mirror of
https://github.com/par274/sharpemu.git
synced 2026-07-30 22:49:53 +08:00
feat(gpu): GPU compute detile for guest tiled textures (Vulkan + Metal) (#592)
* feat(gpu): GPU compute detile for guest tiled textures (Vulkan + Metal) Move RDNA2 exact-XOR deswizzle (swizzle modes 5/9/24/27, 4bpp) off the CPU onto a GPU compute pass. GnmTiling.GetDetileParams resolves the shared addressing into DetileParams; the CPU fallback and both GPU kernels consume the same params so they never disagree. Vulkan (verified bit-exact on NVIDIA): SpirvFixedShaders.CreateDetileCompute hand-emits the SPIR-V kernel; VulkanDetilePass.RecordDetile records the dispatch into the async batch command buffer (never a blocking submit on the render thread) with transients retired via fence; VulkanDetileSelfTest (SHARPEMU_DETILE_SELFTEST=1) checks both entry points against the CPU detile. Metal (Mac-untested): detile_compute.msl (detile_cs) + MetalDetilePass mirror the Vulkan pass. The active Metal path CPU-detiles via the new GnmTiling.DetileWithParams when a texture arrives packaged (empty RgbaPixels + TiledSource/Detile), keeping Metal correct under default-on with no regression; wiring MetalDetilePass live is the remaining on-device step. Flags: GPU detile is default-on (SHARPEMU_GPU_DETILE=0 disables); [GPU-DETILE] diagnostics gated behind SHARPEMU_LOG_GPU_DETILE=1. Tests: 17 detile unit tests pass, incl. DetileWithParams and GetDetileParams each matching TryDetile bit-for-bit across all supported modes/bpp, plus a SPIR-V structural-validity test. * feat(gpu): GPU compute detile for guest tiled textures (Vulkan + Metal) Move RDNA2 exact-XOR deswizzle (swizzle modes 5/9/24/27, 4bpp) off the CPU onto a GPU compute pass. GnmTiling.GetDetileParams resolves the shared addressing into DetileParams that the CPU fallback and both GPU kernels consume, so they never disagree; everything else keeps the CPU path. Vulkan (verified bit-exact on NVIDIA): SpirvFixedShaders.CreateDetileCompute hand-emits the kernel; VulkanDetilePass.RecordDetile records into the async batch command buffer (never a blocking submit on the render thread) with transients retired via fence, falling back to CPU detile on failure. VulkanDetileSelfTest (SHARPEMU_DETILE_SELFTEST=1) checks both entry points. Metal (Mac-untested): detile_compute.msl + MetalDetilePass mirror the Vulkan pass; the active Metal path CPU-detiles via GnmTiling.DetileWithParams so it stays correct under default-on. Wiring MetalDetilePass live is a follow-up. Flags: default-on (SHARPEMU_GPU_DETILE=0 disables); diagnostics behind SHARPEMU_LOG_GPU_DETILE=1. Adds 17 passing detile unit tests. * Fix: added support layered texture support for the GPU-Detiling. * Fix: Added support for BlockTable (1 / 4 / 8 (Morton/Z-order)) * feat: added support for 8 and 16 bpp (bytes per element) * Fixed a build failure specific to this branch ---------
This commit is contained in:
shadowbeat070
@@ -0,0 +1,102 @@
|
||||
// Copyright (C) 2026 SharpEmu Emulator Project
|
||||
// SPDX-License-Identifier: GPL-2.0-or-later
|
||||
|
||||
using System.Buffers.Binary;
|
||||
using SharpEmu.ShaderCompiler.Vulkan;
|
||||
using Xunit;
|
||||
|
||||
namespace SharpEmu.Libs.Tests.Agc;
|
||||
|
||||
// Structural validation of the GPU detile compute kernel. This cannot run the
|
||||
// shader without a Vulkan device, but it pins the SPIR-V is well-formed: the
|
||||
// header is correct, every instruction's word count sums to exactly the module
|
||||
// length (the classic hand-emit bug), and the compute-specific pieces are
|
||||
// present (a GLCompute entry point, a LocalSize execution mode, and a runtime
|
||||
// array for the storage buffers). Full pixel correctness is verified on a GPU.
