Bug 83962

I am looking into this.

The test renders 4 points with a geometry shader, 3 points go to stream 0 and one point goes to stream 1.

The test specs that only the 3 points assigned to stream 0 get rendered, since according to ARB_gpu_shader5, only these should be passed down the rendering pipeline for rasterization, while the point assigned to stream 1 should be removed after stream output.

The Haswell PRM says:

"Rendering Disable
Independent of SOL function enable, if rendering (i.e, 3D pipeline functions past the SOL stage) is enabled (via clearing the Rendering Disable bit), the SOL stage will pass topologies for a specific input stream (as selected by Render Stream Select) down the pipeline, with the exception of PATCHLIST_n
topologies which are never passed downstream."

This is the same language that exists, for example, for IVB, where this test is passing correctly. Unfortunately, the hardware is not doing this, and when SOL stage is not active, it ignores the value of the Render Stream Select bits from the 3DSTATE_STREAMOUT state packet, which means that it renderes primitives coming from all streams, which leads to the piglit error.

Always enabling the SOL unit on Haswell like this:

diff --git a/src/mesa/drivers/dri/i965/gen7_sol_state.c b/src/mesa/drivers/dri/i965/gen7_sol_state.c
index 7e9b285..ea8319e 100644
--- a/src/mesa/drivers/dri/i965/gen7_sol_state.c
+++ b/src/mesa/drivers/dri/i965/gen7_sol_state.c
@@ -223,6 +223,9 @@ upload_3dstate_streamout(struct brw_context *brw, bool active,
    uint32_t dw1 = 0, dw2 = 0;
    int i;
 
+   if (brw->is_haswell)
+      dw1 |= SO_FUNCTION_ENABLE;
+
    if (active) {
       int urb_entry_read_offset = 0;
       int urb_entry_read_length = (vue_map->num_slots + 1) / 2 -

fixes the piglit test, but as we know from past experiences (see 99f8ea295f and 
f976b4c1bf22) this can lead to significant performance degradation, so we probably don't want to do something like this.

I can't think of a another way to fix this though, since after SOL it will be too late to discard geometry assigned to non-zero streams, since we won't have stream information any more and we won't be able to tell which primitives originated from stream 0 and which didn't.

Ok, found a way to work around the hardware bug in Haswell and sent a patch for review to the mailing list:
http://lists.freedesktop.org/archives/mesa-dev/2015-March/078920.html

I fixed this by ignoring vertices bound to non-zero streams in the GS when TF is disabled. In this scenario these vertices are useless (they won't be rendered and they won't be captured by TF) so this should be valid.

I was tempted to enable the patch for all platforms (gen >= 6), at least that seemed to work fine for SNB, IVB and HSW according to piglit, but since the patch might introduce a minor change in behavior when a shader emits more vertices than the maximum it declares I decided to keep it specific to Haswell in the end.

Hi all,

Iago's patch has been merged for quite some time now, so I'm included to set this as resolved. Feel free to reopen/verify as appropriate.

Christophe, 
Does this issue still happen on your machines?