bool VertexShaderCache::CompileVertexShader(VERTEXSHADER& vs, const char* pstrprogram)
{
// Reset GL error before compiling shaders. Yeah, we need to investigate the causes of these.
GLenum err = GL_REPORT_ERROR();
if (err != GL_NO_ERROR)
{
ERROR_LOG(VIDEO, "glError %08x before VS!", err);
}
#if defined HAVE_CG && HAVE_CG
char stropt[64];
sprintf(stropt, "MaxLocalParams=256,MaxInstructions=%d", s_nMaxVertexInstructions);
const char *opts[] = {"-profileopts", stropt, "-O2", "-q", NULL};
CGprogram tempprog = cgCreateProgram(g_cgcontext, CG_SOURCE, pstrprogram, g_cgvProf, "main", opts);
if (!cgIsProgram(tempprog)) {
if (s_displayCompileAlert) {
PanicAlert("Failed to create vertex shader");
s_displayCompileAlert = false;
}
cgDestroyProgram(tempprog);
ERROR_LOG(VIDEO, "Failed to load vs %s:", cgGetLastListing(g_cgcontext));
ERROR_LOG(VIDEO, "%s", pstrprogram);
return false;
}
if (cgGetError() != CG_NO_ERROR)
{
WARN_LOG(VIDEO, "Failed to load vs %s:", cgGetLastListing(g_cgcontext));
WARN_LOG(VIDEO, "%s", pstrprogram);
}
// This looks evil - we modify the program through the const char * we got from cgGetProgramString!
// It SHOULD not have any nasty side effects though - but you never know...
char *pcompiledprog = (char*)cgGetProgramString(tempprog, CG_COMPILED_PROGRAM);
char *plocal = strstr(pcompiledprog, "program.local");
while (plocal != NULL) {
const char* penv = " program.env";
memcpy(plocal, penv, 13);
plocal = strstr(plocal + 13, "program.local");
}
glGenProgramsARB(1, &vs.glprogid);
SetCurrentShader(vs.glprogid);
glProgramStringARB(GL_VERTEX_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, (GLsizei)strlen(pcompiledprog), pcompiledprog);
err = GL_REPORT_ERROR();
if (err != GL_NO_ERROR) {
ERROR_LOG(VIDEO, "%s", pstrprogram);
ERROR_LOG(VIDEO, "%s", pcompiledprog);
}
cgDestroyProgram(tempprog);
#endif
#if defined(_DEBUG) || defined(DEBUGFAST)
vs.strprog = pstrprogram;
#endif
return true;
}
void VertexManager::DestroyDeviceObjects()
{
GL_REPORT_ERRORD();
glBindBuffer(GL_ARRAY_BUFFER, 0 );
glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, 0 );
GL_REPORT_ERROR();
delete s_vertexBuffer;
delete s_indexBuffer;
GL_REPORT_ERROR();
}
GLuint ProgramShaderCache::CompileSingleShader(GLuint type, const char* code)
{
GLuint result = glCreateShader(type);
const char *src[] = {s_glsl_header, code};
glShaderSource(result, 2, src, nullptr);
glCompileShader(result);
GLint compileStatus;
glGetShaderiv(result, GL_COMPILE_STATUS, &compileStatus);
GLsizei length = 0;
glGetShaderiv(result, GL_INFO_LOG_LENGTH, &length);
if (DriverDetails::HasBug(DriverDetails::BUG_BROKENINFOLOG))
length = 1024;
if (compileStatus != GL_TRUE || (length > 1 && DEBUG_GLSL))
{
GLsizei charsWritten;
GLchar* infoLog = new GLchar[length];
glGetShaderInfoLog(result, length, &charsWritten, infoLog);
ERROR_LOG(VIDEO, "%s Shader info log:\n%s", type==GL_VERTEX_SHADER ? "VS" : "PS", infoLog);
std::string filename = StringFromFormat("%sbad_%s_%04i.txt",
File::GetUserPath(D_DUMP_IDX).c_str(),
type==GL_VERTEX_SHADER ? "vs" : "ps",
num_failures++);
std::ofstream file;
OpenFStream(file, filename, std::ios_base::out);
file << s_glsl_header << code << infoLog;
file.close();
if (compileStatus != GL_TRUE)
{
PanicAlert("Failed to compile %s shader!\nThis usually happens when trying to use Dolphin with an outdated GPU or integrated GPU like the Intel GMA series.\n\nIf you're sure this is Dolphin's error anyway, post the contents of %s along with this error message at the forums.\n\nDebug info (%s, %s, %s):\n%s",
type == GL_VERTEX_SHADER ? "vertex" : "pixel",
filename.c_str(),
g_ogl_config.gl_vendor,
g_ogl_config.gl_renderer,
g_ogl_config.gl_version,
infoLog);
}
delete[] infoLog;
}
if (compileStatus != GL_TRUE)
{
// Compile failed
ERROR_LOG(VIDEO, "Shader compilation failed; see info log");
// Don't try to use this shader
glDeleteShader(result);
return 0;
}
(void)GL_REPORT_ERROR();
return result;
}
void PixelShaderCache::Init()
{
glEnable(GL_FRAGMENT_PROGRAM_ARB);
ShaderEnabled = true;
CurrentShader = 0;
last_entry = NULL;
GL_REPORT_ERRORD();
s_displayCompileAlert = true;
glGetProgramivARB(GL_FRAGMENT_PROGRAM_ARB, GL_MAX_PROGRAM_NATIVE_ALU_INSTRUCTIONS_ARB, (GLint *)&s_nMaxPixelInstructions);
if(s_nMaxPixelInstructions == 0) // Some combination of drivers and hardware returns zero for some reason.
