Merge branch 'master' into nvptx

This commit is contained in:
egaburov
2013-11-08 15:42:58 +01:00
8 changed files with 250 additions and 179 deletions

1
.gitignore vendored
View File

@@ -18,5 +18,6 @@ examples/*/objs/*
examples/*/ref examples/*/ref
examples/*/test examples/*/test
*.swp *.swp
check_isa.exe

138
alloy.py
View File

@@ -65,7 +65,12 @@ def try_do_LLVM(text, command, from_validation):
if from_validation == True: if from_validation == True:
text = text + "\n" text = text + "\n"
print_debug("Trying to " + text, from_validation, alloy_build) print_debug("Trying to " + text, from_validation, alloy_build)
if os.system(command + " >> " + alloy_build + " 2>> " + alloy_build) != 0: postfix = ""
if current_OS == "Windows":
postfix = " 1>> " + alloy_build + " 2>&1"
else:
postfix = " >> " + alloy_build + " 2>> " + alloy_build
if os.system(command + postfix) != 0:
print_debug("ERROR.\n", from_validation, alloy_build) print_debug("ERROR.\n", from_validation, alloy_build)
error("can't " + text, 1) error("can't " + text, 1)
print_debug("DONE.\n", from_validation, alloy_build) print_debug("DONE.\n", from_validation, alloy_build)
@@ -144,10 +149,13 @@ def build_LLVM(version_LLVM, revision, folder, tarball, debug, selfbuild, extra,
"tar -xvzf " + tar[1] + " --strip-components 1", from_validation) "tar -xvzf " + tar[1] + " --strip-components 1", from_validation)
os.chdir("../../") os.chdir("../../")
# paching llvm # paching llvm
patches = glob.glob(os.environ["ISPC_HOME"] + "/llvm_patches/*.*") patches = glob.glob(os.environ["ISPC_HOME"] + os.sep + "llvm_patches" + os.sep + "*.*")
for patch in patches: for patch in patches:
if version_LLVM in os.path.basename(patch): if version_LLVM in os.path.basename(patch):
if current_OS != "Windows":
try_do_LLVM("patch LLVM with patch " + patch + " ", "patch -p0 < " + patch, from_validation) try_do_LLVM("patch LLVM with patch " + patch + " ", "patch -p0 < " + patch, from_validation)
else:
try_do_LLVM("patch LLVM with patch " + patch + " ", "C:\\gnuwin32\\bin\\patch.exe -p0 < " + patch, from_validation)
os.chdir("../") os.chdir("../")
# configuring llvm, build first part of selfbuild # configuring llvm, build first part of selfbuild
os.makedirs(LLVM_BUILD) os.makedirs(LLVM_BUILD)
@@ -173,77 +181,65 @@ def build_LLVM(version_LLVM, revision, folder, tarball, debug, selfbuild, extra,
print_debug("Now we have compiler for selfbuild: " + selfbuild_compiler + "\n", from_validation, alloy_build) print_debug("Now we have compiler for selfbuild: " + selfbuild_compiler + "\n", from_validation, alloy_build)
os.chdir(LLVM_BUILD) os.chdir(LLVM_BUILD)
if debug == False: if debug == False:
if current_OS != "Windows":
try_do_LLVM("configure release version ", try_do_LLVM("configure release version ",
"../" + LLVM_SRC + "/configure --prefix=" + llvm_home + "/" + "../" + LLVM_SRC + "/configure --prefix=" + llvm_home + "/" +
LLVM_BIN + " --enable-optimized" + selfbuild_compiler, LLVM_BIN + " --enable-optimized" + selfbuild_compiler,
from_validation) from_validation)
else:
try_do_LLVM("configure release version ",
'cmake -G "Visual Studio 10" -DCMAKE_INSTALL_PREFIX="..\\'+ LLVM_BIN +
'" -DLLVM_LIT_TOOLS_DIR="C:\\gnuwin32\\bin" ..\\' + LLVM_SRC,
from_validation)
else: else:
try_do_LLVM("configure debug version ", try_do_LLVM("configure debug version ",
"../" + LLVM_SRC + "/configure --prefix=" + llvm_home + "/" + LLVM_BIN + "../" + LLVM_SRC + "/configure --prefix=" + llvm_home + "/" + LLVM_BIN +
" --enable-debug-runtime --enable-debug-symbols --enable-keep-symbols" + selfbuild_compiler, " --enable-debug-runtime --enable-debug-symbols --enable-keep-symbols" + selfbuild_compiler,
from_validation) from_validation)
# building llvm # building llvm
if current_OS != "Windows":
try_do_LLVM("build LLVM ", make, from_validation) try_do_LLVM("build LLVM ", make, from_validation)
