← Back to Leaderboard

The AI CUDA Engineer 👷

81_Gemm_Swish_Divide_Clamp_Tanh_Clamp • stride_loop_optimization_base_base

Level 2 • Task 81
import torch
import torch.nn as nn
import torch.nn.functional as F


def module_fn(
    x: torch.Tensor,
    weight: torch.Tensor,
    bias: torch.Tensor,
) -> torch.Tensor:
    """
    Performs gemm, swish, divide, clamp, tanh, and clamp operations.

    Args:
        x (torch.Tensor): Input tensor of shape (batch_size, in_features)
        weight (torch.Tensor): Weight matrix of shape (out_features, in_features)
        bias (torch.Tensor): Bias vector of shape (out_features)

    Returns:
        torch.Tensor: Output tensor of shape (batch_size, out_features)
    """
    x = F.linear(x, weight, bias)
    x = x * torch.sigmoid(x)  # Swish activation
    x = x / 2.0
    x = torch.clamp(x, min=-1.0, max=1.0)  # Clamp between -1 and 1
    x = torch.tanh(x)  # Tanh activation
    x = torch.clamp(x, min=-1.0, max=1.0)  # Clamp between -1 and 1
    return x


class Model(nn.Module):
    """
    Simple model that performs a gemm, swish, divide, clamp, tanh, and clamp operations.
    """

    def __init__(self, in_features, out_features):
        super(Model, self).__init__()
        mm = nn.Linear(in_features, out_features)
        self.weight = nn.Parameter(mm.weight)
        self.bias = nn.Parameter(mm.bias)

    def forward(self, x, fn=module_fn):
        return fn(x, self.weight, self.bias)


batch_size = 128
in_features = 1024
out_features = 512


def get_inputs():
    return [torch.randn(batch_size, in_features)]


def get_init_inputs():
    return [in_features, out_features]

import torch
import torch.nn as nn

class Model(nn.Module):
    """
    Simple model that performs a gemm, swish, divide, clamp, tanh, and clamp operations.
    """
    def __init__(self, in_features, out_features, bias=True):
        super(Model, self).__init__()
        self.gemm = nn.Linear(in_features, out_features, bias=bias)

    def forward(self, x):
        """
        Args:
            x (torch.Tensor): Input tensor of shape (batch_size, in_features).
        Returns:
            torch.Tensor: Output tensor of shape (batch_size, out_features).
        """
        x = self.gemm(x)
        x = x * torch.sigmoid(x)  # Swish activation
        x = x / 2.0
        x = torch.clamp(x, min=-1.0, max=1.0)  # Clamp between -1 and 1
        x = torch.tanh(x)  # Tanh activation
        x = torch.clamp(x, min=-1.0, max=1.0)  # Clamp between -1 and 1
        return x

batch_size = 128
in_features = 1024
out_features = 512

def get_inputs():
    return [torch.randn(batch_size, in_features)]

def get_init_inputs():
    return [in_features, out_features]
Download Evaluation Download PyTorch Download CUDA Download Profiles

Kernel Information

Operation Name 81_Gemm_Swish_Divide_Clamp_Tanh_Clamp
Level ID 2
Task ID 81
Kernel Name stride_loop_optimization_base_base
CUDA Speedup (Native) 2.191x
CUDA Speedup (Compile) 1.898x
CUDA Runtime 0.023 ms
PyTorch Runtime (Native) 0.050 ms
PyTorch Runtime (Compile) 0.044 ms
Correct True
Max Diff (vs. Reference) 0.000000
Model azure-gpt-4o-2024-08-06
Temperature 0.00
View Experiment Progress Details

Related Kernels (Level 2, Task 81 • 81_Gemm_Swish_Divide_Clamp_Tanh_Clamp)

Rank Kernel Name Runtime (ms) Speedup Native Speedup Compile
🥇 gemm_2d_map_base 0.02 2.19 1.90
🥇 gemm_2d_map_improved_base 0.02 2.19 1.90
🥇 hybrid_aligned_2d_kernel_base 0.02 2.19 1.90
🥇 aligned_memory_access_base_base 0.02 2.19 1.90
🥇 optimized_stride_loop_base 0.02 2.19 1.90
🥇 stride_loop_optimization_base_base 0.02 2.19 1.90
🥇 unrolled_2d_map_base_base 0.02 2.19 1.90
🥇 stride_loop_opt_base 0.02 2.19 1.90
🥇 fused_activation_base 0.02 2.19 1.90
🥇 optimized_stride_loop_with_prefetch_base 0.02 2.19 1.90
🥇 optimized_fused_activation_base 0.02 2.19 1.90
🥇 optimized_thread_block_indexing_base 0.02 2.19 1.90
🥇 dim2_grid_activation_base 0.02 2.19 1.90
14 dynamic_block_size_base_base 0.02 2.10 1.82
14 strided_vectorized_base_base 0.02 2.10 1.82
14 even_load_base 0.02 2.10 1.82
14 81_gemm_swish_divide_clamp_tanh_clamp_optimized_blocks_base 0.02 2.10 1.82
14 stride_loop_correct_base 0.02 2.10 1.82
14 adaptive_vectorized_kernel_base 0.02 2.10 1.82
14 uniform_control_flow_base_base 0.02 2.10 1.82 
#include <torch/extension.h>
#include <ATen/ATen.h>
#include <cuda.h>
#include <cuda_runtime.h>

