int EncodeToRamFromTexture(u32 address,GLuint source_texture, bool bFromZBuffer, bool bIsIntensityFmt, u32 copyfmt, int bScaleByHalf, const EFBRectangle& source, u32 writeStride)
{
u32 format = copyfmt;
if (bFromZBuffer)
{
format |= _GX_TF_ZTF;
if (copyfmt == 11)
format = GX_TF_Z16;
else if (format < GX_TF_Z8 || format > GX_TF_Z24X8)
format |= _GX_TF_CTF;
}
else
{
if (copyfmt > GX_TF_RGBA8 || (copyfmt < GX_TF_RGB565 && !bIsIntensityFmt))
format |= _GX_TF_CTF;
}
SHADER& texconv_shader = GetOrCreateEncodingShader(format);
u8 *dest_ptr = Memory::GetPointer(address);
int width = (source.right - source.left) >> bScaleByHalf;
int height = (source.bottom - source.top) >> bScaleByHalf;
int size_in_bytes = TexDecoder_GetTextureSizeInBytes(width, height, format);
u16 blkW = TexDecoder_GetBlockWidthInTexels(format) - 1;
u16 blkH = TexDecoder_GetBlockHeightInTexels(format) - 1;
// only copy on cache line boundaries
// extra pixels are copied but not displayed in the resulting texture
s32 expandedWidth = (width + blkW) & (~blkW);
s32 expandedHeight = (height + blkH) & (~blkH);
texconv_shader.Bind();
glUniform4i(s_encodingUniforms[format],
source.left, source.top,
expandedWidth, bScaleByHalf ? 2 : 1);
unsigned int numBlocksX = expandedWidth / TexDecoder_GetBlockWidthInTexels(format);
unsigned int numBlocksY = expandedHeight / TexDecoder_GetBlockHeightInTexels(format);
unsigned int cacheLinesPerRow;
if ((format & 0x0f) == 6)
cacheLinesPerRow = numBlocksX * 2;
else
cacheLinesPerRow = numBlocksX;
EncodeToRamUsingShader(source_texture,
dest_ptr, cacheLinesPerRow * 32, numBlocksY,
writeStride, bScaleByHalf > 0 && !bFromZBuffer);
return size_in_bytes; // TODO: D3D11 is calculating this value differently!
}
// block dimensions : widthStride, heightStride
// texture dims : width, height, x offset, y offset
static void WriteSwizzler(char*& p, u32 format, API_TYPE ApiType)
{
// left, top, of source rectangle within source texture
// width of the destination rectangle, scale_factor (1 or 2)
WRITE(p, "uniform int4 position;\n");
int blkW = TexDecoder_GetBlockWidthInTexels(format);
int blkH = TexDecoder_GetBlockHeightInTexels(format);
int samples = GetEncodedSampleCount(format);
if (ApiType == API_OPENGL)
{
WRITE(p, "#define samp0 samp9\n");
WRITE(p, "SAMPLER_BINDING(9) uniform sampler2DArray samp0;\n");
WRITE(p, " out vec4 ocol0;\n");
WRITE(p, "void main()\n");
}
else // D3D
{
WRITE(p,"sampler samp0 : register(s0);\n");
WRITE(p, "Texture2D Tex0 : register(t0);\n");
WRITE(p,"void main(\n");
WRITE(p," out float4 ocol0 : SV_Target)\n");
}
WRITE(p, "{\n"
" int2 sampleUv;\n"
" int2 uv1 = int2(gl_FragCoord.xy);\n"
);
WRITE(p, " int y_block_position = uv1.y & %d;\n", ~(blkH - 1));
WRITE(p, " int y_offset_in_block = uv1.y & %d;\n", blkH - 1);
WRITE(p, " int x_virtual_position = (uv1.x << %d) + y_offset_in_block * position.z;\n", IntLog2(samples));
WRITE(p, " int x_block_position = (x_virtual_position >> %d) & %d;\n", IntLog2(blkH), ~(blkW - 1));
if (samples == 1)
{
// 32 bit textures (RGBA8 and Z24) are stored in 2 cache line increments
WRITE(p, " bool first = 0 == (x_virtual_position & %d);\n", 8 * samples); // first cache line, used in the encoders
WRITE(p, " x_virtual_position = x_virtual_position << 1;\n");
}
WRITE(p, " int x_offset_in_block = x_virtual_position & %d;\n", blkW - 1);
WRITE(p, " int y_offset = (x_virtual_position >> %d) & %d;\n", IntLog2(blkW), blkH - 1);
WRITE(p, " sampleUv.x = x_offset_in_block + x_block_position;\n");
WRITE(p, " sampleUv.y = y_block_position + y_offset;\n");
WRITE(p, " float2 uv0 = float2(sampleUv);\n"); // sampleUv is the sample position in (int)gx_coords
WRITE(p, " uv0 += float2(0.5, 0.5);\n"); // move to center of pixel
WRITE(p, " uv0 *= float(position.w);\n"); // scale by two if needed (also move to pixel borders so that linear filtering will average adjacent pixel)
WRITE(p, " uv0 += float2(position.xy);\n"); // move to copied rect
WRITE(p, " uv0 /= float2(%d, %d);\n", EFB_WIDTH, EFB_HEIGHT); // normalize to [0:1]
if (ApiType == API_OPENGL) // ogl has to flip up and down
{
WRITE(p, " uv0.y = 1.0-uv0.y;\n");
}
WRITE(p, " float sample_offset = float(position.w) / float(%d);\n", EFB_WIDTH);
}
int EncodeToRamFromTexture(u32 address,LPDIRECT3DTEXTURE9 source_texture, u32 SourceW, u32 SourceH, bool bFromZBuffer, bool bIsIntensityFmt, u32 copyfmt, int bScaleByHalf, const EFBRectangle& source)
{
u32 format = copyfmt;
if (bFromZBuffer)
{
format |= _GX_TF_ZTF;
if (copyfmt == 11)
format = GX_TF_Z16;
else if (format < GX_TF_Z8 || format > GX_TF_Z24X8)
format |= _GX_TF_CTF;
}
else
if (copyfmt > GX_TF_RGBA8 || (copyfmt < GX_TF_RGB565 && !bIsIntensityFmt))
format |= _GX_TF_CTF;
LPDIRECT3DPIXELSHADER9 texconv_shader = GetOrCreateEncodingShader(format);
if (!texconv_shader)
return 0;
u8 *dest_ptr = Memory::GetPointer(address);
int width = (source.right - source.left) >> bScaleByHalf;
int height = (source.bottom - source.top) >> bScaleByHalf;
int size_in_bytes = TexDecoder_GetTextureSizeInBytes(width, height, format);
u16 blkW = TexDecoder_GetBlockWidthInTexels(format) - 1;
u16 blkH = TexDecoder_GetBlockHeightInTexels(format) - 1;
u16 samples = TextureConversionShader::GetEncodedSampleCount(format);
// only copy on cache line boundaries
// extra pixels are copied but not displayed in the resulting texture
s32 expandedWidth = (width + blkW) & (~blkW);
s32 expandedHeight = (height + blkH) & (~blkH);
float sampleStride = bScaleByHalf ? 2.f : 1.f;
TextureConversionShader::SetShaderParameters(
(float)expandedWidth,
(float)Renderer::EFBToScaledY(expandedHeight), // TODO: Why do we scale this?
(float)Renderer::EFBToScaledX(source.left),
(float)Renderer::EFBToScaledY(source.top),
Renderer::EFBToScaledXf(sampleStride),
Renderer::EFBToScaledYf(sampleStride),
(float)SourceW,
(float)SourceH);
TargetRectangle scaledSource;
scaledSource.top = 0;
scaledSource.bottom = expandedHeight;
scaledSource.left = 0;
scaledSource.right = expandedWidth / samples;
int cacheBytes = 32;
if ((format & 0x0f) == 6)
cacheBytes = 64;
int readStride = (expandedWidth * cacheBytes) / TexDecoder_GetBlockWidthInTexels(format);
EncodeToRamUsingShader(texconv_shader, source_texture, scaledSource, dest_ptr, expandedWidth / samples, expandedHeight, readStride, true, bScaleByHalf > 0,1.0f);
return size_in_bytes; // TODO: D3D11 is calculating this value differently!