try_do_LLVM("install LLVM ", "make install", from_validation) try_do_LLVM("install LLVM ", "make install", from_validation)
else:
try_do_LLVM("build LLVM and than install LLVM ", "msbuild INSTALL.vcxproj /V:m /p:Platform=Win32 /p:Configuration=Release /t:rebuild", from_validation)
os.chdir(current_path) os.chdir(current_path)
def check_targets(): def check_targets():
answer = [] answer = []
answer_sde = [] answer_sde = []
SSE2 = False;
SSE4 = False;
AVX = False;
AVX11 = False;
AVX2 = False;
if current_OS == "Linux":
cpu = open("/proc/cpuinfo")
f_lines = cpu.readlines()
cpu.close()
# check what native targets do we have # check what native targets do we have
for i in range(0,len(f_lines)): if current_OS != "Windows":
if SSE2 == False and "sse2" in f_lines[i]: try_do_LLVM("build check_ISA", "clang check_isa.cpp -o check_isa.exe", True)
SSE2 = True; else:
answer = answer + ["sse2-i32x4", "sse2-i32x8"] try_do_LLVM("build check_ISA", "cl check_isa.cpp", True)
if SSE4 == False and "sse4_1" in f_lines[i]: SSE2 = ["sse2-i32x4", "sse2-i32x8"]
SSE4 = True; SSE4 = ["sse4-i32x4", "sse4-i32x8", "sse4-i16x8", "sse4-i8x16"]
answer = answer + ["sse4-i32x4", "sse4-i32x8", "sse4-i16x8", "sse4-i8x16"] AVX = ["avx1-i32x8", "avx1-i32x16", "avx1-i64x4"]
if AVX == False and "avx" in f_lines[i]: AVX11 = ["avx1.1-i32x8","avx1.1-i32x16","avx1.1-i64x4"]
AVX = True; AVX2 = ["avx2-i32x8", "avx2-i32x16", "avx2-i64x4"]
answer = answer + ["avx1-i32x8", "avx1-i32x16", "avx1-i64x4"] targets = [["AVX2", AVX2, False], ["AVX1.1", AVX11, False], ["AVX", AVX, False], ["SSE4", SSE4, False], ["SSE2", SSE2, False]]
if AVX11 == False and "rdrand" in f_lines[i]: f_lines = take_lines("check_isa.exe", "first")
AVX11 = True; for i in range(0,5):
answer = answer + ["avx1.1-i32x8", "avx1.1-i32x16", "avx1.1-i64x4"] if targets[i][0] in f_lines:
if AVX2 == False and "avx2" in f_lines[i]: for j in range(i,5):
AVX2 = True; answer = targets[j][1] + answer
answer = answer + ["avx2-i32x8", "avx2-i32x16", "avx2-i64x4"] targets[j][2] = True
if current_OS == "MacOS": break
f_lines = take_lines("sysctl machdep.cpu.features", "first") if current_OS != "Windows":
if "SSE2" in f_lines:
SSE2 = True;
answer = answer + ["sse2-i32x4", "sse2-i32x8"]
if "SSE4.1" in f_lines:
SSE4 = True;
answer = answer + ["sse4-i32x4", "sse4-i32x8", "sse4-i16x8", "sse4-i8x16"]
if "AVX1.0" in f_lines:
AVX = True;
answer = answer + ["avx1-i32x8", "avx1-i32x16", "avx1-i64x4"]
if "RDRAND" in f_lines:
AVX11 = True;
answer = answer + ["avx1.1-i32x8", "avx1.1-i32x16", "avx1.1-i64x4"]
if "AVX2.0" in f_lines:
AVX2 = True;
answer = answer + ["avx2-i32x8", "avx2-i32x16", "avx2-i64x4"]
answer = answer + ["generic-4", "generic-16", "generic-8", "generic-1", "generic-32", "generic-64"] answer = answer + ["generic-4", "generic-16", "generic-8", "generic-1", "generic-32", "generic-64"]
# now check what targets we have with the help of SDE # now check what targets we have with the help of SDE
sde_exists = "" sde_exists = ""
PATH_dir = string.split(os.getenv("PATH"), os.pathsep) PATH_dir = string.split(os.getenv("PATH"), os.pathsep)
if current_OS == "Windows":
sde_n = "sde.exe"
else:
sde_n = "sde"
for counter in PATH_dir: for counter in PATH_dir:
if os.path.exists(counter + os.sep + "sde") and sde_exists == "": if os.path.exists(counter + os.sep + sde_n) and sde_exists == "":
sde_exists = counter + os.sep + "sde" sde_exists = counter + os.sep + sde_n
if os.environ.get("SDE_HOME") != None: if os.environ.get("SDE_HOME") != None:
if os.path.exists(os.environ.get("SDE_HOME") + os.sep + "sde"): if os.path.exists(os.environ.get("SDE_HOME") + os.sep + sde_n):
sde_exists = os.environ.get("SDE_HOME") + os.sep + "sde" sde_exists = os.environ.get("SDE_HOME") + os.sep + sde_n
if sde_exists == "": if sde_exists == "":
error("you haven't got sde neither in SDE_HOME nor in your PATH.\n" + error("you haven't got sde neither in SDE_HOME nor in your PATH.\n" +
