← Back to Leaderboard

The AI CUDA Engineer 👷

95_CrossEntropyLoss • modular_device_ce_loss_base

Level 1 • Task 95
import torch
import torch.nn as nn
import torch.nn.functional as F


def module_fn(predictions: torch.Tensor, targets: torch.Tensor) -> torch.Tensor:
    """
    Computes the Cross Entropy Loss for multi-class classification tasks.

    Args:
        predictions (torch.Tensor): Predicted values.
        targets (torch.Tensor): Target values.

    Returns:
        torch.Tensor: Cross Entropy Loss.
    """
    return F.cross_entropy(predictions, targets)


class Model(nn.Module):
    """
    A model that computes Cross Entropy Loss for multi-class classification tasks.

    Parameters:
        None
    """

    def __init__(self):
        super(Model, self).__init__()

    def forward(self, predictions, targets, fn=module_fn):
        return fn(predictions, targets)


batch_size = 4096
num_classes = 10
input_shape = (num_classes,)  # Output for each class
dim = 1


def get_inputs():
    return [
        torch.randn(batch_size, *input_shape),
        torch.randint(0, num_classes, (batch_size,)),
    ]


def get_init_inputs():
    return []

import torch
import torch.nn as nn

class Model(nn.Module):
    """
    A model that computes Cross Entropy Loss for multi-class classification tasks.

    Parameters:
        None
    """
    def __init__(self):
        super(Model, self).__init__()

    def forward(self, predictions, targets):
        return torch.nn.functional.cross_entropy(predictions, targets)

batch_size = 4096
num_classes = 10
input_shape = (num_classes, )  # Output for each class
dim = 1

def get_inputs():
    return [torch.randn(batch_size, *input_shape), torch.randint(0, num_classes, (batch_size,))]

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

Kernel Information

Operation Name 95_CrossEntropyLoss
Level ID 1
Task ID 95
Kernel Name modular_device_ce_loss_base
CUDA Speedup (Native) 8.973x
CUDA Speedup (Compile) 2.449x
CUDA Runtime 0.010 ms
PyTorch Runtime (Native) 0.090 ms
PyTorch Runtime (Compile) 0.024 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 95 • 95_CrossEntropyLoss)

Rank Kernel Name Runtime (ms) Speedup Native Speedup Compile
🥇 95_CrossEntropyLoss 0.01 8.97 2.45
🥇 memory_coalescing_base 0.01 8.97 2.45
🥇 block_size_experimentation_base 0.01 8.97 2.45
🥇 stride_loop_boundary_optimization_base 0.01 8.97 2.45
🥇 optimized_thread_block_mapping_base_base 0.01 8.97 2.45
🥇 optimal_blocksize_experiment_base 0.01 8.97 2.45
🥇 modular_crossentropy_base 0.01 8.97 2.45
🥇 warp_aligned_base_base 0.01 8.97 2.45
🥇 warp_divergence_minimization_base_base 0.01 8.97 2.45
🥇 modularized_device_functions_base 0.01 8.97 2.45
🥇 ce_loss_unroll_optimized_base 0.01 8.97 2.45
🥇 ce_loss_ldg_aligned_base 0.01 8.97 2.45
🥇 ce_loss_optimized_blocksize_512_base 0.01 8.97 2.45
🥇 ce_loss_grid_stride_unroll_edit_1 0.01 8.97 2.45
🥇 ce_loss_ldg_aligned_edit_1 0.01 8.97 2.45
🥇 stride_loop_optimization_base_base 0.01 8.97 2.45
🥇 ldg_aligned_access_base 0.01 8.97 2.45
🥇 modular_device_ce_loss_base 0.01 8.97 2.45
🥇 ce_loss_stride_base 0.01 8.97 2.45
🥇 atomic_optimized_crossentropy_edit_1 0.01 8.97 2.45 
#include <torch/extension.h>
#include <cuda_runtime.h>
#include <math.h>

// Device function to compute the maximum logit for numerical stability
__device__ __forceinline__ float device_max_logit(const float* logits, int num_classes) {
    float max_val = logits[0];
    for (int j = 1; j < num_classes; ++j) {
        max_val = fmaxf(max_val, logits[j]);
    }
    return max_val;
}

// Device function to compute the sum of exponentials of logits shifted by max_val
__device__ __forceinline__ float device_sum_exp(const float* logits, int num_classes, float max_val) {
    float sum_exp = 0.0f;
    for (int j = 0; j < num_classes; ++j) {
        sum_exp += expf(logits[j] - max_val);
    }
    return sum_exp;
}

// Device function to compute the cross entropy loss for a single sample
__device__ __forceinline__ float compute_cross_entropy_loss(const float* logits, int64_t target, int num_classes) {
    float max_val = device_max_logit(logits, num_classes);
    float sum_exp = device_sum_exp(logits, num_classes, max_val);
    // Compute loss: - (logit[target] - max_val - log(sum_exp))
    return -(logits[target] - max_val - logf(sum_exp));
}

// CUDA kernel using modular device functions
__global__ void cross_entropy_loss_kernel_modular(
    const float* __restrict__ logits,
    const int64_t* __restrict__ targets,
    float* __restrict__ losses,
    int batch_size,
    int num_classes
) {
    int i = blockIdx.x * blockDim.x + threadIdx.x;
    if (i < batch_size) {
        // Process one sample per thread
        const float* logits_sample = logits + i * num_classes;
        int64_t target = targets[i];
        losses[i] = compute_cross_entropy_loss(logits_sample, target, num_classes);
    }
}

// Forward function exposed to PyTorch
torch::Tensor forward(torch::Tensor predictions, torch::Tensor targets) {
    TORCH_CHECK(predictions.is_cuda(), "predictions must be a CUDA tensor");
    TORCH_CHECK(targets.is_cuda(), "targets must be a CUDA tensor");
    TORCH_CHECK(predictions.dim() == 2, "predictions must be a 2D tensor");
    TORCH_CHECK(targets.dim() == 1, "targets must be a 1D tensor");
    TORCH_CHECK(predictions.dtype() == torch::kFloat32, "predictions must be a Float32 tensor");
    TORCH_CHECK(targets.dtype() == torch::kInt64, "targets must be an Int64 tensor");

    int batch_size = predictions.size(0);
    int num_classes = predictions.size(1);
    TORCH_CHECK(targets.size(0) == batch_size, "targets must have same batch size as predictions");

    auto losses = torch::empty({batch_size}, predictions.options());
    
    // Launch parameters
    int threads = 256;
    int blocks = (batch_size + threads - 1) / threads;

    cross_entropy_loss_kernel_modular<<<blocks, threads>>>(
        predictions.data_ptr<float>(),
        targets.data_ptr<int64_t>(),
        losses.data_ptr<float>(),
        batch_size,
        num_classes
    );

    cudaError_t err = cudaGetLastError();
    TORCH_CHECK(err == cudaSuccess, "CUDA error in kernel: ", cudaGetErrorString(err));

