← Back to Leaderboard

The AI CUDA Engineer 👷

21_Sigmoid • nondivergent_vectorized_sigmoid_base

Level 1 • Task 21
import torch
import torch.nn as nn
import torch.nn.functional as F


def module_fn(x: torch.Tensor) -> torch.Tensor:
    """
    Applies Sigmoid activation to the input tensor.

    Args:
        x (torch.Tensor): Input tensor of any shape.

    Returns:
        torch.Tensor: Output tensor with Sigmoid applied, same shape as input.
    """
    return torch.sigmoid(x)


class Model(nn.Module):
    """
    Simple model that performs a Sigmoid 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 Sigmoid activation.
    """
    def __init__(self):
        super(Model, self).__init__()
    
    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """
        Applies Sigmoid activation to the input tensor.

        Args:
            x (torch.Tensor): Input tensor of any shape.

        Returns:
            torch.Tensor: Output tensor with Sigmoid applied, same shape as input.
        """
        return torch.sigmoid(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 21_Sigmoid
Level ID 1
Task ID 21
Kernel Name nondivergent_vectorized_sigmoid_base
CUDA Speedup (Native) 1.109x
CUDA Speedup (Compile) 4.825x
CUDA Runtime 0.006 ms
PyTorch Runtime (Native) 0.007 ms
PyTorch Runtime (Compile) 0.029 ms
Correct True
Max Diff (vs. Reference) 0.000000
Model o3-mini-2025-01-31
Temperature 1.00
View Experiment Progress Details

Related Kernels (Level 1, Task 21 • 21_Sigmoid)

Rank Kernel Name Runtime (ms) Speedup Native Speedup Compile
🥇 sigmoid_shared_mem_optimized_base 0.01 1.11 4.82
🥇 21_sigmoid_modular_device_base 0.01 1.11 4.82
🥇 sigmoid_unroll_optimized_base_base 0.01 1.11 4.82
🥇 sigmoid_min_sync_base_base 0.01 1.11 4.82
🥇 optimized_sigmoid_cuda_base 0.01 1.11 4.82
🥇 optimized_sigmoid_limited_sync_base 0.01 1.11 4.82
🥇 sigmoid_ldg_vectorized_base 0.01 1.11 4.82
🥇 optimized_sigmoid_cuda_base 0.01 1.11 4.82
🥇 optimized_sigmoid_vectorized_combined_edit_1 0.01 1.11 4.82
🥇 21_Sigmoid_optimized_memory_base 0.01 1.11 4.82
🥇 sigmoid_minimal_sync_base_base 0.01 1.11 4.82
🥇 vectorized_ldg_aligned_edit_1 0.01 1.11 4.82
🥇 nondivergent_vectorized_sigmoid_base 0.01 1.11 4.82
🥇 vectorized_no_sync_base 0.01 1.11 4.82
🥇 vectorized_sigmoid_base 0.01 1.11 4.82
🥇 syncthreads_minimal_sigmoid_base 0.01 1.11 4.82
🥇 vectorized_ldg_aligned_base 0.01 1.11 4.82
🥇 optimized_sigmoid_vectorized_combined_base 0.01 1.11 4.82
🥇 optimized_sigmoid_blocksize_tuning_edit_1 0.01 1.11 4.82
🥇 optimized_sigmoid_blocksize_tuning_base 0.01 1.11 4.82 
#include <torch/extension.h>
#include <cuda.h>
#include <cuda_runtime.h>
#include <cmath>

// Vectorized kernel for float using float4 with uniform control flow
__global__ void vectorized_sigmoid_kernel_float(const float* __restrict__ input,
                                                  float* __restrict__ output,
                                                  const int n_vec) {
    int idx = blockIdx.x * blockDim.x + threadIdx.x;
    int stride = gridDim.x * blockDim.x;
    // Each thread processes a block of consecutive float4 groups
    for (int i = idx; i < n_vec; i += stride) {
        // Load a group of 4 floats at once
        float4 in_val = reinterpret_cast<const float4*>(input)[i];
        float4 out_val;
        // Compute sigmoid uniformly for each component
        out_val.x = 1.0f / (1.0f + expf(-in_val.x));
        out_val.y = 1.0f / (1.0f + expf(-in_val.y));
        out_val.z = 1.0f / (1.0f + expf(-in_val.z));
        out_val.w = 1.0f / (1.0f + expf(-in_val.w));
        reinterpret_cast<float4*>(output)[i] = out_val;
    }
}

