seems document error for TritonGPUAttrDefs.md

[houqi]

I was reading the https://github.com/triton-lang/triton/blob/9a0a7c2ccc6e6fd5f98c06476a0ca591b65758cf/include/triton/Dialect/TritonGPU/IR/TritonGPUAttrDefs.td and found something confusing:

2. Multiple rows per phase

  #shared<{vec=1, perPhase=2, maxPhase=4, order=[1,0]}>
  [ 0,  1,  2,  3],  // phase 0 (xor with 0)
  [ 4,  5,  6,  7],
  [ 9,  8, 11, 10],  // phase 1 (xor with 1)
  [13, 12, 15, 14]

Elements of row r are xor'ed with r/2.  In other words, perPhase=2
means that pairs of 2 rows get the same swizzling.

3. Max-phase applied

  $shared<{vec=1, perPhase=1, maxPhase=2, order=[1,0]}>
  [ 0,  1,  2,  3],  // phase 0 (xor with 0)
  [ 5,  4,  7,  6],  // phase 1 (xor with 1)
  [ 8,  9, 10, 11],  // phase 0
  [13, 12, 15, 14],  // phase 1
  [16, 17, 18, 19],  // ...
  [21, 20, 23, 22],
  [24, 25, 26, 27],
  [29, 28, 31, 30]

Elements of row r are xor'ed with (r/2) % 2.  In other words, maxPhase=m has the
effect of limiting the maximum value of the xor to m-1.

4. Max-phase and per-phase

  #shared<{vec=1, perPhase=2, maxPhase=2, order=[1,0]}>
  [ 0,  1,  2,  3],  // phase 0 (xor with 0)
  [ 4,  5,  6,  7],  // phase 0
  [ 9,  8, 11, 10],  // phase 1 (xor with 1)
  [13, 12, 15, 14],  // phase 1
  [16, 17, 18, 19],  // phase 0
  [20, 21, 22, 23],  // phase 0
  [25, 24, 27, 26],  // phase 1
  [29, 28, 31, 30]]  // phase 1

Here the xor value (the "phase", I guess?) changes every perPhase rows, up to a
maximum value of maxPhase-1.  In other words, elements of row r are xor'ed with
(r/2) % 2.

3. Max-phase applied said that Elements of row r are xor'ed with (r/2) % 2 .

should it be Elements of row r are xor'ed with r % 2 ?