[shader-decoder] improve RDNA2 shader decoder to support more instructions and fix some issues with the previous implementation.

src/SharpEmu.Libs/Agc/AgcExports.cs

@@ -142,7 +142,7 @@ public static class AgcExports
     private static readonly HashSet<uint> _tracedDcbSizes = new();
     private static readonly HashSet<(ulong Es, ulong Ps, GuestDrawKind Kind)> _tracedShaderTranslations = new();
     private static readonly HashSet<(ulong Es, ulong Ps)> _tracedShaderDecodePairs = new();
-    private static readonly HashSet<(ulong Es, ulong Ps, ulong Target)> _tracedShaderDraws = new();
+    private static readonly HashSet<(ulong Es, ulong Ps, ulong Target, ulong Texture, uint VertexCount)> _tracedShaderDraws = new();
     private static readonly HashSet<(ulong Ps, string Error)> _tracedShaderFailures = new();
     private static readonly HashSet<(int Handle, int Index, ulong Address, string Path)> _tracedDisplayBuffers = new();
     private static readonly HashSet<ulong> _tracedComputeShaders = new();
@@ -2788,7 +2788,7 @@ public static class AgcExports
                 $"ps=0x{(hasPixelShader ? pixelShaderAddress : 0):X16}");
         }
 
-        if (vertexCount is not (3 or 4 or 6))
+        if (vertexCount == 0 || vertexCount > 1_048_576)
         {
             return;
         }
@@ -2872,8 +2872,9 @@ public static class AgcExports
 
             lock (_submitTraceGate)
             {
+                var firstTextureAddress = translatedDraw.Textures.FirstOrDefault()?.Descriptor.Address ?? 0;
                 if (_tracedShaderDraws.Add(
-                        (exportShaderAddress, pixelShaderAddress, firstTarget.Address)))
+                        (exportShaderAddress, pixelShaderAddress, firstTarget.Address, firstTextureAddress, vertexCount)))
                 {
                     TraceTranslatedGuestDraw(
                         ctx,
@@ -3033,6 +3034,11 @@ public static class AgcExports
                 return false;
             }
 
+            TraceAgcShader(
+                $"agc.texture_binding ps=0x{pixelShaderAddress:X16} es=0x{exportShaderAddress:X16} " +
+                $"pc=0x{binding.Pc:X} op={binding.Opcode} storage={(Gen5ShaderTranslator.IsStorageImageOperation(binding.Opcode) ? 1 : 0)} " +
+                $"decoded={FormatTextureDescriptor(texture)} " +
+                $"raw={FormatShaderDwords(binding.ResourceDescriptor)} sampler={FormatShaderDwords(binding.SamplerDescriptor)}");
             textures.Add(
                 new TranslatedImageBinding(
                     texture,
@@ -3477,7 +3483,8 @@ public static class AgcExports
                 ',',
                 draw.VertexInputs.Select(input =>
                     $"{input.Location}:pc=0x{input.Pc:X}:0x{input.BaseAddress:X16}" +
-                    $":stride{input.Stride}:off{input.OffsetBytes}:c{input.ComponentCount}"));
+                    $":stride{input.Stride}:off{input.OffsetBytes}:c{input.ComponentCount}" +
+                    $":fmt{input.DataFormat}/num{input.NumberFormat}"));
         var scissor = draw.RenderState.Scissor is { } drawScissor
             ? $"{drawScissor.X},{drawScissor.Y},{drawScissor.Width}x{drawScissor.Height}"
             : "full";
@@ -3486,6 +3493,29 @@ public static class AgcExports
               $"{drawViewport.Width:0.###}x{drawViewport.Height:0.###}:" +
               $"{drawViewport.MinDepth:0.###}-{drawViewport.MaxDepth:0.###}"
             : "full";
+        var rasterRegisters = new (string Name, uint Offset)[]
+        {
+            ("screen_tl", PaScScreenScissorTl),
+            ("screen_br", PaScScreenScissorBr),
+            ("window_off", PaScWindowOffset),
+            ("window_tl", PaScWindowScissorTl),
+            ("window_br", PaScWindowScissorBr),
+            ("generic_tl", PaScGenericScissorTl),
+            ("generic_br", PaScGenericScissorBr),
+            ("vport_tl", PaScVportScissor0Tl),
+            ("vport_br", PaScVportScissor0Br),
+            ("mode", PaScModeCntl0),
+            ("xscale", PaClVportXScale),
+            ("xoffset", PaClVportXOffset),
+            ("yscale", PaClVportYScale),
+            ("yoffset", PaClVportYOffset),
+        };
+        var raster = string.Join(
+            ',',
+            rasterRegisters.Select(entry =>
+                state.CxRegisters.TryGetValue(entry.Offset, out var value)
+                    ? $"{entry.Name}=0x{value:X8}"
+                    : $"{entry.Name}=missing"));
         var blend = draw.RenderState.Blend;
         TraceAgcShader(
             $"agc.shader_draw es=0x{draw.ExportShaderAddress:X16} " +
@@ -3493,6 +3523,7 @@ public static class AgcExports
             $"primitive=0x{draw.PrimitiveType:X} " +
             $"blend={(blend.Enable ? 1 : 0)}:{blend.ColorSrcFactor}/{blend.ColorDstFactor}/{blend.ColorFunc} " +
             $"write_mask=0x{blend.WriteMask:X} scissor={scissor} viewport={viewport} " +
+            $"raster=[{raster}] " +
             $"ps_ena=0x{psInputEna:X8} ps_addr=0x{psInputAddr:X8} " +
             $"targets=[{targets}] textures=[{textures}] " +
             $"buffers=[{buffers}] vertex=[{vertexInputs}] indices=[{indices}]");
@@ -3550,6 +3581,8 @@ public static class AgcExports
             buffers[index] = new VulkanGuestVertexBuffer(
                 binding.Location,
                 binding.ComponentCount,
+                binding.DataFormat,
+                binding.NumberFormat,
                 binding.BaseAddress,
                 binding.Stride,
                 binding.OffsetBytes,
@@ -3579,7 +3612,10 @@ public static class AgcExports
         }
 
         var sourceWidth = descriptor.TileMode == 0
-            ? Math.Max(descriptor.Width, descriptor.Pitch)
+            ? GetLinearTexturePitch(
+                Math.Max(descriptor.Width, descriptor.Pitch),
+                descriptor.Height,
+                descriptor.Format)
             : descriptor.Width;
         var sourceByteCount = GetTextureByteCount(
             descriptor.Format,
@@ -3617,7 +3653,7 @@ public static class AgcExports
                 IsStorage: true,
                 MipLevels: descriptor.MipLevels,
                 MipLevel: mipLevel,
-                Pitch: descriptor.Pitch,
+                Pitch: sourceWidth,
                 TileMode: descriptor.TileMode,
                 DstSelect: descriptor.DstSelect,
                 Sampler: ToVulkanSampler(samplerDescriptor));
@@ -3663,7 +3699,7 @@ public static class AgcExports
             IsStorage: isStorage,
             MipLevels: descriptor.MipLevels,
             MipLevel: mipLevel,
-            Pitch: descriptor.Pitch,
+            Pitch: sourceWidth,
             TileMode: descriptor.TileMode,
             DstSelect: descriptor.DstSelect,
             Sampler: ToVulkanSampler(samplerDescriptor));
@@ -4326,6 +4362,28 @@ public static class AgcExports
             : checked(((ulong)width + 3) / 4 * (((ulong)height + 3) / 4) * blockBytes);
     }
 
+    private static uint GetLinearTexturePitch(uint pitch, uint height, uint format)
+    {
+        var bytesPerTexel = GetTextureBytesPerTexel(format);
+        if (bytesPerTexel == 0 || height == 0)
+        {
+            return pitch;
+        }
+
+        var pitchAlignment = Math.Max(8UL, 64UL / bytesPerTexel);
+        var alignedPitch = AlignUp(pitch, pitchAlignment);
+        var sliceAlignment = Math.Max(64UL, 256UL / bytesPerTexel);
+        while ((alignedPitch * height) % sliceAlignment != 0)
+        {
+            alignedPitch += pitchAlignment;
+        }
+
+        return checked((uint)alignedPitch);
+    }
+
+    private static ulong AlignUp(ulong value, ulong alignment) =>
+        (value + alignment - 1) & ~(alignment - 1);
+
     private static void TraceShaderTranslationMiss(
         CpuContext ctx,
         SubmittedDcbState state,
@@ -4594,6 +4652,10 @@ public static class AgcExports
             return false;
         }
 
+        // GFX10/RDNA2 T# layout: WIDTH is split across word1[31:30] (lo 2 bits)
+        // and word2[11:0] (hi 12 bits); FORMAT is the combined 9-bit field at
+        // word1[28:20]. Verified against Kyty's decode of the same game
+        // descriptors (fmt=56=8_8_8_8_UNORM, extent 1280x720, sw_mode 27).
         // GNM T# exposes a 38-bit baseaddr256 field, but RPCSX and the
         // Demon's Souls descriptors both show that only the low 32 bits are
         // part of the guest GPU VA. The upper baseaddr bits carry resource
@@ -5367,6 +5429,12 @@ public static class AgcExports
             ? "none"
             : string.Join(',', values.Select(static value => $"{value:X8}"));
 