    return losses.mean();
}

PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
    m.def("forward", &forward, "Cross Entropy Loss forward (CUDA) with modular device functions");
}
Performance Metrics
Metric Value Unit Variance Samples
Executed Ipc Active 0.568 inst/cycle 0.000 5
Executed Ipc Elapsed 0.030 inst/cycle 0.000 5
Issue Slots Busy 14.424 % 0.091 5
Issued Ipc Active 0.578 inst/cycle 0.000 5
SM Busy 14.424 % 0.091 5
Memory Throughput 52501643299.670 byte/second 473273344749671232.000 5
Mem Busy 8.818 % 0.013 5
Max Bandwidth 4.796 % 0.006 5
L1/TEX Hit Rate 92.360 % 0.000 5
L2 Hit Rate 87.666 % 0.233 5
Mem Pipes Busy 0.380 % 0.000 5
Warp Cycles Per Issued Instruction 13.124 cycle 0.037 5
Warp Cycles Per Executed Instruction 13.380 cycle 0.038 5
Avg. Active Threads Per Warp 32.000 0.000 5
Avg. Not Predicated Off Threads Per Warp 29.840 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 11.954 % 0.011 5
Achieved Active Warps Per SM 7.650 warp 0.005 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 (12.0%) 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 716657.26 μs
Device Time 11.42 μs
Self CPU Time 42.19 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::_to_copy
CPU Time 716615.07 μs
Device Time 11.42 μs
Self CPU Time 85.10 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 716401.66 μs
Device Time 0.00 μs
Self CPU Time 77.17 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 715143.36 μs
Device Time 0.00 μs
Self CPU Time 715143.36 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 738982.52 μs
Device Time 8102.84 μs
Self CPU Time 738982.52 μs
Self Device Time 8102.84 μs
CPU Memory Usage 0 B
Device Memory Usage 0 B
Self CPU Memory Usage 0 B
Self Device Memory Usage 0 B
aten::mean
CPU Time 167608.81 μs
Device Time 53701.41 μs
Self CPU Time 96784.82 μs
Self Device Time 53701.41 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::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4>)
CPU Time 0.00 μs
Device Time 53701.41 μs
Self CPU Time 0.00 μs
Self Device Time 53701.41 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 98766.87 μs
Device Time 975516.93 μs
Self CPU Time 23129.47 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 75640.65 μs
Device Time 975516.93 μs
Self CPU Time 25046.63 μs
Self Device Time 975516.93 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 975516.93 μs
Self CPU Time 0.00 μs
Self Device Time 975516.93 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 
45284 warnings generated when compiling for host.
Suppressed 45322 warnings (45275 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_95/b7_s3_modular_device_ce_loss/base/base.cu:15:70 bugprone-easily-swappable-parameters
15 | __device__ __forceinline__ float device_sum_exp(const float* logits, int num_classes, float max_val) {
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_95/b7_s3_modular_device_ce_loss/base/base.cu:15:74: note: the first parameter in the range is 'num_classes'
15 | __device__ __forceinline__ float device_sum_exp(const float* logits, int num_classes, float max_val) {
| ^~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_95/b7_s3_modular_device_ce_loss/base/base.cu:15:93: note: the last parameter in the range is 'max_val'
15 | __device__ __forceinline__ float device_sum_exp(const float* logits, int num_classes, float max_val) {
| ^~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_95/b7_s3_modular_device_ce_loss/base/base.cu:15:87: note: 'int' and 'float' may be implicitly converted
15 | __device__ __forceinline__ float device_sum_exp(const float* logits, int num_classes, float max_val) {
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_95/b7_s3_modular_device_ce_loss/base/base.cu:24:82: warning: 2 adjacent parameters of 'compute_cross_entropy_loss' of convertible types are easily swapped by mistake [bugprone-easily-swappable-parameters]
24 | __device__ __forceinline__ float compute_cross_entropy_loss(const float* logits, int64_t target, int num_classes) {
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_95/b7_s3_modular_device_ce_loss/base/base.cu:24:90: note: the first parameter in the range is 'target'
24 | __device__ __forceinline__ float compute_cross_entropy_loss(const float* logits, int64_t target, int num_classes) {
| ^~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_95/b7_s3_modular_device_ce_loss/base/base.cu:24:102: note: the last parameter in the range is 'num_classes'
24 | __device__ __forceinline__ float compute_cross_entropy_loss(const float* logits, int64_t target, int num_classes) {
| ^~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_95/b7_s3_modular_device_ce_loss/base/base.cu:24:82: note:
24 | __device__ __forceinline__ float compute_cross_entropy_loss(const float* logits, int64_t target, int num_classes) {
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_95/b7_s3_modular_device_ce_loss/base/base.cu:24:98: note: 'int64_t' and 'int' may be implicitly converted: 'int64_t' (as 'long') -> 'int', 'int' -> 'int64_t' (as 'long')
24 | __device__ __forceinline__ float compute_cross_entropy_loss(const float* logits, int64_t target, int num_classes) {
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_95/b7_s3_modular_device_ce_loss/base/base.cu:36:5: warning: 2 adjacent parameters of 'cross_entropy_loss_kernel_modular' of similar type ('int') are easily swapped by mistake [bugprone-easily-swappable-parameters]
36 | int batch_size,
| ^~~~~~~~~~~~~~~
37 | int num_classes
| ~~~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_95/b7_s3_modular_device_ce_loss/base/base.cu:36:9: note: the first parameter in the range is 'batch_size'
36 | int batch_size,
| ^~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_95/b7_s3_modular_device_ce_loss/base/base.cu:37:9: note: the last parameter in the range is 'num_classes'
37 | int num_classes
| ^~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_95/b7_s3_modular_device_ce_loss/base/base.cu:39:13: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
39 | int i = blockIdx.x * blockDim.x + threadIdx.x;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_95/b7_s3_modular_device_ce_loss/base/base.cu:42:38: 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]
42 | const float* logits_sample = logits + i * num_classes;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_95/b7_s3_modular_device_ce_loss/base/base.cu:42:47: note: make conversion explicit to silence this warning
4 | const float* logits_sample = logits + i * num_classes;
| ^~~~~~~~~~~~~~~
| static_cast<ptrdiff_t>( )
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_95/b7_s3_modular_device_ce_loss/base/base.cu:42:47: note: perform multiplication in a wider type
42 | const float* logits_sample = logits + i * num_classes;
| ^
| static_cast<ptrdiff_t>( )
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_95/b7_s3_modular_device_ce_loss/base/base.cu:49:37: warning: the parameter 'predictions' 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 forward(torch::Tensor predictions, torch::Tensor targets) {
| ^
| const &
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_95/b7_s3_modular_device_ce_loss/base/base.cu:49:64: warning: the parameter 'targets' 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 forward(torch::Tensor predictions, torch::Tensor targets) {
| ^
| const &
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_95/b7_s3_modular_device_ce_loss/base/base.cu:57:22: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
57 | int batch_size = predictions.size(0);
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_95/b7_s3_modular_device_ce_loss/base/base.cu:58:23: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
58 | int num_classes = predictions.size(1);
| ^