描述

• 补充tla.mmad中use_hf32参数，使用示例：

# 注意hf32模式需在A, B均为f32时启用
# 可支持的 hf32 模式：
# ``tla.params.Hf32Mode.HF32_NEAREST_ZERO``
# ``tla.params.Hf32Mode.HF32_NEAREST_EVEN``
if tla.const_expr(hf32_mode != tla.params.Hf32Mode.HF32_DISABLE and dtype_a == tla.Float32 and dtype_b == tla.Float32):
   tla.mmad(l0_c, l0_a, l0_b, init_c=init_c, unit_flag=unit_flag, hf32_mode=hf32_mode)
else:
   tla.mmad(l0_c, l0_a, l0_b, init_c=init_c, unit_flag=unit_flag)

• 补充hf32精度测试：

1. to_hf32模拟尾数截断；
2. 精度比较使用混合容差，阈值参考来自：生态算子开源精度标准

关联的Issue

原因

测试

• lit, IR测试

$ lit -sv csrc/mlir/build/tests/lit --filter=mmad-hf32-mode-verify
$ llvm-lit -sv csrc/mlir/build/tests/lit/tla-compile/mmad-hf32.mlir
$ python -m pytest -q tests/test_frontend_lowering.py  -k hf32

• end_to_end 测试

1. RTZ(Round to Zero)

(catlass-pacr) xx@xxxx:~/catlass/python/tla_dsl$ python examples/end_to_end/basic_mmad/basic_matmul.py --device 1 --dtype-a f32 --dtype-b f32 --dtype-c f32
--- mnk=(256,512,1024) layout=row/row dtype=f32/f32/f32 ---
tensor([[ 421.2704,   70.1035,   81.0321,  ..., -161.6765, -442.6078,
         -425.0583],
        ...,
        [-113.9693, -272.3091,  294.2283,  ...,  180.0958, -498.2363,
          184.6163]])
tensor([[ 421.2694,   70.1012,   81.0294,  ..., -161.6766, -442.6080,
         -425.0576],
        ...,
        [-113.9658, -272.3072,  294.2266,  ...,  180.0937, -498.2347,
          184.6140]])
passed=True cache_key=xxx
kernel.o=/xxx/kernel.o

2. RTE(Round to Even)

(catlass-pacr) xx@xxxx:~/catlass/python/tla_dsl$ python examples/end_to_end/basic_mmad/basic_matmul.py  --device 1 --dtype-a f32 --dtype-b f32 --dtype-c f32
--- mnk=(256,512,1024) layout=row/row dtype=f32/f32/f32 ---
tensor([[ 421.2699,   70.1012,   81.0291,  ..., -161.6766, -442.6080,
         -425.0576],
        ...,
        [-113.9693, -272.3091,  294.2283,  ...,  180.0958, -498.2363,
          184.6163]])
tensor([[ 421.2698,   70.1012,   81.0291,  ..., -161.6766, -442.6080,
         -425.0576],
        ...,
        [-113.9693, -272.3092,  294.2284,  ...,  180.0957, -498.2363,
          184.6163]])
passed=True cache_key=xxx
kernel.o=/xxx/kernel.o

• 性能测试：56.577999us 符合预期(M/N/K=3200/4096/256)

passed=True cache_key=ee468b3cae6d14a4
kernel.o=/home/npu_user0/.cache/catlass/ee468b3cae6d14a4/kernel.o
...

2026-08-14 22:21:30 [INFO]  Operator Basic Information:

        Op Name: basic_mmad_kernel
        Op Type: cube
        Task Duration(us): 56.577999
        Block Dim: 28
        Mix Block Dim: 
        Device Id: 6
        Pid: 3521978
        Current Freq: 1650
        Rated Freq: 1650

文档更新

类型标签

• Bug修复
• 新特性
• 性能优化
• 文档更新
• 其他，请描述：