I wanted to test this out on nouveau -- I ran the program, and it created output. What's the expected output?

The expected output:
value[i].x (= gl_GlobalInvocationID.x ;)  should be equal to value[i].y  and value[i].w (= gl_WorkGroupID.x * gl_LocalGroupSizeARB.x + gl_LocalInvocationID.x;)  for all indices i according to spec. With the parameter I gave to glDispatchComputeGroupSizeARB, the divergence of the result happen after the indice i = 63 on my computer. So value[i].x = value[i].z (=gl_LocalInvocationID.x;) instead.

So on my computer, gl_GlobalInvocationID.x appear to be only equal to gl_LocalInvocationID.x notgl_WorkGroupID.x * gl_LocalGroupSizeARB.x + gl_LocalInvocationID.x or I misread the spec.

OK, I see. The relevant shader is fill_cs_450.glsl. It puts stuff into the array with values

(gl_GlobalInvocationID.x, t, gl_LocalInvocationID.x, t)

where t = gl_WorkGroupID.x * gl_LocalGroupSizeARB.x + gl_LocalInvocationID.x

Ideally the values printed should be (x, x, l, x) -- i.e. t == gl_GlobalInvocationID.x.

With nouveau this outputs a bunch of BS... sometimes gl_GlobalInvocationID.x == gl_LocalInvocationID.x, other times gl_GlobalInvocationID.x == randomish value. I suspect this is a combination of a nouveau bug and a core bug [the latter of which is affecting the radeonsi output]. Going to try to take a look.

OK yeah, this is totally broken... for a simplified shader (which I will submit to piglit), this is generating code like

COMP
PROPERTY CS_FIXED_BLOCK_WIDTH 0
PROPERTY CS_FIXED_BLOCK_HEIGHT 0
PROPERTY CS_FIXED_BLOCK_DEPTH 0
DCL SV[0], THREAD_ID
DCL SV[1], BLOCK_ID
DCL SV[2], BLOCK_SIZE
DCL BUFFER[16]
DCL TEMP[0..3], LOCAL
IMM[0] UINT32 {8, 0, 0, 0}
  0: UMAD TEMP[0].x, SV[1].xyzz, TEMP[1].xyzz, SV[0].xyzz
  1: MOV TEMP[1].x, TEMP[0].xxxx
  2: MOV TEMP[1].y, SV[0].xxxx
  3: UMAD TEMP[2].x, SV[1].xxxx, SV[2].xxxx, SV[0].xxxx
  4: UMUL TEMP[3].x, TEMP[2].xxxx, IMM[0].xxxx
  5: STORE BUFFER[16].xy, TEMP[3].xxxx, TEMP[1].xyyy
  6: END

Note the first op which tries to read the uninitialized TEMP[1].xyzz. Oops. Pretty sure this used to work, so probably got broken at some point with the various refactors.

Thank you very much for checking that.

S.C.

https://patchwork.freedesktop.org/patch/184119/

This fixes both your application as well as a small piglit test I wrote on nouveau. The issue was in the core logic. If radeonsi still doesn't work with this, then it has additional issues.

An updated patch fixes the issue in a more complete way (with multiple shaders being linked together, even non-variable local sizes were broken). This is now upstream in

commit 4d24a7cb97641cacecd371d1968f6964785822e4
Author: Ilia Mirkin <imirkin@alum.mit.edu>
Date:   Sat Oct 21 15:15:41 2017 -0400

    glsl: fix derived cs variables

Thanks for reporting and providing a repro! As the original issue was filed against radeonsi, I'm going to leave this open until someone can confirm that this does indeed fix the problem there.

values[i].x == values[i].y == values[i].w for all indices i with radeonsi.

The issue appears to be fixed, thanks guys! Closing.