}
u32 TextureCache::TCacheEntryBase::NumBlocksY() const
{
u32 blockH = TexDecoder_GetBlockHeightInTexels(format);
// Round up source height to multiple of block size
u32 actualHeight = ROUND_UP(native_height, blockH);
return actualHeight / blockH;
}
// block dimensions : widthStride, heightStride
// texture dims : width, height, x offset, y offset
void WriteSwizzler(char*& p, u32 format, API_TYPE ApiType)
{
// left, top, of source rectangle within source texture
// width of the destination rectangle, scale_factor (1 or 2)
WRITE(p, "uniform int4 position;\n");
int blkW = TexDecoder_GetBlockWidthInTexels(format);
int blkH = TexDecoder_GetBlockHeightInTexels(format);
int samples = GetEncodedSampleCount(format);
// 32 bit textures (RGBA8 and Z24) are store in 2 cache line increments
int factor = samples == 1 ? 2 : 1;
if (ApiType == API_OPENGL)
{
WRITE(p, "#define samp0 samp9\n");
WRITE(p, "uniform sampler2D samp0;\n");
WRITE(p, " out vec4 ocol0;\n");
WRITE(p, "void main()\n");
}
else // D3D
{
WRITE(p,"sampler samp0 : register(s0);\n");
WRITE(p, "Texture2D Tex0 : register(t0);\n");
WRITE(p,"void main(\n");
WRITE(p," out float4 ocol0 : SV_Target)\n");
}
WRITE(p, "{\n"
" int2 sampleUv;\n"
" int2 uv1 = int2(gl_FragCoord.xy);\n"
" float2 uv0 = float2(0.0, 0.0);\n"
);
WRITE(p, " uv1.x = uv1.x * %d;\n", samples);
WRITE(p, " int yl = uv1.y / %d;\n", blkH);
WRITE(p, " int yb = yl * %d;\n", blkH);
WRITE(p, " int yoff = uv1.y - yb;\n");
WRITE(p, " int xp = uv1.x + yoff * position.z;\n");
WRITE(p, " int xel = xp / %d;\n", samples == 1 ? factor : blkW);
WRITE(p, " int xb = xel / %d;\n", blkH);
WRITE(p, " int xoff = xel - xb * %d;\n", blkH);
WRITE(p, " int xl = uv1.x * %d / %d;\n", factor, blkW);
WRITE(p, " int xib = uv1.x * %d - xl * %d;\n", factor, blkW);
WRITE(p, " int halfxb = xb / %d;\n", factor);
WRITE(p, " sampleUv.x = xib + halfxb * %d;\n", blkW);
WRITE(p, " sampleUv.y = yb + xoff;\n");
}
// block dimensions : widthStride, heightStride
// texture dims : width, height, x offset, y offset
void Write32BitSwizzler(char*& p, u32 format)
{
// [0] left, top, right, bottom of source rectangle within source texture
// [1] width and height of destination texture in pixels
// Two were merged for GLSL
WRITE(p, "uniform float4 " I_COLORS"[2] %s;\n", WriteRegister("c", C_COLORS));
u32 blkW = TexDecoder_GetBlockWidthInTexels(format);
u32 blkH = TexDecoder_GetBlockHeightInTexels(format);
// 32 bit textures (RGBA8 and Z24) are store in 2 cache line increments
WRITE(p,"uniform sampler samp0 : register(s0);\n");
WRITE(p,"void main(\n");
WRITE(p, " out float4 ocol0 : COLOR0,\n");
WRITE(p, " in float2 uv0 : TEXCOORD0)\n");
WRITE(p, "{\n"
" float2 sampleUv;\n"
" float2 uv1 = floor(uv0);\n");
WRITE(p, " float yl = floor(uv1.y / %u.0);\n", blkH);
WRITE(p, " float yb = yl * %u.0;\n", blkH);
WRITE(p, " float yoff = uv1.y - yb;\n");
WRITE(p, " float xp = uv1.x + (yoff * " I_COLORS"[1].x);\n");
WRITE(p, " float xel = floor(xp / 2.0f);\n");
WRITE(p, " float xb = floor(xel / %u.0);\n", blkH);
WRITE(p, " float xoff = xel - (xb * %u.0);\n", blkH);
WRITE(p, " float x2 = uv1.x * 2.0f;\n");
WRITE(p, " float xl = floor(x2 / %u.0);\n", blkW);
WRITE(p, " float xib = x2 - (xl * %u.0);\n", blkW);
WRITE(p, " float halfxb = floor(xb / 2.0f);\n");
WRITE(p, " sampleUv.x = xib + (halfxb * %u.0);\n", blkW);
WRITE(p, " sampleUv.y = yb + xoff;\n");
WRITE(p, " sampleUv = sampleUv * " I_COLORS"[0].xy;\n");
WRITE(p, " sampleUv = sampleUv + " I_COLORS"[1].zw;\n");
WRITE(p, " sampleUv = sampleUv + float2(0.0f,1.0f);\n");// still to determine the reason for this
WRITE(p, " sampleUv = sampleUv / " I_COLORS"[0].zw;\n");
}
// block dimensions : widthStride, heightStride
// texture dims : width, height, x offset, y offset
void WriteSwizzler(char*& p, u32 format)
{
// [0] left, top, right, bottom of source rectangle within source texture
// [1] width and height of destination texture in pixels
// Two were merged for GLSL
WRITE(p, "uniform float4 " I_COLORS"[2] %s;\n", WriteRegister("c", C_COLORS));
u32 blkW = TexDecoder_GetBlockWidthInTexels(format);
u32 blkH = TexDecoder_GetBlockHeightInTexels(format);
u32 samples = TextureConversionShader::GetEncodedSampleCount(format);
WRITE(p,"uniform sampler samp0 : register(s0);\n");
WRITE(p,"void main(\n");
WRITE(p," out float4 ocol0 : SV_Target,\n");
WRITE(p," in float2 uv0 : TEXCOORD0)\n");
WRITE(p, "{\n"
" float2 sampleUv;\n"
" float2 uv1 = floor(uv0);\n");
WRITE(p, " uv1.x = uv1.x * %u.0;\n", samples);
WRITE(p, " float xl = floor(uv1.x / %u.0);\n", blkW);
WRITE(p, " float xib = uv1.x - (xl * %u.0);\n", blkW);
WRITE(p, " float yl = floor(uv1.y / %u.0);\n", blkH);
WRITE(p, " float yb = yl * %u.0;\n", blkH);
WRITE(p, " float yoff = uv1.y - yb;\n");
WRITE(p, " float xp = uv1.x + (yoff * " I_COLORS"[1].x);\n");
WRITE(p, " float xel = floor(xp / %u.0);\n", blkW);
WRITE(p, " float xb = floor(xel / %u.0);\n", blkH);
WRITE(p, " float xoff = xel - (xb * %u.0);\n", blkH);
WRITE(p, " sampleUv.x = xib + (xb * %u.0);\n", blkW);
WRITE(p, " sampleUv.y = yb + xoff;\n");
WRITE(p, " sampleUv = sampleUv * " I_COLORS"[0].xy;\n");
WRITE(p, " sampleUv = sampleUv + " I_COLORS"[1].zw;\n");
WRITE(p, " sampleUv = sampleUv + float2(0.0f,1.0f);\n");// still to determine the reason for this
WRITE(p, " sampleUv = sampleUv / " I_COLORS"[0].zw;\n");
}
void EncodeToRamFromTexture(u8 *dest_ptr, const TextureCache::TCacheEntryBase *texture_entry, u32 SourceW, u32 SourceH,
LPDIRECT3DTEXTURE9 source_texture, bool bFromZBuffer, bool bIsIntensityFmt, int bScaleByHalf, const EFBRectangle& source)
{
LPDIRECT3DPIXELSHADER9 texconv_shader = GetOrCreateEncodingShader(texture_entry->format);
if (!texconv_shader)
return;
int width = (source.right - source.left) >> bScaleByHalf;
int height = (source.bottom - source.top) >> bScaleByHalf;
u16 blkW = TexDecoder_GetBlockWidthInTexels(texture_entry->format);
u16 blkH = TexDecoder_GetBlockHeightInTexels(texture_entry->format);
u16 samples = TextureConversionShader::GetEncodedSampleCount(texture_entry->format);
// only copy on cache line boundaries
// extra pixels are copied but not displayed in the resulting texture
s32 expandedWidth = ROUND_UP(width, blkW);
s32 expandedHeight = ROUND_UP(height, blkH);
float sampleStride = bScaleByHalf ? 2.f : 1.f;
TextureConversionShaderLegacy::SetShaderParameters(
(float)expandedWidth,
(float)Renderer::EFBToScaledY(expandedHeight), // TODO: Why do we scale this?
(float)Renderer::EFBToScaledX(source.left),
(float)Renderer::EFBToScaledY(source.top),
Renderer::EFBToScaledXf(sampleStride),
Renderer::EFBToScaledYf(sampleStride),
(float)SourceW,
(float)SourceH);
D3D::dev->SetPixelShaderConstantF(C_COLORMATRIX, PixelShaderManager::GetBuffer(), 2);
TargetRectangle scaledSource;
scaledSource.top = 0;
scaledSource.bottom = expandedHeight;
scaledSource.left = 0;
scaledSource.right = expandedWidth / samples;
int cacheBytes = 32;
if ((texture_entry->format & 0x0f) == 6)
cacheBytes = 64;
int readStride = (expandedWidth * cacheBytes) / TexDecoder_GetBlockWidthInTexels(texture_entry->format);
EncodeToRamUsingShader(texconv_shader, source_texture, scaledSource, dest_ptr, (expandedWidth / samples) * 4, expandedHeight, texture_entry->BytesPerRow(), readStride, bScaleByHalf > 0, 1.0f);
}
TextureCache::TCacheEntryBase* TextureCache::Load(unsigned int const stage,
u32 const address, unsigned int width, unsigned int height, int const texformat,
unsigned int const tlutaddr, int const tlutfmt, bool const use_mipmaps, unsigned int maxlevel, bool const from_tmem)
{
if (0 == address)
return nullptr;
// TexelSizeInNibbles(format) * width * height / 16;
const unsigned int bsw = TexDecoder_GetBlockWidthInTexels(texformat) - 1;
const unsigned int bsh = TexDecoder_GetBlockHeightInTexels(texformat) - 1;
unsigned int expandedWidth = (width + bsw) & (~bsw);
unsigned int expandedHeight = (height + bsh) & (~bsh);
const unsigned int nativeW = width;
const unsigned int nativeH = height;
u32 texID = address;
// Hash assigned to texcache entry (also used to generate filenames used for texture dumping and custom texture lookup)
u64 tex_hash = TEXHASH_INVALID;
u64 tlut_hash = TEXHASH_INVALID;
u32 full_format = texformat;
PC_TexFormat pcfmt = PC_TEX_FMT_NONE;
const bool isPaletteTexture = (texformat == GX_TF_C4 || texformat == GX_TF_C8 || texformat == GX_TF_C14X2);
if (isPaletteTexture)
full_format = texformat | (tlutfmt << 16);
const u32 texture_size = TexDecoder_GetTextureSizeInBytes(expandedWidth, expandedHeight, texformat);
const u8* src_data;
if (from_tmem)
src_data = &texMem[bpmem.tex[stage / 4].texImage1[stage % 4].tmem_even * TMEM_LINE_SIZE];
else
src_data = Memory::GetPointer(address);
// TODO: This doesn't hash GB tiles for preloaded RGBA8 textures (instead, it's hashing more data from the low tmem bank than it should)
tex_hash = GetHash64(src_data, texture_size, g_ActiveConfig.iSafeTextureCache_ColorSamples);
if (isPaletteTexture)
{
const u32 palette_size = TexDecoder_GetPaletteSize(texformat);
tlut_hash = GetHash64(&texMem[tlutaddr], palette_size, g_ActiveConfig.iSafeTextureCache_ColorSamples);
// NOTE: For non-paletted textures, texID is equal to the texture address.