"To test all platforms please set SDE_HOME to path containing SDE.\n" + "To test all platforms please set SDE_HOME to path containing SDE.\n" +
@@ -252,24 +248,38 @@ def check_targets():
# here we have SDE # here we have SDE
f_lines = take_lines(sde_exists + " -help", "all") f_lines = take_lines(sde_exists + " -help", "all")
for i in range(0,len(f_lines)): for i in range(0,len(f_lines)):
if SSE4 == False and "wsm" in f_lines[i]: if targets[3][2] == False and "wsm" in f_lines[i]:
answer_sde = answer_sde + [["-wsm", "sse4-i32x4"], ["-wsm", "sse4-i32x8"], ["-wsm", "sse4-i16x8"], ["-wsm", "sse4-i8x16"]] answer_sde = answer_sde + [["-wsm", "sse4-i32x4"], ["-wsm", "sse4-i32x8"], ["-wsm", "sse4-i16x8"], ["-wsm", "sse4-i8x16"]]
if AVX == False and "snb" in f_lines[i]: if targets[2][2] == False and "snb" in f_lines[i]:
answer_sde = answer_sde + [["-snb", "avx1-i32x8"], ["-snb", "avx1-i32x16"], ["-snb", "avx1-i64x4"]] answer_sde = answer_sde + [["-snb", "avx1-i32x8"], ["-snb", "avx1-i32x16"], ["-snb", "avx1-i64x4"]]
if AVX11 == False and "ivb" in f_lines[i]: if targets[1][2] == False and "ivb" in f_lines[i]:
answer_sde = answer_sde + [["-ivb", "avx1.1-i32x8"], ["-ivb", "avx1.1-i32x16"], ["-ivb", "avx1.1-i64x4"]] answer_sde = answer_sde + [["-ivb", "avx1.1-i32x8"], ["-ivb", "avx1.1-i32x16"], ["-ivb", "avx1.1-i64x4"]]
if AVX2 == False and "hsw" in f_lines[i]: if targets[0][2] == False and "hsw" in f_lines[i]:
answer_sde = answer_sde + [["-hsw", "avx2-i32x8"], ["-hsw", "avx2-i32x16"], ["-hsw", "avx2-i64x4"]] answer_sde = answer_sde + [["-hsw", "avx2-i32x8"], ["-hsw", "avx2-i32x16"], ["-hsw", "avx2-i64x4"]]
return [answer, answer_sde] return [answer, answer_sde]
def build_ispc(version_LLVM, make): def build_ispc(version_LLVM, make):
current_path = os.getcwd() current_path = os.getcwd()
os.chdir(os.environ["ISPC_HOME"]) os.chdir(os.environ["ISPC_HOME"])
if current_OS != "Windows":
p_temp = os.getenv("PATH") p_temp = os.getenv("PATH")
os.environ["PATH"] = os.environ["LLVM_HOME"] + "/bin-" + version_LLVM + "/bin:" + os.environ["PATH"] os.environ["PATH"] = os.environ["LLVM_HOME"] + "/bin-" + version_LLVM + "/bin:" + os.environ["PATH"]
try_do_LLVM("clean ISPC for building", "make clean", True) try_do_LLVM("clean ISPC for building", "make clean", True)
try_do_LLVM("build ISPC with LLVM version " + version_LLVM + " ", make, True) try_do_LLVM("build ISPC with LLVM version " + version_LLVM + " ", make, True)
os.environ["PATH"] = p_temp os.environ["PATH"] = p_temp
else:
p_temp = os.getenv("LLVM_INSTALL_DIR")
v_temp = os.getenv("LLVM_VERSION")
os.environ["LLVM_INSTALL_DIR"] = os.environ["LLVM_HOME"] + "\\bin-" + version_LLVM
if version_LLVM == "3.3":
temp = "3_3"
if version_LLVM == "trunk":
temp = "3_4"
os.environ["LLVM_VERSION"] = "LLVM_" + temp
try_do_LLVM("clean ISPC for building", "msbuild ispc.vcxproj /t:clean", True)
try_do_LLVM("build ISPC with LLVM version " + version_LLVM + " ", "msbuild ispc.vcxproj /V:m /p:Platform=Win32 /p:Configuration=Release /t:rebuild", True)
os.environ["LLVM_INSTALL_DIR"] = p_temp
os.environ["LLVM_VERSION"] = v_temp
os.chdir(current_path) os.chdir(current_path)
def execute_stability(stability, R, print_version): def execute_stability(stability, R, print_version):
@@ -305,8 +315,12 @@ def execute_stability(stability, R, print_version):
def run_special_tests(): def run_special_tests():
i = 5 i = 5
class options_for_drivers:
pass
def validation_run(only, only_targets, reference_branch, number, notify, update, speed_number, make, perf_llvm, time): def validation_run(only, only_targets, reference_branch, number, notify, update, speed_number, make, perf_llvm, time):
os.chdir(os.environ["ISPC_HOME"]) os.chdir(os.environ["ISPC_HOME"])
if current_OS != "Windows":
os.environ["PATH"] = os.environ["ISPC_HOME"] + ":" + os.environ["PATH"] os.environ["PATH"] = os.environ["ISPC_HOME"] + ":" + os.environ["PATH"]
if options.notify != "": if options.notify != "":
common.remove_if_exists(os.environ["ISPC_HOME"] + os.sep + "notify_log.log") common.remove_if_exists(os.environ["ISPC_HOME"] + os.sep + "notify_log.log")
@@ -319,8 +333,6 @@ def validation_run(only, only_targets, reference_branch, number, notify, update,
print_debug("Folder: " + os.environ["ISPC_HOME"] + "\n", False, "") print_debug("Folder: " + os.environ["ISPC_HOME"] + "\n", False, "")
date = datetime.datetime.now() date = datetime.datetime.now()
print_debug("Date: " + date.strftime('%H:%M %d/%m/%Y') + "\n", False, "") print_debug("Date: " + date.strftime('%H:%M %d/%m/%Y') + "\n", False, "")
class options_for_drivers:
pass
# *** *** *** # *** *** ***
# Stability validation run # Stability validation run
# *** *** *** # *** *** ***
@@ -475,12 +487,14 @@ def validation_run(only, only_targets, reference_branch, number, notify, update,
performance = options_for_drivers() performance = options_for_drivers()
# performance constant options # performance constant options
performance.number = number performance.number = number
performance.config = "./perf.ini" performance.config = "." + os.sep + "perf.ini"
performance.path = "./" performance.path = "." + os.sep
performance.silent = True performance.silent = True
performance.output = "" performance.output = ""
performance.compiler = "" performance.compiler = ""
performance.ref = "ispc_ref" performance.ref = "ispc_ref"
if current_OS == "Windows":
performance.ref = "ispc_ref.exe"
performance.in_file = "." + os.sep + f_date + os.sep + "performance.log" performance.in_file = "." + os.sep + f_date + os.sep + "performance.log"
# prepare LLVM 3.3 as newest LLVM # prepare LLVM 3.3 as newest LLVM
need_LLVM = check_LLVM(["3.3"]) need_LLVM = check_LLVM(["3.3"])
@@ -502,7 +516,11 @@ def validation_run(only, only_targets, reference_branch, number, notify, update,
sys.stdout.write(".\n") sys.stdout.write(".\n")
build_ispc("3.3", make) build_ispc("3.3", make)
sys.stdout.write(".\n") sys.stdout.write(".\n")
if current_OS != "Windows":
os.rename("ispc", "ispc_ref") os.rename("ispc", "ispc_ref")
else:
common.remove_if_exists("Release\\ispc_ref.exe")
os.rename("Release\\ispc.exe", "Release\\ispc_ref.exe")
try_do_LLVM("checkout test branch " + current_branch + " ", "git checkout " + current_branch, True) try_do_LLVM("checkout test branch " + current_branch + " ", "git checkout " + current_branch, True)
if stashing: if stashing:
try_do_LLVM("return current branch ", "git stash pop", True) try_do_LLVM("return current branch ", "git stash pop", True)
@@ -541,7 +559,6 @@ def Main():
global current_OS global current_OS
if (platform.system() == 'Windows' or 'CYGWIN_NT' in platform.system()) == True: if (platform.system() == 'Windows' or 'CYGWIN_NT' in platform.system()) == True:
current_OS = "Windows" current_OS = "Windows"
error("Windows isn't supported now", 1)
else: else:
if (platform.system() == 'Darwin'): if (platform.system() == 'Darwin'):
current_OS = "MacOS" current_OS = "MacOS"
@@ -566,7 +583,9 @@ def Main():
for iterator in test_only: for iterator in test_only:
if not (" " + iterator + " " in test_only_r): if not (" " + iterator + " " in test_only_r):
error("unknow option for only: " + iterator, 1) error("unknow option for only: " + iterator, 1)
if current_OS == "Windows":
if options.debug == True or options.selfbuild == True or options.tarball != "":
error("Debug, selfbuild and tarball options are unsupported on windows", 1)
global f_date global f_date
f_date = "logs" f_date = "logs"
common.remove_if_exists(f_date) common.remove_if_exists(f_date)
@@ -626,6 +645,7 @@ import common
error = common.error error = common.error
take_lines = common.take_lines take_lines = common.take_lines
print_debug = common.print_debug print_debug = common.print_debug
if __name__ == '__main__':
# parsing options # parsing options
class MyParser(OptionParser): class MyParser(OptionParser):
def format_epilog(self, formatter): def format_epilog(self, formatter):

