← Back to Leaderboard

The AI CUDA Engineer 👷

99_Matmul_GELU_Softmax • modular_device_functions_base

Level 2 • Task 99
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:
    """
    Applies linear transformation, GELU activation, and softmax.

    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 after applying linear, GELU and softmax,
            with shape (batch_size, out_features)
    """
    x = F.linear(x, weight, bias)
    x = F.gelu(x)
    x = F.softmax(x, dim=1)
    return x


class Model(nn.Module):
    """
    Simple model that performs a matrix multiplication, applies GELU, and then applies Softmax.
    """

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

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


batch_size = 128
in_features = 100
out_features = 10


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 matrix multiplication, applies GELU, and then applies Softmax.
    """
    def __init__(self, in_features, out_features):
        super(Model, self).__init__()
        self.linear = nn.Linear(in_features, out_features)

    def forward(self, x):
        x = self.linear(x)
        x = torch.nn.functional.gelu(x)
        x = torch.nn.functional.softmax(x, dim=1)
        return x

batch_size = 128
in_features = 100
out_features = 10

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 99_Matmul_GELU_Softmax
Level ID 2
Task ID 99
Kernel Name modular_device_functions_base
CUDA Speedup (Native) 2.525x
CUDA Speedup (Compile) 2.058x
CUDA Runtime 0.011 ms
PyTorch Runtime (Native) 0.028 ms
PyTorch Runtime (Compile) 0.023 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 2, Task 99 • 99_Matmul_GELU_Softmax)

Rank Kernel Name Runtime (ms) Speedup Native Speedup Compile
🥇 fused_opt_base 0.01 2.78 2.26
🥇 warp_divergence_minimized_kernel_base 0.01 2.78 2.26
🥇 block_size_experiment_fused_kernel_base 0.01 2.78 2.26
🥇 optimized_fused_kernel_base 0.01 2.78 2.26
🥇 fused_shared_mem_kernel_base 0.01 2.78 2.26
🥇 balanced_workload_fused_kernel_base 0.01 2.78 2.26
7 reduced_sync_matmul_gelu_softmax_base 0.01 2.53 2.06
7 aligned_ldg_fused_kernel_base_base 0.01 2.53 2.06
7 warp_optimized_fused_kernel_base_base 0.01 2.53 2.06
7 fused_ldg_vec_kernel_base 0.01 2.53 2.06
7 unrolled_fused_matmul_gelu_softmax_base_base 0.01 2.53 2.06
7 fused_optim_base 0.01 2.53 2.06
7 fused_linear_gelu_softmax_optimized_base 0.01 2.53 2.06
7 warp_reduced_fused_kernel_base 0.01 2.53 2.06
7 fused_nodivergence_kernel_base 0.01 2.53 2.06
7 modular_device_functions_base 0.01 2.53 2.06
7 optimized_linear_gelu_softmax_base 0.01 2.53 2.06
7 optimized_linear_gelu_softmax_edit_1 0.01 2.53 2.06
7 optimized_linear_gelu_softmax_base 0.01 2.53 2.06
7 optimized_linear_gelu_softmax_edit_1 0.01 2.53 2.06 
#include <torch/extension.h>
#include <cuda.h>
#include <cuda_runtime.h>

// Modular device function for GELU
__device__ float gelu(float x) {
    const float sqrt_2_over_pi = 0.7978845608028654f;
    const float coef = 0.044715f;
    float cdf = 0.5f * (1.0f + tanhf(sqrt_2_over_pi * x * (1.0f + coef * x * x)));
    return x * cdf;
}

// Modular device function for linear transformation
__device__ float linear_transform(const float* x, const float* weight, const float* bias, int row, int in_features, int out_feature_idx) {
    float sum = 0.0f;
    for (int k = 0; k < in_features; k++) {
        sum += x[row * in_features + k] * weight[out_feature_idx * in_features + k];
    }
    sum += bias[out_feature_idx];
    return sum;
}