// A paletted texture, however, may have multiple texIDs assigned though depending on the currently used tlut.
// This (changing texID depending on the tlut_hash) is a trick to get around
// an issue with Metroid Prime's fonts (it has multiple sets of fonts on each other
// stored in a single texture and uses the palette to make different characters
// visible or invisible. Thus, unless we want to recreate the textures for every drawn character,
// we must make sure that a paletted texture gets assigned multiple IDs for each tlut used.
//
// TODO: Because texID isn't always the same as the address now, CopyRenderTargetToTexture might be broken now
texID ^= ((u32)tlut_hash) ^(u32)(tlut_hash >> 32);
tex_hash ^= tlut_hash;
}
// D3D doesn't like when the specified mipmap count would require more than one 1x1-sized LOD in the mipmap chain
// e.g. 64x64 with 7 LODs would have the mipmap chain 64x64,32x32,16x16,8x8,4x4,2x2,1x1,1x1, so we limit the mipmap count to 6 there
while (g_ActiveConfig.backend_info.bUseMinimalMipCount && std::max(expandedWidth, expandedHeight) >> maxlevel == 0)
--maxlevel;
TCacheEntryBase *entry = textures[texID];
if (entry)
{
// 1. Calculate reference hash:
// calculated from RAM texture data for normal textures. Hashes for paletted textures are modified by tlut_hash. 0 for virtual EFB copies.
if (g_ActiveConfig.bCopyEFBToTexture && entry->IsEfbCopy())
tex_hash = TEXHASH_INVALID;
// 2. a) For EFB copies, only the hash and the texture address need to match
if (entry->IsEfbCopy() && tex_hash == entry->hash && address == entry->addr)
{
entry->type = TCET_EC_VRAM;
// TODO: Print a warning if the format changes! In this case,
// we could reinterpret the internal texture object data to the new pixel format
// (similar to what is already being done in Renderer::ReinterpretPixelFormat())
return ReturnEntry(stage, entry);
}
// 2. b) For normal textures, all texture parameters need to match
if (address == entry->addr && tex_hash == entry->hash && full_format == entry->format &&
entry->num_mipmaps > maxlevel && entry->native_width == nativeW && entry->native_height == nativeH)
{
return ReturnEntry(stage, entry);
}
// 3. If we reach this line, we'll have to upload the new texture data to VRAM.
// If we're lucky, the texture parameters didn't change and we can reuse the internal texture object instead of destroying and recreating it.
//
// TODO: Don't we need to force texture decoding to RGBA8 for dynamic EFB copies?
// TODO: Actually, it should be enough if the internal texture format matches...
if ((entry->type == TCET_NORMAL &&
width == entry->virtual_width &&
height == entry->virtual_height &&
full_format == entry->format &&
entry->num_mipmaps > maxlevel) ||
(entry->type == TCET_EC_DYNAMIC &&
entry->native_width == width &&
entry->native_height == height))
{
// reuse the texture
}
else
{
// delete the texture and make a new one
delete entry;
entry = nullptr;
}
}
bool using_custom_texture = false;
if (g_ActiveConfig.bHiresTextures)
{
// This function may modify width/height.
pcfmt = LoadCustomTexture(tex_hash, texformat, 0, width, height);
if (pcfmt != PC_TEX_FMT_NONE)
{
if (expandedWidth != width || expandedHeight != height)
{
expandedWidth = width;
expandedHeight = height;
// If we thought we could reuse the texture before, make sure to pool it now!
if (entry)
{
delete entry;
entry = nullptr;
}
}
using_custom_texture = true;
}
}
if (!using_custom_texture)
{
if (!(texformat == GX_TF_RGBA8 && from_tmem))
{
pcfmt = TexDecoder_Decode(temp, src_data, expandedWidth,
expandedHeight, texformat, tlutaddr, tlutfmt, g_ActiveConfig.backend_info.bUseRGBATextures);
}
else
{
u8* src_data_gb = &texMem[bpmem.tex[stage/4].texImage2[stage%4].tmem_odd * TMEM_LINE_SIZE];
pcfmt = TexDecoder_DecodeRGBA8FromTmem(temp, src_data, src_data_gb, expandedWidth, expandedHeight);
}
}
u32 texLevels = use_mipmaps ? (maxlevel + 1) : 1;
const bool using_custom_lods = using_custom_texture && CheckForCustomTextureLODs(tex_hash, texformat, texLevels);
// Only load native mips if their dimensions fit to our virtual texture dimensions
const bool use_native_mips = use_mipmaps && !using_custom_lods && (width == nativeW && height == nativeH);
texLevels = (use_native_mips || using_custom_lods) ? texLevels : 1; // TODO: Should be forced to 1 for non-pow2 textures (e.g. efb copies with automatically adjusted IR)
// create the entry/texture
if (nullptr == entry)
{
textures[texID] = entry = g_texture_cache->CreateTexture(width, height, expandedWidth, texLevels, pcfmt);
// Sometimes, we can get around recreating a texture if only the number of mip levels changes
// e.g. if our texture cache entry got too many mipmap levels we can limit the number of used levels by setting the appropriate render states
// Thus, we don't update this member for every Load, but just whenever the texture gets recreated
// TODO: This is the wrong value. We should be storing the number of levels our actual texture has.
// But that will currently make the above "existing entry" tests fail as "texLevels" is not calculated until after.
// Currently, we might try to reuse a texture which appears to have more levels than actual, maybe..
entry->num_mipmaps = maxlevel + 1;
entry->type = TCET_NORMAL;
GFX_DEBUGGER_PAUSE_AT(NEXT_NEW_TEXTURE, true);
}
else
{
// load texture (CreateTexture also loads level 0)
entry->Load(width, height, expandedWidth, 0);
}
entry->SetGeneralParameters(address, texture_size, full_format, entry->num_mipmaps);
entry->SetDimensions(nativeW, nativeH, width, height);
entry->hash = tex_hash;
if (entry->IsEfbCopy() && !g_ActiveConfig.bCopyEFBToTexture)
entry->type = TCET_EC_DYNAMIC;
else
entry->type = TCET_NORMAL;
if (g_ActiveConfig.bDumpTextures && !using_custom_texture)
DumpTexture(entry, 0);
u32 level = 1;
// load mips - TODO: Loading mipmaps from tmem is untested!
if (pcfmt != PC_TEX_FMT_NONE)
{
if (use_native_mips)
{
src_data += texture_size;
const u8* ptr_even = nullptr;
const u8* ptr_odd = nullptr;
if (from_tmem)
{
ptr_even = &texMem[bpmem.tex[stage/4].texImage1[stage%4].tmem_even * TMEM_LINE_SIZE + texture_size];
ptr_odd = &texMem[bpmem.tex[stage/4].texImage2[stage%4].tmem_odd * TMEM_LINE_SIZE];
}
for (; level != texLevels; ++level)
{
const u32 mip_width = CalculateLevelSize(width, level);
const u32 mip_height = CalculateLevelSize(height, level);
const u32 expanded_mip_width = (mip_width + bsw) & (~bsw);
const u32 expanded_mip_height = (mip_height + bsh) & (~bsh);
const u8*& mip_src_data = from_tmem
? ((level % 2) ? ptr_odd : ptr_even)
: src_data;
TexDecoder_Decode(temp, mip_src_data, expanded_mip_width, expanded_mip_height, texformat, tlutaddr, tlutfmt, g_ActiveConfig.backend_info.bUseRGBATextures);
mip_src_data += TexDecoder_GetTextureSizeInBytes(expanded_mip_width, expanded_mip_height, texformat);
entry->Load(mip_width, mip_height, expanded_mip_width, level);
if (g_ActiveConfig.bDumpTextures)
DumpTexture(entry, level);
}
}
else if (using_custom_lods)
{
for (; level != texLevels; ++level)
{
unsigned int mip_width = CalculateLevelSize(width, level);
unsigned int mip_height = CalculateLevelSize(height, level);
LoadCustomTexture(tex_hash, texformat, level, mip_width, mip_height);
entry->Load(mip_width, mip_height, mip_width, level);
}
}
}
INCSTAT(stats.numTexturesCreated);
SETSTAT(stats.numTexturesAlive, textures.size());
return ReturnEntry(stage, entry);
}
TextureCache::TCacheEntryBase* TextureCache::Load(const u32 stage)
{
const FourTexUnits &tex = bpmem.tex[stage >> 2];
const u32 id = stage & 3;
const u32 address = (tex.texImage3[id].image_base/* & 0x1FFFFF*/) << 5;
u32 width = tex.texImage0[id].width + 1;
u32 height = tex.texImage0[id].height + 1;
const int texformat = tex.texImage0[id].format;
const u32 tlutaddr = tex.texTlut[id].tmem_offset << 9;
const u32 tlutfmt = tex.texTlut[id].tlut_format;
const bool use_mipmaps = (tex.texMode0[id].min_filter & 3) != 0;
u32 tex_levels = use_mipmaps ? ((tex.texMode1[id].max_lod + 0xf) / 0x10 + 1) : 1;
const bool from_tmem = tex.texImage1[id].image_type != 0;
if (0 == address)
return nullptr;
// TexelSizeInNibbles(format) * width * height / 16;
const unsigned int bsw = TexDecoder_GetBlockWidthInTexels(texformat);
const unsigned int bsh = TexDecoder_GetBlockHeightInTexels(texformat);
unsigned int expandedWidth = ROUND_UP(width, bsw);
unsigned int expandedHeight = ROUND_UP(height, bsh);
const unsigned int nativeW = width;
const unsigned int nativeH = height;
// Hash assigned to texcache entry (also used to generate filenames used for texture dumping and custom texture lookup)
u64 base_hash = TEXHASH_INVALID;
u64 full_hash = TEXHASH_INVALID;
u32 full_format = texformat;
const bool isPaletteTexture = (texformat == GX_TF_C4 || texformat == GX_TF_C8 || texformat == GX_TF_C14X2);
// Reject invalid tlut format.
if (isPaletteTexture && tlutfmt > GX_TL_RGB5A3)
return nullptr;
if (isPaletteTexture)
full_format = texformat | (tlutfmt << 16);
const u32 texture_size = TexDecoder_GetTextureSizeInBytes(expandedWidth, expandedHeight, texformat);
u32 additional_mips_size = 0; // not including level 0, which is texture_size
// GPUs don't like when the specified mipmap count would require more than one 1x1-sized LOD in the mipmap chain
// e.g. 64x64 with 7 LODs would have the mipmap chain 64x64,32x32,16x16,8x8,4x4,2x2,1x1,0x0, so we limit the mipmap count to 6 there
tex_levels = std::min<u32>(IntLog2(std::max(width, height)) + 1, tex_levels);
for (u32 level = 1; level != tex_levels; ++level)
{
// We still need to calculate the original size of the mips
const u32 expanded_mip_width = ROUND_UP(CalculateLevelSize(width, level), bsw);
const u32 expanded_mip_height = ROUND_UP(CalculateLevelSize(height, level), bsh);
additional_mips_size += TexDecoder_GetTextureSizeInBytes(expanded_mip_width, expanded_mip_height, texformat);
}
// If we are recording a FifoLog, keep track of what memory we read.
// FifiRecorder does it's own memory modification tracking independant of the texture hashing below.
if (g_bRecordFifoData && !from_tmem)
FifoRecorder::GetInstance().UseMemory(address, texture_size + additional_mips_size, MemoryUpdate::TEXTURE_MAP);
const u8* src_data;
if (from_tmem)
src_data = &texMem[bpmem.tex[stage / 4].texImage1[stage % 4].tmem_even * TMEM_LINE_SIZE];
else
src_data = Memory::GetPointer(address);
// TODO: This doesn't hash GB tiles for preloaded RGBA8 textures (instead, it's hashing more data from the low tmem bank than it should)
base_hash = GetHash64(src_data, texture_size, g_ActiveConfig.iSafeTextureCache_ColorSamples);
u32 palette_size = 0;
if (isPaletteTexture)
{
palette_size = TexDecoder_GetPaletteSize(texformat);
full_hash = base_hash ^ GetHash64(&texMem[tlutaddr], palette_size, g_ActiveConfig.iSafeTextureCache_ColorSamples);
}
else
{
full_hash = base_hash;
}
// Search the texture cache for textures by address
//
// Find all texture cache entries for the current texture address, and decide whether to use one of
// them, or to create a new one
//
// In most cases, the fastest way is to use only one texture cache entry for the same address. Usually,
// when a texture changes, the old version of the texture is unlikely to be used again. If there were
// new cache entries created for normal texture updates, there would be a slowdown due to a huge amount
// of unused cache entries. Also thanks to texture pooling, overwriting an existing cache entry is
// faster than creating a new one from scratch.
//
// Some games use the same address for different textures though. If the same cache entry was used in
// this case, it would be constantly overwritten, and effectively there wouldn't be any caching for
// those textures. Examples for this are Metroid Prime and Castlevania 3. Metroid Prime has multiple
// sets of fonts on each other stored in a single texture and uses the palette to make different
// characters visible or invisible. In Castlevania 3 some textures are used for 2 different things or
// at least in 2 different ways(size 1024x1024 vs 1024x256).
//
// To determine whether to use multiple cache entries or a single entry, use the following heuristic:
// If the same texture address is used several times during the same frame, assume the address is used
// for different purposes and allow creating an additional cache entry. If there's at least one entry
// that hasn't been used for the same frame, then overwrite it, in order to keep the cache as small as
// possible. If the current texture is found in the cache, use that entry.
//
// For efb copies, the entry created in CopyRenderTargetToTexture always has to be used, or else it was
// done in vain.
std::pair<TexCache::iterator, TexCache::iterator> iter_range = textures_by_address.equal_range((u64)address);
TexCache::iterator iter = iter_range.first;
TexCache::iterator oldest_entry = iter;
int temp_frameCount = 0x7fffffff;
TexCache::iterator unconverted_copy = textures_by_address.end();
while (iter != iter_range.second)
{
TCacheEntryBase* entry = iter->second;
// Do not load strided EFB copies, they are not meant to be used directly
if (entry->IsEfbCopy() && entry->native_width == nativeW && entry->native_height == nativeH &&
entry->memory_stride == entry->CacheLinesPerRow() * 32)
{
// EFB copies have slightly different rules as EFB copy formats have different
// meanings from texture formats.
if ((base_hash == entry->hash && (!isPaletteTexture || g_Config.backend_info.bSupportsPaletteConversion)) ||
IsPlayingBackFifologWithBrokenEFBCopies)
{
// TODO: We should check format/width/height/levels for EFB copies. Checking
// format is complicated because EFB copy formats don't exactly match
// texture formats. I'm not sure what effect checking width/height/levels
// would have.
if (!isPaletteTexture || !g_Config.backend_info.bSupportsPaletteConversion)
return ReturnEntry(stage, entry);
// Note that we found an unconverted EFB copy, then continue. We'll
// perform the conversion later. Currently, we only convert EFB copies to
// palette textures; we could do other conversions if it proved to be
// beneficial.
unconverted_copy = iter;
}
else
{
// Aggressively prune EFB copies: if it isn't useful here, it will probably
// never be useful again. It's theoretically possible for a game to do
// something weird where the copy could become useful in the future, but in
// practice it doesn't happen.
iter = FreeTexture(iter);
continue;
}
}
else
{
// For normal textures, all texture parameters need to match
if (entry->hash == full_hash && entry->format == full_format && entry->native_levels >= tex_levels &&
entry->native_width == nativeW && entry->native_height == nativeH)
{
entry = DoPartialTextureUpdates(iter);
return ReturnEntry(stage, entry);
}
}
// Find the texture which hasn't been used for the longest time. Count paletted
// textures as the same texture here, when the texture itself is the same. This
// improves the performance a lot in some games that use paletted textures.
// Example: Sonic the Fighters (inside Sonic Gems Collection)
// Skip EFB copies here, so they can be used for partial texture updates
if (entry->frameCount != FRAMECOUNT_INVALID && entry->frameCount < temp_frameCount &&
!entry->IsEfbCopy() && !(isPaletteTexture && entry->base_hash == base_hash))
{
temp_frameCount = entry->frameCount;
oldest_entry = iter;
}
++iter;
}
if (unconverted_copy != textures_by_address.end())
{
// Perform palette decoding.
TCacheEntryBase *entry = unconverted_copy->second;
TCacheEntryConfig config;
config.rendertarget = true;
config.width = entry->config.width;
config.height = entry->config.height;
config.layers = FramebufferManagerBase::GetEFBLayers();
TCacheEntryBase *decoded_entry = AllocateTexture(config);
decoded_entry->SetGeneralParameters(address, texture_size, full_format);
decoded_entry->SetDimensions(entry->native_width, entry->native_height, 1);
decoded_entry->SetHashes(base_hash, full_hash);
decoded_entry->frameCount = FRAMECOUNT_INVALID;
decoded_entry->is_efb_copy = false;
g_texture_cache->ConvertTexture(decoded_entry, entry, &texMem[tlutaddr], (TlutFormat)tlutfmt);
textures_by_address.emplace((u64)address, decoded_entry);
return ReturnEntry(stage, decoded_entry);
}
// Search the texture cache for normal textures by hash
//
// If the texture was fully hashed, the address does not need to match. Identical duplicate textures cause unnecessary slowdowns
// Example: Tales of Symphonia (GC) uses over 500 small textures in menus, but only around 70 different ones
if (g_ActiveConfig.iSafeTextureCache_ColorSamples == 0 ||
std::max(texture_size, palette_size) <= (u32)g_ActiveConfig.iSafeTextureCache_ColorSamples * 8)
{
iter_range = textures_by_hash.equal_range(full_hash);
iter = iter_range.first;
while (iter != iter_range.second)
{
TCacheEntryBase* entry = iter->second;
// All parameters, except the address, need to match here
if (entry->format == full_format && entry->native_levels >= tex_levels &&
entry->native_width == nativeW && entry->native_height == nativeH)
{
entry = DoPartialTextureUpdates(iter);
return ReturnEntry(stage, entry);
}
++iter;
}
}
// If at least one entry was not used for the same frame, overwrite the oldest one
if (temp_frameCount != 0x7fffffff)
{
// pool this texture and make a new one later
FreeTexture(oldest_entry);
}
std::shared_ptr<HiresTexture> hires_tex;
if (g_ActiveConfig.bHiresTextures)
{
hires_tex = HiresTexture::Search(
src_data, texture_size,
&texMem[tlutaddr], palette_size,
width, height,
texformat, use_mipmaps
);
if (hires_tex)
{
auto& l = hires_tex->m_levels[0];
if (l.width != width || l.height != height)
{
width = l.width;
height = l.height;
}
expandedWidth = l.width;
expandedHeight = l.height;
CheckTempSize(l.data_size);
memcpy(temp, l.data, l.data_size);
}
}
if (!hires_tex)
{
if (!(texformat == GX_TF_RGBA8 && from_tmem))
{
const u8* tlut = &texMem[tlutaddr];
TexDecoder_Decode(temp, src_data, expandedWidth, expandedHeight, texformat, tlut, (TlutFormat)tlutfmt);
}
else
{
u8* src_data_gb = &texMem[bpmem.tex[stage / 4].texImage2[stage % 4].tmem_odd * TMEM_LINE_SIZE];
TexDecoder_DecodeRGBA8FromTmem(temp, src_data, src_data_gb, expandedWidth, expandedHeight);
}
}
// how many levels the allocated texture shall have
const u32 texLevels = hires_tex ? (u32)hires_tex->m_levels.size() : tex_levels;
// create the entry/texture
TCacheEntryConfig config;
config.width = width;
config.height = height;
config.levels = texLevels;
TCacheEntryBase* entry = AllocateTexture(config);
GFX_DEBUGGER_PAUSE_AT(NEXT_NEW_TEXTURE, true);
iter = textures_by_address.emplace((u64)address, entry);
if (g_ActiveConfig.iSafeTextureCache_ColorSamples == 0 ||
std::max(texture_size, palette_size) <= (u32)g_ActiveConfig.iSafeTextureCache_ColorSamples * 8)
{
entry->textures_by_hash_iter = textures_by_hash.emplace(full_hash, entry);
}
entry->SetGeneralParameters(address, texture_size, full_format);
entry->SetDimensions(nativeW, nativeH, tex_levels);
entry->SetHashes(base_hash, full_hash);
entry->is_efb_copy = false;
entry->is_custom_tex = hires_tex != nullptr;
// load texture
entry->Load(width, height, expandedWidth, 0);
std::string basename = "";
if (g_ActiveConfig.bDumpTextures && !hires_tex)
{
basename = HiresTexture::GenBaseName(
src_data, texture_size,
&texMem[tlutaddr], palette_size,
width, height,
texformat, use_mipmaps,
true
);
DumpTexture(entry, basename, 0);
}
if (hires_tex)
{
for (u32 level = 1; level != texLevels; ++level)
{
auto& l = hires_tex->m_levels[level];
CheckTempSize(l.data_size);
memcpy(temp, l.data, l.data_size);
entry->Load(l.width, l.height, l.width, level);
}
}
else
{
// load mips - TODO: Loading mipmaps from tmem is untested!