// Tail kernel for float to process leftover elements with uniform thread count
__global__ void tail_sigmoid_kernel_float(const float* __restrict__ input,
                                            float* __restrict__ output,
                                            const int start,
                                            const int tail_size) {
    int idx = blockIdx.x * blockDim.x + threadIdx.x;
    // Launch exactly 'tail_size' threads to minimize divergence in the tail
    if (idx < tail_size) {
        int i = start + idx;
        float in_val = input[i];
        float out_val = 1.0f / (1.0f + expf(-in_val));
        output[i] = out_val;
    }
}

// Fallback scalar kernel for non-float types
template <typename scalar_t>
__global__ void sigmoid_kernel_scalar(const scalar_t* __restrict__ input,
                                        scalar_t* __restrict__ output,
                                        const int64_t size) {
    int idx = blockIdx.x * blockDim.x + threadIdx.x;
    int stride = blockDim.x * gridDim.x;
    for (int i = idx; i < size; i += stride) {
        float in_val = static_cast<float>(input[i]);
        float out_val = 1.0f / (1.0f + expf(-in_val));
        output[i] = static_cast<scalar_t>(out_val);
    }
}

// Forward function dispatches to vectorized kernels for float and a scalar kernel for others
torch::Tensor forward(torch::Tensor input) {
    auto output = torch::empty_like(input);
    const int64_t size = input.numel();
    const int threads = 256;

    if (input.scalar_type() == at::ScalarType::Float) {
        // Compute the number of groups of 4 elements
        int n_vec = size / 4;
        int tail = size - (n_vec * 4);
        