+    private static string FormatTextureDescriptor(TextureDescriptor descriptor) =>
+        $"addr=0x{descriptor.Address:X16} {descriptor.Width}x{descriptor.Height} " +
+        $"fmt={descriptor.Format} num={descriptor.NumberType} tile={descriptor.TileMode} " +
+        $"type={descriptor.Type} levels={descriptor.BaseLevel}-{descriptor.LastLevel} " +
+        $"pitch={descriptor.Pitch} dst=0x{descriptor.DstSelect:X3}";
+
     private static void DumpSpirv(
         string stage,
         ulong shaderAddress,

src/SharpEmu.Libs/Agc/Gen5ShaderIr.cs

@@ -276,6 +276,8 @@ internal sealed record Gen5VertexInputBinding(
     uint Pc,
     uint Location,
     uint ComponentCount,
+    uint DataFormat,
+    uint NumberFormat,
     ulong BaseAddress,
     uint Stride,
     uint OffsetBytes,

src/SharpEmu.Libs/Agc/Gen5ShaderScalarEvaluator.cs

@@ -20,7 +20,9 @@ internal static class Gen5ShaderScalarEvaluator
         ulong BaseAddress,
         uint Stride,
         uint NumRecords,
-        ulong SizeBytes);
+        ulong SizeBytes,
+        uint NumberFormat,
+        uint DataFormat);
 
     public static bool TryResolveImageBindings(
         CpuContext ctx,
@@ -400,14 +402,21 @@ internal static class Gen5ShaderScalarEvaluator
             return false;
         }
 
+        var bindingData = data;
+        var bindingStride = descriptor.Stride;
+        var bindingOffset = unchecked((uint)control.OffsetBytes + scalarOffset);
+        var bindingDataFormat = descriptor.DataFormat;
+        var bindingNumberFormat = descriptor.NumberFormat;
         binding = new Gen5VertexInputBinding(
             instruction.Pc,
             location,
             control.DwordCount,
+            bindingDataFormat,
+            bindingNumberFormat,
             descriptor.BaseAddress,
-            descriptor.Stride,
-            unchecked((uint)control.OffsetBytes + scalarOffset),
-            data);
+            bindingStride,
+            bindingOffset,
+            bindingData);
         return true;
     }
 
@@ -803,7 +812,11 @@ internal static class Gen5ShaderScalarEvaluator
                 "SFF1I32B32" => left == 0 ? uint.MaxValue : (uint)BitOperations.TrailingZeroCount(left),
                 _ => registers[destination.Value] | (1u << ((int)left & 31)),
             };
-            scalarConditionCode = registers[destination.Value] != 0;
+            if (instruction.Opcode != "SBitset1B32")
+            {
+                scalarConditionCode = registers[destination.Value] != 0;
+            }
+
             return true;
         }
 
@@ -829,13 +842,15 @@ internal static class Gen5ShaderScalarEvaluator
                 }
             case "SSubU32":
                 result = left - right;
-                scalarConditionCode = left >= right;
+                scalarConditionCode = right > left;
                 break;
             case "SAddI32":
                 result = unchecked((uint)((int)left + (int)right));
+                scalarConditionCode = SignedAddOverflow(left, right, result);
                 break;
             case "SSubI32":
                 result = unchecked((uint)((int)left - (int)right));
+                scalarConditionCode = SignedSubOverflow(left, right, result);
                 break;
             case "SAddcU32":
                 {
@@ -846,23 +861,27 @@ internal static class Gen5ShaderScalarEvaluator
                 }
             case "SSubbU32":
                 {
-                    var borrow = scalarConditionCode ? 0UL : 1UL;
+                    var borrow = scalarConditionCode ? 1UL : 0UL;
                     var subtrahend = (ulong)right + borrow;
                     result = unchecked(left - (uint)subtrahend);
-                    scalarConditionCode = left >= subtrahend;
+                    scalarConditionCode = subtrahend > left;
                     break;
                 }
             case "SMinI32":
                 result = unchecked((uint)Math.Min((int)left, (int)right));
+                scalarConditionCode = (int)left < (int)right;
                 break;
             case "SMinU32":
                 result = Math.Min(left, right);
+                scalarConditionCode = left < right;
                 break;
             case "SMaxI32":
                 result = unchecked((uint)Math.Max((int)left, (int)right));
+                scalarConditionCode = (int)left > (int)right;
                 break;
             case "SMaxU32":
                 result = Math.Max(left, right);
+                scalarConditionCode = left > right;
                 break;
             case "SCselectB32":
                 result = scalarConditionCode ? left : right;
@@ -926,6 +945,7 @@ internal static class Gen5ShaderScalarEvaluator
                     var offset = (int)right & 31;
                     var width = Math.Min(((int)right >> 16) & 0x7F, 32 - offset);
                     result = width == 0 ? 0 : left >> offset & (uint.MaxValue >> (32 - width));
+                    scalarConditionCode = result != 0;
                     break;
                 }
             case "SBfeI32":
@@ -935,23 +955,41 @@ internal static class Gen5ShaderScalarEvaluator
                     result = width == 0
                         ? 0
                         : unchecked((uint)(((int)(left << (32 - width - offset))) >> (32 - width)));
+                    scalarConditionCode = result != 0;
                     break;
                 }
             case "SAbsdiffI32":
                 result = unchecked((uint)Math.Abs((long)(int)left - (int)right));
+                scalarConditionCode = result != 0;
                 break;
             case "SLshl1AddU32":
-                result = (left << 1) + right;
-                break;
+                {
+                    var wide = ((ulong)left << 1) + right;
+                    result = (uint)wide;
+                    scalarConditionCode = wide > uint.MaxValue;
+                    break;
+                }
             case "SLshl2AddU32":
-                result = (left << 2) + right;
-                break;
+                {
+                    var wide = ((ulong)left << 2) + right;
+                    result = (uint)wide;
+                    scalarConditionCode = wide > uint.MaxValue;
+                    break;
+                }
             case "SLshl3AddU32":
-                result = (left << 3) + right;
-                break;
+                {
+                    var wide = ((ulong)left << 3) + right;
+                    result = (uint)wide;
+                    scalarConditionCode = wide > uint.MaxValue;
+                    break;
+                }
             case "SLshl4AddU32":
-                result = (left << 4) + right;
-                break;
+                {
+                    var wide = ((ulong)left << 4) + right;
+                    result = (uint)wide;
+                    scalarConditionCode = wide > uint.MaxValue;
+                    break;
+                }
             case "SPackLlB32B16":
                 result = (left & 0xFFFFu) | (right << 16);
                 break;
@@ -993,7 +1031,8 @@ internal static class Gen5ShaderScalarEvaluator
             "SNandSaveexecB64" or
             "SNorSaveexecB64" or
             "SXnorSaveexecB64" or
-            "SAndn1SaveexecB64"))
+            "SAndn1SaveexecB64" or
+            "SOrn1SaveexecB64"))
         {
             return false;
         }
@@ -1021,11 +1060,12 @@ internal static class Gen5ShaderScalarEvaluator
             "SAndSaveexecB64" => oldExec & source,
             "SOrSaveexecB64" => oldExec | source,
             "SXorSaveexecB64" => oldExec ^ source,
-            "SAndn1SaveexecB64" => ~oldExec & source,
-            "SAndn2SaveexecB64" => oldExec & ~source,
-            "SOrn2SaveexecB64" => oldExec | ~source,
-            "SNandSaveexecB64" => ~(oldExec & source),
-            "SNorSaveexecB64" => ~(oldExec | source),
+            "SAndn1SaveexecB64" => ~source & oldExec,
+            "SAndn2SaveexecB64" => source & ~oldExec,
+            "SOrn1SaveexecB64" => ~source | oldExec,
+            "SOrn2SaveexecB64" => source | ~oldExec,
+            "SNandSaveexecB64" => ~(source & oldExec),
+            "SNorSaveexecB64" => ~(source | oldExec),
             _ => ~(oldExec ^ source),
         };
 
@@ -1095,6 +1135,12 @@ internal static class Gen5ShaderScalarEvaluator
 
     private static ulong MaskWaveValue(ulong value) => value & RdnaWaveMask;
 
+    private static bool SignedAddOverflow(uint left, uint right, uint result) =>
+        ((left ^ result) & (right ^ result) & 0x80000000u) != 0;
+
+    private static bool SignedSubOverflow(uint left, uint right, uint result) =>
+        ((left ^ right) & (left ^ result) & 0x80000000u) != 0;
+
     private static bool TryExecuteScalarCompare(
         Gen5ShaderInstruction instruction,
         uint[] registers,
@@ -1416,7 +1462,7 @@ internal static class Gen5ShaderScalarEvaluator
             word2 == 0 &&
             word3 == 0)
         {
-            descriptor = new BufferDescriptor(0, 0, 0, 0);
+            descriptor = new BufferDescriptor(0, 0, 0, 0, 0, 0);
             return true;
         }
 