#define CHECK_CUDA(x) TORCH_CHECK(x.is_cuda(), #x " must be a CUDA tensor")
#define CHECK_CONTIGUOUS(x) TORCH_CHECK(x.is_contiguous(), #x " must be contiguous")
#define CHECK_INPUT(x) \
    CHECK_CUDA(x);     \
    CHECK_CONTIGUOUS(x)

// Kernel using stride loops for efficient workload distribution
template <typename scalar_t>
__global__ void module_kernel_stride(
    const scalar_t* __restrict__ x_in,
    scalar_t* __restrict__ x_out,
    const size_t size) {

    const int tid = blockIdx.x * blockDim.x + threadIdx.x;
    const int stride = blockDim.x * gridDim.x;

    for (size_t i = tid; i < size; i += stride) {
        scalar_t x = x_in[i];

        // Swish activation
        scalar_t sigmoid_x = static_cast<scalar_t>(1) / (static_cast<scalar_t>(1) + exp(-x));
        x = x * sigmoid_x;

        // Divide by 2
        x = x * static_cast<scalar_t>(0.5);

        // Clamp between -1 and 1
        x = max(min(x, static_cast<scalar_t>(1)), static_cast<scalar_t>(-1));

        // Tanh activation
        x = tanh(x);

        // Final clamp
        x = max(min(x, static_cast<scalar_t>(1)), static_cast<scalar_t>(-1));

        x_out[i] = x;
    }
}

// CUDA forward function
torch::Tensor module_forward_cuda(
    torch::Tensor x,
    torch::Tensor weight,
    torch::Tensor bias) {
    auto x_linear = torch::addmm(bias, x, weight.t());
    auto x_out = torch::empty_like(x_linear);

    const int threads = 256;
    const int blocks = (x_linear.numel() + threads - 1) / threads;

    AT_DISPATCH_FLOATING_TYPES_AND_HALF(x_linear.scalar_type(), "module_forward_cuda", ([&] {
        module_kernel_stride<scalar_t><<<blocks, threads>>>(
            x_linear.data_ptr<scalar_t>(),
            x_out.data_ptr<scalar_t>(),
            x_linear.numel());
    }));

    return x_out;
}

// C++ interface
torch::Tensor module_forward(
    torch::Tensor x,
    torch::Tensor weight,
    torch::Tensor bias) {
    CHECK_INPUT(x);
    CHECK_INPUT(weight);
    CHECK_INPUT(bias);
    return module_forward_cuda(x, weight, bias);
}