View File

@@ -719,8 +719,9 @@ static FORCEINLINE __vec16_i32 __shuffle2_i32(__vec16_i32 v0, __vec16_i32 v1, __
template <int ALIGN> static FORCEINLINE __vec16_i32 __load(const __vec16_i32 *p) template <int ALIGN> static FORCEINLINE __vec16_i32 __load(const __vec16_i32 *p)
{ {
#ifdef ISPC_FORCE_ALIGNED_MEMORY__REMOVETHIS_WHEN_FIXED #ifdef ISPC_FORCE_ALIGNED_MEMORY
return __load<64>(p); // return __load<64>(p);
return _mm512_load_epi32(p);
#else #else
__vec16_i32 v; __vec16_i32 v;
v = _mm512_extloadunpacklo_epi32(v, p, _MM_UPCONV_EPI32_NONE, _MM_HINT_NONE); v = _mm512_extloadunpacklo_epi32(v, p, _MM_UPCONV_EPI32_NONE, _MM_HINT_NONE);
@@ -731,8 +732,9 @@ template <int ALIGN> static FORCEINLINE __vec16_i32 __load(const __vec16_i32 *p)
template <int ALIGN> static FORCEINLINE void __store(__vec16_i32 *p, __vec16_i32 v) template <int ALIGN> static FORCEINLINE void __store(__vec16_i32 *p, __vec16_i32 v)
{ {
#ifdef ISPC_FORCE_ALIGNED_MEMORY__REMOVETHIS_WHEN_FIXED #ifdef ISPC_FORCE_ALIGNED_MEMORY
__store<64>(p,v); // __store<64>(p,v);
_mm512_store_epi32(p, v);
#else #else
_mm512_extpackstorelo_epi32( p, v, _MM_DOWNCONV_EPI32_NONE, _MM_HINT_NONE); _mm512_extpackstorelo_epi32( p, v, _MM_DOWNCONV_EPI32_NONE, _MM_HINT_NONE);
_mm512_extpackstorehi_epi32((uint8_t*)p+64, v, _MM_DOWNCONV_EPI32_NONE, _MM_HINT_NONE); _mm512_extpackstorehi_epi32((uint8_t*)p+64, v, _MM_DOWNCONV_EPI32_NONE, _MM_HINT_NONE);
@@ -979,8 +981,11 @@ static FORCEINLINE __vec16_i64 __shuffle2_double(__vec16_i64 _v0, __vec16_i64 _v
template <int ALIGN> static FORCEINLINE __vec16_i64 __load(const __vec16_i64 *p) template <int ALIGN> static FORCEINLINE __vec16_i64 __load(const __vec16_i64 *p)
{ {
#ifdef ISPC_FORCE_ALIGNED_MEMORY__REMOVETHIS_WHEN_FIXED #ifdef ISPC_FORCE_ALIGNED_MEMORY
return __load<128>(p); // return __load<128>(p);
__m512i v2 = _mm512_load_epi32(p);
__m512i v1 = _mm512_load_epi32(((uint8_t*)p)+64);
return __vec16_i64(v2,v1);
#else #else
__vec16_i32 v1; __vec16_i32 v1;
__vec16_i32 v2; __vec16_i32 v2;
@@ -995,8 +1000,12 @@ template <int ALIGN> static FORCEINLINE __vec16_i64 __load(const __vec16_i64 *p)
template <int ALIGN> static FORCEINLINE void __store(__vec16_i64 *p, __vec16_i64 v) template <int ALIGN> static FORCEINLINE void __store(__vec16_i64 *p, __vec16_i64 v)
{ {
#ifdef ISPC_FORCE_ALIGNED_MEMORY__REMOVETHIS_WHEN_FIXED #ifdef ISPC_FORCE_ALIGNED_MEMORY
return __store<128>(p,v); // __store<128>(p,v);
__m512i v1 = v.v2;
__m512i v2 = v.v1;
_mm512_store_epi64(p, v2);
_mm512_store_epi64(((uint8_t*)p)+64, v1);
#else #else
__m512i v1 = v.v2; __m512i v1 = v.v2;
__m512i v2 = v.v1; __m512i v2 = v.v1;
@@ -1099,8 +1108,9 @@ static FORCEINLINE __vec16_f __shuffle2_float(__vec16_f _v0, __vec16_f _v1, __ve
