Imported from wenyi-li/awesome-agent-kernel-skills (kernel-designer/references/dsl-cases/triton-ascend/triton-ascend-case-reduction-sum-large/SKILL.md). Install upstream with npx skills add wenyi-li/awesome-agent-kernel-skills --skill triton-ascend-case-reduction-sum-large. Copyright stays with the author.

大规模 Sum 归约优化

任务特征

• 数据尺寸：(65536, 2048)，非reduce轴非常大，reduce轴中等

优化：reduce轴大切分 + 计算重组

# 简单
total_sum = 0.0
for n_offset in range(0, N, BLOCK_SIZE):
    row_sum += tl.sum(block_vals)

# 正确：优化
acc = tl.zeros([BLOCK_SIZE_M, BLOCK_SIZE_N], dtype=tl.float32)
for n_start in range(0, N, BLOCK_SIZE_N):
    acc += block_vals
row_sum = tl.sum(acc, axis=1)

Autotune 配置

# 1. reduce轴切分较小，UB占满 -> 700.42 us
triton.Config({'BLOCK_SIZE_M': 64, 'BLOCK_SIZE_N': 256})

# 2. reduce轴切分增至512 -> 695.08 us
triton.Config({'BLOCK_SIZE_M': 32, 'BLOCK_SIZE_N': 512})

# 3. reduce轴切分增至1024 -> 685.65 us 最优
triton.Config({'BLOCK_SIZE_M': 16, 'BLOCK_SIZE_N': 1024})

# 4. reduce轴切分增至2048 -> 686.89 us
triton.Config({'BLOCK_SIZE_M': 8, 'BLOCK_SIZE_N': 2048})

# 5. reduce轴切分较大，UB未占满 -> 743.83 us
triton.Config({'BLOCK_SIZE_M': 4, 'BLOCK_SIZE_N': 2048})

总结

在优先占满UB前提下，为reduce轴分配较大切分尺寸，减少循环次数。配置3和4性能最优，共同特征：reduce轴切分值较大且占满UB。