diff options
Diffstat (limited to 'src/video_core/shader')
| -rw-r--r-- | src/video_core/shader/decode/memory.cpp | 70 | ||||
| -rw-r--r-- | src/video_core/shader/shader_ir.h | 9 |
2 files changed, 61 insertions, 18 deletions
diff --git a/src/video_core/shader/decode/memory.cpp b/src/video_core/shader/decode/memory.cpp index 9b579bde1..e006f8138 100644 --- a/src/video_core/shader/decode/memory.cpp +++ b/src/video_core/shader/decode/memory.cpp @@ -104,19 +104,42 @@ u32 ShaderIR::DecodeMemory(NodeBlock& bb, u32 pc) { } case OpCode::Id::LD_L: { UNIMPLEMENTED_IF_MSG(instr.ld_l.unknown == 1, "LD_L Unhandled mode: {}", - static_cast<unsigned>(instr.ld_l.unknown.Value())); - - const Node index = Operation(OperationCode::IAdd, GetRegister(instr.gpr8), - Immediate(static_cast<s32>(instr.smem_imm))); - const Node lmem = GetLocalMemory(index); + static_cast<u32>(instr.ld_l.unknown.Value())); + + const auto GetLmem = [&](s32 offset) { + ASSERT(offset % 4 == 0); + const Node immediate_offset = Immediate(static_cast<s32>(instr.smem_imm) + offset); + const Node address = Operation(OperationCode::IAdd, NO_PRECISE, GetRegister(instr.gpr8), + immediate_offset); + return GetLocalMemory(address); + }; switch (instr.ldst_sl.type.Value()) { - case Tegra::Shader::StoreType::Bytes32: - SetRegister(bb, instr.gpr0, lmem); + case Tegra::Shader::StoreType::Bits32: + case Tegra::Shader::StoreType::Bits64: + case Tegra::Shader::StoreType::Bits128: { + const u32 count = [&]() { + switch (instr.ldst_sl.type.Value()) { + case Tegra::Shader::StoreType::Bits32: + return 1; + case Tegra::Shader::StoreType::Bits64: + return 2; + case Tegra::Shader::StoreType::Bits128: + return 4; + default: + UNREACHABLE(); + return 0; + } + }(); + for (u32 i = 0; i < count; ++i) + SetTemporal(bb, i, GetLmem(i * 4)); + for (u32 i = 0; i < count; ++i) + SetRegister(bb, instr.gpr0.Value() + i, GetTemporal(i)); break; + } default: UNIMPLEMENTED_MSG("LD_L Unhandled type: {}", - static_cast<unsigned>(instr.ldst_sl.type.Value())); + static_cast<u32>(instr.ldst_sl.type.Value())); } break; } @@ -203,12 +226,20 @@ u32 ShaderIR::DecodeMemory(NodeBlock& bb, u32 pc) { UNIMPLEMENTED_IF_MSG(instr.st_l.unknown == 0, "ST_L Unhandled mode: {}", static_cast<u32>(instr.st_l.unknown.Value())); - const Node index = Operation(OperationCode::IAdd, NO_PRECISE, GetRegister(instr.gpr8), - Immediate(static_cast<s32>(instr.smem_imm))); + const auto GetLmemAddr = [&](s32 offset) { + ASSERT(offset % 4 == 0); + const Node immediate = Immediate(static_cast<s32>(instr.smem_imm) + offset); + return Operation(OperationCode::IAdd, NO_PRECISE, GetRegister(instr.gpr8), immediate); + }; switch (instr.ldst_sl.type.Value()) { - case Tegra::Shader::StoreType::Bytes32: - SetLocalMemory(bb, index, GetRegister(instr.gpr0)); + case Tegra::Shader::StoreType::Bits128: + SetLocalMemory(bb, GetLmemAddr(12), GetRegister(instr.gpr0.Value() + 3)); + SetLocalMemory(bb, GetLmemAddr(8), GetRegister(instr.gpr0.Value() + 2)); + case Tegra::Shader::StoreType::Bits64: + SetLocalMemory(bb, GetLmemAddr(4), GetRegister(instr.gpr0.Value() + 1)); + case Tegra::Shader::StoreType::Bits32: + SetLocalMemory(bb, GetLmemAddr(0), GetRegister(instr.gpr0)); break; default: UNIMPLEMENTED_MSG("ST_L Unhandled type: {}", @@ -325,15 +356,18 @@ u32 ShaderIR::DecodeMemory(NodeBlock& bb, u32 pc) { const auto& sampler = GetSampler(instr.sampler, Tegra::Shader::TextureType::Texture2D, false, false); + u32 indexer = 0; switch (instr.txq.query_type) { case Tegra::Shader::TextureQueryType::Dimension: { for (u32 element = 0; element < 4; ++element) { - MetaTexture meta{sampler, element}; - const Node value = Operation(OperationCode::F4TextureQueryDimensions, - std::move(meta), GetRegister(instr.gpr8)); - SetTemporal(bb, element, value); + if (instr.txq.IsComponentEnabled(element)) { + MetaTexture meta{sampler, element}; + const Node value = Operation(OperationCode::F4TextureQueryDimensions, + std::move(meta), GetRegister(instr.gpr8)); + SetTemporal(bb, indexer++, value); + } } - for (u32 i = 0; i < 4; ++i) { + for (u32 i = 0; i < indexer; ++i) { SetRegister(bb, instr.gpr0.Value() + i, GetTemporal(i)); } break; @@ -734,4 +768,4 @@ std::tuple<std::size_t, std::size_t> ShaderIR::ValidateAndGetCoordinateElement( return {coord_count, total_coord_count}; } -} // namespace VideoCommon::Shader
\ No newline at end of file +} // namespace VideoCommon::Shader diff --git a/src/video_core/shader/shader_ir.h b/src/video_core/shader/shader_ir.h index 8f97512ee..1d4fbef53 100644 --- a/src/video_core/shader/shader_ir.h +++ b/src/video_core/shader/shader_ir.h @@ -236,6 +236,11 @@ private: class ConstBuffer { public: + explicit ConstBuffer(u32 max_offset, bool is_indirect) + : max_offset{max_offset}, is_indirect{is_indirect} {} + + ConstBuffer() = default; + void MarkAsUsed(u64 offset) { max_offset = std::max(max_offset, static_cast<u32>(offset)); } @@ -252,6 +257,10 @@ public: return max_offset + sizeof(float); } + u32 GetMaxOffset() const { + return max_offset; + } + private: u32 max_offset{}; bool is_indirect{}; |