ShuriZma
/
xenia
mirror of https://github.com/xenia-project/xenia.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

[D3D12] More texture formats, rename blend map

This commit is contained in:
Triang3l 2018-08-12 22:58:45 +03:00
parent 34a1846859
commit 01a3bc32f7
5 changed files with 84 additions and 15 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
src/xenia/gpu/d3d12/shaders/byte_swap.hlsli
View File

@ -16,4 +16,16 @@ XE_BYTE_SWAP_OVERLOAD(uint2)
XE_BYTE_SWAP_OVERLOAD(uint3)
XE_BYTE_SWAP_OVERLOAD(uint4)
#define XE_BYTE_SWAP_16_OVERLOAD(XeByteSwapType) \
XeByteSwapType XeByteSwap16(XeByteSwapType v, uint endian) { \
[flatten] if (((endian ^ (endian >> 1u)) & 1u) != 0u) { \
v = (v << 8u) | (v >> 8u); \
} \
return v; \
}
XE_BYTE_SWAP_16_OVERLOAD(uint)
XE_BYTE_SWAP_16_OVERLOAD(uint2)
XE_BYTE_SWAP_16_OVERLOAD(uint3)
XE_BYTE_SWAP_16_OVERLOAD(uint4)
#endif // XENIA_GPU_D3D12_SHADERS_BYTE_SWAP_HLSLI_

25
src/xenia/gpu/d3d12/shaders/texture_load_16bpb.cs.hlsl Normal file
View File

@ -0,0 +1,25 @@
#include "texture_copy.hlsli"
[numthreads(8, 32, 1)]
void main(uint3 xe_thread_id : SV_DispatchThreadID) {
// 1 thread = 4 ushort blocks.
uint3 block_index = xe_thread_id;
block_index.x <<= 2u;
[branch] if (any(block_index >= xe_texture_copy_size)) {
return;
}
uint4 block_offsets_guest =
XeTextureCopyGuestBlockOffsets(block_index, 2u, 1u);
uint4 dword_offsets_guest = block_offsets_guest & ~3u;
uint4 blocks = uint4(xe_texture_copy_source.Load(dword_offsets_guest.x),
xe_texture_copy_source.Load(dword_offsets_guest.y),
xe_texture_copy_source.Load(dword_offsets_guest.z),
xe_texture_copy_source.Load(dword_offsets_guest.w));
blocks = (blocks >> ((block_offsets_guest & 2u) << 3u)) & 0xFFFFu;
blocks = XeByteSwap16(blocks, xe_texture_copy_endianness);
uint block_offset_host = XeTextureHostLinearOffset(
block_index, xe_texture_copy_size.y, xe_texture_copy_host_pitch, 2u) +
xe_texture_copy_host_base;
xe_texture_copy_dest.Store2(block_offset_host,
blocks.xz | (blocks.yw << 16u));
}

26
src/xenia/gpu/d3d12/shaders/texture_load_8bpb.cs.hlsl Normal file
View File

@ -0,0 +1,26 @@
#include "texture_copy.hlsli"
[numthreads(8, 32, 1)]
void main(uint3 xe_thread_id : SV_DispatchThreadID) {
// 1 thread = 4 ubyte blocks.
uint3 block_index = xe_thread_id;
block_index.x <<= 2u;
[branch] if (any(block_index >= xe_texture_copy_size)) {
return;
}
uint4 block_offsets_guest =
XeTextureCopyGuestBlockOffsets(block_index, 1u, 0u);
uint4 dword_offsets_guest = block_offsets_guest & ~3u;
uint4 blocks = uint4(xe_texture_copy_source.Load(dword_offsets_guest.x),
xe_texture_copy_source.Load(dword_offsets_guest.y),
xe_texture_copy_source.Load(dword_offsets_guest.z),
xe_texture_copy_source.Load(dword_offsets_guest.w));
blocks = (blocks >> ((block_offsets_guest & 3u) << 3u)) & 0xFFu;
blocks <<= uint4(0u, 8u, 16u, 24u);
blocks.xy |= blocks.zw;
blocks.x |= blocks.y;
uint block_offset_host = XeTextureHostLinearOffset(
block_index, xe_texture_copy_size.y, xe_texture_copy_host_pitch, 1u) +
xe_texture_copy_host_base;
xe_texture_copy_dest.Store(block_offset_host, blocks.x);
}