|
||||
public sealed class DetileComputeSpirvTests
|
||||
{
|
||||
private const uint SpirvMagic = 0x07230203;
|
||||
private const uint SpirvVersion15 = 0x00010500;
|
||||
|
||||
private const ushort OpEntryPoint = 15;
|
||||
private const ushort OpExecutionMode = 16;
|
||||
private const ushort OpTypeRuntimeArray = 29;
|
||||
private const ushort OpFunction = 54;
|
||||
private const ushort OpFunctionEnd = 56;
|
||||
|
||||
private const uint ExecutionModelGLCompute = 5;
|
||||
private const uint ExecutionModeLocalSize = 17;
|
||||
|
||||
[Fact]
|
||||
public void CreateDetileCompute_EmitsWellFormedComputeModule()
|
||||
{
|
||||
var spirv = SpirvFixedShaders.CreateDetileCompute();
|
||||
|
||||
Assert.True(spirv.Length % sizeof(uint) == 0, "SPIR-V must be a whole number of words.");
|
||||
var words = new uint[spirv.Length / sizeof(uint)];
|
||||
for (var i = 0; i < words.Length; i++)
|
||||
{
|
||||
words[i] = BinaryPrimitives.ReadUInt32LittleEndian(spirv.AsSpan(i * sizeof(uint)));
|
||||
}
|
||||
|
||||
Assert.True(words.Length > 5, "Module must have a header plus instructions.");
|
||||
Assert.Equal(SpirvMagic, words[0]);
|
||||
Assert.Equal(SpirvVersion15, words[1]);
|
||||
|
||||
var bound = words[3];
|
||||
Assert.True(bound > 1, "Id bound must be set.");
|
||||
|
||||
var sawComputeEntry = false;
|
||||
var sawLocalSize = false;
|
||||
var sawRuntimeArray = false;
|
||||
var functionCount = 0;
|
||||
var functionEndCount = 0;
|
||||
|
||||
var offset = 5;
|
||||
while (offset < words.Length)
|
||||
{
|
||||
var word = words[offset];
|
||||
var wordCount = (int)(word >> 16);
|
||||
var opcode = (ushort)(word & 0xFFFF);
|
||||
|
||||
Assert.True(wordCount >= 1, $"Instruction at {offset} has a zero word count.");
|
||||
Assert.True(
|
||||
offset + wordCount <= words.Length,
|
||||
$"Instruction at {offset} (op {opcode}, wc {wordCount}) overruns the module.");
|
||||
|
||||
switch (opcode)
|
||||
{
|
||||
case OpEntryPoint when words[offset + 1] == ExecutionModelGLCompute:
|
||||
sawComputeEntry = true;
|
||||
break;
|
||||
case OpExecutionMode
|
||||
when wordCount >= 6 &&
|
||||
words[offset + 2] == ExecutionModeLocalSize &&
|
||||
words[offset + 3] == 8 &&
|
||||
words[offset + 4] == 8 &&
|
||||
words[offset + 5] == 1:
|
||||
sawLocalSize = true;
|
||||
break;
|
||||
case OpTypeRuntimeArray:
|
||||
sawRuntimeArray = true;
|
||||
break;
|
||||
case OpFunction:
|
||||
functionCount++;
|
||||
break;
|
||||
case OpFunctionEnd:
|
||||
functionEndCount++;
|
||||
break;
|
||||
}
|
||||
|
||||
offset += wordCount;
|
||||
}
|
||||
|
||||
// Word counts must tile the module exactly — a wrong length lands here.
|
||||
Assert.Equal(words.Length, offset);
|
||||
Assert.True(sawComputeEntry, "Missing a GLCompute OpEntryPoint.");
|
||||
Assert.True(sawLocalSize, "Missing an 8x8x1 LocalSize execution mode.");
|
||||
Assert.True(sawRuntimeArray, "Missing a runtime array (storage buffers).");
|
||||
Assert.Equal(1, functionCount);
|
||||
Assert.Equal(1, functionEndCount);
|
||||
}
|
||||
}
|
||||
@@ -83,4 +83,152 @@ public sealed class GnmTilingDetileTests
|
||||
Assert.Equal((ushort)i, value);
|
||||
}
|
||||
}
|
||||
|
||||
// GetDetileParams must reproduce TryDetile bit-for-bit: the CPU fallback and
|
||||
// the GPU compute kernel both consume these params, so a detile driven purely
|
||||
// by DetileParams (the shared addressing formula the kernel runs) must equal
|
||||
// the shipped CPU detile for every supported mode/bpp.