// Modular device function for softmax
__device__ void softmax(float* row_data, int out_features, float* shared_mem, int tid) {
    float max_val = -FLT_MAX;
    for (int i = tid; i < out_features; i += blockDim.x) {
        max_val = max(max_val, row_data[i]);
    }
    shared_mem[tid] = max_val;
    __syncthreads();

    if (tid == 0) {
        max_val = -FLT_MAX;
        for (int i = 0; i < blockDim.x; i++) {
            max_val = max(max_val, shared_mem[i]);
        }
        shared_mem[0] = max_val;
    }
    __syncthreads();

    max_val = shared_mem[0];
    float sum = 0.0f;
    for (int i = tid; i < out_features; i += blockDim.x) {
        float val = expf(row_data[i] - max_val);
        row_data[i] = val;
        sum += val;
    }
    shared_mem[tid] = sum;
    __syncthreads();

    if (tid == 0) {
        sum = 0.0f;
        for (int i = 0; i < blockDim.x; i++) {
            sum += shared_mem[i];
        }
        shared_mem[0] = sum;
    }
    __syncthreads();

    sum = shared_mem[0];
    for (int i = tid; i < out_features; i += blockDim.x) {
        row_data[i] /= sum;
    }
}

__global__ void linear_gelu_softmax_kernel(
    const float* x,
    const float* weight,
    const float* bias,
    float* output,
    const int batch_size,
    const int in_features,
    const int out_features
) {
    extern __shared__ float shared_mem[];
    const int row = blockIdx.x;
    const int tid = threadIdx.x;

    if (row < batch_size && tid < out_features) {
        float linear_val = linear_transform(x, weight, bias, row, in_features, tid);
        float gelu_val = gelu(linear_val);
        output[row * out_features + tid] = gelu_val;
    }
    __syncthreads();

    if (row < batch_size) {
        float* row_data = &output[row * out_features];
        softmax(row_data, out_features, shared_mem, tid);
    }
}

torch::Tensor forward(
    torch::Tensor x,
    torch::Tensor weight,
    torch::Tensor bias
) {
    const int batch_size = x.size(0);
    const int in_features = x.size(1);
    const int out_features = weight.size(0);
    
    auto options = torch::TensorOptions()
        .dtype(x.dtype())
        .device(x.device());
    auto output = torch::empty({batch_size, out_features}, options);
    
    const dim3 blocks(batch_size);
    const dim3 threads(out_features);
    size_t shared_mem_size = threads.x * sizeof(float);
    
    linear_gelu_softmax_kernel<<<blocks, threads, shared_mem_size>>>(
        x.data_ptr<float>(),
        weight.data_ptr<float>(),
        bias.data_ptr<float>(),
        output.data_ptr<float>(),
        batch_size,
        in_features, 
        out_features
    );
    
