Skip to content
KernelIndex
Search⌘K
Sign in

submission 68762

koshibat · python · License unknown
Kernel source · 42 lines ↓

Use it

Vendorable · source mirrored · license unknownView source →

No package. Vendor the mirrored source: 42 lines, June 9 Researcher Reciprocity License v1.0.

submission_8192_32.py
curl "https://kernelindex.com/api/v1/implementations/kernelbot-grayscale-v2-68762?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
2.88ms
#37 of 137
2025-11-09
Run →

Reported · How evidence levels are derived →

Source and license

sourceavailable
revision digestsha256:bd8e26fd58225afbc7de87dbac88bbb18a9b5ac34397d9078514b2a574c423c4
license declaredunknown
license concludedunknown
authorskoshibat
imported2026-08-15

Techniques

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

num-warps = 8num_warps=8,
stages = 0num_stages=0,

Kernel source

submission_8192_32.py42 lines
import torch
import triton
import triton.language as tl

@triton.jit
def rgb2gray_fast(
    inp_ptr, out_ptr,
    BLOCK_SIZE: tl.constexpr,
):
    pid  = tl.program_id(0)
    i    = pid * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)   # (B,)
    base = i + (i << 1)

    r = tl.load(inp_ptr + base + 0, cache_modifier=".ca")
    g = tl.load(inp_ptr + base + 1, cache_modifier=".ca")
    b = tl.load(inp_ptr + base + 2, cache_modifier=".ca")
    gray = tl.fma(g, 0.5870, tl.fma(r, 0.2989, b * 0.1140))
    # gray = tl.fma(r, 0.2989, tl.fma(b, 0.1140, g * 0.5870))
    # gray = tl.fma(b, 0.1140, tl.fma(g, 0.5870, r * 0.2989))
    tl.store(out_ptr + i, gray)

def custom_kernel(data: torch.Tensor):
    """
    Input:  (H, W, 3) contiguous, dtype=float32  (元コードと同じ前提)
    Output: (H, W)   dtype=float32
    """
    input, output = data

    H, W, _ = input.shape
    n_pixels = H * W

    BLOCK = 4096
    grid = (triton.cdiv(n_pixels, BLOCK), )

    rgb2gray_fast[grid](
        input, output,
        BLOCK_SIZE=BLOCK,
        num_warps=8,
        num_stages=0,
    )
    return output
scrolls · 42 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 68709.

  ⋯ 13 unchanged lines
      r = tl.load(inp_ptr + base + 0, cache_modifier=".ca")
      g = tl.load(inp_ptr + base + 1, cache_modifier=".ca")
      b = tl.load(inp_ptr + base + 2, cache_modifier=".ca")
-     # gray = tl.fma(g, 0.5870, tl.fma(r, 0.2989, b * 0.1140))
-     gray = tl.fma(r, 0.2989, tl.fma(b, 0.1140, g * 0.5870))
+     gray = tl.fma(g, 0.5870, tl.fma(r, 0.2989, b * 0.1140))
+     # gray = tl.fma(r, 0.2989, tl.fma(b, 0.1140, g * 0.5870))
      # gray = tl.fma(b, 0.1140, tl.fma(g, 0.5870, r * 0.2989))
      tl.store(out_ptr + i, gray)
   
  ⋯ 7 unchanged lines
      H, W, _ = input.shape
      n_pixels = H * W
   
-     BLOCK = 8192
+     BLOCK = 4096
      grid = (triton.cdiv(n_pixels, BLOCK), )
   
      rgb2gray_fast[grid](
          input, output,
          BLOCK_SIZE=BLOCK,
-         num_warps=32,
-         num_stages=2,
+         num_warps=8,
+         num_stages=0,
      )
      return output
scrolls · 28 diff lines total

Best evidence level for this revision: reported

JSON