34
src/xenia/gpu/d3d12/texture_cache.cc
View File

@ -25,26 +25,28 @@ namespace d3d12 {
// Generated with `xb buildhlsl`.
#include "xenia/gpu/d3d12/shaders/bin/texture_load_128bpb_cs.h"
#include "xenia/gpu/d3d12/shaders/bin/texture_load_16bpb_cs.h"
#include "xenia/gpu/d3d12/shaders/bin/texture_load_32bpb_cs.h"
#include "xenia/gpu/d3d12/shaders/bin/texture_load_64bpb_cs.h"
#include "xenia/gpu/d3d12/shaders/bin/texture_load_8bpb_cs.h"
const TextureCache::HostFormat TextureCache::host_formats_[64] = {
{DXGI_FORMAT_UNKNOWN, CopyMode::kUnknown}, // k_1_REVERSE
{DXGI_FORMAT_UNKNOWN, CopyMode::kUnknown}, // k_1
{DXGI_FORMAT_UNKNOWN, CopyMode::kUnknown}, // k_8
{DXGI_FORMAT_UNKNOWN, CopyMode::kUnknown}, // k_1_5_5_5
{DXGI_FORMAT_UNKNOWN, CopyMode::kUnknown}, // k_5_6_5
{DXGI_FORMAT_R8_UNORM, CopyMode::k8bpb}, // k_8
{DXGI_FORMAT_B5G5R5A1_UNORM, CopyMode::k16bpb}, // k_1_5_5_5
{DXGI_FORMAT_B5G6R5_UNORM, CopyMode::k16bpb}, // k_5_6_5
{DXGI_FORMAT_UNKNOWN, CopyMode::kUnknown}, // k_6_5_5
{DXGI_FORMAT_R8G8B8A8_UNORM, CopyMode::k32bpb}, // k_8_8_8_8
{DXGI_FORMAT_R10G10B10A2_UNORM, CopyMode::k32bpb}, // k_2_10_10_10
{DXGI_FORMAT_UNKNOWN, CopyMode::kUnknown}, // k_8_A
{DXGI_FORMAT_R8_UNORM, CopyMode::k8bpb}, // k_8_A
{DXGI_FORMAT_UNKNOWN, CopyMode::kUnknown}, // k_8_B
{DXGI_FORMAT_UNKNOWN, CopyMode::kUnknown}, // k_8_8
{DXGI_FORMAT_R8G8_UNORM, CopyMode::k16bpb}, // k_8_8
{DXGI_FORMAT_UNKNOWN, CopyMode::kUnknown}, // k_Cr_Y1_Cb_Y0
{DXGI_FORMAT_UNKNOWN, CopyMode::kUnknown}, // k_Y1_Cr_Y0_Cb
{DXGI_FORMAT_UNKNOWN, CopyMode::kUnknown}, // k_Shadow
{DXGI_FORMAT_UNKNOWN, CopyMode::kUnknown}, // k_8_8_8_8_A
{DXGI_FORMAT_UNKNOWN, CopyMode::kUnknown}, // k_4_4_4_4
{DXGI_FORMAT_B4G4R4A4_UNORM, CopyMode::k16bpb}, // k_4_4_4_4
{DXGI_FORMAT_UNKNOWN, CopyMode::kUnknown}, // k_10_11_11
{DXGI_FORMAT_UNKNOWN, CopyMode::kUnknown}, // k_11_11_10
{DXGI_FORMAT_BC1_UNORM, CopyMode::k64bpb}, // k_DXT1
@ -53,13 +55,13 @@ const TextureCache::HostFormat TextureCache::host_formats_[64] = {
{DXGI_FORMAT_UNKNOWN, CopyMode::kUnknown}, // k_DXV
{DXGI_FORMAT_UNKNOWN, CopyMode::kUnknown}, // k_24_8
{DXGI_FORMAT_UNKNOWN, CopyMode::kUnknown}, // k_24_8_FLOAT
{DXGI_FORMAT_UNKNOWN, CopyMode::kUnknown}, // k_16
{DXGI_FORMAT_R16_UNORM, CopyMode::k16bpb}, // k_16
{DXGI_FORMAT_R16G16_UNORM, CopyMode::k32bpb}, // k_16_16
{DXGI_FORMAT_R16G16B16A16_UNORM, CopyMode::k64bpb}, // k_16_16_16_16
{DXGI_FORMAT_UNKNOWN, CopyMode::kUnknown}, // k_16_EXPAND
{DXGI_FORMAT_R16_UNORM, CopyMode::k16bpb}, // k_16_EXPAND
{DXGI_FORMAT_R16G16_UNORM, CopyMode::k32bpb}, // k_16_16_EXPAND
{DXGI_FORMAT_R16G16B16A16_UNORM, CopyMode::k64bpb}, // k_16_16_16_16_EXPAND
{DXGI_FORMAT_UNKNOWN, CopyMode::kUnknown}, // k_16_FLOAT
{DXGI_FORMAT_R16_FLOAT, CopyMode::k16bpb}, // k_16_FLOAT
{DXGI_FORMAT_R16G16_FLOAT, CopyMode::k32bpb}, // k_16_16_FLOAT
{DXGI_FORMAT_R16G16B16A16_FLOAT, CopyMode::k64bpb}, // k_16_16_16_16_FLOAT
{DXGI_FORMAT_UNKNOWN, CopyMode::kUnknown}, // k_32
