From da706cad25892de3636ecf92bca26bdf66eedd91 Mon Sep 17 00:00:00 2001 From: ReinUsesLisp Date: Tue, 7 Apr 2020 02:16:51 -0300 Subject: shader/conversion: Implement I2I sign extension, saturation and selection Reimplements I2I adding sign extension, saturation (clamp source value to the destination), selection and destination sizes that are not 32 bits wide. It doesn't implement CC yet. --- src/video_core/engines/shader_bytecode.h | 2 +- src/video_core/shader/decode/conversion.cpp | 113 ++++++++++++++++++++++++---- 2 files changed, 101 insertions(+), 14 deletions(-) (limited to 'src') diff --git a/src/video_core/engines/shader_bytecode.h b/src/video_core/engines/shader_bytecode.h index 930b605af..59d070d7d 100644 --- a/src/video_core/engines/shader_bytecode.h +++ b/src/video_core/engines/shader_bytecode.h @@ -2166,7 +2166,7 @@ private: INST("0011011-11111---", Id::SHF_LEFT_IMM, Type::Shift, "SHF_LEFT_IMM"), INST("0100110011100---", Id::I2I_C, Type::Conversion, "I2I_C"), INST("0101110011100---", Id::I2I_R, Type::Conversion, "I2I_R"), - INST("0011101-11100---", Id::I2I_IMM, Type::Conversion, "I2I_IMM"), + INST("0011100-11100---", Id::I2I_IMM, Type::Conversion, "I2I_IMM"), INST("0100110010111---", Id::I2F_C, Type::Conversion, "I2F_C"), INST("0101110010111---", Id::I2F_R, Type::Conversion, "I2F_R"), INST("0011100-10111---", Id::I2F_IMM, Type::Conversion, "I2F_IMM"), diff --git a/src/video_core/shader/decode/conversion.cpp b/src/video_core/shader/decode/conversion.cpp index c72690b2b..b9989c88c 100644 --- a/src/video_core/shader/decode/conversion.cpp +++ b/src/video_core/shader/decode/conversion.cpp @@ -2,6 +2,10 @@ // Licensed under GPLv2 or any later version // Refer to the license.txt file included. +#include +#include +#include + #include "common/assert.h" #include "common/common_types.h" #include "video_core/engines/shader_bytecode.h" @@ -15,9 +19,49 @@ using Tegra::Shader::OpCode; using Tegra::Shader::Register; namespace { + constexpr OperationCode GetFloatSelector(u64 selector) { return selector == 0 ? OperationCode::FCastHalf0 : OperationCode::FCastHalf1; } + +constexpr u32 SizeInBits(Register::Size size) { + switch (size) { + case Register::Size::Byte: + return 8; + case Register::Size::Short: + return 16; + case Register::Size::Word: + return 32; + case Register::Size::Long: + return 64; + } + return 0; +} + +constexpr std::optional> IntegerSaturateBounds(Register::Size src_size, + Register::Size dst_size, + bool src_signed, + bool dst_signed) { + const u32 dst_bits = SizeInBits(dst_size); + if (src_size == Register::Size::Word && dst_size == Register::Size::Word) { + if (src_signed == dst_signed) { + return std::nullopt; + } + return std::make_pair(0, std::numeric_limits::max()); + } + if (dst_signed) { + // Signed destination, clamp to [-128, 127] for instance + return std::make_pair(-(1 << (dst_bits - 1)), (1 << (dst_bits - 1)) - 1); + } else { + // Unsigned destination + if (dst_bits == 32) { + // Avoid shifting by 32, that is undefined behavior + return std::make_pair(0, s32(std::numeric_limits::max())); + } + return std::make_pair(0, (1 << dst_bits) - 1); + } +} + } // Anonymous namespace u32 ShaderIR::DecodeConversion(NodeBlock& bb, u32 pc) { @@ -28,14 +72,13 @@ u32 ShaderIR::DecodeConversion(NodeBlock& bb, u32 pc) { case OpCode::Id::I2I_R: case OpCode::Id::I2I_C: case OpCode::Id::I2I_IMM: { - UNIMPLEMENTED_IF(instr.conversion.int_src.selector != 0); - UNIMPLEMENTED_IF(instr.conversion.dst_size != Register::Size::Word); - UNIMPLEMENTED_IF(instr.alu.saturate_d); + const bool src_signed = instr.conversion.is_input_signed; + const bool dst_signed = instr.conversion.is_output_signed; + const Register::Size src_size = instr.conversion.src_size; + const Register::Size dst_size = instr.conversion.dst_size; + const u32 selector = static_cast(instr.conversion.int_src.selector); - const bool input_signed = instr.conversion.is_input_signed; - const bool output_signed = instr.conversion.is_output_signed; - - Node value = [&]() { + Node value = [this, instr, opcode] { switch (opcode->get().GetId()) { case OpCode::Id::I2I_R: return GetRegister(instr.gpr20); @@ -48,16 +91,60 @@ u32 ShaderIR::DecodeConversion(NodeBlock& bb, u32 pc) { return Immediate(0); } }(); - value = ConvertIntegerSize(value, instr.conversion.src_size, input_signed); - value = GetOperandAbsNegInteger(value, instr.conversion.abs_a, instr.conversion.negate_a, - input_signed); - if (input_signed != output_signed) { - value = SignedOperation(OperationCode::ICastUnsigned, output_signed, NO_PRECISE, value); + // Ensure the source selector is valid + switch (instr.conversion.src_size) { + case Register::Size::Byte: + break; + case Register::Size::Short: + ASSERT(selector == 0 || selector == 2); + break; + default: + ASSERT(selector == 0); + break; + } + + if (src_size != Register::Size::Word || selector != 0) { + value = SignedOperation(OperationCode::IBitfieldExtract, src_signed, std::move(value), + Immediate(selector * 8), Immediate(SizeInBits(src_size))); + } + + value = GetOperandAbsNegInteger(std::move(value), instr.conversion.abs_a, + instr.conversion.negate_a, src_signed); + + if (instr.alu.saturate_d) { + if (src_signed && !dst_signed) { + Node is_negative = Operation(OperationCode::LogicalUGreaterEqual, value, + Immediate(1 << (SizeInBits(src_size) - 1))); + value = Operation(OperationCode::Select, std::move(is_negative), Immediate(0), + std::move(value)); + + // Simplify generated expressions, this can be removed without semantic impact + SetTemporary(bb, 0, std::move(value)); + value = GetTemporary(0); + + if (dst_size != Register::Size::Word) { + const Node limit = Immediate((1 << SizeInBits(dst_size)) - 1); + Node is_large = + Operation(OperationCode::LogicalUGreaterThan, std::move(value), limit); + value = Operation(OperationCode::Select, std::move(is_large), limit, + std::move(value)); + } + } else if (const std::optional bounds = + IntegerSaturateBounds(src_size, dst_size, src_signed, dst_signed)) { + value = SignedOperation(OperationCode::IMax, src_signed, std::move(value), + Immediate(bounds->first)); + value = SignedOperation(OperationCode::IMin, src_signed, std::move(value), + Immediate(bounds->second)); + } + } else if (dst_size != Register::Size::Word) { + // No saturation, we only have to mask the result + Node mask = Immediate((1 << SizeInBits(dst_size)) - 1); + value = Operation(OperationCode::UBitwiseAnd, std::move(value), std::move(mask)); } SetInternalFlagsFromInteger(bb, value, instr.generates_cc); - SetRegister(bb, instr.gpr0, value); + SetRegister(bb, instr.gpr0, std::move(value)); break; } case OpCode::Id::I2F_R: -- cgit v1.2.3