From c097c9f20c8a246ed5ba37fe37108056e6619c77 Mon Sep 17 00:00:00 2001 From: Samuel Pitoiset Date: Fri, 3 Apr 2020 11:37:56 +0200 Subject: [PATCH] aco: implement 16-bit nir_op_fsqrt/nir_op_frcp/nir_op_frsq MIME-Version: 1.0 Content-Type: text/plain; charset=utf8 Content-Transfer-Encoding: 8bit Signed-off-by: Samuel Pitoiset Reviewed-by: Daniel Schürmann Part-of: --- src/amd/compiler/aco_instruction_selection.cpp | 30 ++++++++++++++++++-------- 1 file changed, 21 insertions(+), 9 deletions(-) diff --git a/src/amd/compiler/aco_instruction_selection.cpp b/src/amd/compiler/aco_instruction_selection.cpp index 4304d02..0befc68 100644 --- a/src/amd/compiler/aco_instruction_selection.cpp +++ b/src/amd/compiler/aco_instruction_selection.cpp @@ -1731,9 +1731,13 @@ void visit_alu_instr(isel_context *ctx, nir_alu_instr *instr) break; } case nir_op_frsq: { - if (dst.size() == 1) { - emit_rsq(ctx, bld, Definition(dst), get_alu_src(ctx, instr->src[0])); - } else if (dst.size() == 2) { + Temp src = get_alu_src(ctx, instr->src[0]); + if (dst.regClass() == v2b) { + Temp tmp = bld.vop1(aco_opcode::v_rsq_f16, bld.def(v1), src); + bld.pseudo(aco_opcode::p_split_vector, Definition(dst), bld.def(v2b), tmp); + } else if (dst.regClass() == v1) { + emit_rsq(ctx, bld, Definition(dst), src); + } else if (dst.regClass() == v2) { emit_vop1_instruction(ctx, instr, aco_opcode::v_rsq_f64, dst); } else { fprintf(stderr, "Unimplemented NIR instr bit size: "); @@ -1814,9 +1818,13 @@ void visit_alu_instr(isel_context *ctx, nir_alu_instr *instr) break; } case nir_op_frcp: { - if (dst.size() == 1) { - emit_rcp(ctx, bld, Definition(dst), get_alu_src(ctx, instr->src[0])); - } else if (dst.size() == 2) { + Temp src = get_alu_src(ctx, instr->src[0]); + if (dst.regClass() == v2b) { + Temp tmp = bld.vop1(aco_opcode::v_rcp_f16, bld.def(v1), src); + bld.pseudo(aco_opcode::p_split_vector, Definition(dst), bld.def(v2b), tmp); + } else if (dst.regClass() == v1) { + emit_rcp(ctx, bld, Definition(dst), src); + } else if (dst.regClass() == v2) { emit_vop1_instruction(ctx, instr, aco_opcode::v_rcp_f64, dst); } else { fprintf(stderr, "Unimplemented NIR instr bit size: "); @@ -1840,9 +1848,13 @@ void visit_alu_instr(isel_context *ctx, nir_alu_instr *instr) break; } case nir_op_fsqrt: { - if (dst.size() == 1) { - emit_sqrt(ctx, bld, Definition(dst), get_alu_src(ctx, instr->src[0])); - } else if (dst.size() == 2) { + Temp src = get_alu_src(ctx, instr->src[0]); + if (dst.regClass() == v2b) { + Temp tmp = bld.vop1(aco_opcode::v_sqrt_f16, bld.def(v1), src); + bld.pseudo(aco_opcode::p_split_vector, Definition(dst), bld.def(v2b), tmp); + } else if (dst.regClass() == v1) { + emit_sqrt(ctx, bld, Definition(dst), src); + } else if (dst.regClass() == v2) { emit_vop1_instruction(ctx, instr, aco_opcode::v_sqrt_f64, dst); } else { fprintf(stderr, "Unimplemented NIR instr bit size: "); -- 2.7.4