@@ -1428,19 +1474,64 @@ internal static class Gen5ShaderScalarEvaluator
                 return false;
             }
 
-            descriptor = new BufferDescriptor(0, 0, 0, 0);
+            descriptor = new BufferDescriptor(0, 0, 0, 0, 0, 0);
             return true;
         }
 
-        var baseAddress = word0 | ((ulong)(word1 & 0x0FFFu) << 32);
+        var baseAddress = word0 | ((ulong)(word1 & 0xFFFFu) << 32);
         var stride = (word1 >> 16) & 0x3FFFu;
+        var unifiedFormat = (word3 >> 12) & 0x7Fu;
+        var (dataFormat, numberFormat) =
+            DecodeGfx10BufferFormat(unifiedFormat);
         var sizeBytes = stride == 0
             ? word2
             : (ulong)stride * word2;
-        descriptor = new BufferDescriptor(baseAddress, stride, word2, sizeBytes);
+        descriptor = new BufferDescriptor(baseAddress, stride, word2, sizeBytes, numberFormat, dataFormat);
         return true;
     }
 
+    private static (uint DataFormat, uint NumberFormat)
+        DecodeGfx10BufferFormat(uint format) =>
+        format switch
+        {
+            0 => (0, 0),
+            >= 1 and <= 6 => (1, format - 1),
+            >= 7 and <= 13 => (2, DecodeUnifiedNumber(format - 7, 7)),
+            >= 14 and <= 19 => (3, format - 14),
+            >= 20 and <= 22 => (4, DecodeIntegerOrFloatNumber(format - 20)),
+            >= 23 and <= 29 => (5, DecodeUnifiedNumber(format - 23, 7)),
+            >= 30 and <= 36 => (6, DecodeUnifiedNumber(format - 30, 7)),
+            >= 37 and <= 43 => (7, DecodeUnifiedNumber(format - 37, 7)),
+            >= 44 and <= 49 => (8, format - 44),
+            >= 50 and <= 55 => (9, format - 50),
+            >= 56 and <= 61 => (10, format - 56),
+            >= 62 and <= 64 => (11, DecodeIntegerOrFloatNumber(format - 62)),
+            >= 65 and <= 71 => (12, DecodeUnifiedNumber(format - 65, 7)),
+            >= 72 and <= 74 => (13, DecodeIntegerOrFloatNumber(format - 72)),
+            >= 75 and <= 77 => (14, DecodeIntegerOrFloatNumber(format - 75)),
+            128 => (1, 9),
+            129 => (3, 9),
+            130 => (10, 9),
+            132 => (34, 7),
+            133 => (16, 0),
+            134 => (17, 0),
+            135 => (18, 0),
+            136 => (19, 0),
+            140 => (4, 7),
+            _ => (0, 0),
+        };
+
+    private static uint DecodeUnifiedNumber(uint offset, uint formatCount) =>
+        offset == formatCount - 1 ? 7u : offset;
+
+    private static uint DecodeIntegerOrFloatNumber(uint offset) =>
+        offset switch
+        {
+            0 => 4,
+            1 => 5,
+            _ => 7,
+        };
+
     private static bool TryReadUserDataScalarLoad(
         Gen5ShaderState state,
         Gen5ShaderInstruction instruction,

src/SharpEmu.Libs/Agc/Gen5ShaderTranslator.cs

@@ -452,6 +452,7 @@ internal static class Gen5ShaderTranslator
             0x2A => "SNorSaveexecB64",
             0x2B => "SXnorSaveexecB64",
             0x37 => "SAndn1SaveexecB64",
+            0x38 => "SOrn1SaveexecB64",
             _ => string.Empty,
         };
 