    return output;
}

PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
    m.def("forward", &forward, "Linear + GELU + Softmax forward");
}
Performance Metrics
Metric Value Unit Variance Samples
Executed Ipc Active 0.060 inst/cycle 0.000 5
Executed Ipc Elapsed 0.040 inst/cycle 0.000 5
Issue Slots Busy 1.410 % 0.000 5
Issued Ipc Active 0.060 inst/cycle 0.000 5
SM Busy 1.410 % 0.000 5
Memory Throughput 6947159286.690 byte/second 75205627124768240.000 5
Mem Busy 4.036 % 0.038 5
Max Bandwidth 3.120 % 0.016 5
L1/TEX Hit Rate 87.250 % 0.000 5
L2 Hit Rate 100.478 % 0.402 5
Mem Pipes Busy 1.458 % 0.004 5
Warp Cycles Per Issued Instruction 17.940 cycle 0.459 5
Warp Cycles Per Executed Instruction 18.054 cycle 0.468 5
Avg. Active Threads Per Warp 8.330 0.000 5
Avg. Not Predicated Off Threads Per Warp 8.060 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 64.000 block 0.000 5
Block Limit Shared Mem 56.000 block 0.000 5
Block Limit Warps 64.000 block 0.000 5
Theoretical Active Warps per SM 32.000 warp 0.000 5
Theoretical Occupancy 50.000 % 0.000 5
Achieved Occupancy 1.560 % 0.000 5
Achieved Active Warps Per SM 1.000 warp 0.000 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 ThreadDivergence Instructions are executed in warps, which are groups of 32 threads. Optimal instruction throughput is achieved if all 32 threads of a warp execute the same instruction. The chosen launch configuration, early thread completion, and divergent flow control can significantly lower the number of active threads in a warp per cycle. This kernel achieves an average of 8.3 threads being active per cycle. This is further reduced to 8.1 threads per warp due to predication. The compiler may use predication to avoid an actual branch. Instead, all instructions are scheduled, but a per-thread condition code or predicate controls which threads execute the instructions. Try to avoid different execution paths within a warp when possible. In addition, ensure your kernel makes use of Independent Thread Scheduling, which allows a warp to reconverge after a data-dependent conditional block by explicitly calling __syncwarp().
WRN Occupancy This kernel's theoretical occupancy (50.0%) is limited by the number of blocks that can fit on the SM. The difference between calculated theoretical (50.0%) and measured achieved occupancy (1.6%) 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 194489.60 μs
Device Time 6.37 μs
Self CPU Time 52.99 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::_to_copy
CPU Time 194436.62 μs
Device Time 6.37 μs
Self CPU Time 122.45 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 194176.08 μs
Device Time 0.00 μs
Self CPU Time 107.49 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 193901.64 μs
Device Time 0.00 μs
Self CPU Time 193901.64 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 516767.57 μs
Device Time 17769.11 μs
Self CPU Time 516767.57 μs
Self Device Time 17769.11 μs
CPU Memory Usage 0 B
Device Memory Usage 0 B
Self CPU Memory Usage 0 B
Self Device Memory Usage 0 B
linear_gelu_softmax_kernel(float const*, float const*, float const*, float*, int, int, int)
CPU Time 0.00 μs
Device Time 60302.54 μs
Self CPU Time 0.00 μs
Self Device Time 60302.54 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 23673.67 μs
Device Time 34108.09 μs
Self CPU Time 23673.67 μs
Self Device Time 34108.09 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 67170.44 μs
Device Time 637929.94 μs
Self CPU Time 14238.90 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 52933.08 μs
Device Time 637929.94 μs
Self CPU Time 16101.12 μs
Self Device Time 637929.94 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 638008.73 μs
Self CPU Time 0.00 μs
Self Device Time 638008.73 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 
45290 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/20250208_optimize_b5_s4_e1_sweep/level_2/task_99/b5_s1_modular_device_functions/base/base.cu:14:51 bugprone-easily-swappable-parameters
14 | __device__ float linear_transform(const float* x, const float* weight, const float* bias, int row, int in_features, int out_feature_idx) {
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250208_optimize_b5_s4_e1_sweep/level_2/task_99/b5_s1_modular_device_functions/base/base.cu:14:64: note: the first parameter in the range is 'weight'
14 | __device__ float linear_transform(const float* x, const float* weight, const float* bias, int row, int in_features, int out_feature_idx) {
| ^~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250208_optimize_b5_s4_e1_sweep/level_2/task_99/b5_s1_modular_device_functions/base/base.cu:14:85: note: the last parameter in the range is 'bias'
