← Back to Leaderboard

The AI CUDA Engineer 👷

98_KLDivLoss • kl_div_modular_reduce_base_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 kl_div_modular_reduce_base_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) 0.004000
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>

namespace {

__device__ float4 load_vector4(const float* ptr, int idx) {
    return __ldg(reinterpret_cast<const float4*>(ptr) + idx);
}

__device__ float process_vector_element(const float4& logp, const float4& targ, int component) {
    const float lp = (&logp.x)[component];
    const float tt = (&targ.x)[component];
    return expf(lp) - tt * lp;
}

__device__ float warp_reduce_sum(float val) {
    for (int offset = 16; offset > 0; offset >>= 1)
        val += __shfl_down_sync(0xffffffff, val, offset);
    return val;
}

__device__ float process_scalar_element(const float* logp, const float* targ, int idx) {
    float lp = __ldg(logp + idx);
    float tt = __ldg(targ + idx);
    return expf(lp) - tt * lp;
}

} // anonymous namespace

__global__ void kl_div_kernel(
    const float* __restrict__ log_predictions,
    const float* __restrict__ targets,
    float* __restrict__ output,
    const int n) {
    
    const int tid = threadIdx.x;
    const int warp_id = tid / 32;
    const int lane = tid % 32;
    const int global_idx = blockIdx.x * blockDim.x + tid;
    
    extern __shared__ float warp_sums[];
    
    float sum = 0.0f;

    // Vectorized processing
    const int n4 = n / 4;
    int vec_idx = global_idx;
    while (vec_idx < n4) {
        float4 logp = load_vector4(log_predictions, vec_idx);
        float4 targ = load_vector4(targets, vec_idx);
        
        for (int i = 0; i < 4; ++i)
            sum += process_vector_element(logp, targ, i);
        
        vec_idx += gridDim.x * blockDim.x;
    }

    // Scalar processing
    int scalar_idx = n4 * 4 + global_idx;
    while (scalar_idx < n) {
        sum += process_scalar_element(log_predictions, targets, scalar_idx);
        scalar_idx += gridDim.x * blockDim.x;
    }

    // Warp-level reduction
    sum = warp_reduce_sum(sum);
    
    // Store warp sums
    if (lane == 0)
        warp_sums[warp_id] = sum;
    __syncthreads();

    // Block-level reduction
    if (warp_id == 0) {
        float block_sum = lane < (blockDim.x / 32) ? warp_sums[lane] : 0.0f;
        block_sum = warp_reduce_sum(block_sum);
        
        if (lane == 0)
            atomicAdd(output, block_sum);
    }
}

torch::Tensor kl_div_cuda_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 warps_per_block = threads / 32;
    const int blocks = min((n + threads*4 - 1) / (threads*4), 1024);
    const int shared_mem = warps_per_block * sizeof(float);
    