src/SharpEmu.Libs/Agc/Gen5SpirvTranslator.Alu.cs

@@ -212,12 +212,26 @@ internal static partial class Gen5SpirvTranslator
                 case "VSinF32":
                     result = EmitFloatResult(
                         instruction,
-                        Ext(13, _floatType, GetFloatSource(instruction, 0)));
+                        Ext(
+                            13,
+                            _floatType,
+                            _module.AddInstruction(
+                                SpirvOp.FMul,
+                                _floatType,
+                                GetFloatSource(instruction, 0),
+                                Float(MathF.Tau))));
                     break;
                 case "VCosF32":
                     result = EmitFloatResult(
                         instruction,
-                        Ext(14, _floatType, GetFloatSource(instruction, 0)));
+                        Ext(
+                            14,
+                            _floatType,
+                            _module.AddInstruction(
+                                SpirvOp.FMul,
+                                _floatType,
+                                GetFloatSource(instruction, 0),
+                                Float(MathF.Tau))));
                     break;
                 case "VAddF32":
                     result = EmitFloatBinary(instruction, SpirvOp.FAdd);
@@ -635,13 +649,29 @@ internal static partial class Gen5SpirvTranslator
                         width);
                     break;
                 }
+                case "VBfiB32":
+                {
+                    var mask = GetRawSource(instruction, 0);
+                    var insert = GetRawSource(instruction, 1);
+                    var source = GetRawSource(instruction, 2);
+                    result = _module.AddInstruction(
+                        SpirvOp.BitwiseOr,
+                        _uintType,
+                        BitwiseAnd(mask, insert),
+                        BitwiseAnd(
+                            _module.AddInstruction(SpirvOp.Not, _uintType, mask),
+                            source));
+                    break;
+                }
                 case "VCvtPkrtzF16F32":
                 {
+                    var first = TruncateFloat32ForPack(GetFloatSource(instruction, 0));
+                    var second = TruncateFloat32ForPack(GetFloatSource(instruction, 1));
                     var vector = _module.AddInstruction(
                         SpirvOp.CompositeConstruct,
                         _vec2Type,
-                        GetFloatSource(instruction, 0),
-                        GetFloatSource(instruction, 1));
+                        first,
+                        second);
                     result = Ext(58, _uintType, vector);
                     break;
                 }
@@ -945,6 +975,16 @@ internal static partial class Gen5SpirvTranslator
                     StoreS(destination, result);
                     Store(_scc, IsNotZero(result));
                     return true;
+                case "SBitset1B32":
+                    result = _module.AddInstruction(
+                        SpirvOp.BitFieldInsert,
+                        _uintType,
+                        LoadS(destination),
+                        UInt(1),
+                        BitwiseAnd(left, UInt(31)),
+                        UInt(1));
+                    StoreS(destination, result);
+                    return true;
                 default:
                 {
                     if (instruction.Sources.Count < 2)
@@ -971,13 +1011,14 @@ internal static partial class Gen5SpirvTranslator
                                 left,
                                 right);
                             Store(_scc, _module.AddInstruction(
-                                SpirvOp.UGreaterThanEqual,
+                                SpirvOp.UGreaterThan,
                                 _boolType,
-                                left,
-                                right));
+                                right,
+                                left));
                             break;
                         case "SAddI32":
                             result = IAdd(left, right);
+                            Store(_scc, SignedAddOverflow(left, right, result));
                             break;
                         case "SSubI32":
                             result = _module.AddInstruction(
@@ -985,6 +1026,7 @@ internal static partial class Gen5SpirvTranslator
                                 _uintType,
                                 left,
                                 right);
+                            Store(_scc, SignedSubOverflow(left, right, result));
                             break;
                         case "SAddcU32":
                         {
@@ -1021,8 +1063,8 @@ internal static partial class Gen5SpirvTranslator
                                 SpirvOp.Select,
                                 _uintType,
                                 Load(_boolType, _scc),
-                                UInt(0),
-                                UInt(1));
+                                UInt(1),
+                                UInt(0));
                             var partial = _module.AddInstruction(
                                 SpirvOp.ISub,
                                 _uintType,
@@ -1033,23 +1075,31 @@ internal static partial class Gen5SpirvTranslator
                                 _uintType,
                                 partial,
                                 borrow);
-                            var firstNoBorrow = _module.AddInstruction(
-                                SpirvOp.UGreaterThanEqual,
+                            var firstBorrow = _module.AddInstruction(
+                                SpirvOp.UGreaterThan,
                                 _boolType,
-                                left,
-                                right);
-                            var secondNoBorrow = _module.AddInstruction(
-                                SpirvOp.UGreaterThanEqual,
+                                right,
+                                left);
+                            var secondBorrow = _module.AddInstruction(
+                                SpirvOp.LogicalAnd,
                                 _boolType,
-                                partial,
-                                borrow);
+                                _module.AddInstruction(
+                                    SpirvOp.IEqual,
+                                    _boolType,
+                                    borrow,
+                                    UInt(1)),
+                                _module.AddInstruction(
+                                    SpirvOp.IEqual,
+                                    _boolType,
+                                    right,
+                                    left));
                             Store(
                                 _scc,
                                 _module.AddInstruction(
-                                    SpirvOp.LogicalAnd,
+                                    SpirvOp.LogicalOr,
                                     _boolType,
-                                    firstNoBorrow,
-                                    secondNoBorrow));
+                                    firstBorrow,
+                                    secondBorrow));
                             break;
                         }
                         case "SMulI32":
@@ -1061,6 +1111,7 @@ internal static partial class Gen5SpirvTranslator
                             break;
                         case "SAndB32":
                             result = BitwiseAnd(left, right);
+                            Store(_scc, IsNotZero(result));
                             break;
                         case "SOrB32":
                             result = _module.AddInstruction(
@@ -1068,6 +1119,7 @@ internal static partial class Gen5SpirvTranslator
                                 _uintType,
                                 left,
                                 right);
+                            Store(_scc, IsNotZero(result));
                             break;
                         case "SXorB32":
                             result = _module.AddInstruction(
@@ -1075,22 +1127,64 @@ internal static partial class Gen5SpirvTranslator
                                 _uintType,
                                 left,
                                 right);
+                            Store(_scc, IsNotZero(result));
                             break;
                         case "SAndn2B32":
                             result = BitwiseAnd(
                                 left,
                                 _module.AddInstruction(SpirvOp.Not, _uintType, right));
+                            Store(_scc, IsNotZero(result));
+                            break;
+                        case "SOrn2B32":
+                            result = _module.AddInstruction(
+                                SpirvOp.BitwiseOr,
+                                _uintType,
+                                left,
+                                _module.AddInstruction(SpirvOp.Not, _uintType, right));
+                            Store(_scc, IsNotZero(result));
+                            break;
+                        case "SNandB32":
+                            result = _module.AddInstruction(
+                                SpirvOp.Not,
+                                _uintType,
+                                BitwiseAnd(left, right));
+                            Store(_scc, IsNotZero(result));
+                            break;
+                        case "SNorB32":
+                            result = _module.AddInstruction(
+                                SpirvOp.Not,
+                                _uintType,
+                                _module.AddInstruction(
+                                    SpirvOp.BitwiseOr,
+                                    _uintType,
+                                    left,
+                                    right));
+                            Store(_scc, IsNotZero(result));
+                            break;
+                        case "SXnorB32":
+                            result = _module.AddInstruction(
+                                SpirvOp.Not,
+                                _uintType,
+                                _module.AddInstruction(
+                                    SpirvOp.BitwiseXor,
+                                    _uintType,
+                                    left,
+                                    right));
+                            Store(_scc, IsNotZero(result));
                             break;
                         case "SLshlB32":
                             result = ShiftLeftLogical(left, right);
+                            Store(_scc, IsNotZero(result));
                             break;
                         case "SLshrB32":
                             result = ShiftRightLogical(
                                 left,
                                 BitwiseAnd(right, UInt(31)));
+                            Store(_scc, IsNotZero(result));
                             break;
                         case "SAshrI32":
                             result = ShiftRightArithmetic(left, right);
+                            Store(_scc, IsNotZero(result));
                             break;
                         case "SBfmB32":
                             result = _module.AddInstruction(
@@ -1133,6 +1227,7 @@ internal static partial class Gen5SpirvTranslator
                                     left,
                                     offset,
                                     width);
+                            Store(_scc, IsNotZero(result));
                             break;
                         }
                         case "SCselectB32":
@@ -1145,9 +1240,47 @@ internal static partial class Gen5SpirvTranslator
                             break;
                         case "SMinU32":
                             result = Ext(38, _uintType, left, right);
+                            Store(
+                                _scc,
+                                _module.AddInstruction(
+                                    SpirvOp.ULessThan,
+                                    _boolType,
+                                    left,
+                                    right));
+                            break;
+                        case "SMinI32":
+                            result = Bitcast(
+                                _uintType,
+                                Ext(39, _intType, Bitcast(_intType, left), Bitcast(_intType, right)));
+                            Store(
+                                _scc,
+                                _module.AddInstruction(
+                                    SpirvOp.SLessThan,
+                                    _boolType,
+                                    Bitcast(_intType, left),
+                                    Bitcast(_intType, right)));
                             break;
                         case "SMaxU32":
                             result = Ext(41, _uintType, left, right);
+                            Store(
+                                _scc,
+                                _module.AddInstruction(
+                                    SpirvOp.UGreaterThan,
+                                    _boolType,
+                                    left,
+                                    right));
+                            break;
+                        case "SMaxI32":
+                            result = Bitcast(
+                                _uintType,
+                                Ext(42, _intType, Bitcast(_intType, left), Bitcast(_intType, right)));
+                            Store(
+                                _scc,
+                                _module.AddInstruction(
+                                    SpirvOp.SGreaterThan,
+                                    _boolType,
+                                    Bitcast(_intType, left),
+                                    Bitcast(_intType, right)));
                             break;
                         case "SLshl1AddU32":
                         case "SLshl2AddU32":
@@ -1292,17 +1425,55 @@ internal static partial class Gen5SpirvTranslator
                     "SXorSaveexecB64" => _module.AddInstruction(
                         SpirvOp.BitwiseXor, _ulongType, oldExec, left),
                     "SAndn2SaveexecB64" => _module.AddInstruction(
-                        SpirvOp.BitwiseAnd, _ulongType, oldExec, notLeft),
-                    "SAndn1SaveexecB64" => _module.AddInstruction(
                         SpirvOp.BitwiseAnd,
                         _ulongType,
+                        left,
                         _module.AddInstruction(
                             SpirvOp.Not,
                             _ulongType,
-                            oldExec),
-                        left),
+                            oldExec)),
+                    "SAndn1SaveexecB64" => _module.AddInstruction(
+                        SpirvOp.BitwiseAnd,
+                        _ulongType,
+                        notLeft,
+                        oldExec),
+                    "SOrn1SaveexecB64" => _module.AddInstruction(
+                        SpirvOp.BitwiseOr,
+                        _ulongType,
+                        notLeft,
+                        oldExec),
                     "SOrn2SaveexecB64" => _module.AddInstruction(
-                        SpirvOp.BitwiseOr, _ulongType, oldExec, notLeft),
+                        SpirvOp.BitwiseOr,
+                        _ulongType,
+                        left,
+                        _module.AddInstruction(
+                            SpirvOp.Not,
+                            _ulongType,
+                            oldExec)),
+                    "SNandSaveexecB64" => _module.AddInstruction(
+                        SpirvOp.Not,
+                        _ulongType,
+                        _module.AddInstruction(
+                            SpirvOp.BitwiseAnd,
+                            _ulongType,
+                            left,
+                            oldExec)),
+                    "SNorSaveexecB64" => _module.AddInstruction(
+                        SpirvOp.Not,
+                        _ulongType,
+                        _module.AddInstruction(
+                            SpirvOp.BitwiseOr,
+                            _ulongType,
+                            left,
+                            oldExec)),
+                    "SXnorSaveexecB64" => _module.AddInstruction(
+                        SpirvOp.Not,
+                        _ulongType,
+                        _module.AddInstruction(
+                            SpirvOp.BitwiseXor,
+                            _ulongType,
+                            left,
+                            oldExec)),
                     _ => 0u,
                 };
                 if (newExec == 0)
@@ -1508,6 +1679,22 @@ internal static partial class Gen5SpirvTranslator
             }
 
             StoreS64(destination, value);
+            if (instruction.Opcode is
+                "SNotB64" or
+                "SAndB64" or
+                "SOrB64" or
+                "SXorB64" or
+                "SAndn1B64" or
+                "SAndn2B64" or
+                "SOrn1B64" or
+                "SOrn2B64" or
+                "SNandB64" or
+                "SNorB64" or
+                "SXnorB64")
+            {
+                Store(_scc, IsNotZero64(value));
+            }
+
             return true;
         }
 
@@ -2067,6 +2254,14 @@ internal static partial class Gen5SpirvTranslator
             return Bitcast(_uintType, value);
         }
 
+        private uint TruncateFloat32ForPack(uint value)
+        {
+            var raw = BitwiseAnd(
+                Bitcast(_uintType, value),
+                UInt(0xFFFF_E000));
+            return Bitcast(_floatType, raw);
+        }
+
         private uint Ext(uint operation, uint resultType, params uint[] operands)
         {
             var values = new uint[2 + operands.Length];
@@ -2086,6 +2281,53 @@ internal static partial class Gen5SpirvTranslator
                 value,
                 _module.Constant64(_ulongType, 0));
 
+        private uint SignBit(uint value) =>
+            ShiftRightLogical(value, UInt(31));
+
+        private uint SignedAddOverflow(uint left, uint right, uint result)
+        {
+            var leftSign = SignBit(left);
+            var rightSign = SignBit(right);
+            var resultSign = SignBit(result);
+            var sameSourceSign = _module.AddInstruction(
+                SpirvOp.IEqual,
+                _boolType,
+                leftSign,
+                rightSign);
+            var resultSignChanged = _module.AddInstruction(
+                SpirvOp.INotEqual,
+                _boolType,
+                leftSign,
+                resultSign);
+            return _module.AddInstruction(
+                SpirvOp.LogicalAnd,
+                _boolType,
+                sameSourceSign,
+                resultSignChanged);
+        }
+
+        private uint SignedSubOverflow(uint left, uint right, uint result)
+        {
+            var leftSign = SignBit(left);
+            var rightSign = SignBit(right);
+            var resultSign = SignBit(result);
+            var differentSourceSign = _module.AddInstruction(
+                SpirvOp.INotEqual,
+                _boolType,
+                leftSign,
+                rightSign);
+            var resultSignChanged = _module.AddInstruction(
+                SpirvOp.INotEqual,
+                _boolType,
+                leftSign,
+                resultSign);
+            return _module.AddInstruction(
+                SpirvOp.LogicalAnd,
+                _boolType,
+                differentSourceSign,
+                resultSignChanged);
+        }
+
         private static bool TryDecodeInlineConstant(uint encoded, out uint value)
         {
             if (encoded == 125)

src/SharpEmu.Libs/VideoOut/VulkanVideoPresenter.cs

@@ -50,6 +50,8 @@ internal sealed record VulkanGuestMemoryBuffer(
 internal sealed record VulkanGuestVertexBuffer(
     uint Location,
     uint ComponentCount,
+    uint DataFormat,
+    uint NumberFormat,
     ulong BaseAddress,
     uint Stride,
     uint OffsetBytes,
@@ -163,6 +165,8 @@ internal static unsafe class VulkanVideoPresenter
     private static readonly object _gate = new();
     private static readonly Queue<object> _pendingGuestWork = new();
     private static readonly Dictionary<ulong, uint> _availableGuestImages = new();
+    private static readonly HashSet<(ulong Address, uint Width, uint Height)>
+        _tracedGuestImageSubmissions = [];
     private static Thread? _thread;
     private static Presentation? _latestPresentation;
     private static byte[]? _copyFragmentSpirv;
@@ -558,44 +562,38 @@ internal static unsafe class VulkanVideoPresenter
         uint height,
         uint pitchInPixel)
     {
-        uint format;
+        var traceSubmission = false;
         lock (_gate)
         {
             if (_closed ||
-                !_availableGuestImages.TryGetValue(address, out format))
+                !_availableGuestImages.ContainsKey(address))
             {
                 return false;
             }
+
+            traceSubmission =
+                _tracedGuestImageSubmissions.Add((address, width, height));
+            var sequence = (_latestPresentation?.Sequence ?? 0) + 1;
+            _latestPresentation = new Presentation(
+                null,
+                width,
+                height,
+                sequence,
+                GuestDrawKind.None,
+                TranslatedDraw: null,
+                RequiredGuestWorkSequence: 0,
+                IsSplash: false,
+                GuestImageAddress: address);
         }
 
-        var effectivePitch = pitchInPixel == 0 ? width : pitchInPixel;
-
-        Console.Error.WriteLine(
-            $"[LOADER][TRACE] vk.submit_guest_image addr=0x{address:X16} " +
-            $"size={width}x{height} pitch={effectivePitch}");
-
-        if (!TryGetCopyFragmentShader(out var fragmentSpirv))
+        if (traceSubmission)
         {
-            return false;
+            var effectivePitch = pitchInPixel == 0 ? width : pitchInPixel;
+            Console.Error.WriteLine(
+                $"[LOADER][TRACE] vk.submit_guest_image addr=0x{address:X16} " +
+                $"size={width}x{height} pitch={effectivePitch}");
         }
 
-        SubmitTranslatedDraw(
-            fragmentSpirv,
-            [
-                new VulkanGuestDrawTexture(
-                    address,
-                    width,
-                    height,
-                    format,
-                    NumberType: 0,
-                    [],
-                    IsFallback: false,
-                    IsStorage: false),
-            ],
-            [],
-            width,
-            height,
-            attributeCount: 1);
         return true;
     }
 
@@ -807,7 +805,8 @@ internal static unsafe class VulkanVideoPresenter
         GuestDrawKind DrawKind,
         VulkanTranslatedGuestDraw? TranslatedDraw,
         long RequiredGuestWorkSequence,
-        bool IsSplash);
+        bool IsSplash,
+        ulong GuestImageAddress = 0);
 
     private sealed class Presenter : IDisposable
     {
@@ -855,11 +854,16 @@ internal static unsafe class VulkanVideoPresenter
         private bool _firstGuestDrawPresented;
         private bool _splashPresented;
         private bool _swapchainRecreateDeferred;
+        private bool _tracedPresentedSwapchain;
+        private bool _swapchainReadbackPending;
+        private int _directPresentationCount;
         private readonly Dictionary<ulong, GuestImageResource> _guestImages = new();
         private readonly HashSet<(ulong Address, uint Width, uint Height, Format Format)> _tracedTextureCacheHits = new();
         private readonly HashSet<(ulong Address, uint Width, uint Height, Format Format)> _tracedTextureUploads = new();
+        private readonly HashSet<(ulong Address, uint Width, uint Height, uint Format)> _dumpedTextures = new();
         private readonly HashSet<(ulong Address, int Size)> _tracedGlobalBuffers = new();
         private readonly HashSet<ulong> _tracedGuestImageContents = new();
+        private readonly Dictionary<ulong, int> _tracedGuestWriteCounts = new();
         private int _tracedVertexBufferCount;
         private readonly Dictionary<byte[], Pipeline> _computePipelines =
             new(ReferenceEqualityComparer.Instance);
@@ -922,6 +926,8 @@ internal static unsafe class VulkanVideoPresenter
             public ulong Size;
             public uint Location;
             public uint ComponentCount;
+            public uint DataFormat;
+            public uint NumberFormat;
             public uint Stride;
             public uint OffsetBytes;
         }
@@ -1361,7 +1367,10 @@ internal static unsafe class VulkanVideoPresenter
                 ImageColorSpace = surfaceFormat.ColorSpace,
                 ImageExtent = _extent,
                 ImageArrayLayers = 1,
-                ImageUsage = ImageUsageFlags.TransferDstBit | ImageUsageFlags.ColorAttachmentBit,
+                ImageUsage =
+                    ImageUsageFlags.TransferDstBit |
+                    ImageUsageFlags.TransferSrcBit |
+                    ImageUsageFlags.ColorAttachmentBit,
                 ImageSharingMode = SharingMode.Exclusive,
                 PreTransform = capabilities.CurrentTransform,
                 CompositeAlpha = compositeAlpha,
@@ -2233,7 +2242,10 @@ internal static unsafe class VulkanVideoPresenter
                     {
                         Location = vertexBuffer.Location,
                         Binding = (uint)index,
-                        Format = ToVkVertexFormat(vertexBuffer.ComponentCount),
+                        Format = ToVkVertexFormat(
+                            vertexBuffer.DataFormat,
+                            vertexBuffer.NumberFormat,
+                            vertexBuffer.ComponentCount),
                         Offset = 0,
                     };
                 }
@@ -2420,8 +2432,7 @@ internal static unsafe class VulkanVideoPresenter
             var vkFormat = GetTextureFormat(texture.Format, texture.NumberType);
             if (texture.Address != 0 &&
                 _guestImages.TryGetValue(texture.Address, out var guestImage) &&
-                guestImage.Width == texture.Width &&
-                guestImage.Height == texture.Height &&
+                IsCompatibleGuestImageAlias(texture, guestImage) &&
                 IsCompatibleViewFormat(guestImage.Format, vkFormat) &&
                 TryGetOrCreateGuestImageView(
                     guestImage,
@@ -2440,6 +2451,16 @@ internal static unsafe class VulkanVideoPresenter
                         $"image_format={guestImage.Format} view_format={vkFormat}");
                 }
 
+                if (guestImage.Width != texture.Width ||
+                    guestImage.Height != texture.Height)
+                {
+                    TraceVulkanShader(
+                        $"vk.texture_cache_alias addr=0x{texture.Address:X16} " +
+                        $"texture={texture.Width}x{texture.Height} " +
+                        $"image={guestImage.Width}x{guestImage.Height} " +
+                        $"tile={texture.TileMode} format={vkFormat}");
+                }
+
                 return new TextureResource
                 {
                     Address = texture.Address,
@@ -2457,6 +2478,27 @@ internal static unsafe class VulkanVideoPresenter
             return CreateTextureResource(texture);
         }
 
