| Mem Busy | 12.004 | % | 0.011 | 5 |
| Max Bandwidth | 11.206 | % | 0.008 | 5 |
| L1/TEX Hit Rate | 0.000 | % | 0.000 | 5 |
| L2 Hit Rate | 66.926 | % | 0.215 | 5 |
| Mem Pipes Busy | 12.736 | % | 0.014 | 5 |
| Warp Cycles Per Issued Instruction | 25.726 | cycle | 0.988 | 5 |
| Warp Cycles Per Executed Instruction | 27.732 | cycle | 1.150 | 5 |
| Avg. Active Threads Per Warp | 32.000 | 0.000 | 5 | |
| Avg. Not Predicated Off Threads Per Warp | 28.900 | 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 | 81.318 | % | 0.001 | 5 |
| Achieved Active Warps Per SM | 52.044 | warp | 0.001 | 5 |
| Rule | Description |
|---|---|
| INF HighPipeUtilization | ALU is the highest-utilized pipeline (31.0%) 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 (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. |
| 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 | 623195.06 | μs |
| Device Time | 40.13 | μs |
| Self CPU Time | 34.79 | 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 | 623160.27 | μs |
| Device Time | 40.13 | μs |
| Self CPU Time | 71.69 | 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 |
| aten::empty_strided | ||
| CPU Time | 640940.92 | μs |
| Device Time | 0.00 | μs |
| Self CPU Time | 18160.38 | 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 | 622582.56 | μs |
| Device Time | 0.00 | μs |
| Self CPU Time | 622582.56 | 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 | 451320.54 | μs |
| Device Time | 20437.66 | μs |
| Self CPU Time | 451320.54 | μs |
| Self Device Time | 20437.66 | 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_unrolled<float>(float const*, float*, int) | ||
| CPU Time | 0.00 | μs |
| Device Time | 25884.42 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 25884.42 | 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 | 20620.14 | μs |
| Device Time | 39394.26 | μs |
| Self CPU Time | 20620.14 | μs |
| Self Device Time | 39394.26 | 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 | 60983.62 | μs |
| Device Time | 584510.29 | μs |
| Self CPU Time | 11080.42 | 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 | 49904.08 | μs |
| Device Time | 584510.29 | μs |
| Self CPU Time | 14157.99 | μs |
| Self Device Time | 584510.29 | 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 | 584588.98 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 584588.98 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
45245 warnings and 1 error generated when compiling for host. Error while processing /home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_29/b7_s3_softplus_unrolled/base/base.cu. Suppressed 45286 warnings (45239 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. Found compiler error(s).