数值检查通过
## 优化后的 HLO
  %lhs.1 = bf16[2048,2048]{1,0:T(8,128)(2,1)} parameter(0)
  %copy-start = (bf16[2048,2048]{1,0:T(8,128)(2,1)S(3)}, bf16[2048,2048]{1,0:T(8,128)(2,1)}, u32[]{:S(2)}) copy-start(%lhs.1), cross_program_prefetch_index=0
  %rhs.1 = bf16[2048,2048]{1,0:T(8,128)(2,1)} parameter(1)
  %copy-done = bf16[2048,2048]{1,0:T(8,128)(2,1)S(3)} copy-done(%copy-start)
  ROOT %fusion = f32[2048,2048]{1,0:T(8,128)} fusion(%copy-done, %rhs.1), kind=kOutput, calls=%fused_computation
## 清单（两个 TensorCore 执行同一份程序，按 core 编号分工）
  cld.8x128 512，dma.strided 2，vdwg.128x128.f16 4，vld.8x128 4487，vmatmul.packed.8x128.f16 1988，vmatmul.packed.dwg.8x128.f16 60，vmatpush.packed.8x128.f16 512，vpop.8x128 4608，vst.8x128 4231
bundle    0: prelude-prefix-start
bundle    9: routine:wait-for-active-dma
bundle   16: routine:program-exit
bundle   19: routine:allocation-table-find
bundle   28: routine:allocation-table-get-or-add
bundle   54: routine:allocation-table-cut-and-add
bundle   71: routine:branch-impl
bundle  122: routine:branch
bundle  128: trampoline:program-start
bundle  549: entry bundle: %fusion = fusion(%copy-done, %rhs.1)
bundle 8863: exit bundle: %fusion = fusion(%copy-done, %rhs.1)
bundle 9000: 清单结束（共 9000 个 bundle）
## 设备上的周期数（LCC），16 对新输入的中位数
  module：TensorCore 0 115504 个周期，TensorCore 1 84208 个周期
  copy-start：TensorCore 0 17 个周期，TensorCore 1 8 个周期
  copy-done：TensorCore 0 9210 个周期，TensorCore 1 9240 个周期
  fusion：TensorCore 0 74787 个周期，TensorCore 1 74737 个周期