← Back to Leaderboard

The AI CUDA Engineer 👷

80_Gemm_Max_Subtract_GELU • minimal_sync_optimized_kernel_base_base

Level 2 • Task 80
import torch
import torch.nn as nn
import torch.nn.functional as F


def module_fn(
    x: torch.Tensor,
    max_dim: int,
    weight: torch.Tensor,
    bias: torch.Tensor,
) -> torch.Tensor:
    """
    Performs a GEMM, followed by a max operation, subtraction, and GELU activation.

    Args:
        x (torch.Tensor): Input tensor of shape (batch_size, in_features)
        max_dim (int): Dimension to perform max operation over
        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 of shape (batch_size, out_features)
    """
    x = F.linear(x, weight, bias)
    x = torch.max(x, dim=max_dim, keepdim=True).values
    x = x - x.mean(dim=1, keepdim=True)
    x = F.gelu(x)
    return x


class Model(nn.Module):
    """
    Model that performs a GEMM, followed by a max operation, subtraction, and GELU activation.
    """

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

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


batch_size = 128
in_features = 512
out_features = 1024
max_dim = 1


def get_inputs():
    return [torch.randn(batch_size, in_features), max_dim]


def get_init_inputs():
    return [in_features, out_features, max_dim]

import torch
import torch.nn as nn

class Model(nn.Module):
    """
    Model that performs a GEMM, followed by a max operation, subtraction, and GELU activation.
    """
    def __init__(self, in_features, out_features, max_dim):
        super(Model, self).__init__()
        self.gemm = nn.Linear(in_features, out_features)
        self.max_dim = max_dim

    def forward(self, x):
        """
        Args:
            x: Input tensor of shape (batch_size, in_features)

        Returns:
            Output tensor of shape (batch_size, out_features)
        """
        x = self.gemm(x)
        x = torch.max(x, dim=self.max_dim, keepdim=True).values
        x = x - x.mean(dim=1, keepdim=True)
        x = torch.nn.functional.gelu(x)
        return x

batch_size = 128
in_features = 512
out_features = 1024
max_dim = 1

def get_inputs():
    return [torch.randn(batch_size, in_features)]

def get_init_inputs():
    return [in_features, out_features, max_dim]
Download Evaluation Download PyTorch Download CUDA Download Profiles

Kernel Information

Operation Name 80_Gemm_Max_Subtract_GELU
Level ID 2
Task ID 80
Kernel Name minimal_sync_optimized_kernel_base_base
CUDA Speedup (Native) 1.182x
CUDA Speedup (Compile) 1.256x
CUDA Runtime 0.036 ms
PyTorch Runtime (Native) 0.043 ms
PyTorch Runtime (Compile) 0.045 ms
Correct True
Max Diff (vs. Reference) 0.000000
Model bedrock/anthropic.claude-3-5-sonnet-20241022-v2:0
Temperature 0.50
View Experiment Progress Details

Related Kernels (Level 2, Task 80 • 80_Gemm_Max_Subtract_GELU)

Rank Kernel Name Runtime (ms) Speedup Native Speedup Compile
🥇 warp_optimized_gemm_max_gelu_base 0.03 1.70 1.81
🥇 warp_optimized_shared_memory_edit_1 0.03 1.70 1.81
🥇 warp_aligned_gemm_base_edit_1 0.03 1.70 1.81
🥇 warp_optimized_shared_memory_base 0.03 1.70 1.81
🥇 warp_balanced_gemm_optimization_base 0.03 1.70 1.81
6 warp_aligned_gemm_base_base 0.03 1.58 1.67
7 warp_aligned_gemm_const_bias_base 0.03 1.47 1.56
8 warp_aligned_gemm_const_bias_edit_1 0.03 1.25 1.33
8 ldg_memory_optimized_kernel_base 0.03 1.25 1.33
10 indexing_optimized_fused_kernel_base 0.04 1.22 1.29
10 workload_balanced_kernel_base_base 0.04 1.22 1.29
10 shared_memory_reduction_warp_optimization_base_base 0.04 1.22 1.29
10 efficient_thread_mapping_kernel_base 0.04 1.22 1.29
14 block_tuned_fused_kernel_base_base 0.04 1.18 1.26
14 minimal_sync_optimized_kernel_base_base 0.04 1.18 1.26
16 warp_balanced_gemm_optimization_edit_1 0.04 1.15 1.22
17 warp_optimized_reduction_base_base 0.04 1.09 1.16
18 evenly_distributed_base 0.04 1.06 1.13
18 fused_gemm_max_reduce_gelu_base 0.04 1.06 1.13
20 fused_stride_loops_base 0.04 1.04 1.10 
#include <torch/extension.h>
#include <cuda.h>
#include <cuda_runtime.h>
#include <math.h>
#include <float.h>

