nv50/ir: do not perform global membar for shared memory

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 <samuel.pitoiset@gmail.com>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>

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 b47fc497c87caac584aee52209aba7f98c115a13..91cef81aa0d0ece1ed247c5331a9f6dd07b63d5c 100644 (file)
--- 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 | TGSI_MEMBAR_SHARED)))
          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: