void VertexShaderCache::Init()
{
const D3D11_INPUT_ELEMENT_DESC simpleelems[2] = {
{"POSITION", 0, DXGI_FORMAT_R32G32B32_FLOAT, 0, 0, D3D11_INPUT_PER_VERTEX_DATA, 0},
{"TEXCOORD", 0, DXGI_FORMAT_R32G32B32_FLOAT, 0, 12, D3D11_INPUT_PER_VERTEX_DATA, 0},
};
const D3D11_INPUT_ELEMENT_DESC clearelems[2] = {
{"POSITION", 0, DXGI_FORMAT_R32G32B32_FLOAT, 0, 0, D3D11_INPUT_PER_VERTEX_DATA, 0},
{"COLOR", 0, DXGI_FORMAT_R8G8B8A8_UNORM, 0, 12, D3D11_INPUT_PER_VERTEX_DATA, 0},
};
unsigned int cbsize = ROUND_UP(sizeof(VertexShaderConstants), 16); // must be a multiple of 16
D3D11_BUFFER_DESC cbdesc = CD3D11_BUFFER_DESC(cbsize, D3D11_BIND_CONSTANT_BUFFER,
D3D11_USAGE_DYNAMIC, D3D11_CPU_ACCESS_WRITE);
HRESULT hr = D3D::device->CreateBuffer(&cbdesc, nullptr, &vscbuf);
CHECK(hr == S_OK, "Create vertex shader constant buffer (size=%u)", cbsize);
D3D::SetDebugObjectName((ID3D11DeviceChild*)vscbuf,
"vertex shader constant buffer used to emulate the GX pipeline");
D3DBlob* blob;
D3D::CompileVertexShader(simple_shader_code, &blob);
D3D::device->CreateInputLayout(simpleelems, 2, blob->Data(), blob->Size(), &SimpleLayout);
SimpleVertexShader = D3D::CreateVertexShaderFromByteCode(blob);
if (SimpleLayout == nullptr || SimpleVertexShader == nullptr)
PanicAlert("Failed to create simple vertex shader or input layout at %s %d\n", __FILE__,
__LINE__);
blob->Release();
D3D::SetDebugObjectName((ID3D11DeviceChild*)SimpleVertexShader, "simple vertex shader");
D3D::SetDebugObjectName((ID3D11DeviceChild*)SimpleLayout, "simple input layout");
D3D::CompileVertexShader(clear_shader_code, &blob);
D3D::device->CreateInputLayout(clearelems, 2, blob->Data(), blob->Size(), &ClearLayout);
ClearVertexShader = D3D::CreateVertexShaderFromByteCode(blob);
if (ClearLayout == nullptr || ClearVertexShader == nullptr)
PanicAlert("Failed to create clear vertex shader or input layout at %s %d\n", __FILE__,
__LINE__);
blob->Release();
D3D::SetDebugObjectName((ID3D11DeviceChild*)ClearVertexShader, "clear vertex shader");
D3D::SetDebugObjectName((ID3D11DeviceChild*)ClearLayout, "clear input layout");
Clear();
if (!File::Exists(File::GetUserPath(D_SHADERCACHE_IDX)))
File::CreateDir(File::GetUserPath(D_SHADERCACHE_IDX));
SETSTAT(stats.numVertexShadersCreated, 0);
SETSTAT(stats.numVertexShadersAlive, 0);
std::string cache_filename =
StringFromFormat("%sdx11-%s-vs.cache", File::GetUserPath(D_SHADERCACHE_IDX).c_str(),
SConfig::GetInstance().m_strUniqueID.c_str());
VertexShaderCacheInserter inserter;
g_vs_disk_cache.OpenAndRead(cache_filename, inserter);
if (g_Config.bEnableShaderDebugging)
Clear();
last_entry = nullptr;
}
void PixelShaderCache::Init()
{
last_entry = NULL;
//program used for clear screen
{
char pprog[3072];
sprintf(pprog, "void main(\n"
"out float4 ocol0 : COLOR0,\n"
" in float4 incol0 : COLOR0){\n"
"ocol0 = incol0;\n"
"}\n");
s_ClearProgram = D3D::CompileAndCreatePixelShader(pprog, (int)strlen(pprog));
}
int shaderModel = ((D3D::GetCaps().PixelShaderVersion >> 8) & 0xFF);
int maxConstants = (shaderModel < 3) ? 32 : ((shaderModel < 4) ? 224 : 65536);
// other screen copy/convert programs
for(int copyMatrixType = 0; copyMatrixType < NUM_COPY_TYPES; copyMatrixType++)
{
for(int depthType = 0; depthType < NUM_DEPTH_CONVERSION_TYPES; depthType++)
{
for(int ssaaMode = 0; ssaaMode < MAX_SSAA_SHADERS; ssaaMode++)
{
if(ssaaMode && !s_CopyProgram[copyMatrixType][depthType][ssaaMode-1]
|| depthType && !s_CopyProgram[copyMatrixType][depthType-1][ssaaMode]
|| copyMatrixType && !s_CopyProgram[copyMatrixType-1][depthType][ssaaMode])
{
// if it failed at a lower setting, it's going to fail here for the same reason it did there,
// so skip this attempt to avoid duplicate error messages.