14 | __device__ float linear_transform(const float* x, const float* weight, const float* bias, int row, int in_features, int out_feature_idx) {
| ^~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250208_optimize_b5_s4_e1_sweep/level_2/task_99/b5_s1_modular_device_functions/base/base.cu:26:46: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
26 | for (int i = tid; i < out_features; i += blockDim.x) {
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250208_optimize_b5_s4_e1_sweep/level_2/task_99/b5_s1_modular_device_functions/base/base.cu:43:46: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
43 | for (int i = tid; i < out_features; i += blockDim.x) {
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250208_optimize_b5_s4_e1_sweep/level_2/task_99/b5_s1_modular_device_functions/base/base.cu:61:46: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
61 | for (int i = tid; i < out_features; i += blockDim.x) {
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250208_optimize_b5_s4_e1_sweep/level_2/task_99/b5_s1_modular_device_functions/base/base.cu:71:5: warning: 2 adjacent parameters of 'linear_gelu_softmax_kernel' of similar type ('const int') are easily swapped by mistake [bugprone-easily-swappable-parameters]
71 | const int batch_size,
| ^~~~~~~~~~~~~~~~~~~~~
72 | const int in_features,
| ~~~~~~~~~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250208_optimize_b5_s4_e1_sweep/level_2/task_99/b5_s1_modular_device_functions/base/base.cu:71:15: note: the first parameter in the range is 'batch_size'
71 | const int batch_size,
| ^~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250208_optimize_b5_s4_e1_sweep/level_2/task_99/b5_s1_modular_device_functions/base/base.cu:72:15: note: the last parameter in the range is 'in_features'
72 | const int in_features,
| ^~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250208_optimize_b5_s4_e1_sweep/level_2/task_99/b5_s1_modular_device_functions/base/base.cu:76:21: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
76 | const int row = blockIdx.x;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250208_optimize_b5_s4_e1_sweep/level_2/task_99/b5_s1_modular_device_functions/base/base.cu:77:21: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
77 | const int tid = threadIdx.x;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250208_optimize_b5_s4_e1_sweep/level_2/task_99/b5_s1_modular_device_functions/base/base.cu:87:28: warning: result of multiplication in type 'int' is used as a pointer offset after an implicit widening conversion to type 'ptrdiff_t' [bugprone-implicit-widening-of-multiplication-result]
87 | float* row_data = &output[row * out_features];
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250208_optimize_b5_s4_e1_sweep/level_2/task_99/b5_s1_modular_device_functions/base/base.cu:87:35: note: make conversion explicit to silence this warning
4 | float* row_data = &output[row * out_features];
| ^~~~~~~~~~~~~~~~~~
| static_cast<ptrdiff_t>( )
/home/robert_sakana_ai/llm_cuda/experiments/20250208_optimize_b5_s4_e1_sweep/level_2/task_99/b5_s1_modular_device_functions/base/base.cu:87:35: note: perform multiplication in a wider type
87 | float* row_data = &output[row * out_features];
| ^~~
| static_cast<ptrdiff_t>( )
/home/robert_sakana_ai/llm_cuda/experiments/20250208_optimize_b5_s4_e1_sweep/level_2/task_99/b5_s1_modular_device_functions/base/base.cu:93:19: 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]
93 | torch::Tensor x,
| ^
| const &
/home/robert_sakana_ai/llm_cuda/experiments/20250208_optimize_b5_s4_e1_sweep/level_2/task_99/b5_s1_modular_device_functions/base/base.cu:94:19: 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]
94 | torch::Tensor weight,
| ^
| const &
/home/robert_sakana_ai/llm_cuda/experiments/20250208_optimize_b5_s4_e1_sweep/level_2/task_99/b5_s1_modular_device_functions/base/base.cu:95:19: 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]
95 | torch::Tensor bias
| ^
| const &
/home/robert_sakana_ai/llm_cuda/experiments/20250208_optimize_b5_s4_e1_sweep/level_2/task_99/b5_s1_modular_device_functions/base/base.cu:97:28: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
97 | const int batch_size = x.size(0);
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250208_optimize_b5_s4_e1_sweep/level_2/task_99/b5_s1_modular_device_functions/base/base.cu:98:29: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
98 | const int in_features = x.size(1);
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250208_optimize_b5_s4_e1_sweep/level_2/task_99/b5_s1_modular_device_functions/base/base.cu:99:30: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
99 | const int out_features = weight.size(0);
| ^