← Back to Leaderboard

The AI CUDA Engineer 👷

98_KLDivLoss • coalesced_chunked_kl_base

Level 1 • Task 98
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 Kullback-Leibler Divergence for comparing two distributions.

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

    Returns:
        torch.Tensor: Kullback-Leibler Divergence.
    """
    return F.kl_div(torch.log(predictions), targets, reduction="batchmean")


class Model(nn.Module):
    """
    A model that computes Kullback-Leibler Divergence for comparing two distributions.

    Parameters:
        None
    """

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

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


batch_size = 128
input_shape = (4096,)
dim = 1


def get_inputs():
    return [
        torch.randn(batch_size, *input_shape).softmax(dim=-1),
        torch.randn(batch_size, *input_shape).softmax(dim=-1),
    ]


def get_init_inputs():
    return []

import torch
import torch.nn as nn

class Model(nn.Module):
    """
    A model that computes Kullback-Leibler Divergence for comparing two distributions.

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

    def forward(self, predictions, targets):
        return torch.nn.functional.kl_div(torch.log(predictions), targets, reduction='batchmean')

batch_size = 128
input_shape = (4096, )
dim = 1

def get_inputs():
    return [torch.randn(batch_size, *input_shape).softmax(dim=-1), torch.randn(batch_size, *input_shape).softmax(dim=-1)]

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

Kernel Information

Operation Name 98_KLDivLoss
Level ID 1
Task ID 98
Kernel Name coalesced_chunked_kl_base
CUDA Speedup (Native) 2.595x
CUDA Speedup (Compile) 2.931x
CUDA Runtime 0.012 ms
PyTorch Runtime (Native) 0.031 ms
PyTorch Runtime (Compile) 0.035 ms
Correct True
Max Diff (vs. Reference) 0.001000
Model openrouter-deepseek-reasoner
Temperature 1.00
View Experiment Progress Details

Related Kernels (Level 1, Task 98 • 98_KLDivLoss)

Rank Kernel Name Runtime (ms) Speedup Native Speedup Compile
🥇 optimized_kl_div_cuda_base 0.01 2.83 3.20
🥈 kl_div_sync_optimized_base 0.01 2.59 2.93
🥈 optimized_kl_div_kernel_base 0.01 2.59 2.93
🥈 kl_div_balanced_workload_base 0.01 2.59 2.93
🥈 kl_div_warp_reduce_base_base 0.01 2.59 2.93
🥈 optimized_kl_div_base 0.01 2.59 2.93
🥈 kl_div_modular_reduce_base_base 0.01 2.59 2.93
🥈 kldiv_optimized_stride_base_base_base 0.01 2.59 2.93
🥈 vectorized_aligned_kl_base 0.01 2.59 2.93
🥈 98_KLDivLoss_optimal_reduce_edit_1 0.01 2.59 2.93
🥈 strided_warp_kl_base_base 0.01 2.59 2.93
🥈 fast_strided_kl_base 0.01 2.59 2.93
🥈 coalesced_chunked_kl_base 0.01 2.59 2.93
🥈 kldiv_modular_per_thread_base_base 0.01 2.59 2.93
🥈 kldiv_unrolled_reduction_base_base 0.01 2.59 2.93
🥈 kl_div_unrolled_reduce_base_base 0.01 2.59 2.93
🥈 warp_block_vec4_opt_base 0.01 2.59 2.93
🥈 vectorized_kldiv_base_base 0.01 2.59 2.93
🥈 kl_div_even_workload_distribution_base 0.01 2.59 2.93
🥈 adaptive_kl_div_cuda_base 0.01 2.59 2.93 
#include <torch/extension.h>
#include <cuda.h>
#include <cuda_runtime.h>

constexpr int WARP_SIZE = 32;
constexpr int ELEMENTS_PER_THREAD = 4;

__global__ void coalesced_chunked_kl_kernel(
    const float* __restrict__ log_predictions,
    const float* __restrict__ targets,
    float* __restrict__ output,
    const int n) {
    
    const int tid = blockIdx.x * blockDim.x + threadIdx.x;
    const int element_stride = blockDim.x * gridDim.x * ELEMENTS_PER_THREAD;
    const int lane_id = threadIdx.x % WARP_SIZE;
    const int warp_id = threadIdx.x / WARP_SIZE;
    float thread_sum = 0.0f;