s_nMaxPixelInstructions = 4096;
if (strstr((const char*)glGetString(GL_VENDOR), "Humper") != NULL) s_nMaxPixelInstructions = 4096;
#if CG_VERSION_NUM == 2100
if (strstr((const char*)glGetString(GL_VENDOR), "ATI") != NULL)
{
s_nMaxPixelInstructions = 4096;
}
#endif
int maxinst, maxattribs;
glGetProgramivARB(GL_FRAGMENT_PROGRAM_ARB, GL_MAX_PROGRAM_NATIVE_INSTRUCTIONS_ARB, (GLint *)&maxinst);
glGetProgramivARB(GL_FRAGMENT_PROGRAM_ARB, GL_MAX_PROGRAM_NATIVE_ATTRIBS_ARB, (GLint *)&maxattribs);
INFO_LOG(VIDEO, "pixel max_alu=%d, max_inst=%d, max_attrib=%d", s_nMaxPixelInstructions, maxinst, maxattribs);
char pmatrixprog[2048];
sprintf(pmatrixprog, "!!ARBfp1.0"
"TEMP R0;\n"
"TEMP R1;\n"
"PARAM K0 = { 0.5, 0.5, 0.5, 0.5};\n"
"TEX R0, fragment.texcoord[0], texture[0], RECT;\n"
"MUL R0, R0, program.env[%d];\n"
"ADD R0, R0, K0;\n"
"FLR R0, R0;\n"
"MUL R0, R0, program.env[%d];\n"
"DP4 R1.x, R0, program.env[%d];\n"
"DP4 R1.y, R0, program.env[%d];\n"
"DP4 R1.z, R0, program.env[%d];\n"
"DP4 R1.w, R0, program.env[%d];\n"
"ADD result.color, R1, program.env[%d];\n"
"END\n",C_COLORMATRIX+5,C_COLORMATRIX+6, C_COLORMATRIX, C_COLORMATRIX+1, C_COLORMATRIX+2, C_COLORMATRIX+3, C_COLORMATRIX+4);
glGenProgramsARB(1, &s_ColorMatrixProgram);
SetCurrentShader(s_ColorMatrixProgram);
glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, (GLsizei)strlen(pmatrixprog), pmatrixprog);
GLenum err = GL_REPORT_ERROR();
if (err != GL_NO_ERROR) {
ERROR_LOG(VIDEO, "Failed to create color matrix fragment program");
glDeleteProgramsARB(1, &s_ColorMatrixProgram);
s_ColorMatrixProgram = 0;
}
sprintf(pmatrixprog, "!!ARBfp1.0\n"
"TEMP R0;\n"
"TEMP R1;\n"
"TEMP R2;\n"
//16777215/16777216*256, 1/255, 256, 0
"PARAM K0 = { 255.99998474121, 0.003921568627451, 256.0, 0.0};\n"
"PARAM K1 = { 15.0, 0.066666666666, 0.0, 0.0};\n"
//sample the depth value
"TEX R2, fragment.texcoord[0], texture[0], RECT;\n"
//scale from [0*16777216..1*16777216] to
//[0*16777215..1*16777215], multiply by 256
"MUL R0, R2.x, K0.x;\n" // *16777215/16777216*256
//It is easy to get bad results due to low precision
//here, for example converting like this:
//MUL R0,R0,{ 65536, 256, 1, 16777216 }
//FRC R0,R0
//gives {?, 128/255, 254/255, ?} for depth value 254/255
//on some gpus
"FLR R0.x,R0;\n" //bits 31..24
"SUB R0.yzw,R0,R0.x;\n" //subtract bits 31..24 from rest
"MUL R0.yzw,R0,K0.z;\n" // *256
"FLR R0.y,R0;\n" //bits 23..16
"SUB R0.zw,R0,R0.y;\n" //subtract bits 23..16 from rest
"MUL R0.zw,R0,K0.z;\n" // *256
"FLR R0.z,R0;\n" //bits 15..8
"MOV R0.w,R0.x;\n" //duplicate bit 31..24
"MUL R0,R0,K0.y;\n" // /255
"MUL R0.w,R0,K1.x;\n" // *15
"FLR R0.w,R0;\n" //bits 31..28
"MUL R0.w,R0,K1.y;\n" // /15
"DP4 R1.x, R0, program.env[%d];\n"
"DP4 R1.y, R0, program.env[%d];\n"
"DP4 R1.z, R0, program.env[%d];\n"
"DP4 R1.w, R0, program.env[%d];\n"
"ADD result.color, R1, program.env[%d];\n"
"END\n", C_COLORMATRIX, C_COLORMATRIX+1, C_COLORMATRIX+2, C_COLORMATRIX+3, C_COLORMATRIX+4);
glGenProgramsARB(1, &s_DepthMatrixProgram);
SetCurrentShader(s_DepthMatrixProgram);
glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, (GLsizei)strlen(pmatrixprog), pmatrixprog);
err = GL_REPORT_ERROR();
if (err != GL_NO_ERROR) {
ERROR_LOG(VIDEO, "Failed to create depth matrix fragment program");
glDeleteProgramsARB(1, &s_DepthMatrixProgram);
s_DepthMatrixProgram = 0;
}
}
bool PixelShaderCache::CompilePixelShader(FRAGMENTSHADER& ps, const char* pstrprogram)
{
GLenum err = GL_REPORT_ERROR();
if (err != GL_NO_ERROR)
{
ERROR_LOG(VIDEO, "glError %08x before PS!", err);
}
#if defined HAVE_CG && HAVE_CG
char stropt[128];
sprintf(stropt, "MaxLocalParams=224,NumInstructionSlots=%d", s_nMaxPixelInstructions);
const char *opts[] = {"-profileopts", stropt, "-O2", "-q", NULL};
CGprogram tempprog = cgCreateProgram(g_cgcontext, CG_SOURCE, pstrprogram, g_cgfProf, "main", opts);
// handle errors
if (!cgIsProgram(tempprog))
{
cgDestroyProgram(tempprog);
static int num_failures = 0;
char szTemp[MAX_PATH];
sprintf(szTemp, "%sbad_ps_%04i.txt", File::GetUserPath(D_DUMP_IDX).c_str(), num_failures++);
std::ofstream file(szTemp);
file << pstrprogram;
file.close();
PanicAlert("Failed to compile pixel shader!\nThis usually happens when trying to use Dolphin with an outdated GPU or integrated GPU like the Intel GMA series.\n\nIf you're sure this is Dolphin's error anyway, post the contents of %s along with this error message at the forums.\n\nDebug info (%d):\n%s",
szTemp,
g_cgfProf,
cgGetLastListing(g_cgcontext));
return false;
}
// handle warnings
if (cgGetError() != CG_NO_ERROR)
{
WARN_LOG(VIDEO, "Warnings on compile ps %s:", cgGetLastListing(g_cgcontext));
WARN_LOG(VIDEO, "%s", pstrprogram);
}
// This looks evil - we modify the program through the const char * we got from cgGetProgramString!
// It SHOULD not have any nasty side effects though - but you never know...
char *pcompiledprog = (char*)cgGetProgramString(tempprog, CG_COMPILED_PROGRAM);
char *plocal = strstr(pcompiledprog, "program.local");
while (plocal != NULL)
{
const char *penv = " program.env";
memcpy(plocal, penv, 13);
plocal = strstr(plocal+13, "program.local");
}
glGenProgramsARB(1, &ps.glprogid);
SetCurrentShader(ps.glprogid);
glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, (GLsizei)strlen(pcompiledprog), pcompiledprog);
err = GL_REPORT_ERROR();
if (err != GL_NO_ERROR)
{
GLint error_pos, native_limit;
glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &error_pos);
glGetProgramivARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_UNDER_NATIVE_LIMITS_ARB, &native_limit);
// Error occur
if (error_pos != -1) {
const char *program_error = (const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB);
char line[256];
strncpy(line, (const char *)pcompiledprog + error_pos, 255);
line[255] = 0;
ERROR_LOG(VIDEO, "Error at %i: %s", error_pos, program_error);
ERROR_LOG(VIDEO, "Line dump: \n%s", line);
} else if (native_limit != -1) {
ERROR_LOG(VIDEO, "Hit limit? %i", native_limit);
// TODO
}
ERROR_LOG(VIDEO, "%s", pstrprogram);
ERROR_LOG(VIDEO, "%s", pcompiledprog);
}
cgDestroyProgram(tempprog);
#endif
return true;
}