/* Get string of options to pass to the CL compiler. Result must be freed */
char* getCompilerFlags(const CLInfo* ci, const AstronomyParameters* ap, cl_bool useILKernel)
{
const DevInfo* di = &ci->di;
std::string flagStr;
const char* str;
std::stringstream flags(std::stringstream::out);
flags.precision(15);
if (DOUBLEPREC)
{
flags << "-D DOUBLEPREC=1 ";
}
else
{
flags << "-D DOUBLEPREC=0 -cl-single-precision-constant ";
}
/* Math options for CL compiler */
flags << "-cl-mad-enable ";
flags << "-cl-no-signed-zeros ";
flags << "-cl-strict-aliasing ";
flags << "-cl-finite-math-only ";
/* Get constant definitions */
flags << "-D BACKGROUND_PROFILE=" << ap->background_profile << " ";
flags << "-D AUX_BG_PROFILE=" << ap->aux_bg_profile << " ";
flags << "-D NSTREAM=" << ap->number_streams << " ";
flags << "-D CONVOLVE=" << ap->convolve << " ";
flags << "-D R0=" << ap->r0 << " ";
flags << "-D SUN_R0=" << ap->sun_r0 << " ";
flags << "-D Q_INV_SQR=" << ap->q_inv_sqr << " ";
flags << "-D BG_A=" << ap->bg_a << " ";
flags << "-D BG_B=" << ap->bg_b << " ";
flags << "-D BG_C=" << ap->bg_c << " ";
flags << "-D ALPHA=" << ap->alpha << " ";
flags << "-D ALPHA_DELTA_3=" << ap->alpha_delta3 << " ";
/* FIXME: Device vendor not necessarily the platform vendor */
if (mwHasNvidiaCompilerFlags(di))
{
flags << "-cl-nv-verbose ";
flags << getNvidiaRegCount(di);
}
if (useILKernel)
{
/* Options to only emit AMD IL in the binary */
flags << "-fno-bin-exe -fno-bin-llvmir -fno-bin-source -fbin-amdil ";
}
flagStr = flags.str();
str = flagStr.c_str();
return str ? strdup(str) : NULL;
}
/* Get string of options to pass to the CL compiler. */
static char* getCompilerFlags(const AstronomyParameters* ap, const DevInfo* di, cl_bool useImages)
{
char* compileFlags = NULL;
char cwd[1024] = "";
char extraFlags[1024] = "";
char includeFlags[1024] = "";
char precBuf[1024] = "";
char kernelDefBuf[4096] = "";
/* Math options for CL compiler */
const char mathFlags[] = "-cl-mad-enable "
"-cl-no-signed-zeros "
"-cl-strict-aliasing "
"-cl-finite-math-only ";
/* Extra flags for different compilers */
const char nvidiaOptFlags[] = "-cl-nv-verbose ";
const char atiOptFlags[] = "";
const char kernelDefStr[] = "-DFAST_H_PROB=%d "
"-DAUX_BG_PROFILE=%d "
"-DNSTREAM=%u "
"-DCONVOLVE=%u "
"-DR0=%.15f "
"-DSUN_R0=%.15f "
"-DQ_INV_SQR=%.15f "
"-DUSE_IMAGES=%d ";
const char includeStr[] = "-I%s "
"-I%s/../include ";
if (snprintf(kernelDefBuf, sizeof(kernelDefBuf), kernelDefStr,
ap->fast_h_prob,
ap->aux_bg_profile,
ap->number_streams,
ap->convolve,
ap->r0,
ap->sun_r0,
ap->q_inv_sqr,
useImages) < 0)
{
warn("Error getting kernel constant definitions\n");
return NULL;
}
snprintf(precBuf, sizeof(precBuf), "-DDOUBLEPREC=%d %s ",
DOUBLEPREC, DOUBLEPREC ? "" : "-cl-single-precision-constant");
/* FIXME: Device vendor not necessarily the platform vendor */
if (di->vendorID == MW_NVIDIA)
{
if (snprintf(extraFlags, sizeof(extraFlags),
"%s %s ",
nvidiaOptFlags, getNvidiaRegCount(di)) < 0)
{
warn("Error getting extra Nvidia flags\n");
return NULL;
}
}
else if (di->vendorID == MW_AMD_ATI)
{
if (snprintf(extraFlags, sizeof(extraFlags), "%s", atiOptFlags) < 0)
{
warn("Error getting extra ATI flags\n");
return NULL;
}
}
else
warn("Unknown vendor ID: 0x%x\n", di->vendorID);
if (!inlinedIntegralKernelSrc)
{
if (!getcwd(cwd, sizeof(cwd)))
{
warn("Failed to get header directory\n");
return NULL;
}
if (snprintf(includeFlags, sizeof(includeFlags), includeStr, cwd, cwd) < 0)
{
warn("Failed to get include flags\n");
return NULL;
}
}
if (asprintf(&compileFlags, "%s%s%s%s%s ",
includeFlags,
mathFlags,
extraFlags,
precBuf,
kernelDefBuf) < 0)
{
warn("Failed to get compile flags\n");
return NULL;
}
return compileFlags;
}