    // Process ELEMENTS_PER_THREAD consecutive elements per iteration
    for (int idx_base = tid * ELEMENTS_PER_THREAD; 
         idx_base < n; 
         idx_base += element_stride) {
        
        #pragma unroll
        for (int i = 0; i < ELEMENTS_PER_THREAD; ++i) {
            const int idx = idx_base + i;
            if (idx < n) {
                const float log_pred = __ldg(log_predictions + idx);
                const float target = __ldg(targets + idx);
                thread_sum += expf(log_pred) - target * log_pred;
            }
        }
    }

    // Warp-level reduction
    #pragma unroll
    for (int offset = WARP_SIZE/2; offset > 0; offset >>= 1) {
        thread_sum += __shfl_down_sync(0xffffffff, thread_sum, offset);
    }

    // Shared memory buffer for warp sums
    extern __shared__ float warp_sums[];
    if (lane_id == 0) {
        warp_sums[warp_id] = thread_sum;
    }
    __syncthreads();

    // First warp reduces all warp contributions
    if (warp_id == 0) {
        float sum = (lane_id < (blockDim.x / WARP_SIZE)) ? warp_sums[lane_id] : 0.0f;
        
        #pragma unroll
        for (int offset = WARP_SIZE/2; offset > 0; offset >>= 1) {
            sum += __shfl_down_sync(0xffffffff, sum, offset);
        }

        if (lane_id == 0) {
            atomicAdd(output, sum);
        }
    }
}

torch::Tensor coalesced_chunked_kl_forward(
    torch::Tensor log_predictions,
    torch::Tensor targets) {
    
    const int n = log_predictions.numel();
    auto output = torch::zeros({1}, log_predictions.options());

    const int threads = 256;
    const int elements_per_block = threads * ELEMENTS_PER_THREAD;
    const int desired_blocks = (n + elements_per_block - 1) / elements_per_block;
    const int max_blocks = 512;
    const int blocks = min(desired_blocks, max_blocks);
    const int shared_mem = (threads / WARP_SIZE) * sizeof(float);

    coalesced_chunked_kl_kernel<<<blocks, threads, shared_mem>>>(
        log_predictions.data_ptr<float>(),
        targets.data_ptr<float>(),
        output.data_ptr<float>(),
        n
    );

