Thanks for sending an issue! Please fill in the following template to help quickly solve your problem.
matmul/transpose_quant_batch_mat_mul/docs/aclnnTransposeQuantBatchMatMul.md
return (sign ? -1.0f : 1.0f) * (float)mant * (1.0f / (1 << 7)) * (1.0f / (1, 127));
修复了多个C++测试文件中BF16(Brain Floating Point 16)到单精度浮点数(float)转换函数的逻辑错误。原代码在处理非规格化(denormalized)BF16数值时,使用了错误的计算公式 (1.0f / (1, 127))。该表达式误用了C++逗号运算符,导致分母计算错误,使得非规格化数的转换结果不准确,影响了相关测试的可靠性。
/assign
Thanks for sending an issue! Please fill in the following template to help quickly solve your problem.
Document Link(文档链接)
matmul/transpose_quant_batch_mat_mul/docs/aclnnTransposeQuantBatchMatMul.md
Issues Section(问题文档片段)
return (sign ? -1.0f : 1.0f) * (float)mant * (1.0f / (1 << 7)) * (1.0f / (1, 127));
Existing Issues(存在的问题)
修复了多个C++测试文件中BF16(Brain Floating Point 16)到单精度浮点数(float)转换函数的逻辑错误。原代码在处理非规格化(denormalized)BF16数值时,使用了错误的计算公式 (1.0f / (1, 127))。该表达式误用了C++逗号运算符,导致分母计算错误,使得非规格化数的转换结果不准确,影响了相关测试的可靠性。