+        private static bool IsCompatibleGuestImageAlias(
+            VulkanGuestDrawTexture texture,
+            GuestImageResource guestImage)
+        {
+            if (guestImage.Width == texture.Width &&
+                guestImage.Height == texture.Height)
+            {
+                return true;
+            }
+
+            if (texture.TileMode == 0 ||
+                texture.Width == 0 ||
+                texture.Height == 0)
+            {
+                return false;
+            }
+
+            return texture.Width <= guestImage.Width &&
+                texture.Height <= guestImage.Height;
+        }
+
         [MethodImpl(MethodImplOptions.NoInlining)]
         private TextureResource ResolveStorageImageResource(VulkanGuestDrawTexture texture)
         {
@@ -2673,6 +2715,7 @@ internal static unsafe class VulkanVideoPresenter
             var pixels = texture.RgbaPixels.Length == (int)expectedSize
                 ? texture.RgbaPixels
                 : CreateFallbackTexturePixels(texture.Format, rowLength, height, expectedSize);
+            DumpTextureUpload(texture, pixels, rowLength, width, height);
             var uploadPixels = texture.Format == 13
                 ? ExpandRgb32Pixels(pixels)
                 : pixels;
@@ -2754,6 +2797,102 @@ internal static unsafe class VulkanVideoPresenter
             };
         }
 
+        private void DumpTextureUpload(
+            VulkanGuestDrawTexture texture,
+            byte[] pixels,
+            uint rowLength,
+            uint width,
+            uint height)
+        {
+            if (!string.Equals(
+                    Environment.GetEnvironmentVariable("SHARPEMU_DUMP_TEXTURES"),
+                    "1",
+                    StringComparison.Ordinal) ||
+                texture.IsFallback ||
+                texture.IsStorage ||
+                GetTextureBytesPerPixel(texture.Format) != 4 ||
+                width == 0 ||
+                height == 0 ||
+                !_dumpedTextures.Add((texture.Address, width, height, texture.Format)))
+            {
+                return;
+            }
+
+            var rowBytes = checked((int)rowLength * 4);
+            var visibleRowBytes = checked((int)width * 4);
+            if (pixels.Length < checked(rowBytes * (int)height))
+            {
+                return;
+            }
+
+            var directory = Path.Combine(AppContext.BaseDirectory, "texture-dumps");
+            Directory.CreateDirectory(directory);
+            var path = Path.Combine(
+                directory,
+                $"tex-{texture.Address:X16}-{width}x{height}-fmt{texture.Format}-row{rowLength}.bmp");
+            WriteRgbaBmp(path, pixels, rowBytes, visibleRowBytes, (int)width, (int)height);
+        }
+
+        private static void WriteRgbaBmp(
+            string path,
+            byte[] rgba,
+            int sourceRowBytes,
+            int visibleRowBytes,
+            int width,
+            int height)
+        {
+            const int fileHeaderSize = 14;
+            const int infoHeaderSize = 40;
+            const int bytesPerPixel = 4;
+            var pixelBytes = checked(width * height * bytesPerPixel);
+            var fileSize = fileHeaderSize + infoHeaderSize + pixelBytes;
+            var output = new byte[fileSize];
+
+            output[0] = (byte)'B';
+            output[1] = (byte)'M';
+            WriteUInt32(output, 2, (uint)fileSize);
+            WriteUInt32(output, 10, fileHeaderSize + infoHeaderSize);
+            WriteUInt32(output, 14, infoHeaderSize);
+            WriteInt32(output, 18, width);
+            WriteInt32(output, 22, -height);
+            WriteUInt16(output, 26, 1);
+            WriteUInt16(output, 28, 32);
+            WriteUInt32(output, 34, (uint)pixelBytes);
+
+            var destinationOffset = fileHeaderSize + infoHeaderSize;
+            for (var y = 0; y < height; y++)
+            {
+                var sourceOffset = y * sourceRowBytes;
+                for (var x = 0; x < visibleRowBytes; x += bytesPerPixel)
+                {
+                    var destination = destinationOffset + y * visibleRowBytes + x;
+                    output[destination + 0] = rgba[sourceOffset + x + 2];
+                    output[destination + 1] = rgba[sourceOffset + x + 1];
+                    output[destination + 2] = rgba[sourceOffset + x + 0];
+                    output[destination + 3] = rgba[sourceOffset + x + 3];
+                }
+            }
+
+            File.WriteAllBytes(path, output);
+        }
+
+        private static void WriteUInt16(byte[] output, int offset, ushort value)
+        {
+            output[offset + 0] = (byte)value;
+            output[offset + 1] = (byte)(value >> 8);
+        }
+
+        private static void WriteUInt32(byte[] output, int offset, uint value)
+        {
+            output[offset + 0] = (byte)value;
+            output[offset + 1] = (byte)(value >> 8);
+            output[offset + 2] = (byte)(value >> 16);
+            output[offset + 3] = (byte)(value >> 24);
+        }
+
+        private static void WriteInt32(byte[] output, int offset, int value) =>
+            WriteUInt32(output, offset, unchecked((uint)value));
+
         private Sampler CreateSampler(VulkanGuestSampler sampler)
         {
             var minLod = DecodeSamplerMipFilter(sampler) == 0
@@ -2872,6 +3011,7 @@ internal static unsafe class VulkanVideoPresenter
                     $"vk.vertex_buffer loc={guestBuffer.Location} " +
                     $"base=0x{guestBuffer.BaseAddress:X16} stride={guestBuffer.Stride} " +
                     $"offset={guestBuffer.OffsetBytes} comps={guestBuffer.ComponentCount} " +
+                    $"fmt={guestBuffer.DataFormat}/num={guestBuffer.NumberFormat} " +
                     $"bytes={guestBuffer.Data.Length}");
             }
 
@@ -2882,6 +3022,8 @@ internal static unsafe class VulkanVideoPresenter
                 Size = size,
                 Location = guestBuffer.Location,
                 ComponentCount = guestBuffer.ComponentCount,
+                DataFormat = guestBuffer.DataFormat,
+                NumberFormat = guestBuffer.NumberFormat,
                 Stride = guestBuffer.Stride,
                 OffsetBytes = guestBuffer.OffsetBytes,
             };
@@ -2931,7 +3073,83 @@ internal static unsafe class VulkanVideoPresenter
                 _ => PrimitiveTopology.TriangleList,
             };
 