// PyBind11 module definition
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
    m.def("forward", &module_forward, "Custom module forward function (CUDA)");
}
Performance Metrics
Metric Value Unit Variance Samples
Executed Ipc Active 0.400 inst/cycle 0.000 5
Executed Ipc Elapsed 0.148 inst/cycle 0.000 5
Issue Slots Busy 10.962 % 0.156 5
Issued Ipc Active 0.440 inst/cycle 0.000 5
SM Busy 10.962 % 0.156 5
Memory Throughput 77154561790.410 byte/second 4126736425410776064.000 5
Mem Busy 10.668 % 0.073 5
Max Bandwidth 6.934 % 0.032 5
L1/TEX Hit Rate 0.000 % 0.000 5
L2 Hit Rate 82.930 % 0.088 5
Mem Pipes Busy 2.300 % 0.004 5
Warp Cycles Per Issued Instruction 32.128 cycle 0.356 5
Warp Cycles Per Executed Instruction 35.228 cycle 0.428 5
Avg. Active Threads Per Warp 30.870 0.000 5
Avg. Not Predicated Off Threads Per Warp 29.640 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 16.000 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 21.932 % 0.082 5
Achieved Active Warps Per SM 14.036 warp 0.034 5
Analysis Rules
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 (21.7%) 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 449262.84 μs
Device Time 176.61 μs
Self CPU Time 50.81 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 449212.04 μs
Device Time 176.61 μs
Self CPU Time 93.71 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 471974.82 μs
Device Time 0.00 μs
Self CPU Time 23409.04 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 448196.51 μs
Device Time 0.00 μs
Self CPU Time 448196.51 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::addmm
CPU Time 602797.27 μs
Device Time 148527.05 μs
Self CPU Time 207335.24 μs
Self Device Time 148527.05 μs
CPU Memory Usage 0 B
Device Memory Usage 0 B
Self CPU Memory Usage 0 B
Self Device Memory Usage 0 B
sm80_xmma_gemm_f32f32_f32f32_f32_tn_n_tilesize32x32x8_stage3_warpsize1x2x1_ffma_aligna4_alignc4_execute_kernel__51_cublas
CPU Time 0.00 μs
Device Time 133669.21 μs
Self CPU Time 0.00 μs
Self Device Time 133669.21 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 75784.26 μs
Device Time 693819.82 μs
Self CPU Time 14069.32 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 61716.22 μs
Device Time 693819.82 μs
Self CPU Time 20993.26 μs
Self Device Time 693819.82 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 693819.82 μs
Self CPU Time 0.00 μs
Self Device Time 693819.82 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 
45291 warnings generated when compiling for host.
Suppressed 45323 warnings (45276 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/20250203_optimize_b10_s4_e0_sweep/level_2/task_81/b5_s3_stride_loop_optimization_base/base/base.cu:6:35 bugprone-macro-parentheses
6 | #define CHECK_CUDA(x) TORCH_CHECK(x.is_cuda(), #x " must be a CUDA tensor")
| ^
| ()
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_81/b5_s3_stride_loop_optimization_base/base/base.cu:7:41: warning: macro argument should be enclosed in parentheses [bugprone-macro-parentheses]
7 | #define CHECK_CONTIGUOUS(x) TORCH_CHECK(x.is_contiguous(), #x " must be contiguous")
| ^
| ()
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_81/b5_s3_stride_loop_optimization_base/base/base.cu:19:21: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
19 | const int tid = blockIdx.x * blockDim.x + threadIdx.x;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_81/b5_s3_stride_loop_optimization_base/base/base.cu:20:24: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
20 | const int stride = blockDim.x * gridDim.x;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_81/b5_s3_stride_loop_optimization_base/base/base.cu:47:19: warning: the parameter 'x' is copied for each invocation but only used as a const reference; consider making it a const reference [performance-unnecessary-value-param]
47 | torch::Tensor x,
| ^
| const &
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_81/b5_s3_stride_loop_optimization_base/base/base.cu:48:19: warning: the parameter 'weight' is copied for each invocation but only used as a const reference; consider making it a const reference [performance-unnecessary-value-param]
48 | torch::Tensor weight,
| ^
| const &
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_81/b5_s3_stride_loop_optimization_base/base/base.cu:49:19: warning: the parameter 'bias' is copied for each invocation but only used as a const reference; consider making it a const reference [performance-unnecessary-value-param]
49 | torch::Tensor bias) {
| ^
| const &
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_81/b5_s3_stride_loop_optimization_base/base/base.cu:54:24: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
54 | const int blocks = (x_linear.numel() + threads - 1) / threads;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_81/b5_s3_stride_loop_optimization_base/base/base.cu:56:5: 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]
56 | AT_DISPATCH_FLOATING_TYPES_AND_HALF(x_linear.scalar_type(), "module_forward_cuda", ([&] {
| ^
/home/robert_sakana_ai/miniconda3/envs/llm2cuda/lib/python3.11/site-packages/torch/include/ATen/Dispatch.h:246:19: note: expanded from macro 'AT_DISPATCH_FLOATING_TYPES_AND_HALF'
246 | TYPE, NAME, AT_DISPATCH_CASE_FLOATING_TYPES_AND_HALF(__VA_ARGS__))
| ^
/home/robert_sakana_ai/miniconda3/envs/llm2cuda/lib/python3.11/site-packages/torch/include/ATen/Dispatch.h:240:3: note: expanded from macro 'AT_DISPATCH_CASE_FLOATING_TYPES_AND_HALF'
240 | 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/20250203_optimize_b10_s4_e0_sweep/level_2/task_81/b5_s3_stride_loop_optimization_base/base/base.cu:68:19: warning: the parameter 'x' is copied for each invocation but only used as a const reference; consider making it a const reference [performance-unnecessary-value-param]
68 | torch::Tensor x,
| ^
| const &
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_81/b5_s3_stride_loop_optimization_base/base/base.cu:69:19: warning: the parameter 'weight' is copied for each invocation but only used as a const reference; consider making it a const reference [performance-unnecessary-value-param]
69 | torch::Tensor weight,
| ^
| const &
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_81/b5_s3_stride_loop_optimization_base/base/base.cu:70:19: warning: the parameter 'bias' is copied for each invocation but only used as a const reference; consider making it a const reference [performance-unnecessary-value-param]
70 | torch::Tensor bias) {
| ^
| const &