template <int ALIGN> static FORCEINLINE __vec16_f __load(const __vec16_f *p) template <int ALIGN> static FORCEINLINE __vec16_f __load(const __vec16_f *p)
{ {
#ifdef ISPC_FORCE_ALIGNED_MEMORY__REMOVETHIS_WHEN_FIXED #ifdef ISPC_FORCE_ALIGNED_MEMORY
return __load<64>(p); // return __load<64>(p);
return _mm512_load_ps(p);
#else #else
__vec16_f v; __vec16_f v;
v = _mm512_extloadunpacklo_ps(v, p, _MM_UPCONV_PS_NONE, _MM_HINT_NONE); v = _mm512_extloadunpacklo_ps(v, p, _MM_UPCONV_PS_NONE, _MM_HINT_NONE);
@@ -1111,8 +1121,9 @@ template <int ALIGN> static FORCEINLINE __vec16_f __load(const __vec16_f *p)
template <int ALIGN> static FORCEINLINE void __store(__vec16_f *p, __vec16_f v) template <int ALIGN> static FORCEINLINE void __store(__vec16_f *p, __vec16_f v)
{ {
#ifdef ISPC_FORCE_ALIGNED_MEMORY__REMOVETHIS_WHEN_FIXED #ifdef ISPC_FORCE_ALIGNED_MEMORY
__store<64>(p,v); // __store<64>(p,v);
_mm512_store_ps(p, v);
#else #else
_mm512_extpackstorelo_ps( p, v, _MM_DOWNCONV_PS_NONE, _MM_HINT_NONE); _mm512_extpackstorelo_ps( p, v, _MM_DOWNCONV_PS_NONE, _MM_HINT_NONE);
_mm512_extpackstorehi_ps((uint8_t*)p+64, v, _MM_DOWNCONV_PS_NONE, _MM_HINT_NONE); _mm512_extpackstorehi_ps((uint8_t*)p+64, v, _MM_DOWNCONV_PS_NONE, _MM_HINT_NONE);
@@ -1372,8 +1383,9 @@ static FORCEINLINE __vec16_d __shuffle2_double(__vec16_d _v0, __vec16_d _v1, con
template <int ALIGN> static FORCEINLINE __vec16_d __load(const __vec16_d *p) \ template <int ALIGN> static FORCEINLINE __vec16_d __load(const __vec16_d *p) \
{ {
#ifdef ISPC_FORCE_ALIGNED_MEMORY__REMOVETHIS_WHEN_FIXED #ifdef ISPC_FORCE_ALIGNED_MEMORY
return __load<128>(p); // return __load<128>(p);
return __vec16_d(_mm512_load_pd(p), _mm512_load_pd(((uint8_t*)p)+64));
#else #else
__vec16_d ret; __vec16_d ret;
ret.v1 = _mm512_extloadunpacklo_pd(ret.v1, p, _MM_UPCONV_PD_NONE, _MM_HINT_NONE); ret.v1 = _mm512_extloadunpacklo_pd(ret.v1, p, _MM_UPCONV_PD_NONE, _MM_HINT_NONE);
@@ -1386,8 +1398,10 @@ template <int ALIGN> static FORCEINLINE __vec16_d __load(const __vec16_d *p) \
template <int ALIGN> static FORCEINLINE void __store(__vec16_d *p, __vec16_d v) template <int ALIGN> static FORCEINLINE void __store(__vec16_d *p, __vec16_d v)
{ {
#ifdef ISPC_FORCE_ALIGNED_MEMORY__REMOVETHIS_WHEN_FIXED #ifdef ISPC_FORCE_ALIGNED_MEMORY
return __store<128>(p,v); // return __store<128>(p,v);
_mm512_store_pd(p, v.v1);
_mm512_store_pd(((uint8_t*)p)+64, v.v2);
#else #else
_mm512_extpackstorelo_pd(p, v.v1, _MM_DOWNCONV_PD_NONE, _MM_HINT_NONE); _mm512_extpackstorelo_pd(p, v.v1, _MM_DOWNCONV_PD_NONE, _MM_HINT_NONE);
_mm512_extpackstorehi_pd((uint8_t*)p+64, v.v1, _MM_DOWNCONV_PD_NONE, _MM_HINT_NONE); _mm512_extpackstorehi_pd((uint8_t*)p+64, v.v1, _MM_DOWNCONV_PD_NONE, _MM_HINT_NONE);

