From 1956a34ee554d641b8a066c4882bc3b0ae77e104 Mon Sep 17 00:00:00 2001 From: Nguyen Dac Nam Date: Fri, 21 Feb 2020 19:08:07 +0700 Subject: shader: implement LOP3 fast replace for old function ref: https://devtalk.nvidia.com/default/topic/1070081/cuda-programming-and-performance/reverse-lut-for-lop3-lut/ --- .../shader/decode/arithmetic_integer.cpp | 94 +++++++++++++--------- 1 file changed, 58 insertions(+), 36 deletions(-) diff --git a/src/video_core/shader/decode/arithmetic_integer.cpp b/src/video_core/shader/decode/arithmetic_integer.cpp index 21366869d..83d3944ac 100644 --- a/src/video_core/shader/decode/arithmetic_integer.cpp +++ b/src/video_core/shader/decode/arithmetic_integer.cpp @@ -293,44 +293,66 @@ u32 ShaderIR::DecodeArithmeticInteger(NodeBlock& bb, u32 pc) { void ShaderIR::WriteLop3Instruction(NodeBlock& bb, Register dest, Node op_a, Node op_b, Node op_c, Node imm_lut, bool sets_cc) { - constexpr u32 lop_iterations = 32; - const Node one = Immediate(1); - const Node two = Immediate(2); - - Node value; - for (u32 i = 0; i < lop_iterations; ++i) { - const Node shift_amount = Immediate(i); - - const Node a = Operation(OperationCode::ILogicalShiftRight, NO_PRECISE, op_c, shift_amount); - const Node pack_0 = Operation(OperationCode::IBitwiseAnd, NO_PRECISE, a, one); - - const Node b = Operation(OperationCode::ILogicalShiftRight, NO_PRECISE, op_b, shift_amount); - const Node c = Operation(OperationCode::IBitwiseAnd, NO_PRECISE, b, one); - const Node pack_1 = Operation(OperationCode::ILogicalShiftLeft, NO_PRECISE, c, one); - - const Node d = Operation(OperationCode::ILogicalShiftRight, NO_PRECISE, op_a, shift_amount); - const Node e = Operation(OperationCode::IBitwiseAnd, NO_PRECISE, d, one); - const Node pack_2 = Operation(OperationCode::ILogicalShiftLeft, NO_PRECISE, e, two); - - const Node pack_01 = Operation(OperationCode::IBitwiseAnd, NO_PRECISE, pack_0, pack_1); - const Node pack_012 = Operation(OperationCode::IBitwiseAnd, NO_PRECISE, pack_01, pack_2); - - const Node shifted_bit = - Operation(OperationCode::ILogicalShiftRight, NO_PRECISE, imm_lut, pack_012); - const Node bit = Operation(OperationCode::IBitwiseAnd, NO_PRECISE, shifted_bit, one); - - const Node right = - Operation(OperationCode::ILogicalShiftLeft, NO_PRECISE, bit, shift_amount); - - if (i > 0) { - value = Operation(OperationCode::IBitwiseOr, NO_PRECISE, value, right); - } else { - value = right; + const Node lop3_fast = [&](Node na, Node nb, Node nc, Node ttbl) { + Node value = Immediate(0); + ImmediateNode imm = std::get(*ttbl); + if (imm.GetValue() & 0x01) { + Node a = Operation(OperationCode::IBitwiseNot, na); + Node b = Operation(OperationCode::IBitwiseNot, nb); + Node c = Operation(OperationCode::IBitwiseNot, nc); + Node r = Operation(OperationCode::IBitwiseAnd, NO_PRECISE, a, b); + r = Operation(OperationCode::IBitwiseAnd, NO_PRECISE, r, c); + value = Operation(OperationCode::IBitwiseOr, value, r); } - } + if (imm.GetValue() & 0x02) { + Node a = Operation(OperationCode::IBitwiseNot, na); + Node b = Operation(OperationCode::IBitwiseNot, nb); + Node r = Operation(OperationCode::IBitwiseAnd, NO_PRECISE, a, b); + r = Operation(OperationCode::IBitwiseAnd, NO_PRECISE, r, nc); + value = Operation(OperationCode::IBitwiseOr, value, r); + } + if (imm.GetValue() & 0x04) { + Node a = Operation(OperationCode::IBitwiseNot, na); + Node c = Operation(OperationCode::IBitwiseNot, nc); + Node r = Operation(OperationCode::IBitwiseAnd, NO_PRECISE, a, nb); + r = Operation(OperationCode::IBitwiseAnd, NO_PRECISE, r, c); + value = Operation(OperationCode::IBitwiseOr, value, r); + } + if (imm.GetValue() & 0x08) { + Node a = Operation(OperationCode::IBitwiseNot, na); + Node r = Operation(OperationCode::IBitwiseAnd, NO_PRECISE, a, nb); + r = Operation(OperationCode::IBitwiseAnd, NO_PRECISE, r, nc); + value = Operation(OperationCode::IBitwiseOr, value, r); + } + if (imm.GetValue() & 0x10) { + Node b = Operation(OperationCode::IBitwiseNot, nb); + Node c = Operation(OperationCode::IBitwiseNot, nc); + Node r = Operation(OperationCode::IBitwiseAnd, NO_PRECISE, na, b); + r = Operation(OperationCode::IBitwiseAnd, NO_PRECISE, r, c); + value = Operation(OperationCode::IBitwiseOr, value, r); + } + if (imm.GetValue() & 0x20) { + Node b = Operation(OperationCode::IBitwiseNot, nb); + Node r = Operation(OperationCode::IBitwiseAnd, NO_PRECISE, na, b); + r = Operation(OperationCode::IBitwiseAnd, NO_PRECISE, r, nc); + value = Operation(OperationCode::IBitwiseOr, value, r); + } + if (imm.GetValue() & 0x40) { + Node c = Operation(OperationCode::IBitwiseNot, nc); + Node r = Operation(OperationCode::IBitwiseAnd, NO_PRECISE, na, nb); + r = Operation(OperationCode::IBitwiseAnd, NO_PRECISE, r, c); + value = Operation(OperationCode::IBitwiseOr, value, r); + } + if (imm.GetValue() & 0x80) { + Node r = Operation(OperationCode::IBitwiseAnd, NO_PRECISE, na, nb); + r = Operation(OperationCode::IBitwiseAnd, NO_PRECISE, r, nc); + value = Operation(OperationCode::IBitwiseOr, value, r); + } + return value; + }(op_a, op_b, op_c, imm_lut); - SetInternalFlagsFromInteger(bb, value, sets_cc); - SetRegister(bb, dest, value); + SetInternalFlagsFromInteger(bb, lop3_fast, sets_cc); + SetRegister(bb, dest, lop3_fast); } } // namespace VideoCommon::Shader -- cgit v1.2.3