#define GEMM_BLOCK_DIM 32
#define REDUCE_BLOCK_SIZE 512
#define WARP_SIZE 32

__device__ inline float gelu(float x) {
    return 0.5f * x * (1.0f + erf(x * 0.70710678118654752440f));
}

__global__ void minimal_sync_gemm_kernel(const float* __restrict__ x,
                                         const float* __restrict__ weight,
                                         const float* __restrict__ bias,
                                         float* __restrict__ y,
                                         int batch, int in_features, int out_features) {
    __shared__ float tile_x[GEMM_BLOCK_DIM][GEMM_BLOCK_DIM];
    __shared__ float tile_w[GEMM_BLOCK_DIM][GEMM_BLOCK_DIM];

    const int row = blockIdx.y * GEMM_BLOCK_DIM + threadIdx.y;
    const int col = blockIdx.x * GEMM_BLOCK_DIM + threadIdx.x;
    float sum = 0.0f;

    #pragma unroll 4
    for (int t = 0; t < (in_features + GEMM_BLOCK_DIM - 1) / GEMM_BLOCK_DIM; t++) {
        const int idx = t * GEMM_BLOCK_DIM + threadIdx.x;
        const int idy = t * GEMM_BLOCK_DIM + threadIdx.y;

        // Load tiles with a single synchronization point
        tile_x[threadIdx.y][threadIdx.x] = (row < batch && idx < in_features) ?
            x[row * in_features + idx] : 0.0f;
        tile_w[threadIdx.y][threadIdx.x] = (col < out_features && idy < in_features) ?
            weight[col * in_features + idy] : 0.0f;
        
        // Single sync after both loads
        __syncthreads();

        #pragma unroll
        for (int k = 0; k < GEMM_BLOCK_DIM; k++) {
            sum += tile_x[threadIdx.y][k] * tile_w[k][threadIdx.x];
        }
        
        // Single sync at end of tile processing
        __syncthreads();
    }

    if (row < batch && col < out_features) {
        y[row * out_features + col] = sum + bias[col];
    }
}

__device__ inline float warp_reduce_max(float val) {
    #pragma unroll
    for (int offset = WARP_SIZE/2; offset > 0; offset >>= 1) {
        val = fmaxf(val, __shfl_down_sync(0xffffffff, val, offset));
    }
    return val;
}

__global__ void minimal_sync_reduce_gelu_kernel(float* data,
                                                int batch,
                                                int out_features,
                                                int max_dim) {
    __shared__ float sdata[REDUCE_BLOCK_SIZE];
    const int tid = threadIdx.x;
    const int lane_id = tid % WARP_SIZE;
    const int warp_id = tid / WARP_SIZE;
    float max_val = -FLT_MAX;

    if (max_dim == 0) {
        const int col = blockIdx.x;
        
        // Grid-stride loop with minimal synchronization
        #pragma unroll 2
        for (int i = tid; i < batch; i += REDUCE_BLOCK_SIZE) {
            max_val = fmaxf(max_val, data[i * out_features + col]);
        }
    } else {
        const int row = blockIdx.x;
        
        #pragma unroll 2
        for (int j = tid; j < out_features; j += REDUCE_BLOCK_SIZE) {
            max_val = fmaxf(max_val, data[row * out_features + j]);
        }
    }

    // Warp-level reduction first (no sync needed within warp)
    max_val = warp_reduce_max(max_val);

    // Only the first thread in each warp writes to shared memory
    if (lane_id == 0) {
        sdata[warp_id] = max_val;
    }

    // Single sync before final reduction
    __syncthreads();

    // Final reduction using only the first warp
    if (warp_id == 0) {
        max_val = (tid < (REDUCE_BLOCK_SIZE / WARP_SIZE)) ? sdata[tid] : -FLT_MAX;
        max_val = warp_reduce_max(max_val);

        if (tid == 0) {
            sdata[0] = max_val;
            sdata[1] = max_val / (max_dim == 0 ? batch : out_features);
        }
    }