s_CopyProgram[copyMatrixType][depthType][ssaaMode] = NULL;
}
else
{
s_CopyProgram[copyMatrixType][depthType][ssaaMode] = CreateCopyShader(copyMatrixType, depthType, ssaaMode);
}
}
}
}
Clear();
if (!File::Exists(File::GetUserPath(D_SHADERCACHE_IDX)))
File::CreateDir(File::GetUserPath(D_SHADERCACHE_IDX).c_str());
SETSTAT(stats.numPixelShadersCreated, 0);
SETSTAT(stats.numPixelShadersAlive, 0);
char cache_filename[MAX_PATH];
sprintf(cache_filename, "%sdx9-%s-ps.cache", File::GetUserPath(D_SHADERCACHE_IDX).c_str(),
SConfig::GetInstance().m_LocalCoreStartupParameter.m_strUniqueID.c_str());
PixelShaderCacheInserter inserter;
g_ps_disk_cache.OpenAndRead(cache_filename, inserter);
if (g_Config.bEnableShaderDebugging)
Clear();
}
void VertexShaderCache::Init()
{
const D3D11_INPUT_ELEMENT_DESC simpleelems[2] = {
{"POSITION", 0, DXGI_FORMAT_R32G32B32_FLOAT, 0, 0, D3D11_INPUT_PER_VERTEX_DATA, 0},
{"TEXCOORD", 0, DXGI_FORMAT_R32G32B32_FLOAT, 0, 12, D3D11_INPUT_PER_VERTEX_DATA, 0},
};
const D3D11_INPUT_ELEMENT_DESC clearelems[2] = {
{"POSITION", 0, DXGI_FORMAT_R32G32B32_FLOAT, 0, 0, D3D11_INPUT_PER_VERTEX_DATA, 0},
{"COLOR", 0, DXGI_FORMAT_R8G8B8A8_UNORM, 0, 12, D3D11_INPUT_PER_VERTEX_DATA, 0},
};
unsigned int cbsize = Common::AlignUp(static_cast<unsigned int>(sizeof(VertexShaderConstants)),
16); // must be a multiple of 16
D3D11_BUFFER_DESC cbdesc = CD3D11_BUFFER_DESC(cbsize, D3D11_BIND_CONSTANT_BUFFER,
D3D11_USAGE_DYNAMIC, D3D11_CPU_ACCESS_WRITE);
HRESULT hr = D3D::device->CreateBuffer(&cbdesc, nullptr, &vscbuf);
CHECK(hr == S_OK, "Create vertex shader constant buffer (size=%u)", cbsize);
D3D::SetDebugObjectName(vscbuf, "vertex shader constant buffer used to emulate the GX pipeline");
D3DBlob* blob;
D3D::CompileVertexShader(simple_shader_code, &blob);
D3D::device->CreateInputLayout(simpleelems, 2, blob->Data(), blob->Size(), &SimpleLayout);
SimpleVertexShader = D3D::CreateVertexShaderFromByteCode(blob);
if (SimpleLayout == nullptr || SimpleVertexShader == nullptr)
PanicAlert("Failed to create simple vertex shader or input layout at %s %d\n", __FILE__,
__LINE__);
blob->Release();
D3D::SetDebugObjectName(SimpleVertexShader, "simple vertex shader");
D3D::SetDebugObjectName(SimpleLayout, "simple input layout");
D3D::CompileVertexShader(clear_shader_code, &blob);
D3D::device->CreateInputLayout(clearelems, 2, blob->Data(), blob->Size(), &ClearLayout);
ClearVertexShader = D3D::CreateVertexShaderFromByteCode(blob);
if (ClearLayout == nullptr || ClearVertexShader == nullptr)
PanicAlert("Failed to create clear vertex shader or input layout at %s %d\n", __FILE__,
__LINE__);
blob->Release();
D3D::SetDebugObjectName(ClearVertexShader, "clear vertex shader");
D3D::SetDebugObjectName(ClearLayout, "clear input layout");
Clear();
SETSTAT(stats.numVertexShadersCreated, 0);
SETSTAT(stats.numVertexShadersAlive, 0);
if (g_ActiveConfig.bShaderCache)
LoadShaderCache();
g_async_compiler = std::make_unique<VideoCommon::AsyncShaderCompiler>();
g_async_compiler->ResizeWorkerThreads(g_ActiveConfig.CanPrecompileUberShaders() ?
g_ActiveConfig.GetShaderPrecompilerThreads() :
g_ActiveConfig.GetShaderCompilerThreads());
if (g_ActiveConfig.CanPrecompileUberShaders())
QueueUberShaderCompiles();
}
void PixelShaderCache::Init()
{
Clear();
SETSTAT(stats.numPixelShadersCreated, 0);
SETSTAT(stats.numPixelShadersAlive, 0);
last_entry = nullptr;
}
ssize_t
_sys_read(
struct fdinfo *fio,
bitptr bufptr,
size_t nbytes,
struct ffsw *retstat,
int fulp,
int *ubc)
{
ssize_t ret = 0;
char *buf;
buf = BPTR2CP(bufptr);
if ((BPBITOFF(bufptr) & 7) != 0 || *ubc != 0)
ERETURN(retstat, FDC_ERR_UBC, 0);
#ifdef __mips
/*
* If our last i/o was asynchronous, then our file position
* won't be what we expect. Seek to the right position. We
* could use a pread instead of seeking, but that would also
* not update the file position. I'm doing this because it seems
* to me most 'expected' for the system call layer.
*/
if (((struct sys_f *)fio->lyr_info)->needpos) {
if (lseek( fio->realfd, ((struct sys_f *)fio->lyr_info)->curpos,
0) < 0)
ERETURN(retstat, errno, 0);
((struct sys_f *)fio->lyr_info)->needpos = 0;
}
#endif
if (nbytes > 0)
{
if (((struct sys_f *)fio->lyr_info)->nointrio)
ret = read(fio->realfd, buf, nbytes);
else {
LOOP_SYSCALL(ret, read(fio->realfd, buf, nbytes));
}
if (ret < 0)
ERETURN(retstat, errno, 0);
}
if (ret == 0 && nbytes > 0) {
SETSTAT(retstat, FFEOD, ret);
}
else {
SETSTAT(retstat, FFCNT, ret);
#ifdef __mips
((struct sys_f *)(fio->lyr_info))->curpos += ret;
#endif
}
return (ret);
}
void PixelShaderCache::Init()
{
unsigned int cbsize = Common::AlignUp(static_cast<unsigned int>(sizeof(PixelShaderConstants)),
16); // must be a multiple of 16
D3D11_BUFFER_DESC cbdesc = CD3D11_BUFFER_DESC(cbsize, D3D11_BIND_CONSTANT_BUFFER,
D3D11_USAGE_DYNAMIC, D3D11_CPU_ACCESS_WRITE);
D3D::device->CreateBuffer(&cbdesc, nullptr, &pscbuf);
CHECK(pscbuf != nullptr, "Create pixel shader constant buffer");
D3D::SetDebugObjectName((ID3D11DeviceChild*)pscbuf,
"pixel shader constant buffer used to emulate the GX pipeline");
// used when drawing clear quads
s_ClearProgram = D3D::CompileAndCreatePixelShader(clear_program_code);
CHECK(s_ClearProgram != nullptr, "Create clear pixel shader");
D3D::SetDebugObjectName((ID3D11DeviceChild*)s_ClearProgram, "clear pixel shader");
// used for anaglyph stereoscopy
s_AnaglyphProgram = D3D::CompileAndCreatePixelShader(anaglyph_program_code);
CHECK(s_AnaglyphProgram != nullptr, "Create anaglyph pixel shader");
D3D::SetDebugObjectName((ID3D11DeviceChild*)s_AnaglyphProgram, "anaglyph pixel shader");
// used when copying/resolving the color buffer
s_ColorCopyProgram[0] = D3D::CompileAndCreatePixelShader(color_copy_program_code);
CHECK(s_ColorCopyProgram[0] != nullptr, "Create color copy pixel shader");
D3D::SetDebugObjectName((ID3D11DeviceChild*)s_ColorCopyProgram[0], "color copy pixel shader");
// used for color conversion
s_ColorMatrixProgram[0] = D3D::CompileAndCreatePixelShader(color_matrix_program_code);
CHECK(s_ColorMatrixProgram[0] != nullptr, "Create color matrix pixel shader");
D3D::SetDebugObjectName((ID3D11DeviceChild*)s_ColorMatrixProgram[0], "color matrix pixel shader");
// used for depth copy
s_DepthMatrixProgram[0] = D3D::CompileAndCreatePixelShader(depth_matrix_program);
CHECK(s_DepthMatrixProgram[0] != nullptr, "Create depth matrix pixel shader");
D3D::SetDebugObjectName((ID3D11DeviceChild*)s_DepthMatrixProgram[0], "depth matrix pixel shader");
Clear();
if (!File::Exists(File::GetUserPath(D_SHADERCACHE_IDX)))
File::CreateDir(File::GetUserPath(D_SHADERCACHE_IDX));
SETSTAT(stats.numPixelShadersCreated, 0);
SETSTAT(stats.numPixelShadersAlive, 0);
std::string cache_filename =
StringFromFormat("%sdx11-%s-ps.cache", File::GetUserPath(D_SHADERCACHE_IDX).c_str(),
SConfig::GetInstance().m_strGameID.c_str());
PixelShaderCacheInserter inserter;
g_ps_disk_cache.OpenAndRead(cache_filename, inserter);
last_entry = nullptr;
}
static ssize_t
_tmf_eovseen(
struct tmfio *xfinfo,
size_t nbytes,
char *dataptr,
ssize_t ret,
struct ffsw *stat)
{
xfinfo->tmf_eovhit = 1;
if (xfinfo->eovbuf == NULL) {
xfinfo->eovbuf = malloc(xfinfo->tmf_bufsiz);
if (xfinfo->eovbuf == NULL) {
ERETURN(stat,FDC_ERR_NOMEM, 0);
}
xfinfo->eovbase = xfinfo->eovbuf;
}
if (ret < 0) {
(void) memcpy(xfinfo->eovbuf, dataptr, nbytes);
xfinfo->eovbytes = nbytes;
xfinfo->tmf_partblk = 0;
}
else {
(void) memcpy(xfinfo->eovbuf, dataptr + ret, nbytes - ret);
xfinfo->eovbytes = nbytes - ret;
xfinfo->tmf_partblk = 1;
}
xfinfo->tmf_cnt = 0;
xfinfo->tmf_bufptr = xfinfo->tmf_base;
SETSTAT(stat, FFEOR, nbytes);
return(nbytes);
}
bool PixelShaderCache::InsertByteCode(const PixelShaderUid &uid, const void* bytecode, unsigned int bytecodelen)
{
ID3D11PixelShader* shader = D3D::CreatePixelShaderFromByteCode(bytecode, bytecodelen);
if (shader == nullptr)
return false;
// TODO: Somehow make the debug name a bit more specific
D3D::SetDebugObjectName((ID3D11DeviceChild*)shader, "a pixel shader of PixelShaderCache");
// Make an entry in the table
PSCacheEntry newentry;
newentry.shader = shader;
PixelShaders[uid] = newentry;
last_entry = &PixelShaders[uid];
if (!shader)
{
// INCSTAT(stats.numPixelShadersFailed);
return false;
}
INCSTAT(stats.numPixelShadersCreated);
SETSTAT(stats.numPixelShadersAlive, PixelShaders.size());
return true;
}
FRAGMENTSHADER* PixelShaderCache::SetShader(DSTALPHA_MODE dstAlphaMode, u32 components)
{
PIXELSHADERUID uid;
GetPixelShaderId(&uid, dstAlphaMode, components);
// Check if the shader is already set
if (last_entry)
{
if (uid == last_uid)
{
GFX_DEBUGGER_PAUSE_AT(NEXT_PIXEL_SHADER_CHANGE, true);
ValidatePixelShaderIDs(API_OPENGL, last_entry->safe_uid, last_entry->shader.strprog, dstAlphaMode, components);
return &last_entry->shader;
}
}
last_uid = uid;
PSCache::iterator iter = PixelShaders.find(uid);
if (iter != PixelShaders.end())
{
PSCacheEntry &entry = iter->second;
last_entry = &entry;
GFX_DEBUGGER_PAUSE_AT(NEXT_PIXEL_SHADER_CHANGE, true);
ValidatePixelShaderIDs(API_OPENGL, entry.safe_uid, entry.shader.strprog, dstAlphaMode, components);
return &last_entry->shader;
}
// Make an entry in the table
PSCacheEntry& newentry = PixelShaders[uid];
last_entry = &newentry;
const char *code = GeneratePixelShaderCode(dstAlphaMode, API_OPENGL, components);
if (g_ActiveConfig.bEnableShaderDebugging && code)
{
GetSafePixelShaderId(&newentry.safe_uid, dstAlphaMode, components);
newentry.shader.strprog = code;
}
#if defined(_DEBUG) || defined(DEBUGFAST)
if (g_ActiveConfig.iLog & CONF_SAVESHADERS && code) {
static int counter = 0;
char szTemp[MAX_PATH];
sprintf(szTemp, "%sps_%04i.txt", File::GetUserPath(D_DUMP_IDX).c_str(), counter++);
SaveData(szTemp, code);
}
#endif
if (!code || !CompilePixelShader(newentry.shader, code)) {
GFX_DEBUGGER_PAUSE_AT(NEXT_ERROR, true);
return NULL;
}
INCSTAT(stats.numPixelShadersCreated);
SETSTAT(stats.numPixelShadersAlive, PixelShaders.size());
GFX_DEBUGGER_PAUSE_AT(NEXT_PIXEL_SHADER_CHANGE, true);
return &last_entry->shader;
}
void ShaderCache::HandlePSUIDChange(PixelShaderUid ps_uid, DSTALPHA_MODE ps_dst_alpha_mode)
{
s_last_pixel_shader_uid = ps_uid;
auto ps_iterator = s_ps_bytecode_cache.find(ps_uid);
if (ps_iterator != s_ps_bytecode_cache.end())
{
s_last_pixel_shader_bytecode = ps_iterator->second;
GFX_DEBUGGER_PAUSE_AT(NEXT_PIXEL_SHADER_CHANGE, true);
}
else
{
ShaderCode ps_code = GeneratePixelShaderCode(APIType::D3D, ps_uid.GetUidData());
ID3DBlob* ps_bytecode = nullptr;
if (!D3D::CompilePixelShader(ps_code.GetBuffer(), &ps_bytecode))
{
GFX_DEBUGGER_PAUSE_AT(NEXT_ERROR, true);
return;
}
s_last_pixel_shader_bytecode = InsertByteCode(ps_uid, &s_ps_bytecode_cache, ps_bytecode);
s_ps_disk_cache.Append(ps_uid, reinterpret_cast<u8*>(ps_bytecode->GetBufferPointer()),
static_cast<u32>(ps_bytecode->GetBufferSize()));
SETSTAT(stats.numPixelShadersAlive, static_cast<int>(s_ps_bytecode_cache.size()));
INCSTAT(stats.numPixelShadersCreated);
}
}
void ShaderCache::HandleVSUIDChange(VertexShaderUid vs_uid)
{
s_last_vertex_shader_uid = vs_uid;
auto vs_iterator = s_vs_bytecode_cache.find(vs_uid);
if (vs_iterator != s_vs_bytecode_cache.end())
{
s_last_vertex_shader_bytecode = vs_iterator->second;
GFX_DEBUGGER_PAUSE_AT(NEXT_VERTEX_SHADER_CHANGE, true);
}
else
{
ShaderCode vs_code = GenerateVertexShaderCode(APIType::D3D, vs_uid.GetUidData());
ID3DBlob* vs_bytecode = nullptr;
if (!D3D::CompileVertexShader(vs_code.GetBuffer(), &vs_bytecode))
{
GFX_DEBUGGER_PAUSE_AT(NEXT_ERROR, true);
return;
}
s_last_vertex_shader_bytecode = InsertByteCode(vs_uid, &s_vs_bytecode_cache, vs_bytecode);
s_vs_disk_cache.Append(vs_uid, reinterpret_cast<u8*>(vs_bytecode->GetBufferPointer()),
static_cast<u32>(vs_bytecode->GetBufferSize()));
SETSTAT(stats.numVertexShadersAlive, static_cast<int>(s_vs_bytecode_cache.size()));
INCSTAT(stats.numVertexShadersCreated);
}
}
/*
* Write an EOF to a buffer-layer file.
