On 10/24/2016 04:35 PM, Ilia Mirkin wrote:
On Mon, Oct 24, 2016 at 10:29 AM, Samuel Pitoiset
<[email protected]> wrote:
Shared memory is local to CTA, thus we should only wait for
prior memory writes which are visible to other threads in
the same CTA, and not at global level. This should speedup
compute shaders which use shared memory.
Signed-off-by: Samuel Pitoiset <[email protected]>
---
src/gallium/drivers/nouveau/codegen/nv50_ir_from_tgsi.cpp | 5 ++++-
1 file changed, 4 insertions(+), 1 deletion(-)
diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_from_tgsi.cpp
b/src/gallium/drivers/nouveau/codegen/nv50_ir_from_tgsi.cpp
index b47fc49..621a468 100644
--- a/src/gallium/drivers/nouveau/codegen/nv50_ir_from_tgsi.cpp
+++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_from_tgsi.cpp
@@ -3561,12 +3561,15 @@ Converter::handleInstruction(const struct
tgsi_full_instruction *insn)
geni->subOp = tgsi::opcodeToSubOp(tgsi.getOpcode());
break;
case TGSI_OPCODE_MEMBAR:
+ {
+ uint32_t level = tgsi.getSrc(0).getValueU32(0, info);
geni = mkOp(OP_MEMBAR, TYPE_NONE, NULL);
geni->fixed = 1;
- if (tgsi.getSrc(0).getValueU32(0, info) & TGSI_MEMBAR_THREAD_GROUP)
+ if ((level & TGSI_MEMBAR_THREAD_GROUP) || level == TGSI_MEMBAR_SHARED)
Probably just level & (A | B)
No, because TGSI_MEMBAR_SHARED is also defined for memoryBarrier() and
we should perform a global synchronization for that one. We want to only
do it for memoryBarrierShared().
geni->subOp = NV50_IR_SUBOP_MEMBAR(M, CTA);
else
geni->subOp = NV50_IR_SUBOP_MEMBAR(M, GL);
+ }
break;
case TGSI_OPCODE_ATOMUADD:
case TGSI_OPCODE_ATOMXCHG:
--
2.10.1
_______________________________________________
mesa-dev mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/mesa-dev
--
-Samuel
_______________________________________________
mesa-dev mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/mesa-dev
