← Back to Leaderboard

The AI CUDA Engineer 👷

29_Softplus • softplus_coalesced_memory_access_base

Level 1 • Task 29
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
Download Evaluation Download PyTorch Download CUDA Download Profiles

Kernel Information

Operation Name 29_Softplus
Level ID 1
Task ID 29
Kernel Name softplus_coalesced_memory_access_base
CUDA Speedup (Native) 0.993x
CUDA Speedup (Compile) 4.180x
CUDA Runtime 0.007 ms
PyTorch Runtime (Native) 0.007 ms
PyTorch Runtime (Compile) 0.029 ms
Correct True
Max Diff (vs. Reference) 0.000000
Model azure-gpt-4o-2024-08-06
Temperature 0.50
View Experiment Progress Details

Related Kernels (Level 1, Task 29 • 29_Softplus)

Rank Kernel Name Runtime (ms) Speedup Native Speedup Compile
🥇 softplus_modular_base_base 0.01 1.16 4.88
🥇 warp_and_alignment_optimized_softplus_edit_1 0.01 1.16 4.88
🥇 branchless_softplus_edit_1 0.01 1.16 4.88
🥇 warp_optimized_softplus_base 0.01 1.16 4.88
5 softplus_unrolled_base_base 0.01 0.99 4.18
5 softplus_coalesced_base 0.01 0.99 4.18
5 softplus_2d_block_thread_base 0.01 0.99 4.18
5 optimized_softplus_cuda_base 0.01 0.99 4.18
5 softplus_coalesced_memory_access_base 0.01 0.99 4.18
5 softplus_tuned_indexing_base_base 0.01 0.99 4.18
5 softplus_blockstride_base 0.01 0.99 4.18
5 softplus_loop_unroll_base_base 0.01 0.99 4.18
5 softplus_branchless_base 0.01 0.99 4.18
5 softplus_blocksize_experiment_base 0.01 0.99 4.18
5 softplus_unrolled_base 0.01 0.99 4.18
5 softplus_constant_memory_base_base 0.01 0.99 4.18
5 optimized_softplus_cuda_base 0.01 0.99 4.18
5 29_Softplus 0.01 0.99 4.18
5 softplus_constant_memory_base 0.01 0.99 4.18
5 optimized_softplus_base 0.01 0.99 4.18 
#include <torch/extension.h>
#include <cuda.h>
#include <cuda_runtime.h>

// Device function for Softplus computation

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));
}

// Kernel with coalesced memory access

template <typename scalar_t>
__global__ void softplus_kernel_coalesced(
    const scalar_t* __restrict__ input,
    scalar_t* __restrict__ output,
    const int size) {
    
    const int idx = blockIdx.x * blockDim.x + threadIdx.x;
    const int stride = blockDim.x * gridDim.x;
    