*
*/
int
_sqb_weof(struct fdinfo *fio, struct ffsw *stat)
{
struct fdinfo *llfio;
if (fio->rwflag == WRITIN) {
if (_sqb_flush(fio, stat) < 0) {
return(ERR);
}
}
else if (fio->rwflag == READIN || fio->rwflag == POSITIN) {
/* synchronize physical position with logical */
if (_sqb_sync(fio, stat, 1) < 0) {
return(ERR);
}
}
fio->rwflag = WRITIN;
/*
* Tell the underlying layer to write an EOF
*/
llfio = fio->fioptr;
if (XRCALL(llfio,weofrtn) llfio, stat) == ERR)
return(ERR);
SETSTAT(stat, FFEOF, 0);
return(0);
}
void PixelShaderCache::Init()
{
unsigned int cbsize = ((sizeof(PixelShaderConstants))&(~0xf))+0x10; // must be a multiple of 16
D3D11_BUFFER_DESC cbdesc = CD3D11_BUFFER_DESC(cbsize, D3D11_BIND_CONSTANT_BUFFER, D3D11_USAGE_DYNAMIC, D3D11_CPU_ACCESS_WRITE);
D3D::device->CreateBuffer(&cbdesc, NULL, &pscbuf);
CHECK(pscbuf!=NULL, "Create pixel shader constant buffer");
D3D::SetDebugObjectName((ID3D11DeviceChild*)pscbuf, "pixel shader constant buffer used to emulate the GX pipeline");
// used when drawing clear quads
s_ClearProgram = D3D::CompileAndCreatePixelShader(clear_program_code, sizeof(clear_program_code));
CHECK(s_ClearProgram!=NULL, "Create clear pixel shader");
D3D::SetDebugObjectName((ID3D11DeviceChild*)s_ClearProgram, "clear pixel shader");
// used when copying/resolving the color buffer
s_ColorCopyProgram[0] = D3D::CompileAndCreatePixelShader(color_copy_program_code, sizeof(color_copy_program_code));
CHECK(s_ColorCopyProgram[0]!=NULL, "Create color copy pixel shader");
D3D::SetDebugObjectName((ID3D11DeviceChild*)s_ColorCopyProgram[0], "color copy pixel shader");
// used for color conversion
s_ColorMatrixProgram[0] = D3D::CompileAndCreatePixelShader(color_matrix_program_code, sizeof(color_matrix_program_code));
CHECK(s_ColorMatrixProgram[0]!=NULL, "Create color matrix pixel shader");
D3D::SetDebugObjectName((ID3D11DeviceChild*)s_ColorMatrixProgram[0], "color matrix pixel shader");
// used for depth copy
s_DepthMatrixProgram[0] = D3D::CompileAndCreatePixelShader(depth_matrix_program, sizeof(depth_matrix_program));
CHECK(s_DepthMatrixProgram[0]!=NULL, "Create depth matrix pixel shader");
D3D::SetDebugObjectName((ID3D11DeviceChild*)s_DepthMatrixProgram[0], "depth matrix pixel shader");
Clear();
if (!File::Exists(File::GetUserPath(D_SHADERCACHE_IDX)))
File::CreateDir(File::GetUserPath(D_SHADERCACHE_IDX).c_str());
SETSTAT(stats.numPixelShadersCreated, 0);
SETSTAT(stats.numPixelShadersAlive, 0);
char cache_filename[MAX_PATH];
sprintf(cache_filename, "%sdx11-%s-ps.cache", File::GetUserPath(D_SHADERCACHE_IDX).c_str(),
SConfig::GetInstance().m_LocalCoreStartupParameter.m_strUniqueID.c_str());
PixelShaderCacheInserter inserter;
g_ps_disk_cache.OpenAndRead(cache_filename, inserter);
if (g_Config.bEnableShaderDebugging)
Clear();
last_entry = NULL;
}
int
_er90b_write(struct fdinfo *fio, bitptr bufptr, int nbytes,
struct ffsw *retstat, int fulp, int *ubc)
{
int ret;
int nbt = 0; /* number of bytes transferred so far */
int nbreq; /* number of bytes requested this request */
char *buf;
ER90BYT *f;
int zero = 0;
struct ffsw dumstat;
buf = BPTR2CP(bufptr);
if ((BPBITOFF(bufptr) & 7) != 0 || *ubc != 0)
ERETURN(retstat, FDC_ERR_UBC, 0);
nbreq = nbytes;
if (fio->rwflag == POSITIN) {
f = (ER90BYT *)fio->lyr_info;
if (f->tpos) {
ret = _tape_tpwait(f->fd, &(f->tpos));
if (ret < 0)
ERETURN(retstat, errno, 0);
}
}
else if (fio->rwflag == READIN) {
/* write after read requires position to zero */
ret = _er90b_pos(fio, FP_RSEEK, &zero, 1,
&dumstat);
if (ret < 0) {
*retstat = dumstat;
return(ERR);
}
}
if (nbreq > 0) {
again:
/* It is not safe to reissue the write if it fails */
/* with EINTR. Some data may have been transferred */
ret= write(fio->realfd, buf, nbreq);
if (ret < 0)
ERETURN(retstat, errno, nbt);
nbt += ret;
/*
* The assumption is made here that the system will never return
* zero bytes on a non-zero request without an error!