|
||||
[Theory]
|
||||
[InlineData(27u, 2, 384, 200)] // 64 KiB RB+ R_X (exact-XOR)
|
||||
[InlineData(27u, 4, 256, 256)] // 64 KiB RB+ R_X (exact-XOR)
|
||||
[InlineData(9u, 4, 300, 300)] // 64 KiB standard (exact-XOR)
|
||||
[InlineData(24u, 4, 128, 256)] // 64 KiB RB+ Z_X (exact-XOR)
|
||||
[InlineData(5u, 4, 200, 120)] // 4 KiB standard (exact-XOR)
|
||||
[InlineData(8u, 4, 128, 128)] // 64 KiB Z (block-table path)
|
||||
[InlineData(1u, 4, 64, 64)] // 256 B standard (block-table path)
|
||||
public void GetDetileParams_ReproducesTryDetile(uint mode, int bpp, int w, int h)
|
||||
{
|
||||
var p = GnmTiling.GetDetileParams(mode, bpp, w, h);
|
||||
Assert.True(p.IsSupported);
|
||||
|
||||
// Whole-block tiled buffer (block addressing overshoots the linear extent),
|
||||
// filled with a deterministic non-trivial pattern.
|
||||
var blocksHigh = (h + p.BlockHeight - 1) / p.BlockHeight;
|
||||
var tiled = new byte[(long)p.BlocksPerRow * blocksHigh * p.BlockBytes];
|
||||
for (var i = 0; i < tiled.Length; i++)
|
||||
{
|
||||
tiled[i] = (byte)((i * 31 + 7) & 0xFF);
|
||||
}
|
||||
|
||||
var expected = new byte[w * h * bpp];
|
||||
Assert.True(GnmTiling.TryDetile(tiled, expected, mode, w, h, bpp));
|
||||
|
||||
var actual = DetileViaParams(tiled, p, w, h, bpp);
|
||||
Assert.Equal(expected, actual);
|
||||
}
|
||||
|
||||
// The production GnmTiling.DetileWithParams (the active CPU fallback used by
|
||||
// the Metal path under default-on GPU detile) must equal TryDetile for every
|
||||
// supported mode/bpp — same DetileParams addressing, no re-derived swizzle.
|
||||
[Theory]
|
||||
[InlineData(27u, 2, 384, 200)]
|
||||
[InlineData(27u, 4, 256, 256)]
|
||||
[InlineData(9u, 4, 300, 300)]
|
||||
[InlineData(24u, 4, 128, 256)]
|
||||
[InlineData(5u, 4, 200, 120)]
|
||||
[InlineData(8u, 4, 128, 128)]
|
||||
[InlineData(1u, 4, 64, 64)]
|
||||
[InlineData(27u, 8, 256, 256)] // 8bpp (GPU: 2 words/element)
|
||||
[InlineData(27u, 16, 128, 128)] // 16bpp (GPU: 4 words/element)
|
||||
[InlineData(9u, 8, 128, 96)]
|
||||
[InlineData(8u, 16, 64, 64)] // block-table, 16bpp
|
||||
public void DetileWithParams_MatchesTryDetile(uint mode, int bpp, int w, int h)
|
||||
{
|
||||
var p = GnmTiling.GetDetileParams(mode, bpp, w, h);
|
||||
Assert.True(p.IsSupported);
|
||||
|
||||
var blocksHigh = (h + p.BlockHeight - 1) / p.BlockHeight;
|
||||
var tiled = new byte[(long)p.BlocksPerRow * blocksHigh * p.BlockBytes];
|
||||
for (var i = 0; i < tiled.Length; i++)
|
||||
{
|
||||
tiled[i] = (byte)((i * 31 + 7) & 0xFF);
|
||||
}
|
||||
|
||||
var expected = new byte[w * h * bpp];
|
||||
Assert.True(GnmTiling.TryDetile(tiled, expected, mode, w, h, bpp));
|
||||
|
||||
var actual = new byte[w * h * bpp];
|
||||
Assert.True(GnmTiling.DetileWithParams(p, tiled, actual));
|
||||
Assert.Equal(expected, actual);
|
||||
}
|
||||
|
||||
// Array textures are packed as contiguous tiled slices and detiled one slice
|
||||
// per layer (dispatch-Z on the GPU; a per-layer loop in the CPU fallbacks)
|
||||
// into a layer-major linear buffer. This pins that packing: each slice must
|
||||
// deswizzle into its own region and match a per-slice TryDetile, and a
|
||||
// per-layer-distinct pattern catches any slice cross-talk.