    return output / static_cast<float>(n);
}

PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
    m.def("forward", &coalesced_chunked_kl_forward, "Coalesced chunked KL divergence (CUDA)");
}
Performance Metrics
Metric Value Unit Variance Samples
Executed Ipc Active 0.680 inst/cycle 0.000 5
Executed Ipc Elapsed 0.356 inst/cycle 0.000 5
Issue Slots Busy 17.910 % 0.075 5
Issued Ipc Active 0.714 inst/cycle 0.000 5
SM Busy 17.910 % 0.075 5
Memory Throughput 852350940405.990 byte/second 292162823933056352256.000 5
Mem Busy 16.092 % 0.092 5
Max Bandwidth 25.646 % 0.249 5
L1/TEX Hit Rate 74.930 % 0.000 5
L2 Hit Rate 18.560 % 0.003 5
Mem Pipes Busy 12.042 % 0.053 5
Warp Cycles Per Issued Instruction 36.874 cycle 0.505 5
Warp Cycles Per Executed Instruction 38.834 cycle 0.532 5
Avg. Active Threads Per Warp 31.830 0.000 5
Avg. Not Predicated Off Threads Per Warp 29.140 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 28.000 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 41.136 % 0.015 5
Achieved Active Warps Per SM 26.326 warp 0.006 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 (41.1%) 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::zeros
CPU Time 5125309.02 μs
Device Time 242932.09 μs
Self CPU Time 187240.58 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::zero_
CPU Time 5507583.49 μs
Device Time 7596311.14 μs
Self CPU Time 259722.20 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 5247862.31 μs
Device Time 7596311.14 μs
Self CPU Time 377294.21 μs
Self Device Time 7596151.43 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 5837631.35 μs
Device Time 688944.96 μs
Self CPU Time 5837631.35 μs
Self Device Time 688944.96 μs
CPU Memory Usage 0 B
Device Memory Usage 0 B
Self CPU Memory Usage 0 B
Self Device Memory Usage 0 B
coalesced_chunked_kl_kernel(float const*, float const*, float*, int)
CPU Time 0.00 μs
Device Time 438841.25 μs
Self CPU Time 0.00 μs
Self Device Time 438841.25 μs
CPU Memory Usage 0 B
Device Memory Usage 0 B
Self CPU Memory Usage 0 B
Self Device Memory Usage 0 B
aten::div
CPU Time 998644.13 μs
Device Time 239179.66 μs
Self CPU Time 502777.11 μs
Self Device Time 239179.66 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 7353300.53 μs
Self CPU Time 0.00 μs
Self Device Time 7353300.53 μs
CPU Memory Usage 0 B
Device Memory Usage 0 B
Self CPU Memory Usage 0 B
Self Device Memory Usage 0 B
Status: Failed 
45248 warnings and 1 error generated when compiling for host.
Error while processing /home/robert_sakana_ai/llm_cuda/experiments/20250213_optimize_b10_s4_e0_sweep_rag_translate/level_1/task_98/b8_s3_coalesced_chunked_kl/base/base.cu.
Suppressed 45287 warnings (45240 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.
Found compiler error(s).
/home/robert_sakana_ai/llm_cuda/experiments/20250213_optimize_b10_s4_e0_sweep_rag_translate/level_1/task_98/b8_s3_coalesced_chunked_kl/base/base.cu:9:5 bugprone-easily-swappable-parameters
9 | const float* __restrict__ log_predictions,
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
10 | const float* __restrict__ targets,
| ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250213_optimize_b10_s4_e0_sweep_rag_translate/level_1/task_98/b8_s3_coalesced_chunked_kl/base/base.cu:9:31: note: the first parameter in the range is 'log_predictions'
9 | const float* __restrict__ log_predictions,
| ^~~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250213_optimize_b10_s4_e0_sweep_rag_translate/level_1/task_98/b8_s3_coalesced_chunked_kl/base/base.cu:10:31: note: the last parameter in the range is 'targets'
10 | const float* __restrict__ targets,
| ^~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250213_optimize_b10_s4_e0_sweep_rag_translate/level_1/task_98/b8_s3_coalesced_chunked_kl/base/base.cu:14:21: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
14 | const int tid = blockIdx.x * blockDim.x + threadIdx.x;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250213_optimize_b10_s4_e0_sweep_rag_translate/level_1/task_98/b8_s3_coalesced_chunked_kl/base/base.cu:15:32: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
15 | const int element_stride = blockDim.x * gridDim.x * ELEMENTS_PER_THREAD;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250213_optimize_b10_s4_e0_sweep_rag_translate/level_1/task_98/b8_s3_coalesced_chunked_kl/base/base.cu:16:25: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
16 | const int lane_id = threadIdx.x % WARP_SIZE;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250213_optimize_b10_s4_e0_sweep_rag_translate/level_1/task_98/b8_s3_coalesced_chunked_kl/base/base.cu:17:25: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
17 | const int warp_id = threadIdx.x / WARP_SIZE;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250213_optimize_b10_s4_e0_sweep_rag_translate/level_1/task_98/b8_s3_coalesced_chunked_kl/base/base.cu:65:19: warning: the parameter 'log_predictions' is copied for each invocation but only used as a const reference; consider making it a const reference [performance-unnecessary-value-param]
