submission 551534 · KernelIndex

Use it

Vendorable · source mirrored · license unknownView source →

No package. Vendor the mirrored source: 29 lines, June 9 Researcher Reciprocity License v1.0.

v_compile.py

curl "https://kernelindex.com/api/v1/implementations/kernelbot-grayscale-v2-551534?include=source"

interfacepython

Compatibility

measured onNVIDIA A100

declared hardwareNVIDIA A100

architecturessm_80

dtypesfp32

Benchmark evidence

1 measurement across 1 GPU, fastest first.

Operation / workload

Hardware

Latency

Rank

Observed

RGB to grayscalesuite of 6 cases

NVIDIA A100

8.10ms

#51 of 137

2026-03-14

Run →

Reported · How evidence levels are derived →

Source and license

sourceavailable

revision digestsha256:16dad2233a0f9489f986309efc278f2711bff2b2fd45b1f9f69ee5561077e355

license declaredunknown

license concludedunknown

authorsidanbeck

imported2026-08-15

Techniques

Extracted from the mirrored source by pattern, never inferred. Each row cites its line.

autotunereturn torch.compile(f, fullgraph=True, mode='max-autotune')

Kernel source

v_compile.py29 lines

#!POPCORN leaderboard grayscale_v2
#!POPCORN gpu A100
import torch
_W=None
_F=None

def _w(device,dtype):
    global _W
    if _W is None or _W.device!=device or _W.dtype!=dtype:
        _W=torch.tensor([0.2989,0.5870,0.1140],device=device,dtype=dtype)
    return _W

def _build(dtype):
    def f(x):
        return x[...,0]*0.2989 + x[...,1]*0.5870 + x[...,2]*0.1140
    try:
        return torch.compile(f, fullgraph=True, mode='max-autotune')
    except Exception:
        return f

@torch.inference_mode()
def custom_kernel(data):
    global _F
    rgb,out=data
    if _F is None:
        _F = _build(rgb.dtype)
    out[...] = _F(rgb)
    return out

scrolls · 29 lines total

Source code from GPU Mode and the KernelBot dataset · June 9 Researcher Reciprocity License v1.0

Changes from previous submission

Against this author's previous submission submission 551260.

  #!POPCORN leaderboard grayscale_v2
  #!POPCORN gpu A100
-  
  import torch
+ _W=None
+ _F=None
   
- _WEIGHTS = None
+ def _w(device,dtype):
+     global _W
+     if _W is None or _W.device!=device or _W.dtype!=dtype:
+         _W=torch.tensor([0.2989,0.5870,0.1140],device=device,dtype=dtype)
+     return _W
   
+ def _build(dtype):
+     def f(x):
+         return x[...,0]*0.2989 + x[...,1]*0.5870 + x[...,2]*0.1140
+     try:
+         return torch.compile(f, fullgraph=True, mode='max-autotune')
+     except Exception:
+         return f
   
- def _weights(device: torch.device, dtype: torch.dtype) -> torch.Tensor:
-     global _WEIGHTS
-     if _WEIGHTS is None or _WEIGHTS.device != device or _WEIGHTS.dtype != dtype:
-         _WEIGHTS = torch.tensor([0.2989, 0.5870, 0.1140], device=device, dtype=dtype)
-     return _WEIGHTS
-  
-  
  @torch.inference_mode()
  def custom_kernel(data):
-     rgb, out = data
-     # Reference-equivalent weighted sum over channel dimension.
-     w = _weights(rgb.device, rgb.dtype)
-     gray = torch.sum(rgb * w, dim=-1)
-     out[...] = gray
+     global _F
+     rgb,out=data
+     if _F is None:
+         _F = _build(rgb.dtype)
+     out[...] = _F(rgb)
      return out

scrolls · 42 diff lines total

Best evidence level for this revision: reported

JSON