opengl 4 - GLSL coherent imageBuffer access in single-stage (fragment shader), single-pass scenario -

i have single fragment shader performs processing on imagebuffer using image load/store operations. exclusively concerned following scenario:

• i have single fragment shader (no multistage (eg. vertex fragment shaders) considerations, , no multipass rendering)
• imagebuffer variables declared coherent. exclusively interested in coherent imagebuffers.

to make things clear, scenario following:

// source code of sole , unique fragment shader: coherent layout(1x32) uniform uimagebuffer data;  void main() {   ...   various calls imageload(data, ..., ...);   ...   various calls imagestore(data, ..., ...);   ... } 

i have largely looked @ spec

arb_shader_image_load_store

especially paragraph:

"using variables declared "coherent" guarantees results of stores visible shader invocations using similarly-declared variables; calling memorybarrier required ensure stores visible other operations."

note: "coherent uniform imagebuffer data;" declaration precisely "similarly-declared" variable. scenario single-pass, single-stage (fragment shader).

now, have looked @ various web sites , stumbled (like people think) upon thread on stackoverflow.com:

how glsl's "coherent" memory qualifier interpreted gpu drivers multi-pass rendering?

and more specifically, paragraph:

"your shaders cannot make assumption issuing load right after store memory stored in shader (yes really. have put memorybarrier in pull 1 off)."

my question following:

with coherent qualifier specified, in single-shader, single-pass processing scenario, can yes or no sure imagestore()'s visible invocations of fragment shader (eg. current invocation other concurrent invocations)?

by reading arb_shader_image_load_store spec, seems me that:

• the answer question yes,
• i don't need kind of memorybarrier(),
• the quoted sentence in above referenced thread in stackoverflow may indeed misleading , wrong.

thanks insight.

use memory barrier.

for 1 thing gpu may optimize , fetch whole blocks of memeory read from, , have separate memory write to.

in other words if shader modify single location once ok, if relay on neighbors values after computation applied, need memory barrier.