Otherwise LLVM can sink them and their texture coordinate calculations
into divergent branches.
v2: simplify the conditions on which the intrinsic is marked as convergent
v3: only mark as convergent in FS and CS with derivative groups
Cc: <mesa-sta...@lists.freedesktop.org>
Signed-off-by: Rhys Perry <pendingchao...@gmail.com>
---
src/amd/common/ac_nir_to_llvm.c | 18 ++++++++++++++++++
1 file changed, 18 insertions(+)
diff --git a/src/amd/common/ac_nir_to_llvm.c b/src/amd/common/ac_nir_to_llvm.c
index 265e3b636c4..9e9fade7227 100644
--- a/src/amd/common/ac_nir_to_llvm.c
+++ b/src/amd/common/ac_nir_to_llvm.c
@@ -38,6 +38,7 @@ struct ac_nir_context {
struct ac_shader_abi *abi;
gl_shader_stage stage;
+ shader_info *info;
LLVMValueRef *ssa_defs;
@@ -1394,6 +1395,22 @@ static LLVMValueRef build_tex_intrinsic(struct
ac_nir_context *ctx,
}
args->attributes = AC_FUNC_ATTR_READNONE;
+ bool cs_derivs = ctx->stage == MESA_SHADER_COMPUTE &&
+ ctx->info->cs.derivative_group !=
DERIVATIVE_GROUP_NONE;
+ if (ctx->stage == MESA_SHADER_FRAGMENT || cs_derivs) {
+ /* Prevent texture instructions with implicit derivatives from
being
+ * sinked into branches. */
+ switch (instr->op) {
+ case nir_texop_tex:
+ case nir_texop_txb:
+ case nir_texop_lod:
+ args->attributes |= AC_FUNC_ATTR_CONVERGENT;
+ break;
+ default:
+ break;
+ }
+ }
+
return ac_build_image_opcode(&ctx->ac, args);
}
@@ -4350,6 +4367,7 @@ void ac_nir_translate(struct ac_llvm_context *ac, struct
ac_shader_abi *abi,
ctx.abi = abi;
ctx.stage = nir->info.stage;
+ ctx.info = &nir->info;
ctx.main_function =
LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx.ac.builder));
--
2.21.0
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/mesa-dev