    // Single sync before reading results
    __syncthreads();

    const float mean = sdata[1];

    // Apply GELU without additional synchronization
    if (max_dim == 0) {
        const int col = blockIdx.x;
        #pragma unroll 2
        for (int i = tid; i < batch; i += REDUCE_BLOCK_SIZE) {
            const int idx = i * out_features + col;
            data[idx] = gelu(data[idx] - mean);
        }
    } else {
        const int row = blockIdx.x;
        #pragma unroll 2
        for (int j = tid; j < out_features; j += REDUCE_BLOCK_SIZE) {
            const int idx = row * out_features + j;
            data[idx] = gelu(data[idx] - mean);
        }
    }
}

torch::Tensor forward(torch::Tensor x, int max_dim, torch::Tensor weight, torch::Tensor bias) {
    const int batch = x.size(0);
    const int in_features = x.size(1);
    const int out_features = weight.size(0);

    auto y = torch::empty({batch, out_features}, x.options());

    dim3 blockDimGEMM(GEMM_BLOCK_DIM, GEMM_BLOCK_DIM);
    dim3 gridDimGEMM((out_features + GEMM_BLOCK_DIM - 1) / GEMM_BLOCK_DIM,
                     (batch + GEMM_BLOCK_DIM - 1) / GEMM_BLOCK_DIM);
    
    minimal_sync_gemm_kernel<<<gridDimGEMM, blockDimGEMM>>>(
        x.data_ptr<float>(),
        weight.data_ptr<float>(),
        bias.data_ptr<float>(),
        y.data_ptr<float>(),
        batch, in_features, out_features
    );

    auto max_out = torch::empty({max_dim == 0 ? 1 : batch, max_dim == 0 ? out_features : 1}, y.options());

    const int gridDim = max_dim == 0 ? out_features : batch;
    const int sharedMem = REDUCE_BLOCK_SIZE * sizeof(float);

    minimal_sync_reduce_gelu_kernel<<<gridDim, REDUCE_BLOCK_SIZE, sharedMem>>>(
        max_out.data_ptr<float>(),
        batch,
        out_features,
        max_dim
    );