65 | torch::Tensor log_predictions,
| ^
| const &
/home/robert_sakana_ai/llm_cuda/experiments/20250213_optimize_b10_s4_e0_sweep_rag_translate/level_1/task_98/b8_s3_coalesced_chunked_kl/base/base.cu:66:19: 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]
66 | torch::Tensor targets) {
| ^
| const &
/home/robert_sakana_ai/llm_cuda/experiments/20250213_optimize_b10_s4_e0_sweep_rag_translate/level_1/task_98/b8_s3_coalesced_chunked_kl/base/base.cu:68:19: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
68 | const int n = log_predictions.numel();
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250213_optimize_b10_s4_e0_sweep_rag_translate/level_1/task_98/b8_s3_coalesced_chunked_kl/base/base.cu:75:24: error: no matching function for call to 'min' [clang-diagnostic-error]
75 | const int blocks = min(desired_blocks, max_blocks);
| ^~~
/home/common_modules/clang-tidy/20.0.0git/lib/clang/20/include/__clang_cuda_math.h:201:16: note: candidate function not viable: call to __device__ function from __host__ function
201 | __DEVICE__ int min(int __a, int __b) { return __nv_min(__a, __b); }
| ^
/usr/local/cuda/include/crt/math_functions.hpp:868:38: note: candidate function not viable: call to __device__ function from __host__ function
868 | __MATH_FUNCTIONS_DECL__ unsigned int min(const unsigned int a, const unsigned int b)
| ^
/usr/local/cuda/include/crt/math_functions.hpp:873:38: note: candidate function not viable: call to __device__ function from __host__ function
873 | __MATH_FUNCTIONS_DECL__ unsigned int min(const int a, const unsigned int b)
| ^
/usr/local/cuda/include/crt/math_functions.hpp:878:38: note: candidate function not viable: call to __device__ function from __host__ function
878 | __MATH_FUNCTIONS_DECL__ unsigned int min(const unsigned int a, const int b)
| ^
/usr/local/cuda/include/crt/math_functions.hpp:883:34: note: candidate function not viable: call to __device__ function from __host__ function
883 | __MATH_FUNCTIONS_DECL__ long int min(const long int a, const long int b)
| ^
/usr/local/cuda/include/crt/math_functions.hpp:902:43: note: candidate function not viable: call to __device__ function from __host__ function
902 | __MATH_FUNCTIONS_DECL__ unsigned long int min(const unsigned long int a, const unsigned long int b)
| ^
/usr/local/cuda/include/crt/math_functions.hpp:919:43: note: candidate function not viable: call to __device__ function from __host__ function
919 | __MATH_FUNCTIONS_DECL__ unsigned long int min(const long int a, const unsigned long int b)
| ^
/usr/local/cuda/include/crt/math_functions.hpp:936:43: note: candidate function not viable: call to __device__ function from __host__ function
936 | __MATH_FUNCTIONS_DECL__ unsigned long int min(const unsigned long int a, const long int b)
| ^
/usr/local/cuda/include/crt/math_functions.hpp:953:39: note: candidate function not viable: call to __device__ function from __host__ function
953 | __MATH_FUNCTIONS_DECL__ long long int min(const long long int a, const long long int b)
| ^
/usr/local/cuda/include/crt/math_functions.hpp:958:48: note: candidate function not viable: call to __device__ function from __host__ function
958 | __MATH_FUNCTIONS_DECL__ unsigned long long int min(const unsigned long long int a, const unsigned long long int b)
| ^
/usr/local/cuda/include/crt/math_functions.hpp:963:48: note: candidate function not viable: call to __device__ function from __host__ function
963 | __MATH_FUNCTIONS_DECL__ unsigned long long int min(const long long int a, const unsigned long long int b)
| ^
/usr/local/cuda/include/crt/math_functions.hpp:968:48: note: candidate function not viable: call to __device__ function from __host__ function
968 | __MATH_FUNCTIONS_DECL__ unsigned long long int min(const unsigned long long int a, const long long int b)
| ^
/usr/local/cuda/include/crt/math_functions.hpp:973:31: note: candidate function not viable: call to __device__ function from __host__ function
973 | __MATH_FUNCTIONS_DECL__ float min(const float a, const float b)
| ^
/usr/local/cuda/include/crt/math_functions.hpp:978:32: note: candidate function not viable: call to __device__ function from __host__ function
978 | __MATH_FUNCTIONS_DECL__ double min(const double a, const double b)
| ^
/usr/local/cuda/include/crt/math_functions.hpp:983:32: note: candidate function not viable: call to __device__ function from __host__ function
983 | __MATH_FUNCTIONS_DECL__ double min(const float a, const double b)
| ^
/usr/local/cuda/include/crt/math_functions.hpp:988:32: note: candidate function not viable: call to __device__ function from __host__ function
988 | __MATH_FUNCTIONS_DECL__ double min(const double a, const float b)
| ^