ac/nir/radeonsi: add ARB_shader_ballot support

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>

2 files changed, 46 insertions, 0 deletions

diff --git a/src/amd/common/ac_nir_to_llvm.c b/src/amd/common/ac_nir_to_llvm.c
index 53940e2c9e8..82f29c05880 100644
--- a/src/amd/common/ac_nir_to_llvm.c
+++ b/src/amd/common/ac_nir_to_llvm.c
@@ -4247,6 +4247,43 @@ static void visit_intrinsic(struct ac_nir_context *ctx,
 	LLVMValueRef result = NULL;
 
 	switch (instr->intrinsic) {
+	case nir_intrinsic_ballot:
+		result = ac_build_ballot(&ctx->ac, get_src(ctx, instr->src[0]));
+		break;
+	case nir_intrinsic_read_invocation:
+	case nir_intrinsic_read_first_invocation: {
+		LLVMValueRef args[2];
+
+		/* Value */
+		args[0] = get_src(ctx, instr->src[0]);
+
+		unsigned num_args;
+		const char *intr_name;
+		if (instr->intrinsic == nir_intrinsic_read_invocation) {
+			num_args = 2;
+			intr_name = "llvm.amdgcn.readlane";
+
+			/* Invocation */
+			args[1] = get_src(ctx, instr->src[1]);
+		} else {
+			num_args = 1;
+			intr_name = "llvm.amdgcn.readfirstlane";
+		}
+
+		/* We currently have no other way to prevent LLVM from lifting the icmp
+		 * calls to a dominating basic block.
+		 */
+		ac_build_optimization_barrier(&ctx->ac, &args[0]);
+
+		result = ac_build_intrinsic(&ctx->ac, intr_name,
+					    ctx->ac.i32, args, num_args,
+					    AC_FUNC_ATTR_READNONE |
+					    AC_FUNC_ATTR_CONVERGENT);
+		break;
+	}
+	case nir_intrinsic_load_subgroup_invocation:
+		result = ac_get_thread_id(&ctx->ac);
+		break;
 	case nir_intrinsic_load_work_group_id: {
 		LLVMValueRef values[3];
 
diff --git a/src/gallium/drivers/radeonsi/si_shader_nir.c b/src/gallium/drivers/radeonsi/si_shader_nir.c
index e26994c06b9..97b647434b6 100644
--- a/src/gallium/drivers/radeonsi/si_shader_nir.c
+++ b/src/gallium/drivers/radeonsi/si_shader_nir.c
@@ -505,6 +505,15 @@ si_lower_nir(struct si_shader_selector* sel)
 	};
 	NIR_PASS_V(sel->nir, nir_lower_tex, &lower_tex_options);
 
+	const nir_lower_subgroups_options subgroups_options = {
+		.subgroup_size = 64,
+		.ballot_bit_size = 32,
+		.lower_to_scalar = true,
+		.lower_subgroup_masks = true,
+		.lower_vote_trivial = false,
+	};
+	NIR_PASS_V(sel->nir, nir_lower_subgroups, &subgroups_options);
+
 	bool progress;
 	do {
 		progress = false;