src_data += texture_size;
const u8* ptr_even = nullptr;
const u8* ptr_odd = nullptr;
if (from_tmem)
{
ptr_even = &texMem[bpmem.tex[stage / 4].texImage1[stage % 4].tmem_even * TMEM_LINE_SIZE + texture_size];
ptr_odd = &texMem[bpmem.tex[stage / 4].texImage2[stage % 4].tmem_odd * TMEM_LINE_SIZE];
}
for (u32 level = 1; level != texLevels; ++level)
{
const u32 mip_width = CalculateLevelSize(width, level);
const u32 mip_height = CalculateLevelSize(height, level);
const u32 expanded_mip_width = ROUND_UP(mip_width, bsw);
const u32 expanded_mip_height = ROUND_UP(mip_height, bsh);
const u8*& mip_src_data = from_tmem
? ((level % 2) ? ptr_odd : ptr_even)
: src_data;
const u8* tlut = &texMem[tlutaddr];
TexDecoder_Decode(temp, mip_src_data, expanded_mip_width, expanded_mip_height, texformat, tlut, (TlutFormat)tlutfmt);
mip_src_data += TexDecoder_GetTextureSizeInBytes(expanded_mip_width, expanded_mip_height, texformat);
entry->Load(mip_width, mip_height, expanded_mip_width, level);
if (g_ActiveConfig.bDumpTextures)
DumpTexture(entry, basename, level);
}
}
INCSTAT(stats.numTexturesUploaded);
SETSTAT(stats.numTexturesAlive, textures_by_address.size());
entry = DoPartialTextureUpdates(iter);
return ReturnEntry(stage, entry);
}
TextureCache::TCacheEntryBase* TextureCache::DoPartialTextureUpdates(TexCache::iterator iter_t)
{
TCacheEntryBase* entry_to_update = iter_t->second;
const bool isPaletteTexture = (entry_to_update->format == GX_TF_C4
|| entry_to_update->format == GX_TF_C8
|| entry_to_update->format == GX_TF_C14X2
|| entry_to_update->format >= 0x10000);
// Efb copies and paletted textures are excluded from these updates, until there's an example where a game would
// benefit from this. Both would require more work to be done.
// TODO: Implement upscaling support for normal textures, and then remove the efb to ram and the scaled efb restrictions
if (entry_to_update->IsEfbCopy()
|| isPaletteTexture)
return entry_to_update;
u32 block_width = TexDecoder_GetBlockWidthInTexels(entry_to_update->format);
u32 block_height = TexDecoder_GetBlockHeightInTexels(entry_to_update->format);
u32 block_size = block_width * block_height * TexDecoder_GetTexelSizeInNibbles(entry_to_update->format) / 2;
u32 numBlocksX = (entry_to_update->native_width + block_width - 1) / block_width;
TexCache::iterator iter = textures_by_address.lower_bound(entry_to_update->addr);
TexCache::iterator iterend = textures_by_address.upper_bound(entry_to_update->addr + entry_to_update->size_in_bytes);
bool entry_need_scaling = true;
while (iter != iterend)
{
TCacheEntryBase* entry = iter->second;
if (entry != entry_to_update
&& entry->IsEfbCopy()
&& entry_to_update->addr <= entry->addr
&& entry->addr + entry->size_in_bytes <= entry_to_update->addr + entry_to_update->size_in_bytes
&& entry->frameCount == FRAMECOUNT_INVALID
&& entry->memory_stride == numBlocksX * block_size)
{
u32 block_offset = (entry->addr - entry_to_update->addr) / block_size;
u32 block_x = block_offset % numBlocksX;
u32 block_y = block_offset / numBlocksX;
u32 x = block_x * block_width;
u32 y = block_y * block_height;
MathUtil::Rectangle<int> srcrect, dstrect;
srcrect.left = 0;
srcrect.top = 0;
dstrect.left = 0;
dstrect.top = 0;
if (entry_need_scaling)
{
entry_need_scaling = false;
u32 w = entry_to_update->native_width * entry->config.width / entry->native_width;
u32 h = entry_to_update->native_height * entry->config.height / entry->native_height;
u32 max = g_renderer->GetMaxTextureSize();
if (max < w || max < h)
{
iter++;
continue;
}
if (entry_to_update->config.width != w || entry_to_update->config.height != h)
{
TextureCache::TCacheEntryConfig newconfig;
newconfig.width = w;
newconfig.height = h;
newconfig.rendertarget = true;
TCacheEntryBase* newentry = AllocateTexture(newconfig);
newentry->SetGeneralParameters(entry_to_update->addr, entry_to_update->size_in_bytes, entry_to_update->format);
newentry->SetDimensions(entry_to_update->native_width, entry_to_update->native_height, 1);
newentry->SetHashes(entry_to_update->base_hash, entry_to_update->hash);
newentry->frameCount = frameCount;
newentry->is_efb_copy = false;
srcrect.right = entry_to_update->config.width;
srcrect.bottom = entry_to_update->config.height;
dstrect.right = w;
dstrect.bottom = h;
newentry->CopyRectangleFromTexture(entry_to_update, srcrect, dstrect);
entry_to_update = newentry;
u64 key = iter_t->first;
iter_t = FreeTexture(iter_t);
textures_by_address.emplace(key, entry_to_update);
}
}
srcrect.right = entry->config.width;
srcrect.bottom = entry->config.height;
dstrect.left = x * entry_to_update->config.width / entry_to_update->native_width;
dstrect.top = y * entry_to_update->config.height / entry_to_update->native_height;
dstrect.right = (x + entry->native_width) * entry_to_update->config.width / entry_to_update->native_width;
dstrect.bottom = (y + entry->native_height) * entry_to_update->config.height / entry_to_update->native_height;
entry_to_update->CopyRectangleFromTexture(entry, srcrect, dstrect);
// Mark the texture update as used, so it isn't applied more than once
entry->frameCount = frameCount;
}
++iter;
}
return entry_to_update;
}
// block dimensions : widthStride, heightStride
// texture dims : width, height, x offset, y offset
void WriteSwizzler(char*& p, u32 format, API_TYPE ApiType)
{
WRITE(p, "uniform float4 blkDims : register(c%d);\n", C_COLORMATRIX);
WRITE(p, "uniform float4 textureDims : register(c%d);\n", C_COLORMATRIX + 1);
float blkW = (float)TexDecoder_GetBlockWidthInTexels(format);
float blkH = (float)TexDecoder_GetBlockHeightInTexels(format);
float samples = (float)GetEncodedSampleCount(format);
if(ApiType == API_OPENGL)
{
WRITE(p,"uniform samplerRECT samp0 : register(s0);\n");
}
else if (ApiType & API_D3D9)
{
WRITE(p,"uniform sampler samp0 : register(s0);\n");
}
else
{
WRITE(p,"sampler samp0 : register(s0);\n");
WRITE(p, "Texture2D Tex0 : register(t0);\n");
}
WRITE(p,"void main(\n");
if(ApiType != API_D3D11)
{
WRITE(p," out float4 ocol0 : COLOR0,\n");
}
else
{
WRITE(p," out float4 ocol0 : SV_Target,\n");
}
WRITE(p," in float2 uv0 : TEXCOORD0)\n"
"{\n"
" float2 sampleUv;\n"
" float2 uv1 = floor(uv0);\n");
WRITE(p, " uv1.x = uv1.x * %f;\n", samples);
WRITE(p, " float xl = floor(uv1.x / %f);\n", blkW);
WRITE(p, " float xib = uv1.x - (xl * %f);\n", blkW);
WRITE(p, " float yl = floor(uv1.y / %f);\n", blkH);
WRITE(p, " float yb = yl * %f;\n", blkH);
WRITE(p, " float yoff = uv1.y - yb;\n");
WRITE(p, " float xp = uv1.x + (yoff * textureDims.x);\n");
WRITE(p, " float xel = floor(xp / %f);\n", blkW);
WRITE(p, " float xb = floor(xel / %f);\n", blkH);
WRITE(p, " float xoff = xel - (xb * %f);\n", blkH);
WRITE(p, " sampleUv.x = xib + (xb * %f);\n", blkW);
WRITE(p, " sampleUv.y = yb + xoff;\n");
WRITE(p, " sampleUv = sampleUv * blkDims.xy;\n");
if(ApiType == API_OPENGL)
WRITE(p," sampleUv.y = textureDims.y - sampleUv.y;\n");
WRITE(p, " sampleUv = sampleUv + textureDims.zw;\n");
if(ApiType != API_OPENGL)
{
WRITE(p, " sampleUv = sampleUv + float2(0.0f,1.0f);\n");// still to determine the reason for this
WRITE(p, " sampleUv = sampleUv / blkDims.zw;\n");
}
}
void TextureCacheBase::CopyRenderTargetToTexture(u32 dstAddr, unsigned int dstFormat, u32 dstStride, PEControl::PixelFormat srcFormat,
const EFBRectangle& srcRect, bool isIntensity, bool scaleByHalf)
{
// Emulation methods:
//
// - EFB to RAM:
// Encodes the requested EFB data at its native resolution to the emulated RAM using shaders.