*/
if (nbt < nbytes) {
buf += ret;
nbreq -= ret;
goto again;
}
}
else if (nbytes < 0)
ERETURN(retstat, FDC_ERR_REQ, 0);
SETSTAT(retstat, FFCNT, nbt);
fio->rwflag = WRITIN;
return (nbt);
}
void Init()
{
MarkAllDirty();
for (auto& map_entry : g_main_cp_state.vertex_loaders)
map_entry = nullptr;
for (auto& map_entry : g_preprocess_cp_state.vertex_loaders)
map_entry = nullptr;
SETSTAT(stats.numVertexLoaders, 0);
}
_ffseek_t
_sys_lseek(struct fdinfo *fio, off_t pos, int whence, struct ffsw *stat)
{
off_t ret;
#ifdef __mips
if ((((struct sys_f *)(fio->lyr_info))->needpos) &&
whence == SEEK_CUR)
ret = lseek(fio->realfd,
pos + ((struct sys_f *)(fio->lyr_info))->curpos,
SEEK_SET);
else
ret = lseek(fio->realfd, pos, whence);
#else
ret = lseek(fio->realfd, pos, whence);
#endif
if (ret < 0)
{
ERETURN(stat, errno, 0);
}
#ifdef __mips
((struct sys_f *)(fio->lyr_info))->curpos = ret;
((struct sys_f *)(fio->lyr_info))->needpos = 0;
#endif
fio->rwflag = POSITIN;
if (whence == SEEK_SET && pos == 0)
{
fio->ateof = 0;
fio->ateod = 0;
SETSTAT(stat, FFBOD, 0);
}
else if (whence == SEEK_END && pos == 0)
{
fio->ateof = 1;
fio->ateod = 1; /* seek to end, set stat */
SETSTAT(stat, FFEOD, 0);
}
else
SETSTAT(stat, FFCNT, 0);
fio->last_recbits = fio->recbits;
fio->recbits = 0;
return (ret);
}
void VertexShaderCache::PushByteCode(const VertexShaderUid &uid, const u8 *bytecode, int bytecodelen, VertexShaderCache::VSCacheEntry* entry)
{
entry->shader = D3D::CreateVertexShaderFromByteCode(bytecode, bytecodelen);
entry->compiled = true;
if (entry->shader)
{
INCSTAT(stats.numVertexShadersCreated);
SETSTAT(stats.numVertexShadersAlive, (int)vshaders.size());
}
}
void PixelShaderCache::PushByteCode(const PixelShaderUid &uid, const u8 *bytecode, int bytecodelen, PixelShaderCache::PSCacheEntry* entry)
{
entry->shader = D3D::CreatePixelShaderFromByteCode(bytecode, bytecodelen);
entry->compiled = true;
if (entry->shader)
{
INCSTAT(stats.numPixelShadersCreated);
SETSTAT(stats.numPixelShadersAlive, PixelShaders.size());
}
}
void ProgramShaderCache::Init(void)
{
// We have to get the UBO alignment here because
// if we generate a buffer that isn't aligned
// then the UBO will fail.
if (g_ActiveConfig.backend_info.bSupportsGLSLUBO)
{
GLint Align;
glGetIntegerv(GL_UNIFORM_BUFFER_OFFSET_ALIGNMENT, &Align);
s_ps_data_size = C_PENVCONST_END * sizeof(float) * 4;
s_vs_data_size = C_VENVCONST_END * sizeof(float) * 4;
s_vs_data_offset = ROUND_UP(s_ps_data_size, Align);
s_ubo_buffer_size = ROUND_UP(s_ps_data_size, Align) + ROUND_UP(s_vs_data_size, Align);
// We multiply by *4*4 because we need to get down to basic machine units.
// So multiply by four to get how many floats we have from vec4s
// Then once more to get bytes
s_buffer = new StreamBuffer(GL_UNIFORM_BUFFER, UBO_LENGTH);
s_ubo_buffer = new u8[s_ubo_buffer_size];
memset(s_ubo_buffer, 0, s_ubo_buffer_size);
s_ubo_dirty = true;
}
// Read our shader cache, only if supported
if (g_ogl_config.bSupportsGLSLCache)
{
GLint Supported;
glGetIntegerv(GL_NUM_PROGRAM_BINARY_FORMATS, &Supported);
if(!Supported)
{
ERROR_LOG(VIDEO, "GL_ARB_get_program_binary is supported, but no binary format is known. So disable shader cache.");
g_ogl_config.bSupportsGLSLCache = false;
}
else
{
if (!File::Exists(File::GetUserPath(D_SHADERCACHE_IDX)))
File::CreateDir(File::GetUserPath(D_SHADERCACHE_IDX).c_str());
char cache_filename[MAX_PATH];
sprintf(cache_filename, "%sogl-%s-shaders.cache", File::GetUserPath(D_SHADERCACHE_IDX).c_str(),
SConfig::GetInstance().m_LocalCoreStartupParameter.m_strUniqueID.c_str());
ProgramShaderCacheInserter inserter;
g_program_disk_cache.OpenAndRead(cache_filename, inserter);
}
SETSTAT(stats.numPixelShadersAlive, pshaders.size());
}
CreateHeader();
CurrentProgram = 0;
last_entry = NULL;
}
void ShaderCache::Init()
{
// This class intentionally shares its shader cache files with DX11, as the shaders are (right
// now) identical.