        // Launch vectorized kernel if there is a complete float4 block
        if (n_vec > 0) {
            int blocks = (n_vec + threads - 1) / threads;
            vectorized_sigmoid_kernel_float<<<blocks, threads>>>(
                input.data_ptr<float>(),
                output.data_ptr<float>(),
                n_vec
            );
        }
        // Launch a separate kernel to handle leftover elements
        if (tail > 0) {
            int blocks_tail = (tail + threads - 1) / threads;
            tail_sigmoid_kernel_float<<<blocks_tail, threads>>>(
                input.data_ptr<float>(),
                output.data_ptr<float>(),
                n_vec * 4,
                tail
            );
        }
    } else {
        int blocks = (size + threads - 1) / threads;
        AT_DISPATCH_FLOATING_TYPES(input.scalar_type(), "sigmoid_kernel_scalar", ([&] {
            sigmoid_kernel_scalar<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", &forward, "Nondivergent Vectorized Sigmoid forward (CUDA)");
}
Performance Metrics
Metric Value Unit Variance Samples
Analysis Rules
Rule Description
Operation / Metric Value Unit
aten::to
CPU Time 547032.69 μs
Device Time 40.13 μs
Self CPU Time 42.03 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 546990.66 μs
Device Time 40.13 μs
Self CPU Time 108.87 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 553938.07 μs
Device Time 0.00 μs
Self CPU Time 7434.70 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 546269.09 μs
Device Time 0.00 μs
Self CPU Time 546269.09 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 149193.17 μs
Device Time 6839.77 μs
Self CPU Time 149193.17 μs
Self Device Time 6839.77 μs
CPU Memory Usage 0 B
Device Memory Usage 0 B
Self CPU Memory Usage 0 B
Self Device Memory Usage 0 B
vectorized_sigmoid_kernel_float(float const*, float*, int)
CPU Time 0.00 μs
Device Time 7743.70 μs
Self CPU Time 0.00 μs
Self Device Time 7743.70 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 7230.24 μs
Device Time 13126.69 μs
Self CPU Time 7230.24 μs
Self Device Time 13126.69 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 37944.06 μs
Device Time 202842.31 μs
Self CPU Time 4132.99 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 33812.89 μs
Device Time 202842.31 μs
Self CPU Time 4931.00 μs
Self Device Time 202842.31 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 202842.31 μs
Self CPU Time 0.00 μs
Self Device Time 202842.31 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 
45287 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/20250212_optimize_b5_s4_e1_v2/level_1/task_21/b5_s2_nondivergent_vectorized_sigmoid/base/base.cu:10:15 bugprone-narrowing-conversions
10 | int idx = blockIdx.x * blockDim.x + threadIdx.x;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250212_optimize_b5_s4_e1_v2/level_1/task_21/b5_s2_nondivergent_vectorized_sigmoid/base/base.cu:11:18: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
11 | int stride = gridDim.x * blockDim.x;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250212_optimize_b5_s4_e1_v2/level_1/task_21/b5_s2_nondivergent_vectorized_sigmoid/base/base.cu:29:45: warning: 2 adjacent parameters of 'tail_sigmoid_kernel_float' of similar type ('const int') are easily swapped by mistake [bugprone-easily-swappable-parameters]
29 | const int start,
| ^~~~~~~~~~~~~~~~
30 | const int tail_size) {
| ~~~~~~~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250212_optimize_b5_s4_e1_v2/level_1/task_21/b5_s2_nondivergent_vectorized_sigmoid/base/base.cu:29:55: note: the first parameter in the range is 'start'
29 | const int start,
| ^~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250212_optimize_b5_s4_e1_v2/level_1/task_21/b5_s2_nondivergent_vectorized_sigmoid/base/base.cu:30:55: note: the last parameter in the range is 'tail_size'
30 | const int tail_size) {
| ^~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250212_optimize_b5_s4_e1_v2/level_1/task_21/b5_s2_nondivergent_vectorized_sigmoid/base/base.cu:31:15: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
31 | int idx = blockIdx.x * blockDim.x + threadIdx.x;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250212_optimize_b5_s4_e1_v2/level_1/task_21/b5_s2_nondivergent_vectorized_sigmoid/base/base.cu:46:15: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
46 | int idx = blockIdx.x * blockDim.x + threadIdx.x;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250212_optimize_b5_s4_e1_v2/level_1/task_21/b5_s2_nondivergent_vectorized_sigmoid/base/base.cu:47:18: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
47 | int stride = blockDim.x * gridDim.x;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250212_optimize_b5_s4_e1_v2/level_1/task_21/b5_s2_nondivergent_vectorized_sigmoid/base/base.cu:63:21: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
63 | int n_vec = size / 4;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250212_optimize_b5_s4_e1_v2/level_1/task_21/b5_s2_nondivergent_vectorized_sigmoid/base/base.cu:64:20: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
64 | int tail = size - (n_vec * 4);
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250212_optimize_b5_s4_e1_v2/level_1/task_21/b5_s2_nondivergent_vectorized_sigmoid/base/base.cu:64:28: warning: performing an implicit widening conversion to type 'int64_t' (aka 'long') of a multiplication performed in type 'int' [bugprone-implicit-widening-of-multiplication-result]
64 | int tail = size - (n_vec * 4);
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250212_optimize_b5_s4_e1_v2/level_1/task_21/b5_s2_nondivergent_vectorized_sigmoid/base/base.cu:64:28: note: make conversion explicit to silence this warning
5 | int tail = size - (n_vec * 4);
| ^~~~~~~~~
| static_cast<int64_t>( )
/home/robert_sakana_ai/llm_cuda/experiments/20250212_optimize_b5_s4_e1_v2/level_1/task_21/b5_s2_nondivergent_vectorized_sigmoid/base/base.cu:64:28: note: perform multiplication in a wider type
64 | int tail = size - (n_vec * 4);
| ^~~~~
| static_cast<int64_t>( )
/home/robert_sakana_ai/llm_cuda/experiments/20250212_optimize_b5_s4_e1_v2/level_1/task_21/b5_s2_nondivergent_vectorized_sigmoid/base/base.cu:86:22: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
86 | int blocks = (size + threads - 1) / threads;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250212_optimize_b5_s4_e1_v2/level_1/task_21/b5_s2_nondivergent_vectorized_sigmoid/base/base.cu:87:9: 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]
87 | AT_DISPATCH_FLOATING_TYPES(input.scalar_type(), "sigmoid_kernel_scalar", ([&] {
| ^
/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__, \
| ^