From e77a4a9b1f66de383043df95aada40fd5a004913 Mon Sep 17 00:00:00 2001 From: Francisco Jerez Date: Tue, 4 Aug 2015 19:08:45 +0300 Subject: [PATCH] i965/fs: Implement nir_op_imul/umul_high in terms of MULH. And get rid of another no16() call. Reviewed-by: Matt Turner --- src/mesa/drivers/dri/i965/brw_fs_nir.cpp | 33 ++---------------------- 1 file changed, 2 insertions(+), 31 deletions(-) diff --git a/src/mesa/drivers/dri/i965/brw_fs_nir.cpp b/src/mesa/drivers/dri/i965/brw_fs_nir.cpp index e922a85573c..ee964a0f45d 100644 --- a/src/mesa/drivers/dri/i965/brw_fs_nir.cpp +++ b/src/mesa/drivers/dri/i965/brw_fs_nir.cpp @@ -768,38 +768,9 @@ fs_visitor::nir_emit_alu(const fs_builder &bld, nir_alu_instr *instr) break; case nir_op_imul_high: - case nir_op_umul_high: { - if (devinfo->gen >= 7) - no16("SIMD16 explicit accumulator operands unsupported\n"); - - struct brw_reg acc = retype(brw_acc_reg(dispatch_width), result.type); - - fs_inst *mul = bld.MUL(acc, op[0], op[1]); - bld.MACH(result, op[0], op[1]); - - /* Until Gen8, integer multiplies read 32-bits from one source, and - * 16-bits from the other, and relying on the MACH instruction to - * generate the high bits of the result. - * - * On Gen8, the multiply instruction does a full 32x32-bit multiply, - * but in order to do a 64x64-bit multiply we have to simulate the - * previous behavior and then use a MACH instruction. - * - * FINISHME: Don't use source modifiers on src1. - */ - if (devinfo->gen >= 8) { - assert(mul->src[1].type == BRW_REGISTER_TYPE_D || - mul->src[1].type == BRW_REGISTER_TYPE_UD); - if (mul->src[1].type == BRW_REGISTER_TYPE_D) { - mul->src[1].type = BRW_REGISTER_TYPE_W; - mul->src[1].stride = 2; - } else { - mul->src[1].type = BRW_REGISTER_TYPE_UW; - mul->src[1].stride = 2; - } - } + case nir_op_umul_high: + bld.emit(SHADER_OPCODE_MULH, result, op[0], op[1]); break; - } case nir_op_idiv: case nir_op_udiv: -- 2.30.2