// Reduces unnecessary compilation when switching between APIs.
s_last_geometry_shader_bytecode = {};
s_last_pixel_shader_bytecode = {};
s_last_vertex_shader_bytecode = {};
s_last_geometry_shader_uid = {};
s_last_pixel_shader_uid = {};
s_last_vertex_shader_uid = {};
// Ensure shader cache directory exists..
std::string shader_cache_path = File::GetUserPath(D_SHADERCACHE_IDX);
if (!File::Exists(shader_cache_path))
File::CreateDir(File::GetUserPath(D_SHADERCACHE_IDX));
std::string title_game_id = SConfig::GetInstance().m_strGameID.c_str();
std::string gs_cache_filename =
StringFromFormat("%sdx11-%s-gs.cache", shader_cache_path.c_str(), title_game_id.c_str());
std::string ps_cache_filename =
StringFromFormat("%sdx11-%s-ps.cache", shader_cache_path.c_str(), title_game_id.c_str());
std::string vs_cache_filename =
StringFromFormat("%sdx11-%s-vs.cache", shader_cache_path.c_str(), title_game_id.c_str());
ShaderCacheInserter<GeometryShaderUid, GsBytecodeCache, &s_gs_bytecode_cache> gs_inserter;
s_gs_disk_cache.OpenAndRead(gs_cache_filename, gs_inserter);
ShaderCacheInserter<PixelShaderUid, PsBytecodeCache, &s_ps_bytecode_cache> ps_inserter;
s_ps_disk_cache.OpenAndRead(ps_cache_filename, ps_inserter);
ShaderCacheInserter<VertexShaderUid, VsBytecodeCache, &s_vs_bytecode_cache> vs_inserter;
s_vs_disk_cache.OpenAndRead(vs_cache_filename, vs_inserter);
SETSTAT(stats.numPixelShadersAlive, static_cast<int>(s_ps_bytecode_cache.size()));
SETSTAT(stats.numPixelShadersCreated, static_cast<int>(s_ps_bytecode_cache.size()));
SETSTAT(stats.numVertexShadersAlive, static_cast<int>(s_vs_bytecode_cache.size()));
SETSTAT(stats.numVertexShadersCreated, static_cast<int>(s_vs_bytecode_cache.size()));
}
void GeometryShaderCache::PushByteCode(const void* bytecode, unsigned int bytecodelen, GeometryShaderCache::GSCacheEntry* entry)
{
entry->shader = std::move(D3D::CreateGeometryShaderFromByteCode(bytecode, bytecodelen));
entry->compiled = true;
if (entry->shader != nullptr)
{
// TODO: Somehow make the debug name a bit more specific
D3D::SetDebugObjectName(entry->shader.get(), "a Geometry shader of GeometryShaderCache");
INCSTAT(stats.numGeometryShadersCreated);
SETSTAT(stats.numGeometryShadersAlive, static_cast<int>(s_geometry_shaders->size()));
}
}
void VertexShaderCache::PushByteCode(D3DBlob&& bcodeblob, VSCacheEntry* entry)
{
entry->shader = std::move(D3D::CreateVertexShaderFromByteCode(bcodeblob));
entry->compiled = true;
entry->SetByteCode(std::move(bcodeblob));
if (entry->shader)
{
// TODO: Somehow make the debug name a bit more specific
D3D::SetDebugObjectName(entry->shader.get(), "a vertex shader of VertexShaderCache");
INCSTAT(stats.numVertexShadersCreated);
SETSTAT(stats.numVertexShadersAlive, static_cast<int>(s_vshaders->size()));
}
}
VERTEXSHADER* VertexShaderCache::SetShader(u32 components)
{
VERTEXSHADERUID uid;
GetVertexShaderId(&uid, components);
if (uid == last_vertex_shader_uid && vshaders[uid].frameCount == frameCount)
{
GFX_DEBUGGER_PAUSE_AT(NEXT_VERTEX_SHADER_CHANGE, true);
return pShaderLast;
}
memcpy(&last_vertex_shader_uid, &uid, sizeof(VERTEXSHADERUID));
VSCache::iterator iter = vshaders.find(uid);
if (iter != vshaders.end())
{
iter->second.frameCount = frameCount;
VSCacheEntry &entry = iter->second;
if (&entry.shader != pShaderLast) {
pShaderLast = &entry.shader;
}
GFX_DEBUGGER_PAUSE_AT(NEXT_VERTEX_SHADER_CHANGE, true);
return pShaderLast;
}
// Make an entry in the table
VSCacheEntry& entry = vshaders[uid];
entry.frameCount = frameCount;
pShaderLast = &entry.shader;
const char *code = GenerateVertexShaderCode(components, API_OPENGL);
#if defined(_DEBUG) || defined(DEBUGFAST)
if (g_ActiveConfig.iLog & CONF_SAVESHADERS && code) {
static int counter = 0;
char szTemp[MAX_PATH];
sprintf(szTemp, "%svs_%04i.txt", File::GetUserPath(D_DUMP_IDX).c_str(), counter++);
SaveData(szTemp, code);
}
#endif
if (!code || !VertexShaderCache::CompileVertexShader(entry.shader, code)) {
ERROR_LOG(VIDEO, "failed to create vertex shader");
GFX_DEBUGGER_PAUSE_AT(NEXT_ERROR, true);
return NULL;
}
INCSTAT(stats.numVertexShadersCreated);
SETSTAT(stats.numVertexShadersAlive, vshaders.size());
GFX_DEBUGGER_PAUSE_AT(NEXT_VERTEX_SHADER_CHANGE, true);
return pShaderLast;
}
void ProgramShaderCache::Init()
{
// We have to get the UBO alignment here because
// if we generate a buffer that isn't aligned
// then the UBO will fail.
glGetIntegerv(GL_UNIFORM_BUFFER_OFFSET_ALIGNMENT, &s_ubo_align);
s_ubo_buffer_size = ROUND_UP(PixelShaderManager::ConstantBufferSize * sizeof(float), s_ubo_align)
+ ROUND_UP(VertexShaderManager::ConstantBufferSize * sizeof(float), s_ubo_align)
+ ROUND_UP(sizeof(GeometryShaderConstants), s_ubo_align);
// We multiply by *4*4 because we need to get down to basic machine units.