-        private static Format ToVkVertexFormat(uint componentCount) =>
+        private static Format ToVkVertexFormat(
+            uint dataFormat,
+            uint numberFormat,
+            uint componentCount) =>
+            (dataFormat, numberFormat) switch
+            {
+                (1, 0) => Format.R8Unorm,
+                (1, 1) => Format.R8SNorm,
+                (1, 4) => Format.R8Uint,
+                (1, 5) => Format.R8Sint,
+                (1, 9) => Format.R8Srgb,
+                (2, 0) => Format.R16Unorm,
+                (2, 1) => Format.R16SNorm,
+                (2, 4) => Format.R16Uint,
+                (2, 5) => Format.R16Sint,
+                (2, 7) => Format.R16Sfloat,
+                (3, 0) => Format.R8G8Unorm,
+                (3, 1) => Format.R8G8SNorm,
+                (3, 4) => Format.R8G8Uint,
+                (3, 5) => Format.R8G8Sint,
+                (3, 9) => Format.R8G8Srgb,
+                (4, 4) => Format.R32Uint,
+                (4, 5) => Format.R32Sint,
+                (4, 7) => Format.R32Sfloat,
+                (5, 0) => Format.R16G16Unorm,
+                (5, 1) => Format.R16G16SNorm,
+                (5, 2) => Format.R16G16Uscaled,
+                (5, 3) => Format.R16G16Sscaled,
+                (5, 4) => Format.R16G16Uint,
+                (5, 5) => Format.R16G16Sint,
+                (5, 7) => Format.R16G16Sfloat,
+                (6, 7) => Format.B10G11R11UfloatPack32,
+                (7, 7) => Format.B10G11R11UfloatPack32,
+                (8, 0) => Format.A2B10G10R10UnormPack32,
+                (8, 1) => Format.A2B10G10R10SNormPack32,
+                (8, 2) => Format.A2B10G10R10UscaledPack32,
+                (8, 3) => Format.A2B10G10R10SscaledPack32,
+                (8, 4) => Format.A2B10G10R10UintPack32,
+                (8, 5) => Format.A2B10G10R10SintPack32,
+                (9, 0) => Format.A2R10G10B10UnormPack32,
+                (9, 1) => Format.A2R10G10B10SNormPack32,
+                (9, 2) => Format.A2R10G10B10UscaledPack32,
+                (9, 3) => Format.A2R10G10B10SscaledPack32,
+                (9, 4) => Format.A2R10G10B10UintPack32,
+                (9, 5) => Format.A2R10G10B10SintPack32,
+                (10, 0) => Format.R8G8B8A8Unorm,
+                (10, 1) => Format.R8G8B8A8SNorm,
+                (10, 2) => Format.R8G8B8A8Uscaled,
+                (10, 3) => Format.R8G8B8A8Sscaled,
+                (10, 4) => Format.R8G8B8A8Uint,
+                (10, 5) => Format.R8G8B8A8Sint,
+                (10, 9) => Format.R8G8B8A8Srgb,
+                (11, 4) => Format.R32G32Uint,
+                (11, 5) => Format.R32G32Sint,
+                (11, 7) => Format.R32G32Sfloat,
+                (12, 0) => Format.R16G16B16A16Unorm,
+                (12, 1) => Format.R16G16B16A16SNorm,
+                (12, 2) => Format.R16G16B16A16Uscaled,
+                (12, 3) => Format.R16G16B16A16Sscaled,
+                (12, 4) => Format.R16G16B16A16Uint,
+                (12, 5) => Format.R16G16B16A16Sint,
+                (12, 6) => Format.R16G16B16A16SNorm,
+                (12, 7) => Format.R16G16B16A16Sfloat,
+                (13, 4) => Format.R32G32B32Uint,
+                (13, 5) => Format.R32G32B32Sint,
+                (13, 7) => Format.R32G32B32Sfloat,
+                (14, 4) => Format.R32G32B32A32Uint,
+                (14, 5) => Format.R32G32B32A32Sint,
+                (14, 7) => Format.R32G32B32A32Sfloat,
+                (16, 0) => Format.B5G6R5UnormPack16,
+                (17, 0) => Format.R5G5B5A1UnormPack16,
+                (19, 0) => Format.R4G4B4A4UnormPack16,
+                (34, 7) => Format.E5B9G9R9UfloatPack32,
+                _ => ToVkFloatVertexFormat(componentCount),
+            };
+
+        private static Format ToVkFloatVertexFormat(uint componentCount) =>
             componentCount switch
             {
                 1 => Format.R32Sfloat,
@@ -3317,7 +3535,7 @@ internal static unsafe class VulkanVideoPresenter
         {
             _stagingBuffer = CreateBuffer(
                 size,
-                BufferUsageFlags.TransferSrcBit,
+                BufferUsageFlags.TransferSrcBit | BufferUsageFlags.TransferDstBit,
                 MemoryPropertyFlags.HostVisibleBit | MemoryPropertyFlags.HostCoherentBit,
                 out _stagingMemory);
             _stagingSize = size;
@@ -3588,6 +3806,27 @@ internal static unsafe class VulkanVideoPresenter
                         _availableGuestImages[target.Address] = guestTextureFormat;
                     }
                 }
+                if (ShouldTraceGuestImageWriteForDiagnostics(target.Address))
+                {
+                    var writeCount = _tracedGuestWriteCounts.TryGetValue(
+                        target.Address,
+                        out var previousCount)
+                        ? previousCount + 1
+                        : 1;
+                    _tracedGuestWriteCounts[target.Address] = writeCount;
+                    if (writeCount <= 3)
+                    {
+                        _commandBuffer = _presentationCommandBuffer;
+                        Check(
+                            _vk.QueueWaitIdle(_queue),
+                            "vkQueueWaitIdle(guest write trace)");
+                        Console.Error.WriteLine(
+                            $"[LOADER][TRACE] vk.guest_write_sample " +
+                            $"addr=0x{target.Address:X16} write={writeCount} " +
+                            $"ps_bytes={work.Draw.PixelSpirv.Length}");
+                        TraceGuestImageContents(target);
+                    }
+                }
                 TraceVulkanShader(
                     $"vk.offscreen_draw addr=0x{target.Address:X16} " +
                     $"size={target.Width}x{target.Height} format={target.Format} " +
@@ -4051,7 +4290,8 @@ internal static unsafe class VulkanVideoPresenter
 
             if (presentation.Pixels is null &&
                 presentation.DrawKind != GuestDrawKind.FullscreenBarycentric &&
-                presentation.TranslatedDraw is null)
+                presentation.TranslatedDraw is null &&
+                presentation.GuestImageAddress == 0)
             {
                 return;
             }
@@ -4074,6 +4314,28 @@ internal static unsafe class VulkanVideoPresenter
             }
 
             TranslatedDrawResources? translatedResources = null;
+            GuestImageResource? presentedGuestImage = null;
+            if (presentation.GuestImageAddress != 0 &&
+                (!_guestImages.TryGetValue(
+                    presentation.GuestImageAddress,
+                    out presentedGuestImage) ||
+                 !presentedGuestImage.Initialized))
+            {
+                return;
+            }
+            if (presentedGuestImage is not null)
+            {
+                _directPresentationCount++;
+                if (ShouldTracePresentedGuestImageContentsForDiagnostics() &&
+                    _directPresentationCount is 1 or 30 or 120)
+                {
+                    Console.Error.WriteLine(
+                        $"[LOADER][TRACE] vk.present_sample frame={_directPresentationCount} " +
+                        $"addr=0x{presentedGuestImage.Address:X16}");
+                    TraceGuestImageContents(presentedGuestImage);
+                }
+            }
+
             if (presentation.TranslatedDraw is { } translatedDraw)
             {
                 try
@@ -4175,6 +4437,11 @@ internal static unsafe class VulkanVideoPresenter
                 _vk.CmdEndRenderPass(_commandBuffer);
                 waitStage = PipelineStageFlags.ColorAttachmentOutputBit;
             }
+            else if (presentedGuestImage is not null)
+            {
+                RecordGuestImageBlit(imageIndex, presentedGuestImage);
+                waitStage = PipelineStageFlags.TransferBit;
+            }
             else if (translatedResources is not null)
             {
                 RecordTranslatedDraw(imageIndex, translatedResources);
@@ -4231,6 +4498,10 @@ internal static unsafe class VulkanVideoPresenter
             CheckSwapchainResult(presentResult, "vkQueuePresentKHR");
             recreateAfterPresent |= presentResult == Result.SuboptimalKhr;
             Check(_vk.QueueWaitIdle(_queue), "vkQueueWaitIdle");
+            if (_swapchainReadbackPending)
+            {
+                TraceSwapchainReadback();
+            }
             CollectCompletedGuestSubmissions(waitForOldest: false);
             if (translatedResources is not null)
             {
@@ -4259,8 +4530,11 @@ internal static unsafe class VulkanVideoPresenter
             {
                 _firstGuestDrawPresented = true;
                 Console.Error.WriteLine(
-                    $"[LOADER][INFO] Vulkan VideoOut presented translated guest draw: " +
-                    (presentation.TranslatedDraw is null
+                    $"[LOADER][INFO] Vulkan VideoOut presented guest frame: " +
+                    (presentedGuestImage is not null
+                        ? $"image=0x{presentedGuestImage.Address:X16} " +
+                          $"{presentedGuestImage.Width}x{presentedGuestImage.Height}"
+                        : presentation.TranslatedDraw is null
                         ? $"{presentation.DrawKind}"
                         : $"shader textures={presentation.TranslatedDraw.Textures.Count}"));
             }
@@ -4760,7 +5034,23 @@ internal static unsafe class VulkanVideoPresenter
 
         private static bool ShouldTraceGuestImageAddressForDiagnostics(ulong address)
         {
-            var addresses = Environment.GetEnvironmentVariable("SHARPEMU_TRACE_GUEST_IMAGE_ADDRS");
+            return AddressListContains(
+                "SHARPEMU_TRACE_GUEST_IMAGE_ADDRS",
+                address);
+        }
+
+        private static bool ShouldTraceGuestImageWriteForDiagnostics(ulong address)
+        {
+            return AddressListContains(
+                "SHARPEMU_TRACE_GUEST_WRITES",
+                address);
+        }
+
+        private static bool AddressListContains(
+            string environmentVariable,
+            ulong address)
+        {
+            var addresses = Environment.GetEnvironmentVariable(environmentVariable);
             if (string.IsNullOrWhiteSpace(addresses))
             {
                 return false;
@@ -4770,6 +5060,11 @@ internal static unsafe class VulkanVideoPresenter
                          [',', ';', ' ', '\t'],
                          StringSplitOptions.RemoveEmptyEntries | StringSplitOptions.TrimEntries))
             {
+                if (token == "*")
+                {
+                    return true;
+                }
+
                 var span = token.AsSpan();
                 if (span.StartsWith("0x", StringComparison.OrdinalIgnoreCase))
                 {
@@ -5105,6 +5400,234 @@ internal static unsafe class VulkanVideoPresenter
                 &toPresent);
         }
 