    kl_div_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", &kl_div_cuda_forward, "KL divergence forward (CUDA Modular Reduction)");
}
Performance Metrics
Metric Value Unit Variance Samples
Executed Ipc Active 0.560 inst/cycle 0.003 5
Executed Ipc Elapsed 0.306 inst/cycle 0.000 5
Issue Slots Busy 14.972 % 1.630 5
Issued Ipc Active 0.600 inst/cycle 0.003 5
SM Busy 14.972 % 1.630 5
Memory Throughput 890218798075.450 byte/second 109571912319626444800.000 5
Mem Busy 15.486 % 0.044 5
Max Bandwidth 26.698 % 0.103 5
L1/TEX Hit Rate 0.000 % 0.000 5
L2 Hit Rate 18.550 % 0.006 5
Mem Pipes Busy 7.966 % 0.010 5
Warp Cycles Per Issued Instruction 42.900 cycle 0.024 5
Warp Cycles Per Executed Instruction 45.858 cycle 0.027 5
Avg. Active Threads Per Warp 31.790 0.000 5
Avg. Not Predicated Off Threads Per Warp 28.140 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 41.592 % 0.119 5
Achieved Active Warps Per SM 26.618 warp 0.048 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 (40.9%) 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 5073336.20 μs
Device Time 231260.59 μs
Self CPU Time 132726.75 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 5475434.35 μs
Device Time 7552863.72 μs
Self CPU Time 340309.53 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 5135126.29 μs
Device Time 7552863.72 μs
Self CPU Time 390169.53 μs
Self Device Time 7552863.72 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 5472854.28 μs
Device Time 2128.46 μs
Self CPU Time 5472854.28 μs
Self Device Time 2128.46 μs
CPU Memory Usage 0 B
Device Memory Usage 0 B
Self CPU Memory Usage 0 B
Self Device Memory Usage 0 B
kl_div_kernel(float const*, float const*, float*, int)
CPU Time 0.00 μs
Device Time 433247.55 μs
Self CPU Time 0.00 μs
Self Device Time 433247.55 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 886257.60 μs
Device Time 237564.42 μs
Self CPU Time 513390.78 μs
Self Device Time 237564.42 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 263128.40 μs
Device Time 682666.64 μs
Self CPU Time 263128.40 μs
Self Device Time 682666.64 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 7321603.13 μs
Self CPU Time 0.00 μs
Self Device Time 7321603.13 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_random/level_1/task_98/b7_s2_kl_div_modular_reduce_base/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_random/level_1/task_98/b7_s2_kl_div_modular_reduce_base/base/base.cu:23:41 bugprone-easily-swappable-parameters
23 | __device__ float process_scalar_element(const float* logp, const float* targ, int idx) {
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250213_optimize_b10_s4_e0_sweep_rag_random/level_1/task_98/b7_s2_kl_div_modular_reduce_base/base/base.cu:23:54: note: the first parameter in the range is 'logp'
23 | __device__ float process_scalar_element(const float* logp, const float* targ, int idx) {
| ^~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250213_optimize_b10_s4_e0_sweep_rag_random/level_1/task_98/b7_s2_kl_div_modular_reduce_base/base/base.cu:23:73: note: the last parameter in the range is 'targ'
23 | __device__ float process_scalar_element(const float* logp, const float* targ, int idx) {
| ^~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250213_optimize_b10_s4_e0_sweep_rag_random/level_1/task_98/b7_s2_kl_div_modular_reduce_base/base/base.cu:37:21: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
37 | const int tid = threadIdx.x;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250213_optimize_b10_s4_e0_sweep_rag_random/level_1/task_98/b7_s2_kl_div_modular_reduce_base/base/base.cu:40:28: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
40 | const int global_idx = blockIdx.x * blockDim.x + tid;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250213_optimize_b10_s4_e0_sweep_rag_random/level_1/task_98/b7_s2_kl_div_modular_reduce_base/base/base.cu:56:20: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
56 | vec_idx += gridDim.x * blockDim.x;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250213_optimize_b10_s4_e0_sweep_rag_random/level_1/task_98/b7_s2_kl_div_modular_reduce_base/base/base.cu:63:23: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
63 | scalar_idx += gridDim.x * blockDim.x;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250213_optimize_b10_s4_e0_sweep_rag_random/level_1/task_98/b7_s2_kl_div_modular_reduce_base/base/base.cu:85:19: 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]
85 | torch::Tensor log_predictions,
| ^
| const &
/home/robert_sakana_ai/llm_cuda/experiments/20250213_optimize_b10_s4_e0_sweep_rag_random/level_1/task_98/b7_s2_kl_div_modular_reduce_base/base/base.cu:86:19: 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]
86 | torch::Tensor targets) {
| ^
| const &
/home/robert_sakana_ai/llm_cuda/experiments/20250213_optimize_b10_s4_e0_sweep_rag_random/level_1/task_98/b7_s2_kl_div_modular_reduce_base/base/base.cu:88:19: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
88 | const int n = log_predictions.numel();
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250213_optimize_b10_s4_e0_sweep_rag_random/level_1/task_98/b7_s2_kl_div_modular_reduce_base/base/base.cu:93:24: error: no matching function for call to 'min' [clang-diagnostic-error]
93 | const int blocks = min((n + threads*4 - 1) / (threads*4), 1024);
| ^~~
/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)
| ^