// Load() decodes the data from there again (using TextureDecoder) if the EFB copy is being used as a texture again.
// Advantage: CPU can read data from the EFB copy and we don't lose any important updates to the texture
// Disadvantage: Encoding+decoding steps often are redundant because only some games read or modify EFB copies before using them as textures.
//
// - EFB to texture:
// Copies the requested EFB data to a texture object in VRAM, performing any color conversion using shaders.
// Advantage: Works for many games, since in most cases EFB copies aren't read or modified at all before being used as a texture again.
// Since we don't do any further encoding or decoding here, this method is much faster.
// It also allows enhancing the visual quality by doing scaled EFB copies.
//
// - Hybrid EFB copies:
// 1a) Whenever this function gets called, encode the requested EFB data to RAM (like EFB to RAM)
// 1b) Set type to TCET_EC_DYNAMIC for all texture cache entries in the destination address range.
// If EFB copy caching is enabled, further checks will (try to) prevent redundant EFB copies.
// 2) Check if a texture cache entry for the specified dstAddr already exists (i.e. if an EFB copy was triggered to that address before):
// 2a) Entry doesn't exist:
// - Also copy the requested EFB data to a texture object in VRAM (like EFB to texture)
// - Create a texture cache entry for the target (type = TCET_EC_VRAM)
// - Store a hash of the encoded RAM data in the texcache entry.
// 2b) Entry exists AND type is TCET_EC_VRAM:
// - Like case 2a, but reuse the old texcache entry instead of creating a new one.
// 2c) Entry exists AND type is TCET_EC_DYNAMIC:
// - Only encode the texture to RAM (like EFB to RAM) and store a hash of the encoded data in the existing texcache entry.
// - Do NOT copy the requested EFB data to a VRAM object. Reason: the texture is dynamic, i.e. the CPU is modifying it. Storing a VRAM copy is useless, because we'd always end up deleting it and reloading the data from RAM anyway.
// 3) If the EFB copy gets used as a texture, compare the source RAM hash with the hash you stored when encoding the EFB data to RAM.
// 3a) If the two hashes match AND type is TCET_EC_VRAM, reuse the VRAM copy you created
// 3b) If the two hashes differ AND type is TCET_EC_VRAM, screw your existing VRAM copy. Set type to TCET_EC_DYNAMIC.
// Redecode the source RAM data to a VRAM object. The entry basically behaves like a normal texture now.
// 3c) If type is TCET_EC_DYNAMIC, treat the EFB copy like a normal texture.
// Advantage: Non-dynamic EFB copies can be visually enhanced like with EFB to texture.
// Compatibility is as good as EFB to RAM.
// Disadvantage: Slower than EFB to texture and often even slower than EFB to RAM.
// EFB copy cache depends on accurate texture hashing being enabled. However, with accurate hashing you end up being as slow as without a copy cache anyway.
//
// Disadvantage of all methods: Calling this function requires the GPU to perform a pipeline flush which stalls any further CPU processing.
//
// For historical reasons, Dolphin doesn't actually implement "pure" EFB to RAM emulation, but only EFB to texture and hybrid EFB copies.
float colmat[28] = { 0 };
float *const fConstAdd = colmat + 16;
float *const ColorMask = colmat + 20;
ColorMask[0] = ColorMask[1] = ColorMask[2] = ColorMask[3] = 255.0f;
ColorMask[4] = ColorMask[5] = ColorMask[6] = ColorMask[7] = 1.0f / 255.0f;
unsigned int cbufid = -1;
bool efbHasAlpha = bpmem.zcontrol.pixel_format == PEControl::RGBA6_Z24;
if (srcFormat == PEControl::Z24)
{
switch (dstFormat)
{
case 0: // Z4
colmat[3] = colmat[7] = colmat[11] = colmat[15] = 1.0f;
cbufid = 0;
dstFormat |= _GX_TF_CTF;
break;
case 8: // Z8H
dstFormat |= _GX_TF_CTF;
case 1: // Z8
colmat[0] = colmat[4] = colmat[8] = colmat[12] = 1.0f;
cbufid = 1;
break;
case 3: // Z16
colmat[1] = colmat[5] = colmat[9] = colmat[12] = 1.0f;
cbufid = 2;
break;
case 11: // Z16 (reverse order)
colmat[0] = colmat[4] = colmat[8] = colmat[13] = 1.0f;
cbufid = 3;
dstFormat |= _GX_TF_CTF;
break;
case 6: // Z24X8
colmat[0] = colmat[5] = colmat[10] = 1.0f;
cbufid = 4;
break;
case 9: // Z8M
colmat[1] = colmat[5] = colmat[9] = colmat[13] = 1.0f;
cbufid = 5;
dstFormat |= _GX_TF_CTF;
break;
case 10: // Z8L
colmat[2] = colmat[6] = colmat[10] = colmat[14] = 1.0f;
cbufid = 6;
dstFormat |= _GX_TF_CTF;
break;
case 12: // Z16L - copy lower 16 depth bits
// expected to be used as an IA8 texture (upper 8 bits stored as intensity, lower 8 bits stored as alpha)
// Used e.g. in Zelda: Skyward Sword
colmat[1] = colmat[5] = colmat[9] = colmat[14] = 1.0f;
cbufid = 7;
dstFormat |= _GX_TF_CTF;
break;
default:
ERROR_LOG(VIDEO, "Unknown copy zbuf format: 0x%x", dstFormat);
colmat[2] = colmat[5] = colmat[8] = 1.0f;
cbufid = 8;
break;
}
dstFormat |= _GX_TF_ZTF;
}
else if (isIntensity)
{
fConstAdd[0] = fConstAdd[1] = fConstAdd[2] = 16.0f / 255.0f;
switch (dstFormat)
{
case 0: // I4
case 1: // I8
case 2: // IA4
case 3: // IA8
case 8: // I8
// TODO - verify these coefficients
colmat[0] = 0.257f; colmat[1] = 0.504f; colmat[2] = 0.098f;
colmat[4] = 0.257f; colmat[5] = 0.504f; colmat[6] = 0.098f;
colmat[8] = 0.257f; colmat[9] = 0.504f; colmat[10] = 0.098f;
if (dstFormat < 2 || dstFormat == 8)
{
colmat[12] = 0.257f; colmat[13] = 0.504f; colmat[14] = 0.098f;
fConstAdd[3] = 16.0f / 255.0f;
if (dstFormat == 0)
{
ColorMask[0] = ColorMask[1] = ColorMask[2] = 15.0f;
ColorMask[4] = ColorMask[5] = ColorMask[6] = 1.0f / 15.0f;
cbufid = 9;
}
else
{
cbufid = 10;
}
}
else// alpha
{
colmat[15] = 1;
if (dstFormat == 2)
{
ColorMask[0] = ColorMask[1] = ColorMask[2] = ColorMask[3] = 15.0f;
ColorMask[4] = ColorMask[5] = ColorMask[6] = ColorMask[7] = 1.0f / 15.0f;
cbufid = 11;
}
else
{
cbufid = 12;
}
}
break;
default:
ERROR_LOG(VIDEO, "Unknown copy intensity format: 0x%x", dstFormat);
colmat[0] = colmat[5] = colmat[10] = colmat[15] = 1.0f;
cbufid = 13;
break;
}
}
else
{
switch (dstFormat)
{
case 0: // R4
colmat[0] = colmat[4] = colmat[8] = colmat[12] = 1;
ColorMask[0] = 15.0f;
ColorMask[4] = 1.0f / 15.0f;
cbufid = 14;
dstFormat |= _GX_TF_CTF;
break;
case 1: // R8
case 8: // R8
colmat[0] = colmat[4] = colmat[8] = colmat[12] = 1;
cbufid = 15;
dstFormat = GX_CTF_R8;
break;
case 2: // RA4
colmat[0] = colmat[4] = colmat[8] = colmat[15] = 1.0f;
ColorMask[0] = ColorMask[3] = 15.0f;
ColorMask[4] = ColorMask[7] = 1.0f / 15.0f;
cbufid = 16;
if (!efbHasAlpha)
{
ColorMask[3] = 0.0f;
fConstAdd[3] = 1.0f;
cbufid = 17;
}
dstFormat |= _GX_TF_CTF;
break;
case 3: // RA8
colmat[0] = colmat[4] = colmat[8] = colmat[15] = 1.0f;
cbufid = 18;
if (!efbHasAlpha)
{
ColorMask[3] = 0.0f;
fConstAdd[3] = 1.0f;
cbufid = 19;
}
dstFormat |= _GX_TF_CTF;
break;
case 7: // A8
colmat[3] = colmat[7] = colmat[11] = colmat[15] = 1.0f;
cbufid = 20;
if (!efbHasAlpha)
{
ColorMask[3] = 0.0f;
fConstAdd[0] = 1.0f;
fConstAdd[1] = 1.0f;
fConstAdd[2] = 1.0f;
fConstAdd[3] = 1.0f;
cbufid = 21;
}
dstFormat |= _GX_TF_CTF;
break;
case 9: // G8
colmat[1] = colmat[5] = colmat[9] = colmat[13] = 1.0f;
cbufid = 22;
dstFormat |= _GX_TF_CTF;
break;
case 10: // B8
colmat[2] = colmat[6] = colmat[10] = colmat[14] = 1.0f;
cbufid = 23;
dstFormat |= _GX_TF_CTF;
break;
case 11: // RG8
colmat[0] = colmat[4] = colmat[8] = colmat[13] = 1.0f;
cbufid = 24;
dstFormat |= _GX_TF_CTF;
break;
case 12: // GB8
colmat[1] = colmat[5] = colmat[9] = colmat[14] = 1.0f;
cbufid = 25;
dstFormat |= _GX_TF_CTF;
break;
case 4: // RGB565
colmat[0] = colmat[5] = colmat[10] = 1.0f;
ColorMask[0] = ColorMask[2] = 31.0f;
ColorMask[4] = ColorMask[6] = 1.0f / 31.0f;
ColorMask[1] = 63.0f;
ColorMask[5] = 1.0f / 63.0f;
fConstAdd[3] = 1.0f; // set alpha to 1
cbufid = 26;
break;
case 5: // RGB5A3
colmat[0] = colmat[5] = colmat[10] = colmat[15] = 1.0f;
ColorMask[0] = ColorMask[1] = ColorMask[2] = 31.0f;
ColorMask[4] = ColorMask[5] = ColorMask[6] = 1.0f / 31.0f;
ColorMask[3] = 7.0f;
ColorMask[7] = 1.0f / 7.0f;
cbufid = 27;
if (!efbHasAlpha)
{
ColorMask[3] = 0.0f;
fConstAdd[3] = 1.0f;
cbufid = 28;
}
break;
case 6: // RGBA8
colmat[0] = colmat[5] = colmat[10] = colmat[15] = 1.0f;
cbufid = 29;
if (!efbHasAlpha)
{
ColorMask[3] = 0.0f;
fConstAdd[3] = 1.0f;
cbufid = 30;
}
break;
default:
ERROR_LOG(VIDEO, "Unknown copy color format: 0x%x", dstFormat);
colmat[0] = colmat[5] = colmat[10] = colmat[15] = 1.0f;
cbufid = 31;
break;
}
}
u8* dst = Memory::GetPointer(dstAddr);
if (dst == nullptr)
{
ERROR_LOG(VIDEO, "Trying to copy from EFB to invalid address 0x%8x", dstAddr);
return;
}
const unsigned int tex_w = scaleByHalf ? srcRect.GetWidth() / 2 : srcRect.GetWidth();
const unsigned int tex_h = scaleByHalf ? srcRect.GetHeight() / 2 : srcRect.GetHeight();
unsigned int scaled_tex_w = g_ActiveConfig.bCopyEFBScaled ? Renderer::EFBToScaledX(tex_w) : tex_w;
unsigned int scaled_tex_h = g_ActiveConfig.bCopyEFBScaled ? Renderer::EFBToScaledY(tex_h) : tex_h;
// Remove all texture cache entries at dstAddr
// It's not possible to have two EFB copies at the same address, this makes sure any old efb copies
// (or normal textures) are removed from texture cache. They are also un-linked from any partially
// updated textures, which forces that partially updated texture to be updated.