@ -78,7 +80,7 @@ const TextureCache::HostFormat TextureCache::host_formats_[64] = {
{DXGI_FORMAT_UNKNOWN, CopyMode::kUnknown}, // k_16_INTERLACED
{DXGI_FORMAT_UNKNOWN, CopyMode::kUnknown}, // k_16_MPEG_INTERLACED
{DXGI_FORMAT_UNKNOWN, CopyMode::kUnknown}, // k_16_16_MPEG_INTERLACED
{DXGI_FORMAT_UNKNOWN, CopyMode::kUnknown}, // k_DXN
{DXGI_FORMAT_BC5_UNORM, CopyMode::k128bpb}, // k_DXN
{DXGI_FORMAT_R8G8B8A8_UNORM, CopyMode::k32bpb}, // k_8_8_8_8_AS_16_16_16_16
{DXGI_FORMAT_BC1_UNORM, CopyMode::k64bpb}, // k_DXT1_AS_16_16_16_16
{DXGI_FORMAT_BC2_UNORM, CopyMode::k128bpb}, // k_DXT2_3_AS_16_16_16_16
@ -89,7 +91,7 @@ const TextureCache::HostFormat TextureCache::host_formats_[64] = {
{DXGI_FORMAT_UNKNOWN, CopyMode::kUnknown}, // k_11_11_10_AS_16_16_16_16
{DXGI_FORMAT_UNKNOWN, CopyMode::kUnknown}, // k_32_32_32_FLOAT
{DXGI_FORMAT_UNKNOWN, CopyMode::kUnknown}, // k_DXT3A
{DXGI_FORMAT_UNKNOWN, CopyMode::kUnknown}, // k_DXT5A
{DXGI_FORMAT_BC4_UNORM, CopyMode::k64bpb}, // k_DXT5A
{DXGI_FORMAT_UNKNOWN, CopyMode::kUnknown}, // k_CTX1
{DXGI_FORMAT_UNKNOWN, CopyMode::kUnknown}, // k_DXT3A_AS_1_1_1_1
{DXGI_FORMAT_R8G8B8A8_UNORM, CopyMode::k32bpb}, // k_8_8_8_8_GAMMA
@ -100,6 +102,8 @@ const char* const TextureCache::dimension_names_[4] = {"1D", "2D", "3D",
"cube"};
const TextureCache::CopyModeInfo TextureCache::copy_mode_info_[] = {
{texture_load_8bpb_cs, sizeof(texture_load_8bpb_cs)},
{texture_load_16bpb_cs, sizeof(texture_load_16bpb_cs)},
{texture_load_32bpb_cs, sizeof(texture_load_32bpb_cs)},
{texture_load_64bpb_cs, sizeof(texture_load_64bpb_cs)},
{texture_load_128bpb_cs, sizeof(texture_load_128bpb_cs)},
@ -400,7 +404,7 @@ void TextureCache::WriteSampler(uint32_t fetch_constant,
desc.MaxAnisotropy = 1;
}
// FIXME(Triang3l): Halfway and mirror clamp to border aren't mapped properly.
static const D3D12_TEXTURE_ADDRESS_MODE address_mode_map[] = {
static const D3D12_TEXTURE_ADDRESS_MODE kAddressModeMap[] = {
/* kRepeat */ D3D12_TEXTURE_ADDRESS_MODE_WRAP,
/* kMirroredRepeat */ D3D12_TEXTURE_ADDRESS_MODE_MIRROR,
/* kClampToEdge */ D3D12_TEXTURE_ADDRESS_MODE_CLAMP,
@ -410,9 +414,9 @@ void TextureCache::WriteSampler(uint32_t fetch_constant,
/* kClampToBorder */ D3D12_TEXTURE_ADDRESS_MODE_BORDER,
/* kMirrorClampToBorder */ D3D12_TEXTURE_ADDRESS_MODE_MIRROR_ONCE,
};
desc.AddressU = address_mode_map[fetch.clamp_x];
desc.AddressV = address_mode_map[fetch.clamp_y];
desc.AddressW = address_mode_map[fetch.clamp_z];
desc.AddressU = kAddressModeMap[fetch.clamp_x];
desc.AddressV = kAddressModeMap[fetch.clamp_y];
desc.AddressW = kAddressModeMap[fetch.clamp_z];
desc.MipLODBias = fetch.lod_bias * (1.0f / 32.0f);
desc.ComparisonFunc = D3D12_COMPARISON_FUNC_NEVER;
// TODO(Triang3l): Border colors k_ACBYCR_BLACK and k_ACBCRY_BLACK.

2
src/xenia/gpu/d3d12/texture_cache.h
View File

@ -85,6 +85,8 @@ class TextureCache {
};
enum class CopyMode {
k8bpb,
k16bpb,
k32bpb,
k64bpb,
k128bpb,