|
||||
[Theory]
|
||||
[InlineData(27u, 4, 256, 256, 3)]
|
||||
[InlineData(9u, 4, 128, 96, 2)]
|
||||
[InlineData(24u, 4, 64, 128, 4)]
|
||||
public void DetileWithParams_MultiLayer_MatchesPerSliceTryDetile(uint mode, int bpp, int w, int h, int layers)
|
||||
{
|
||||
var p = GnmTiling.GetDetileParams(mode, bpp, w, h);
|
||||
Assert.True(p.IsSupported);
|
||||
|
||||
var blocksHigh = (h + p.BlockHeight - 1) / p.BlockHeight;
|
||||
var sliceTiledBytes = (int)((long)p.BlocksPerRow * blocksHigh * p.BlockBytes);
|
||||
var sliceLinearBytes = w * h * bpp;
|
||||
|
||||
var tiled = new byte[sliceTiledBytes * layers];
|
||||
for (var layer = 0; layer < layers; layer++)
|
||||
{
|
||||
for (var i = 0; i < sliceTiledBytes; i++)
|
||||
{
|
||||
tiled[layer * sliceTiledBytes + i] = (byte)((i * 31 + 7 + layer * 101) & 0xFF);
|
||||
}
|
||||
}
|
||||
|
||||
// Expected: each slice detiled independently via the shipped CPU detile.
|
||||
var expected = new byte[sliceLinearBytes * layers];
|
||||
for (var layer = 0; layer < layers; layer++)
|
||||
{
|
||||
Assert.True(GnmTiling.TryDetile(
|
||||
tiled.AsSpan(layer * sliceTiledBytes, sliceTiledBytes),
|
||||
expected.AsSpan(layer * sliceLinearBytes, sliceLinearBytes),
|
||||
mode, w, h, bpp));
|
||||
}
|
||||
|
||||
// Actual: the layer-major loop the Vulkan/Metal CPU fallbacks run.
|
||||
var actual = new byte[sliceLinearBytes * layers];
|
||||
for (var layer = 0; layer < layers; layer++)
|
||||
{
|
||||
Assert.True(GnmTiling.DetileWithParams(
|
||||
p,
|
||||
tiled.AsSpan(layer * sliceTiledBytes, sliceTiledBytes),
|
||||
actual.AsSpan(layer * sliceLinearBytes, sliceLinearBytes)));
|
||||
}
|
||||
|
||||
Assert.Equal(expected, actual);
|
||||
}
|
||||
|
||||
// Reference detile driven entirely by DetileParams — the single shared
|
||||
// addressing formula the Vulkan/Metal compute kernel will run per texel.
|
||||
private static byte[] DetileViaParams(byte[] tiled, DetileParams p, int w, int h, int bpp)
|
||||
{
|
||||
var linear = new byte[w * h * bpp];
|
||||
for (var y = 0; y < h; y++)
|
||||
{
|
||||
for (var x = 0; x < w; x++)
|
||||
{
|
||||
var blockX = x / p.BlockWidth;
|
||||
var blockY = y / p.BlockHeight;
|
||||
var inX = x % p.BlockWidth;
|
||||
var inY = y % p.BlockHeight;
|
||||
var inBlockByte = p.Equation == DetileEquation.ExactXor
|
||||
? p.XByteTerm[x & p.XMask] ^ p.YByteTerm[y & p.YMask]
|
||||
: p.BlockTable[inY * p.BlockWidth + inX] * p.BytesPerElement;
|
||||
var srcByte = ((long)blockY * p.BlocksPerRow + blockX) * p.BlockBytes + inBlockByte;
|
||||
var dstByte = ((long)y * w + x) * bpp;
|
||||
if (srcByte < 0 || srcByte + bpp > tiled.Length)
|
||||
{
|
||||
continue;
|
||||
}
|
||||
|
||||
Array.Copy(tiled, srcByte, linear, dstByte, bpp);
|
||||
}
|
||||
}
|
||||
|
||||
return linear;
|
||||
}
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user