+        private void RecordGuestImageBlit(
+            uint imageIndex,
+            GuestImageResource source)
+        {
+            var traceDestination =
+                ShouldTracePresentedGuestImageContentsForDiagnostics() &&
+                !_tracedPresentedSwapchain;
+            _tracedPresentedSwapchain |= traceDestination;
+            BeginDebugLabel(
+                _commandBuffer,
+                $"SharpEmu present image 0x{source.Address:X16}");
+
+            var sourceToTransfer = new ImageMemoryBarrier
+            {
+                SType = StructureType.ImageMemoryBarrier,
+                SrcAccessMask = AccessFlags.ShaderReadBit,
+                DstAccessMask = AccessFlags.TransferReadBit,
+                OldLayout = ImageLayout.ShaderReadOnlyOptimal,
+                NewLayout = ImageLayout.TransferSrcOptimal,
+                SrcQueueFamilyIndex = Vk.QueueFamilyIgnored,
+                DstQueueFamilyIndex = Vk.QueueFamilyIgnored,
+                Image = source.Image,
+                SubresourceRange = ColorSubresourceRange(),
+            };
+            var destinationToTransfer = new ImageMemoryBarrier
+            {
+                SType = StructureType.ImageMemoryBarrier,
+                SrcAccessMask = _imageInitialized[imageIndex]
+                    ? AccessFlags.MemoryReadBit
+                    : 0,
+                DstAccessMask = AccessFlags.TransferWriteBit,
+                OldLayout = _imageInitialized[imageIndex]
+                    ? ImageLayout.PresentSrcKhr
+                    : ImageLayout.Undefined,
+                NewLayout = ImageLayout.TransferDstOptimal,
+                SrcQueueFamilyIndex = Vk.QueueFamilyIgnored,
+                DstQueueFamilyIndex = Vk.QueueFamilyIgnored,
+                Image = _swapchainImages[imageIndex],
+                SubresourceRange = ColorSubresourceRange(),
+            };
+            var barriers = stackalloc ImageMemoryBarrier[2];
+            barriers[0] = sourceToTransfer;
+            barriers[1] = destinationToTransfer;
+            _vk.CmdPipelineBarrier(
+                _commandBuffer,
+                PipelineStageFlags.AllCommandsBit,
+                PipelineStageFlags.TransferBit,
+                0,
+                0,
+                null,
+                0,
+                null,
+                2,
+                barriers);
+
+            var sourceOffsets = new ImageBlit.SrcOffsetsBuffer
+            {
+                Element0 = new Offset3D(0, 0, 0),
+                Element1 = new Offset3D(
+                    checked((int)source.Width),
+                    checked((int)source.Height),
+                    1),
+            };
+            var destinationOffsets = new ImageBlit.DstOffsetsBuffer
+            {
+                Element0 = new Offset3D(0, 0, 0),
+                Element1 = new Offset3D(
+                    checked((int)_extent.Width),
+                    checked((int)_extent.Height),
+                    1),
+            };
+            var region = new ImageBlit
+            {
+                SrcSubresource = new ImageSubresourceLayers(
+                    ImageAspectFlags.ColorBit,
+                    0,
+                    0,
+                    1),
+                SrcOffsets = sourceOffsets,
+                DstSubresource = new ImageSubresourceLayers(
+                    ImageAspectFlags.ColorBit,
+                    0,
+                    0,
+                    1),
+                DstOffsets = destinationOffsets,
+            };
+            _vk.CmdBlitImage(
+                _commandBuffer,
+                source.Image,
+                ImageLayout.TransferSrcOptimal,
+                _swapchainImages[imageIndex],
+                ImageLayout.TransferDstOptimal,
+                1,
+                &region,
+                Filter.Nearest);
+
+            if (traceDestination)
+            {
+                var destinationToReadback = new ImageMemoryBarrier
+                {
+                    SType = StructureType.ImageMemoryBarrier,
+                    SrcAccessMask = AccessFlags.TransferWriteBit,
+                    DstAccessMask = AccessFlags.TransferReadBit,
+                    OldLayout = ImageLayout.TransferDstOptimal,
+                    NewLayout = ImageLayout.TransferSrcOptimal,
+                    SrcQueueFamilyIndex = Vk.QueueFamilyIgnored,
+                    DstQueueFamilyIndex = Vk.QueueFamilyIgnored,
+                    Image = _swapchainImages[imageIndex],
+                    SubresourceRange = ColorSubresourceRange(),
+                };
+                _vk.CmdPipelineBarrier(
+                    _commandBuffer,
+                    PipelineStageFlags.TransferBit,
+                    PipelineStageFlags.TransferBit,
+                    0,
+                    0,
+                    null,
+                    0,
+                    null,
+                    1,
+                    &destinationToReadback);
+
+                var copyRegion = new BufferImageCopy
+                {
+                    ImageSubresource = new ImageSubresourceLayers
+                    {
+                        AspectMask = ImageAspectFlags.ColorBit,
+                        LayerCount = 1,
+                    },
+                    ImageExtent = new Extent3D(_extent.Width, _extent.Height, 1),
+                };
+                _vk.CmdCopyImageToBuffer(
+                    _commandBuffer,
+                    _swapchainImages[imageIndex],
+                    ImageLayout.TransferSrcOptimal,
+                    _stagingBuffer,
+                    1,
+                    &copyRegion);
+                _swapchainReadbackPending = true;
+            }
+
+            var sourceToShaderRead = new ImageMemoryBarrier
+            {
+                SType = StructureType.ImageMemoryBarrier,
+                SrcAccessMask = AccessFlags.TransferReadBit,
+                DstAccessMask = AccessFlags.ShaderReadBit,
+                OldLayout = ImageLayout.TransferSrcOptimal,
+                NewLayout = ImageLayout.ShaderReadOnlyOptimal,
+                SrcQueueFamilyIndex = Vk.QueueFamilyIgnored,
+                DstQueueFamilyIndex = Vk.QueueFamilyIgnored,
+                Image = source.Image,
+                SubresourceRange = ColorSubresourceRange(),
+            };
+            var destinationToPresent = new ImageMemoryBarrier
+            {
+                SType = StructureType.ImageMemoryBarrier,
+                SrcAccessMask = traceDestination
+                    ? AccessFlags.TransferReadBit
+                    : AccessFlags.TransferWriteBit,
+                DstAccessMask = AccessFlags.MemoryReadBit,
+                OldLayout = traceDestination
+                    ? ImageLayout.TransferSrcOptimal
+                    : ImageLayout.TransferDstOptimal,
+                NewLayout = ImageLayout.PresentSrcKhr,
+                SrcQueueFamilyIndex = Vk.QueueFamilyIgnored,
+                DstQueueFamilyIndex = Vk.QueueFamilyIgnored,
+                Image = _swapchainImages[imageIndex],
+                SubresourceRange = ColorSubresourceRange(),
+            };
+            barriers[0] = sourceToShaderRead;
+            barriers[1] = destinationToPresent;
+            _vk.CmdPipelineBarrier(
+                _commandBuffer,
+                PipelineStageFlags.TransferBit,
+                PipelineStageFlags.AllCommandsBit,
+                0,
+                0,
+                null,
+                0,
+                null,
+                2,
+                barriers);
+            EndDebugLabel(_commandBuffer);
+        }
+
+        private void TraceSwapchainReadback()
+        {
+            _swapchainReadbackPending = false;
+            var byteCount = checked((ulong)_extent.Width * _extent.Height * 4);
+            void* mapped;
+            Check(
+                _vk.MapMemory(_device, _stagingMemory, 0, byteCount, 0, &mapped),
+                "vkMapMemory(swapchain readback)");
+            try
+            {
+                var bytes = new ReadOnlySpan<byte>(mapped, checked((int)byteCount));
+                var nonzeroBytes = 0L;
+                var nonblackPixels = 0L;
+                ulong hash = 14695981039346656037UL;
+                for (var offset = 0; offset < bytes.Length; offset += 4)
+                {
+                    var b0 = bytes[offset];
+                    var b1 = bytes[offset + 1];
+                    var b2 = bytes[offset + 2];
+                    var b3 = bytes[offset + 3];
+                    nonzeroBytes += b0 == 0 ? 0 : 1;
+                    nonzeroBytes += b1 == 0 ? 0 : 1;
+                    nonzeroBytes += b2 == 0 ? 0 : 1;
+                    nonzeroBytes += b3 == 0 ? 0 : 1;
+                    nonblackPixels += b0 != 0 || b1 != 0 || b2 != 0 ? 1 : 0;
+                    hash = (hash ^ b0) * 1099511628211UL;
+                    hash = (hash ^ b1) * 1099511628211UL;
+                    hash = (hash ^ b2) * 1099511628211UL;
+                    hash = (hash ^ b3) * 1099511628211UL;
+                }
+
+                Console.Error.WriteLine(
+                    $"[LOADER][TRACE] vk.swapchain_image size={_extent.Width}x{_extent.Height} " +
+                    $"format={_swapchainFormat} nonzero_bytes={nonzeroBytes}/{byteCount} " +
+                    $"nonblack_pixels={nonblackPixels}/{(ulong)_extent.Width * _extent.Height} " +
+                    $"hash=0x{hash:X16}");
+            }
+            finally
+            {
+                _vk.UnmapMemory(_device, _stagingMemory);
+            }
+        }
+
         private Extent2D ChooseExtent(SurfaceCapabilitiesKHR capabilities)
         {
             if (capabilities.CurrentExtent.Width != uint.MaxValue)