    for (int i = idx; i < size; i += stride) {
        const scalar_t x = input[i];
        output[i] = compute_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.type(), "softplus_forward_cuda", ([&] {
        softplus_kernel_coalesced<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)");
}
Performance Metrics
Metric Value Unit Variance Samples
Executed Ipc Active 1.780 inst/cycle 0.004 5
Executed Ipc Elapsed 0.834 inst/cycle 0.000 5
Issue Slots Busy 46.692 % 2.679 5
Issued Ipc Active 1.866 inst/cycle 0.004 5
SM Busy 46.692 % 2.679 5
Memory Throughput 256507273314.540 byte/second 15187539094445119488.000 5
Mem Busy 12.138 % 0.027 5
Max Bandwidth 11.310 % 0.028 5
L1/TEX Hit Rate 0.000 % 0.000 5
L2 Hit Rate 66.422 % 0.123 5
Mem Pipes Busy 12.812 % 0.040 5
Warp Cycles Per Issued Instruction 26.552 cycle 0.037 5
Warp Cycles Per Executed Instruction 27.886 cycle 0.040 5
Avg. Active Threads Per Warp 32.000 0.000 5
Avg. Not Predicated Off Threads Per Warp 28.690 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 10.000 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 80.964 % 0.384 5
Achieved Active Warps Per SM 51.818 warp 0.158 5
Analysis Rules
Rule Description
INF HighPipeUtilization ALU is the highest-utilized pipeline (31.8%) 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 (80.2%) 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 234324.82 μs
Device Time 40.10 μs
Self CPU Time 57.12 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 234267.69 μs
Device Time 40.10 μs
Self CPU Time 128.19 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 235789.72 μs
Device Time 0.00 μs
Self CPU Time 2070.04 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 233493.78 μs
Device Time 0.00 μs
Self CPU Time 233493.78 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 109053.33 μs
Device Time 2046.84 μs
Self CPU Time 109053.33 μs
Self Device Time 2046.84 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_coalesced<float>(float const*, float*, int)
CPU Time 0.00 μs
Device Time 3707.20 μs
Self CPU Time 0.00 μs
Self Device Time 3707.20 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 2392.09 μs
Device Time 3853.07 μs
Self CPU Time 2392.09 μs
Self Device Time 3853.07 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 63651.36 μs
Device Time 68224.31 μs
Self CPU Time 1266.52 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 62386.53 μs
Device Time 68224.31 μs
Self CPU Time 1655.14 μs
Self Device Time 68224.31 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 68224.31 μs
Self CPU Time 0.00 μs
Self Device Time 68224.31 μs
CPU Memory Usage 0 B
Device Memory Usage 0 B
Self CPU Memory Usage 0 B
Self Device Memory Usage 0 B
Status: Completed 
45282 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.
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_29/b6_s2_softplus_coalesced_memory_access/base/base.cu:25:21 bugprone-narrowing-conversions
25 | const int idx = blockIdx.x * blockDim.x + threadIdx.x;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_29/b6_s2_softplus_coalesced_memory_access/base/base.cu:26:24: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
26 | const int stride = blockDim.x * gridDim.x;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_29/b6_s2_softplus_coalesced_memory_access/base/base.cu:38:22: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
38 | const int size = input.numel();
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_29/b6_s2_softplus_coalesced_memory_access/base/base.cu:42:5: warning: inside a lambda, '__func__' expands to the name of the function call operator; consider capturing the name of the enclosing function explicitly [bugprone-lambda-function-name]
42 | AT_DISPATCH_FLOATING_TYPES(input.type(), "softplus_forward_cuda", ([&] {
| ^
/home/robert_sakana_ai/miniconda3/envs/llm2cuda/lib/python3.11/site-packages/torch/include/ATen/Dispatch.h:237:34: note: expanded from macro 'AT_DISPATCH_FLOATING_TYPES'