    return max_out;
}

PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
    m.def("forward", &forward, "Minimal sync optimized GEMM and reduction");
}
Performance Metrics
Metric Value Unit Variance Samples
Executed Ipc Active 0.932 inst/cycle 0.001 5
Executed Ipc Elapsed 0.402 inst/cycle 0.000 5
Issue Slots Busy 23.828 % 0.880 5
Issued Ipc Active 0.952 inst/cycle 0.001 5
SM Busy 23.828 % 0.880 5
Memory Throughput 129687694143.194 byte/second 4235837368560357376.000 5
Mem Busy 10.154 % 0.022 5
Max Bandwidth 7.738 % 0.013 5
L1/TEX Hit Rate 66.670 % 0.000 5
L2 Hit Rate 75.626 % 0.074 5
Mem Pipes Busy 5.114 % 0.006 5
Warp Cycles Per Issued Instruction 16.420 cycle 0.405 5
Warp Cycles Per Executed Instruction 16.820 cycle 0.425 5
Avg. Active Threads Per Warp 31.880 0.000 5
Avg. Not Predicated Off Threads Per Warp 27.690 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 5.000 block 0.000 5
Block Limit Shared Mem 12.000 block 0.000 5
Block Limit Warps 4.000 block 0.000 5
Theoretical Active Warps per SM 64.000 warp 0.000 5
Theoretical Occupancy 100.000 % 0.000 5
Achieved Occupancy 24.612 % 0.000 5
Achieved Active Warps Per SM 15.752 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.
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 (24.6%) 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 260279.09 μs
Device Time 175.61 μs
Self CPU Time 54.33 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 260224.76 μs
Device Time 175.61 μs
Self CPU Time 109.66 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 259624.05 μs
Device Time 0.00 μs
Self CPU Time 123.49 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 259052.88 μs
Device Time 0.00 μs
Self CPU Time 259052.88 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 597703.42 μs
Device Time 37084.56 μs
Self CPU Time 597703.42 μs
Self Device Time 37084.56 μs
CPU Memory Usage 0 B
Device Memory Usage 0 B
Self CPU Memory Usage 0 B
Self Device Memory Usage 0 B
minimal_sync_gemm_kernel(float const*, float const*, float const*, float*, int, int, int)
CPU Time 0.00 μs
Device Time 230863.30 μs
Self CPU Time 0.00 μs
Self Device Time 230863.30 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 78162.56 μs
Device Time 636794.10 μs
Self CPU Time 12772.89 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 65392.21 μs
Device Time 636794.10 μs
Self CPU Time 16443.67 μs
Self Device Time 636794.10 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 636794.10 μs
Self CPU Time 0.00 μs
Self Device Time 636794.10 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 
45294 warnings generated when compiling for host.
Suppressed 45324 warnings (45277 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/20250203_optimize_b10_s4_e0_sweep/level_2/task_80/b9_s3_minimal_sync_optimized_kernel_base/base/base.cu:15:42 bugprone-easily-swappable-parameters
15 | __global__ void minimal_sync_gemm_kernel(const float* __restrict__ x,
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~
16 | const float* __restrict__ weight,
| ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
17 | const float* __restrict__ bias,
| ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_80/b9_s3_minimal_sync_optimized_kernel_base/base/base.cu:15:68: note: the first parameter in the range is 'x'
15 | __global__ void minimal_sync_gemm_kernel(const float* __restrict__ x,
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_80/b9_s3_minimal_sync_optimized_kernel_base/base/base.cu:17:68: note: the last parameter in the range is 'bias'
17 | const float* __restrict__ bias,
| ^~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_80/b9_s3_minimal_sync_optimized_kernel_base/base/base.cu:23:21: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
23 | const int row = blockIdx.y * GEMM_BLOCK_DIM + threadIdx.y;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_80/b9_s3_minimal_sync_optimized_kernel_base/base/base.cu:24:21: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
24 | const int col = blockIdx.x * GEMM_BLOCK_DIM + threadIdx.x;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_80/b9_s3_minimal_sync_optimized_kernel_base/base/base.cu:29:25: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
29 | const int idx = t * GEMM_BLOCK_DIM + threadIdx.x;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_80/b9_s3_minimal_sync_optimized_kernel_base/base/base.cu:30:25: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
30 | const int idy = t * GEMM_BLOCK_DIM + threadIdx.y;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_80/b9_s3_minimal_sync_optimized_kernel_base/base/base.cu:68:21: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
68 | const int tid = threadIdx.x;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_80/b9_s3_minimal_sync_optimized_kernel_base/base/base.cu:74:25: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
74 | const int col = blockIdx.x;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_80/b9_s3_minimal_sync_optimized_kernel_base/base/base.cu:82:25: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
82 | const int row = blockIdx.x;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_80/b9_s3_minimal_sync_optimized_kernel_base/base/base.cu:108:34: warning: narrowing conversion from 'int' to 'float' [bugprone-narrowing-conversions]
108 | sdata[1] = max_val / (max_dim == 0 ? batch : out_features);
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_80/b9_s3_minimal_sync_optimized_kernel_base/base/base.cu:119:25: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
119 | const int col = blockIdx.x;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_80/b9_s3_minimal_sync_optimized_kernel_base/base/base.cu:126:25: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
126 | const int row = blockIdx.x;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_80/b9_s3_minimal_sync_optimized_kernel_base/base/base.cu:135:37: 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]
135 | torch::Tensor forward(torch::Tensor x, int max_dim, torch::Tensor weight, torch::Tensor bias) {
| ^
| const &
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_80/b9_s3_minimal_sync_optimized_kernel_base/base/base.cu:135:67: 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]
135 | torch::Tensor forward(torch::Tensor x, int max_dim, torch::Tensor weight, torch::Tensor bias) {
| ^
| const &
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_80/b9_s3_minimal_sync_optimized_kernel_base/base/base.cu:135:89: 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]
135 | torch::Tensor forward(torch::Tensor x, int max_dim, torch::Tensor weight, torch::Tensor bias) {
| ^
| const &
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_80/b9_s3_minimal_sync_optimized_kernel_base/base/base.cu:136:23: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
136 | const int batch = x.size(0);
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_80/b9_s3_minimal_sync_optimized_kernel_base/base/base.cu:137:29: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
137 | const int in_features = x.size(1);
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_80/b9_s3_minimal_sync_optimized_kernel_base/base/base.cu:138:30: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
138 | const int out_features = weight.size(0);
| ^