// TODO: This also wipes out non-efb copies, which is counterproductive.
{
std::pair<TexCache::iterator, TexCache::iterator> iter_range = textures_by_address.equal_range((u64)dstAddr);
TexCache::iterator iter = iter_range.first;
while (iter != iter_range.second)
{
iter = InvalidateTexture(iter);
}
}
// Get the base (in memory) format of this efb copy.
int baseFormat = TexDecoder_GetEfbCopyBaseFormat(dstFormat);
u32 blockH = TexDecoder_GetBlockHeightInTexels(baseFormat);
const u32 blockW = TexDecoder_GetBlockWidthInTexels(baseFormat);
// Round up source height to multiple of block size
u32 actualHeight = ROUND_UP(tex_h, blockH);
const u32 actualWidth = ROUND_UP(tex_w, blockW);
u32 num_blocks_y = actualHeight / blockH;
const u32 num_blocks_x = actualWidth / blockW;
// RGBA takes two cache lines per block; all others take one
const u32 bytes_per_block = baseFormat == GX_TF_RGBA8 ? 64 : 32;
u32 bytes_per_row = num_blocks_x * bytes_per_block;
bool copy_to_ram = !g_ActiveConfig.bSkipEFBCopyToRam;
bool copy_to_vram = true;
if (copy_to_ram)
{
g_texture_cache->CopyEFB(
dst,
dstFormat,
tex_w,
bytes_per_row,
num_blocks_y,
dstStride,
srcFormat,
srcRect,
isIntensity,
scaleByHalf);
}
else
{
// Hack: Most games don't actually need the correct texture data in RAM
// and we can just keep a copy in VRAM. We zero the memory so we
// can check it hasn't changed before using our copy in VRAM.
u8* ptr = dst;
for (u32 i = 0; i < num_blocks_y; i++)
{
memset(ptr, 0, bytes_per_row);
ptr += dstStride;
}
}
if (g_bRecordFifoData)
{
// Mark the memory behind this efb copy as dynamicly generated for the Fifo log
u32 address = dstAddr;
for (u32 i = 0; i < num_blocks_y; i++)
{
FifoRecorder::GetInstance().UseMemory(address, bytes_per_row, MemoryUpdate::TEXTURE_MAP, true);
address += dstStride;
}
}
if (dstStride < bytes_per_row)
{
// This kind of efb copy results in a scrambled image.
// I'm pretty sure no game actually wants to do this, it might be caused by a
// programming bug in the game, or a CPU/Bounding box emulation issue with dolphin.
// The copy_to_ram code path above handles this "correctly" and scrambles the image
// but the copy_to_vram code path just saves and uses unscrambled texture instead.
// To avoid a "incorrect" result, we simply skip doing the copy_to_vram code path
// so if the game does try to use the scrambled texture, dolphin will grab the scrambled
// texture (or black if copy_to_ram is also disabled) out of ram.
ERROR_LOG(VIDEO, "Memory stride too small (%i < %i)", dstStride, bytes_per_row);
copy_to_vram = false;
}
// Invalidate all textures that overlap the range of our efb copy.
// Unless our efb copy has a weird stride, then we want avoid invalidating textures which
// we might be able to do a partial texture update on.
// TODO: This also invalidates partial overlaps, which we currently don't have a better way
// of dealing with.
if (dstStride == bytes_per_row || !copy_to_vram)
{
TexCache::iterator iter = textures_by_address.begin();
while (iter != textures_by_address.end())
{
if (iter->second->addr + iter->second->size_in_bytes <= dstAddr || iter->second->addr >= dstAddr + num_blocks_y * dstStride)
++iter;
else
iter = InvalidateTexture(iter);
}
}
if (copy_to_vram)
{
// create the texture
TCacheEntryConfig config;
config.rendertarget = true;
config.width = scaled_tex_w;
config.height = scaled_tex_h;
config.layers = FramebufferManagerBase::GetEFBLayers();
TCacheEntryBase* entry = AllocateTexture(config);
if (entry)
{
entry->SetGeneralParameters(dstAddr, 0, baseFormat);
entry->SetDimensions(tex_w, tex_h, 1);
entry->frameCount = FRAMECOUNT_INVALID;
entry->SetEfbCopy(dstStride);
entry->is_custom_tex = false;
entry->FromRenderTarget(dst, srcFormat, srcRect, scaleByHalf, cbufid, colmat);
u64 hash = entry->CalculateHash();
entry->SetHashes(hash, hash);
if (g_ActiveConfig.bDumpEFBTarget)
{
static int count = 0;
entry->Save(StringFromFormat("%sefb_frame_%i.png", File::GetUserPath(D_DUMPTEXTURES_IDX).c_str(),
count++), 0);
}
textures_by_address.emplace((u64)dstAddr, entry);
}
}
}
TextureCacheBase::TCacheEntryBase* TextureCacheBase::DoPartialTextureUpdates(TexCache::iterator iter_t, u8* palette, u32 tlutfmt)
{
TCacheEntryBase* entry_to_update = iter_t->second;
const bool isPaletteTexture = (entry_to_update->format == GX_TF_C4
|| entry_to_update->format == GX_TF_C8
|| entry_to_update->format == GX_TF_C14X2
|| entry_to_update->format >= 0x10000);
// EFB copies are excluded from these updates, until there's an example where a game would
// benefit from updating. This would require more work to be done.
