← Back to Leaderboard

The AI CUDA Engineer 👷

95_CrossEntropyLoss • optimal_blocksize_experiment_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 optimal_blocksize_experiment_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.h>
#include <cuda_runtime.h>
#include <math.h>

// Kernel using grid-stride loop, allowing dynamic block size selection
__global__ void cross_entropy_loss_kernel_experiment(
    const float* __restrict__ logits,
    const int64_t* __restrict__ targets,
    float* __restrict__ losses,
    int batch_size,
    int num_classes
) {
    int tid = blockIdx.x * blockDim.x + threadIdx.x;
    int stride = gridDim.x * blockDim.x;
    for (int i = tid; i < batch_size; i += stride) {
        const float* logits_i = logits + i * num_classes;
        int64_t target = targets[i];

        // Compute maximum logit for numerical stability
        float max_logit = logits_i[0];
        for (int j = 1; j < num_classes; ++j) {
            max_logit = fmaxf(max_logit, logits_i[j]);
        }

        // Compute sum of exponentials
        float sum_exp = 0.0f;
        for (int j = 0; j < num_classes; ++j) {
            sum_exp += expf(logits_i[j] - max_logit);
        }

        // Compute cross entropy loss for the sample
        losses[i] = -(logits_i[target] - max_logit - logf(sum_exp));
    }
}

// Forward function that selects an optimal block size from candidate values
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);
    auto losses = torch::empty({batch_size}, predictions.options());

    // Heuristic for block size: experiment with different candidates
    int block_size;
    if (batch_size < 32)
        block_size = 32;
    else if (batch_size < 64)
        block_size = 64;
    else if (batch_size < 128)
        block_size = 128;
    else if (batch_size < 256)
        block_size = 256;
    else if (batch_size < 512)
        block_size = 512;
    else
        block_size = 256; // Default value for large batch sizes

    int blocks = (batch_size + block_size - 1) / block_size;

    cross_entropy_loss_kernel_experiment<<<blocks, block_size>>>(
        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: ", cudaGetErrorString(err));

    return losses.mean();
}

PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
    m.def("forward", &forward, "Cross Entropy Loss forward (CUDA) with optimal blocksize experimentation");
}
Performance Metrics
Metric Value Unit Variance Samples
Executed Ipc Active 0.514 inst/cycle 0.000 5
Executed Ipc Elapsed 0.030 inst/cycle 0.000 5
Issue Slots Busy 13.216 % 0.232 5
Issued Ipc Active 0.530 inst/cycle 0.000 5
SM Busy 13.216 % 0.232 5
Memory Throughput 52562652450.776 byte/second 971047119708930816.000 5
Mem Busy 8.616 % 0.032 5
Max Bandwidth 4.752 % 0.010 5
L1/TEX Hit Rate 92.360 % 0.000 5
L2 Hit Rate 89.644 % 0.272 5
Mem Pipes Busy 0.384 % 0.000 5
Warp Cycles Per Issued Instruction 14.068 cycle 0.065 5
Warp Cycles Per Executed Instruction 14.460 cycle 0.067 5
Avg. Active Threads Per Warp 32.000 0.000 5
Avg. Not Predicated Off Threads Per Warp 28.980 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 11.936 % 0.006 5
Achieved Active Warps Per SM 7.638 warp 0.002 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 500691.65 μs
Device Time 11.52 μs
Self CPU Time 45.35 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 500646.30 μs
Device Time 11.52 μs
Self CPU Time 92.08 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 500415.75 μs
Device Time 0.00 μs
Self CPU Time 91.22 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 500110.65 μs
Device Time 0.00 μs
Self CPU Time 500110.65 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 732223.55 μs
Device Time 137973.61 μs
Self CPU Time 732223.55 μs
Self Device Time 137973.61 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 168282.09 μs
Device Time 51648.66 μs
Self CPU Time 94604.00 μs
Self Device Time 51648.66 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 97624.86 μs
Device Time 977812.64 μs
Self CPU Time 23399.63 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 74226.74 μs
Device Time 977812.64 μs
Self CPU Time 24708.47 μs
Self Device Time 977812.64 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 977812.64 μs
Self CPU Time 0.00 μs
Self Device Time 977812.64 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_s2_optimal_blocksize_experiment/base/base.cu:11:5 bugprone-easily-swappable-parameters
11 | int batch_size,
| ^~~~~~~~~~~~~~~
12 | int num_classes
| ~~~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_95/b7_s2_optimal_blocksize_experiment/base/base.cu:11:9: note: the first parameter in the range is 'batch_size'
11 | int batch_size,
| ^~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_95/b7_s2_optimal_blocksize_experiment/base/base.cu:12:9: note: the last parameter in the range is 'num_classes'
12 | int num_classes
| ^~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_95/b7_s2_optimal_blocksize_experiment/base/base.cu:14:15: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
14 | int tid = blockIdx.x * blockDim.x + threadIdx.x;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_95/b7_s2_optimal_blocksize_experiment/base/base.cu:15:18: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
15 | int stride = gridDim.x * blockDim.x;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_95/b7_s2_optimal_blocksize_experiment/base/base.cu:17:33: 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]
17 | const float* logits_i = logits + i * num_classes;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_95/b7_s2_optimal_blocksize_experiment/base/base.cu:17:42: note: make conversion explicit to silence this warning
5 |
6 | // Kernel using grid-stride loop, allowing dynamic block size selection
7 | __global__ void cross_entropy_loss_kernel_experiment(
8 | const float* __restrict__ logits,
9 | const int64_t* __restrict__ targets,
10 | float* __restrict__ losses,
11 | int batch_size,
12 | int num_classes
13 | ) {
14 | int tid = blockIdx.x * blockDim.x + threadIdx.x;
15 | int stride = gridDim.x * blockDim.x;
16 | for (int i = tid; i < batch_size; i += stride) {
17 | const float* logits_i = 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_s2_optimal_blocksize_experiment/base/base.cu:17:42: note: perform multiplication in a wider type
17 | const float* logits_i = 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_s2_optimal_blocksize_experiment/base/base.cu:38:37: 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]
38 | 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_s2_optimal_blocksize_experiment/base/base.cu:38:64: 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]
38 | 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_s2_optimal_blocksize_experiment/base/base.cu:46:22: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
46 | int batch_size = predictions.size(0);
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_95/b7_s2_optimal_blocksize_experiment/base/base.cu:47:23: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
47 | int num_classes = predictions.size(1);
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_95/b7_s2_optimal_blocksize_experiment/base/base.cu:59:9: warning: repeated branch body in conditional chain [bugprone-branch-clone]
59 | block_size = 256;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_95/b7_s2_optimal_blocksize_experiment/base/base.cu:59:25: note: end of the original
59 | block_size = 256;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_95/b7_s2_optimal_blocksize_experiment/base/base.cu:63:9: note: clone 1 starts here
63 | block_size = 256; // Default value for large batch sizes
| ^