// So multiply by four to get how many floats we have from vec4s
// Then once more to get bytes
s_buffer = StreamBuffer::Create(GL_UNIFORM_BUFFER, UBO_LENGTH);
// Read our shader cache, only if supported
if (g_ogl_config.bSupportsGLSLCache && !g_Config.bEnableShaderDebugging)
{
GLint Supported;
glGetIntegerv(GL_NUM_PROGRAM_BINARY_FORMATS, &Supported);
if (!Supported)
{
ERROR_LOG(VIDEO, "GL_ARB_get_program_binary is supported, but no binary format is known. So disable shader cache.");
g_ogl_config.bSupportsGLSLCache = false;
}
else
{
if (!File::Exists(File::GetUserPath(D_SHADERCACHE_IDX)))
File::CreateDir(File::GetUserPath(D_SHADERCACHE_IDX));
std::string cache_filename = StringFromFormat("%sIOGL-%s-shaders.cache", File::GetUserPath(D_SHADERCACHE_IDX).c_str(),
SConfig::GetInstance().m_strUniqueID.c_str());
ProgramShaderCacheInserter inserter;
g_program_disk_cache.OpenAndRead(cache_filename, inserter);
}
SETSTAT(stats.numPixelShadersAlive, pshaders.size());
}
CreateHeader();
CurrentProgram = 0;
last_entry = nullptr;
}
_er90b_lseek(struct fdinfo *fio, int pos, int whence, struct ffsw *stat)
{
int ret;
if (whence == 0 && pos == 0) {
ret = _er90b_rewd((ER90BYT *)fio->lyr_info);
if (ret < 0)
ERETURN(stat, errno, 0);
fio->ateof = 0;
fio->ateod = 0;
SETSTAT(stat, FFBOD, 0);
}
else {
ERETURN(stat, FDC_ERR_NOSUP, 0);
}
fio->rwflag = POSITIN;
return (0);
}
bool VertexShaderCache::InsertShader(const VertexShaderUid& uid, ID3D11VertexShader* shader,
D3DBlob* blob)
{
auto iter = vshaders.find(uid);
if (iter != vshaders.end() && !iter->second.pending)
return false;
VSCacheEntry& newentry = vshaders[uid];
newentry.pending = false;
if (!shader || !blob)
return false;
shader->AddRef();
newentry.SetByteCode(blob);
newentry.shader = shader;
INCSTAT(stats.numPixelShadersCreated);
SETSTAT(stats.numPixelShadersAlive, static_cast<int>(vshaders.size()));
return true;
}
/*
* Write an EOD.
* Truncate the file. This is the 'real' end of the data.
* We cannot do this if we were reading or if we just positioned.
*/
int
_tmf_weod(struct fdinfo *fio, struct ffsw *stat)
{
register int ret;
struct tmfio *xfinfo;
xfinfo = (struct tmfio *)fio->lyr_info;
if (xfinfo->rwflag != WRITIN) {
ERETURN(stat, FENOENDF, 0);
}
ret = XRCALL(fio, flushrtn) fio, stat);
if (ret < 0) return(ret);
SETSTAT(stat, FFEOD, 0);
return(0);
}
bool VertexShaderCache::InsertByteCode(const VertexShaderUid &uid, const u8 *bytecode, int bytecodelen, bool activate) {
LPDIRECT3DVERTEXSHADER9 shader = D3D::CreateVertexShaderFromByteCode(bytecode, bytecodelen);
// Make an entry in the table
VSCacheEntry entry;
entry.shader = shader;
vshaders[uid] = entry;
last_entry = &vshaders[uid];
if (!shader)
return false;
INCSTAT(stats.numVertexShadersCreated);
SETSTAT(stats.numVertexShadersAlive, (int)vshaders.size());
if (activate)
{
D3D::SetVertexShader(shader);
return true;
}
return false;
}
bool VertexShaderCache::InsertByteCode(const VertexShaderUid& uid, D3DBlob* bcodeblob)
{
ID3D11VertexShader* shader = D3D::CreateVertexShaderFromByteCode(bcodeblob);
if (shader == nullptr)
return false;
// TODO: Somehow make the debug name a bit more specific
D3D::SetDebugObjectName((ID3D11DeviceChild*)shader, "a vertex shader of VertexShaderCache");
// Make an entry in the table
VSCacheEntry entry;
entry.shader = shader;
entry.SetByteCode(bcodeblob);
vshaders[uid] = entry;
last_entry = &vshaders[uid];
INCSTAT(stats.numVertexShadersCreated);
SETSTAT(stats.numVertexShadersAlive, (int)vshaders.size());
return true;
}
bool PixelShaderCache::InsertByteCode(const PIXELSHADERUID &uid, const u8 *bytecode, int bytecodelen, bool activate)
{
LPDIRECT3DPIXELSHADER9 shader = D3D::CreatePixelShaderFromByteCode(bytecode, bytecodelen);
// Make an entry in the table
PSCacheEntry newentry;
newentry.shader = shader;
PixelShaders[uid] = newentry;
last_entry = &PixelShaders[uid];
if (!shader) {
// INCSTAT(stats.numPixelShadersFailed);
return false;
}
INCSTAT(stats.numPixelShadersCreated);
SETSTAT(stats.numPixelShadersAlive, PixelShaders.size());
if (activate)
{
D3D::SetPixelShader(shader);
}
return true;
}