if (entry_to_update->IsEfbCopy())
return entry_to_update;
u32 block_width = TexDecoder_GetBlockWidthInTexels(entry_to_update->format & 0xf);
u32 block_height = TexDecoder_GetBlockHeightInTexels(entry_to_update->format & 0xf);
u32 block_size = block_width * block_height * TexDecoder_GetTexelSizeInNibbles(entry_to_update->format & 0xf) / 2;
u32 numBlocksX = (entry_to_update->native_width + block_width - 1) / block_width;
TexCache::iterator iter = textures_by_address.lower_bound(entry_to_update->addr > MAX_TEXTURE_BINARY_SIZE ? entry_to_update->addr - MAX_TEXTURE_BINARY_SIZE : 0);
TexCache::iterator iterend = textures_by_address.upper_bound(entry_to_update->addr + entry_to_update->size_in_bytes);
while (iter != iterend)
{
TCacheEntryBase* entry = iter->second;
if (entry != entry_to_update
&& entry->IsEfbCopy()
&& entry->references.count(entry_to_update) == 0
&& entry->OverlapsMemoryRange(entry_to_update->addr, entry_to_update->size_in_bytes)
&& entry->memory_stride == numBlocksX * block_size)
{
if (entry->hash == entry->CalculateHash())
{
if (isPaletteTexture)
{
TCacheEntryBase *decoded_entry = entry->ApplyPalette(palette, tlutfmt);
if (decoded_entry)
{
// Link the efb copy with the partially updated texture, so we won't apply this partial update again
entry->CreateReference(entry_to_update);
// Mark the texture update as used, as if it was loaded directly
entry->frameCount = FRAMECOUNT_INVALID;
entry = decoded_entry;
}
else
{
++iter;
continue;
}
}
u32 src_x, src_y, dst_x, dst_y;
// Note for understanding the math:
// Normal textures can't be strided, so the 2 missing cases with src_x > 0 don't exist
if (entry->addr >= entry_to_update->addr)
{
u32 block_offset = (entry->addr - entry_to_update->addr) / block_size;
u32 block_x = block_offset % numBlocksX;
u32 block_y = block_offset / numBlocksX;
src_x = 0;
src_y = 0;
dst_x = block_x * block_width;
dst_y = block_y * block_height;
}
else
{
u32 block_offset = (entry_to_update->addr - entry->addr) / block_size;
u32 block_x = (~block_offset + 1) % numBlocksX;
u32 block_y = (block_offset + block_x) / numBlocksX;
src_x = 0;
src_y = block_y * block_height;
dst_x = block_x * block_width;
dst_y = 0;
}
u32 copy_width = std::min(entry->native_width - src_x, entry_to_update->native_width - dst_x);
u32 copy_height = std::min(entry->native_height - src_y, entry_to_update->native_height - dst_y);
// If one of the textures is scaled, scale both with the current efb scaling factor
if (entry_to_update->native_width != entry_to_update->config.width
|| entry_to_update->native_height != entry_to_update->config.height
|| entry->native_width != entry->config.width || entry->native_height != entry->config.height)
{
ScaleTextureCacheEntryTo(&entry_to_update, Renderer::EFBToScaledX(entry_to_update->native_width), Renderer::EFBToScaledY(entry_to_update->native_height));
ScaleTextureCacheEntryTo(&entry, Renderer::EFBToScaledX(entry->native_width), Renderer::EFBToScaledY(entry->native_height));
src_x = Renderer::EFBToScaledX(src_x);
src_y = Renderer::EFBToScaledY(src_y);
dst_x = Renderer::EFBToScaledX(dst_x);
dst_y = Renderer::EFBToScaledY(dst_y);
copy_width = Renderer::EFBToScaledX(copy_width);
copy_height = Renderer::EFBToScaledY(copy_height);
}
MathUtil::Rectangle<int> srcrect, dstrect;
srcrect.left = src_x;
srcrect.top = src_y;
srcrect.right = (src_x + copy_width);
srcrect.bottom = (src_y + copy_height);
dstrect.left = dst_x;
dstrect.top = dst_y;
dstrect.right = (dst_x + copy_width);
dstrect.bottom = (dst_y + copy_height);
entry_to_update->CopyRectangleFromTexture(entry, srcrect, dstrect);
if (isPaletteTexture)
{
// Remove the temporary converted texture, it won't be used anywhere else
// TODO: It would be nice to convert and copy in one step, but this code path isn't common
InvalidateTexture(GetTexCacheIter(entry));
}
else
{
// Link the two textures together, so we won't apply this partial update again
entry->CreateReference(entry_to_update);
// Mark the texture update as used, as if it was loaded directly
entry->frameCount = FRAMECOUNT_INVALID;
}
}
else
{
// If the hash does not match, this EFB copy will not be used for anything, so remove it
iter = InvalidateTexture(iter);
continue;
}
}
++iter;
}
return entry_to_update;
}
// block dimensions : widthStride, heightStride
// texture dims : width, height, x offset, y offset
static void WriteSwizzler(char*& p, u32 format, APIType ApiType)
{
// left, top, of source rectangle within source texture
// width of the destination rectangle, scale_factor (1 or 2)
WRITE(p, "uniform int4 position;\n");
int blkW = TexDecoder_GetBlockWidthInTexels(format);
int blkH = TexDecoder_GetBlockHeightInTexels(format);
int samples = GetEncodedSampleCount(format);
if (ApiType == APIType::OpenGL)
{
WRITE(p, "#define samp0 samp9\n");
WRITE(p, "SAMPLER_BINDING(9) uniform sampler2DArray samp0;\n");
WRITE(p, " out vec4 ocol0;\n");
WRITE(p, "void main()\n");
WRITE(p, "{\n"
" int2 sampleUv;\n"
" int2 uv1 = int2(gl_FragCoord.xy);\n");
}
else // D3D
{
WRITE(p, "sampler samp0 : register(s0);\n");
WRITE(p, "Texture2DArray Tex0 : register(t0);\n");
WRITE(p, "void main(\n");
WRITE(p, " out float4 ocol0 : SV_Target, in float4 rawpos : SV_Position)\n");
WRITE(p, "{\n"
" int2 sampleUv;\n"
" int2 uv1 = int2(rawpos.xy);\n");
}
WRITE(p, " int x_block_position = (uv1.x >> %d) << %d;\n", IntLog2(blkH * blkW / samples),
IntLog2(blkW));
WRITE(p, " int y_block_position = uv1.y << %d;\n", IntLog2(blkH));
if (samples == 1)
{
// With samples == 1, we write out pairs of blocks; one A8R8, one G8B8.
WRITE(p, " bool first = (uv1.x & %d) == 0;\n", blkH * blkW / 2);
samples = 2;
}
WRITE(p, " int offset_in_block = uv1.x & %d;\n", (blkH * blkW / samples) - 1);
WRITE(p, " int y_offset_in_block = offset_in_block >> %d;\n", IntLog2(blkW / samples));
WRITE(p, " int x_offset_in_block = (offset_in_block & %d) << %d;\n", (blkW / samples) - 1,
IntLog2(samples));
WRITE(p, " sampleUv.x = x_block_position + x_offset_in_block;\n");
WRITE(p, " sampleUv.y = y_block_position + y_offset_in_block;\n");
WRITE(p,
" float2 uv0 = float2(sampleUv);\n"); // sampleUv is the sample position in (int)gx_coords
WRITE(p, " uv0 += float2(0.5, 0.5);\n"); // move to center of pixel
WRITE(p, " uv0 *= float(position.w);\n"); // scale by two if needed (also move to pixel borders
// so that linear filtering will average adjacent
// pixel)
WRITE(p, " uv0 += float2(position.xy);\n"); // move to copied rect
WRITE(p, " uv0 /= float2(%d, %d);\n", EFB_WIDTH, EFB_HEIGHT); // normalize to [0:1]
if (ApiType == APIType::OpenGL) // ogl has to flip up and down
{
WRITE(p, " uv0.y = 1.0-uv0.y;\n");
}
WRITE(p, " float sample_offset = float(position.w) / float(%d);\n", EFB_WIDTH);
}
TextureCacheBase::TCacheEntryBase* TextureCacheBase::DoPartialTextureUpdates(TexCache::iterator iter_t)
{
TCacheEntryBase* entry_to_update = iter_t->second;
const bool isPaletteTexture = (entry_to_update->format == GX_TF_C4
|| entry_to_update->format == GX_TF_C8
|| entry_to_update->format == GX_TF_C14X2
|| entry_to_update->format >= 0x10000);
// Efb copies and paletted textures are excluded from these updates, until there's an example where a game would
// benefit from this. Both would require more work to be done.
if (entry_to_update->IsEfbCopy()
|| isPaletteTexture)
return entry_to_update;
u32 block_width = TexDecoder_GetBlockWidthInTexels(entry_to_update->format & 0xf);
u32 block_height = TexDecoder_GetBlockHeightInTexels(entry_to_update->format & 0xf);
u32 block_size = block_width * block_height * TexDecoder_GetTexelSizeInNibbles(entry_to_update->format & 0xf) / 2;
u32 numBlocksX = (entry_to_update->native_width + block_width - 1) / block_width;
TexCache::iterator iter = textures_by_address.lower_bound(entry_to_update->addr);
TexCache::iterator iterend = textures_by_address.upper_bound(entry_to_update->addr + entry_to_update->size_in_bytes);
while (iter != iterend)
{
TCacheEntryBase* entry = iter->second;
if (entry != entry_to_update
&& entry->IsEfbCopy()
&& entry->OverlapsMemoryRange(entry_to_update->addr, entry_to_update->size_in_bytes)
&& entry->frameCount == FRAMECOUNT_INVALID
&& entry->memory_stride == numBlocksX * block_size)
{
if (entry->hash == entry->CalculateHash())
{
u32 src_x, src_y, dst_x, dst_y;
// Note for understanding the math:
// Normal textures can't be strided, so the 2 missing cases with src_x > 0 don't exist
if (entry->addr >= entry_to_update->addr)
{
u32 block_offset = (entry->addr - entry_to_update->addr) / block_size;
u32 block_x = block_offset % numBlocksX;
u32 block_y = block_offset / numBlocksX;
src_x = 0;
src_y = 0;
dst_x = block_x * block_width;
dst_y = block_y * block_height;
}
else
{
u32 block_offset = (entry_to_update->addr - entry->addr) / block_size;
u32 block_x = (~block_offset + 1) % numBlocksX;
u32 block_y = (block_offset + block_x) / numBlocksX;
src_x = 0;
src_y = block_y * block_height;
dst_x = block_x * block_width;
dst_y = 0;
}
u32 copy_width = std::min(entry->native_width - src_x, entry_to_update->native_width - dst_x);
u32 copy_height = std::min(entry->native_height - src_y, entry_to_update->native_height - dst_y);
// If one of the textures is scaled, scale both with the current efb scaling factor
if (entry_to_update->native_width != entry_to_update->config.width
|| entry_to_update->native_height != entry_to_update->config.height
|| entry->native_width != entry->config.width || entry->native_height != entry->config.height)
{
ScaleTextureCacheEntryTo(&entry_to_update, Renderer::EFBToScaledX(entry_to_update->native_width), Renderer::EFBToScaledY(entry_to_update->native_height));
ScaleTextureCacheEntryTo(&entry, Renderer::EFBToScaledX(entry->native_width), Renderer::EFBToScaledY(entry->native_height));
src_x = Renderer::EFBToScaledX(src_x);
src_y = Renderer::EFBToScaledY(src_y);
dst_x = Renderer::EFBToScaledX(dst_x);
dst_y = Renderer::EFBToScaledY(dst_y);
copy_width = Renderer::EFBToScaledX(copy_width);
copy_height = Renderer::EFBToScaledY(copy_height);
}
MathUtil::Rectangle<int> srcrect, dstrect;
srcrect.left = src_x;
srcrect.top = src_y;
srcrect.right = (src_x + copy_width);
srcrect.bottom = (src_y + copy_height);
dstrect.left = dst_x;
dstrect.top = dst_y;
dstrect.right = (dst_x + copy_width);
dstrect.bottom = (dst_y + copy_height);
entry_to_update->CopyRectangleFromTexture(entry, srcrect, dstrect);
// Mark the texture update as used, so it isn't applied more than once
entry->frameCount = frameCount;
}
else
{
// If the hash does not match, this EFB copy will not be used for anything, so remove it
iter = FreeTexture(iter);
continue;
}
}
++iter;
}
return entry_to_update;
}