237 | AT_DISPATCH_SWITCH(TYPE, NAME, AT_DISPATCH_CASE_FLOATING_TYPES(__VA_ARGS__))
| ^
/home/robert_sakana_ai/miniconda3/envs/llm2cuda/lib/python3.11/site-packages/torch/include/ATen/Dispatch.h:233:3: note: expanded from macro 'AT_DISPATCH_CASE_FLOATING_TYPES'
233 | AT_DISPATCH_CASE(at::ScalarType::Double, __VA_ARGS__) \
| ^
/home/robert_sakana_ai/miniconda3/envs/llm2cuda/lib/python3.11/site-packages/torch/include/ATen/Dispatch.h:74:3: note: expanded from macro 'AT_DISPATCH_CASE'
74 | AT_PRIVATE_CASE_TYPE_USING_HINT(enum_type, scalar_t, __VA_ARGS__)
| ^
note: (skipping 1 expansions in backtrace; use -fmacro-backtrace-limit=0 to see all)
/home/robert_sakana_ai/miniconda3/envs/llm2cuda/lib/python3.11/site-packages/torch/include/ATen/Dispatch.h:58:7: note: expanded from macro 'AT_PRIVATE_CHECK_SELECTIVE_BUILD'
58 | AT_ERROR( \
| ^
/home/robert_sakana_ai/miniconda3/envs/llm2cuda/lib/python3.11/site-packages/torch/include/c10/util/Exception.h:711:32: note: expanded from macro 'AT_ERROR'
711 | C10_EXPAND_MSVC_WORKAROUND(TORCH_CHECK(false, ::c10::str(__VA_ARGS__))); \
| ^
/home/robert_sakana_ai/miniconda3/envs/llm2cuda/lib/python3.11/site-packages/torch/include/c10/util/Exception.h:536:9: note: expanded from macro 'TORCH_CHECK'
536 | __func__, \
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_29/b6_s2_softplus_coalesced_memory_access/base/base.cu:42:5: warning: 'scalar_type' is deprecated: passing at::DeprecatedTypeProperties to an AT_DISPATCH macro is deprecated, pass an at::ScalarType instead [clang-diagnostic-deprecated-declarations]
42 | AT_DISPATCH_FLOATING_TYPES(input.type(), "softplus_forward_cuda", ([&] {
| ^
/home/robert_sakana_ai/miniconda3/envs/llm2cuda/lib/python3.11/site-packages/torch/include/ATen/Dispatch.h:237:3: note: expanded from macro 'AT_DISPATCH_FLOATING_TYPES'
237 | AT_DISPATCH_SWITCH(TYPE, NAME, AT_DISPATCH_CASE_FLOATING_TYPES(__VA_ARGS__))
| ^
/home/robert_sakana_ai/miniconda3/envs/llm2cuda/lib/python3.11/site-packages/torch/include/ATen/Dispatch.h:218:36: note: expanded from macro 'AT_DISPATCH_SWITCH'
218 | at::ScalarType _st = ::detail::scalar_type(the_type); \
| ^
/home/robert_sakana_ai/miniconda3/envs/llm2cuda/lib/python3.11/site-packages/torch/include/ATen/Dispatch.h:106:1: note: 'scalar_type' has been explicitly marked deprecated here
106 | C10_DEPRECATED_MESSAGE(
| ^
/home/robert_sakana_ai/miniconda3/envs/llm2cuda/lib/python3.11/site-packages/torch/include/c10/util/Deprecated.h:24:43: note: expanded from macro 'C10_DEPRECATED_MESSAGE'
24 | #define C10_DEPRECATED_MESSAGE(message) [[deprecated(message)]]
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_29/b6_s2_softplus_coalesced_memory_access/base/base.cu:42:38: warning: 'type' is deprecated: Tensor.type() is deprecated. Instead use Tensor.options(), which in many cases (e.g. in a constructor) is a drop-in replacement. If you were using data from type(), that is now available from Tensor itself, so instead of tensor.type().scalar_type(), use tensor.scalar_type() instead and instead of tensor.type().backend() use tensor.device(). [clang-diagnostic-deprecated-declarations]
42 | AT_DISPATCH_FLOATING_TYPES(input.type(), "softplus_forward_cuda", ([&] {
| ^
/home/robert_sakana_ai/miniconda3/envs/llm2cuda/lib/python3.11/site-packages/torch/include/ATen/core/TensorBody.h:224:3: note: 'type' has been explicitly marked deprecated here
224 | C10_DEPRECATED_MESSAGE("Tensor.type() is deprecated. Instead use Tensor.options(), which in many cases (e.g. in a constructor) is a drop-in replacement. If you were using data from type(), that is now available from Tensor itself, so instead of tensor.type().scalar_type(), use tensor.scalar_type() instead and instead of tensor.type().backend() use tensor.device().")
| ^
/home/robert_sakana_ai/miniconda3/envs/llm2cuda/lib/python3.11/site-packages/torch/include/c10/util/Deprecated.h:24:43: note: expanded from macro 'C10_DEPRECATED_MESSAGE'
24 | #define C10_DEPRECATED_MESSAGE(message) [[deprecated(message)]]
| ^