[Mesa-dev] [PATCH 4/4] radeonsi: postponed KILL isn't postponed anymore, but maintains WQM

[Marek Olšák]

From: Marek Olšák <marek.olsak at amd.com>

This restores performance for the drirc workaround, i.e.
KILL_IF does:
   visible = src0 >= 0;
   kill_flag &= visible; // accumulate kills
   amdgcn_kill(wqm_vote(visible)); // kill fully dead quads only

And all helper pixels are killed at the end of the shader:
   amdgcn_kill(kill_flag);
---
 src/amd/common/ac_llvm_build.c                    | 7 +++++++
 src/amd/common/ac_llvm_build.h                    | 1 +
 src/gallium/drivers/radeonsi/si_shader_tgsi_alu.c | 6 ++++++
 3 files changed, 14 insertions(+)

diff --git a/src/amd/common/ac_llvm_build.c b/src/amd/common/ac_llvm_build.c
index 0550f80..2f8f59f 100644
--- a/src/amd/common/ac_llvm_build.c
+++ b/src/amd/common/ac_llvm_build.c
@@ -1398,20 +1398,27 @@ LLVMValueRef ac_build_cvt_pkrtz_f16(struct ac_llvm_context *ctx,
 					   v2f16, args, 2,
 					   AC_FUNC_ATTR_READNONE);
 		return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
 	}
 
 	return ac_build_intrinsic(ctx, "llvm.SI.packf16", ctx->i32, args, 2,
 				  AC_FUNC_ATTR_READNONE |
 				  AC_FUNC_ATTR_LEGACY);
 }
 
+LLVMValueRef ac_build_wqm_vote(struct ac_llvm_context *ctx, LLVMValueRef i1)
+{
+	assert(HAVE_LLVM >= 0x0600);
+	return ac_build_intrinsic(ctx, "llvm.amdgcn.wqm.vote", ctx->i1,
+				  &i1, 1, AC_FUNC_ATTR_READNONE);
+}
+
 void ac_build_kill_if_false(struct ac_llvm_context *ctx, LLVMValueRef i1)
 {
 	if (HAVE_LLVM >= 0x0600) {
 		ac_build_intrinsic(ctx, "llvm.amdgcn.kill", ctx->voidt,
 				   &i1, 1, 0);
 		return;
 	}
 
 	LLVMValueRef value = LLVMBuildSelect(ctx->builder, i1,
 					     LLVMConstReal(ctx->f32, 1),
diff --git a/src/amd/common/ac_llvm_build.h b/src/amd/common/ac_llvm_build.h
index b721782..088a986 100644
--- a/src/amd/common/ac_llvm_build.h
+++ b/src/amd/common/ac_llvm_build.h
@@ -258,20 +258,21 @@ struct ac_image_args {
 	LLVMValueRef addr;
 	unsigned dmask;
 	bool unorm;
 	bool da;
 };
 
 LLVMValueRef ac_build_image_opcode(struct ac_llvm_context *ctx,
 				   struct ac_image_args *a);
 LLVMValueRef ac_build_cvt_pkrtz_f16(struct ac_llvm_context *ctx,
 				    LLVMValueRef args[2]);
+LLVMValueRef ac_build_wqm_vote(struct ac_llvm_context *ctx, LLVMValueRef i1);
 void ac_build_kill_if_false(struct ac_llvm_context *ctx, LLVMValueRef i1);
 LLVMValueRef ac_build_bfe(struct ac_llvm_context *ctx, LLVMValueRef input,
 			  LLVMValueRef offset, LLVMValueRef width,
 			  bool is_signed);
 
 void ac_get_image_intr_name(const char *base_name,
 			    LLVMTypeRef data_type,
 			    LLVMTypeRef coords_type,
 			    LLVMTypeRef rsrc_type,
 			    char *out_name, unsigned out_len);
diff --git a/src/gallium/drivers/radeonsi/si_shader_tgsi_alu.c b/src/gallium/drivers/radeonsi/si_shader_tgsi_alu.c
index 283a889..913b6c3 100644
--- a/src/gallium/drivers/radeonsi/si_shader_tgsi_alu.c
+++ b/src/gallium/drivers/radeonsi/si_shader_tgsi_alu.c
@@ -62,20 +62,26 @@ static void kil_emit(const struct lp_build_tgsi_action *action,
 	LLVMValueRef visible;
 
 	if (emit_data->inst->Instruction.Opcode == TGSI_OPCODE_KILL_IF) {
 		visible = emit_data->args[0];
 	} else {
 		assert(emit_data->inst->Instruction.Opcode == TGSI_OPCODE_KILL);
 		visible = LLVMConstInt(ctx->i1, false, 0);
 	}
 
 	if (ctx->shader->selector->force_correct_derivs_after_kill) {
+		/* LLVM 6.0 can kill immediately while maintaining WQM. */
+		if (HAVE_LLVM >= 0x0600) {
+			ac_build_kill_if_false(&ctx->ac,
+					       ac_build_wqm_vote(&ctx->ac, visible));
+		}
+
 		LLVMValueRef mask = LLVMBuildLoad(builder, ctx->postponed_kill, "");
 		mask = LLVMBuildAnd(builder, mask, visible, "");
 		LLVMBuildStore(builder, mask, ctx->postponed_kill);
 		return;
 	}
 
 	ac_build_kill_if_false(&ctx->ac, visible);
 }
 
 static void emit_icmp(const struct lp_build_tgsi_action *action,
-- 
2.7.4