diff options
Diffstat (limited to 'src/video_core')
| -rw-r--r-- | src/video_core/renderer_opengl/gl_shader_decompiler.cpp | 10 | ||||
| -rw-r--r-- | src/video_core/renderer_opengl/gl_shader_manager.cpp | 4 | ||||
| -rw-r--r-- | src/video_core/renderer_vulkan/maxwell_to_vk.cpp | 4 | ||||
| -rw-r--r-- | src/video_core/renderer_vulkan/vk_device.cpp | 1 | ||||
| -rw-r--r-- | src/video_core/renderer_vulkan/vk_shader_decompiler.cpp | 17 | ||||
| -rw-r--r-- | src/video_core/shader/decode/memory.cpp | 3 | ||||
| -rw-r--r-- | src/video_core/shader/decode/other.cpp | 5 | ||||
| -rw-r--r-- | src/video_core/shader/node.h | 1 |
8 files changed, 41 insertions, 4 deletions
diff --git a/src/video_core/renderer_opengl/gl_shader_decompiler.cpp b/src/video_core/renderer_opengl/gl_shader_decompiler.cpp index 84ca830f4..253484968 100644 --- a/src/video_core/renderer_opengl/gl_shader_decompiler.cpp +++ b/src/video_core/renderer_opengl/gl_shader_decompiler.cpp @@ -2335,6 +2335,15 @@ private: return {fmt::format("readInvocationARB({}, {})", value, index), Type::Float}; } + Expression Barrier(Operation) { + if (!ir.IsDecompiled()) { + LOG_ERROR(Render_OpenGL, "barrier() used but shader is not decompiled"); + return {}; + } + code.AddLine("barrier();"); + return {}; + } + Expression MemoryBarrierGL(Operation) { code.AddLine("memoryBarrier();"); return {}; @@ -2581,6 +2590,7 @@ private: &GLSLDecompiler::ThreadMask<Func::Lt>, &GLSLDecompiler::ShuffleIndexed, + &GLSLDecompiler::Barrier, &GLSLDecompiler::MemoryBarrierGL, }; static_assert(operation_decompilers.size() == static_cast<std::size_t>(OperationCode::Amount)); diff --git a/src/video_core/renderer_opengl/gl_shader_manager.cpp b/src/video_core/renderer_opengl/gl_shader_manager.cpp index 96605db84..8e754fa90 100644 --- a/src/video_core/renderer_opengl/gl_shader_manager.cpp +++ b/src/video_core/renderer_opengl/gl_shader_manager.cpp @@ -47,6 +47,10 @@ void ProgramManager::BindHostPipeline(GLuint pipeline) { old_state.geometry = 0; glDisable(GL_GEOMETRY_PROGRAM_NV); } + } else { + if (!is_graphics_bound) { + glUseProgram(0); + } } glBindProgramPipeline(pipeline); } diff --git a/src/video_core/renderer_vulkan/maxwell_to_vk.cpp b/src/video_core/renderer_vulkan/maxwell_to_vk.cpp index 12be691a5..2871035f5 100644 --- a/src/video_core/renderer_vulkan/maxwell_to_vk.cpp +++ b/src/video_core/renderer_vulkan/maxwell_to_vk.cpp @@ -142,7 +142,7 @@ struct FormatTuple { {VK_FORMAT_BC6H_UFLOAT_BLOCK}, // BC6H_UF16 {VK_FORMAT_BC6H_SFLOAT_BLOCK}, // BC6H_SF16 {VK_FORMAT_ASTC_4x4_UNORM_BLOCK}, // ASTC_2D_4X4 - {VK_FORMAT_B8G8R8A8_UNORM}, // BGRA8 + {VK_FORMAT_B8G8R8A8_UNORM, Attachable}, // BGRA8 {VK_FORMAT_R32G32B32A32_SFLOAT, Attachable | Storage}, // RGBA32F {VK_FORMAT_R32G32_SFLOAT, Attachable | Storage}, // RG32F {VK_FORMAT_R32_SFLOAT, Attachable | Storage}, // R32F @@ -168,7 +168,7 @@ struct FormatTuple { {VK_FORMAT_ASTC_8x8_UNORM_BLOCK}, // ASTC_2D_8X8 {VK_FORMAT_UNDEFINED}, // ASTC_2D_8X5 {VK_FORMAT_UNDEFINED}, // ASTC_2D_5X4 - {VK_FORMAT_UNDEFINED}, // BGRA8_SRGB + {VK_FORMAT_B8G8R8A8_SRGB, Attachable}, // BGRA8_SRGB {VK_FORMAT_BC1_RGBA_SRGB_BLOCK}, // DXT1_SRGB {VK_FORMAT_BC2_SRGB_BLOCK}, // DXT23_SRGB {VK_FORMAT_BC3_SRGB_BLOCK}, // DXT45_SRGB diff --git a/src/video_core/renderer_vulkan/vk_device.cpp b/src/video_core/renderer_vulkan/vk_device.cpp index f0c491d00..750e5a0ca 100644 --- a/src/video_core/renderer_vulkan/vk_device.cpp +++ b/src/video_core/renderer_vulkan/vk_device.cpp @@ -104,6 +104,7 @@ std::unordered_map<VkFormat, VkFormatProperties> GetFormatProperties( VK_FORMAT_R16_SFLOAT, VK_FORMAT_R16G16B16A16_SFLOAT, VK_FORMAT_B8G8R8A8_UNORM, + VK_FORMAT_B8G8R8A8_SRGB, VK_FORMAT_R4G4B4A4_UNORM_PACK16, VK_FORMAT_D32_SFLOAT, VK_FORMAT_D16_UNORM, diff --git a/src/video_core/renderer_vulkan/vk_shader_decompiler.cpp b/src/video_core/renderer_vulkan/vk_shader_decompiler.cpp index b8169d832..890f34a2c 100644 --- a/src/video_core/renderer_vulkan/vk_shader_decompiler.cpp +++ b/src/video_core/renderer_vulkan/vk_shader_decompiler.cpp @@ -2199,6 +2199,22 @@ private: return {OpSubgroupReadInvocationKHR(t_float, value, index), Type::Float}; } + Expression Barrier(Operation) { + if (!ir.IsDecompiled()) { + LOG_ERROR(Render_Vulkan, "OpBarrier used by shader is not decompiled"); + return {}; + } + + const auto scope = spv::Scope::Workgroup; + const auto memory = spv::Scope::Workgroup; + const auto semantics = + spv::MemorySemanticsMask::WorkgroupMemory | spv::MemorySemanticsMask::AcquireRelease; + OpControlBarrier(Constant(t_uint, static_cast<u32>(scope)), + Constant(t_uint, static_cast<u32>(memory)), + Constant(t_uint, static_cast<u32>(semantics))); + return {}; + } + Expression MemoryBarrierGL(Operation) { const auto scope = spv::Scope::Device; const auto semantics = @@ -2664,6 +2680,7 @@ private: &SPIRVDecompiler::ThreadMask<4>, // Lt &SPIRVDecompiler::ShuffleIndexed, + &SPIRVDecompiler::Barrier, &SPIRVDecompiler::MemoryBarrierGL, }; static_assert(operation_decompilers.size() == static_cast<std::size_t>(OperationCode::Amount)); diff --git a/src/video_core/shader/decode/memory.cpp b/src/video_core/shader/decode/memory.cpp index 9392f065b..63adbc4a3 100644 --- a/src/video_core/shader/decode/memory.cpp +++ b/src/video_core/shader/decode/memory.cpp @@ -387,7 +387,6 @@ u32 ShaderIR::DecodeMemory(NodeBlock& bb, u32 pc) { } case OpCode::Id::RED: { UNIMPLEMENTED_IF_MSG(instr.red.type != GlobalAtomicType::U32); - UNIMPLEMENTED_IF_MSG(instr.red.operation != AtomicOp::Add); const auto [real_address, base_address, descriptor] = TrackGlobalMemory(bb, instr, true, true); if (!real_address || !base_address) { @@ -396,7 +395,7 @@ u32 ShaderIR::DecodeMemory(NodeBlock& bb, u32 pc) { } Node gmem = MakeNode<GmemNode>(real_address, base_address, descriptor); Node value = GetRegister(instr.gpr0); - bb.push_back(Operation(OperationCode::ReduceIAdd, move(gmem), move(value))); + bb.push_back(Operation(GetAtomOperation(instr.red.operation), move(gmem), move(value))); break; } case OpCode::Id::ATOM: { diff --git a/src/video_core/shader/decode/other.cpp b/src/video_core/shader/decode/other.cpp index 399a455c4..694b325e1 100644 --- a/src/video_core/shader/decode/other.cpp +++ b/src/video_core/shader/decode/other.cpp @@ -293,6 +293,11 @@ u32 ShaderIR::DecodeOther(NodeBlock& bb, u32 pc) { SetRegister(bb, instr.gpr0, GetRegister(instr.gpr8)); break; } + case OpCode::Id::BAR: { + UNIMPLEMENTED_IF_MSG(instr.value != 0xF0A81B8000070000ULL, "BAR is not BAR.SYNC 0x0"); + bb.push_back(Operation(OperationCode::Barrier)); + break; + } case OpCode::Id::MEMBAR: { UNIMPLEMENTED_IF(instr.membar.type != Tegra::Shader::MembarType::GL); UNIMPLEMENTED_IF(instr.membar.unknown != Tegra::Shader::MembarUnknown::Default); diff --git a/src/video_core/shader/node.h b/src/video_core/shader/node.h index cce8aeebe..c06512413 100644 --- a/src/video_core/shader/node.h +++ b/src/video_core/shader/node.h @@ -233,6 +233,7 @@ enum class OperationCode { ThreadLtMask, /// () -> uint ShuffleIndexed, /// (uint value, uint index) -> uint + Barrier, /// () -> void MemoryBarrierGL, /// () -> void Amount, |