View File

@@ -2798,6 +2798,17 @@ BinaryExpr::TypeCheck() {
} }
} }
const Type *
BinaryExpr::GetLValueType() const {
const Type *t = GetType();
if (CastType<PointerType>(t) != NULL) {
// Are we doing something like (basePtr + offset)[...] = ...
return t;
}
else {
return NULL;
}
}
int int
BinaryExpr::EstimateCost() const { BinaryExpr::EstimateCost() const {
@@ -4275,8 +4286,9 @@ IndexExpr::GetValue(FunctionEmitContext *ctx) const {
} }
else { else {
Symbol *baseSym = GetBaseSymbol(); Symbol *baseSym = GetBaseSymbol();
if (dynamic_cast<FunctionCallExpr *>(baseExpr) == NULL) { if (dynamic_cast<FunctionCallExpr *>(baseExpr) == NULL &&
// Only check for non-function calls dynamic_cast<BinaryExpr *>(baseExpr) == NULL) {
// Don't check if we're doing a function call or pointer arith
AssertPos(pos, baseSym != NULL); AssertPos(pos, baseSym != NULL);
} }
mask = lMaskForSymbol(baseSym, ctx); mask = lMaskForSymbol(baseSym, ctx);

1
expr.h
View File

@@ -155,6 +155,7 @@ public:
llvm::Value *GetValue(FunctionEmitContext *ctx) const; llvm::Value *GetValue(FunctionEmitContext *ctx) const;
const Type *GetType() const; const Type *GetType() const;
const Type *GetLValueType() const;
void Print() const; void Print() const;
Expr *Optimize(); Expr *Optimize();

View File

@@ -4930,8 +4930,6 @@ CreatePeepholePass() {
return new PeepholePass; return new PeepholePass;
} }
#include <iostream>
/** Given an llvm::Value known to be an integer, return its value as /** Given an llvm::Value known to be an integer, return its value as
an int64_t. an int64_t.
*/ */
@@ -4978,13 +4976,13 @@ ReplaceStdlibShiftPass::runOnBasicBlock(llvm::BasicBlock &bb) {
if (llvm::CallInst *ci = llvm::dyn_cast<llvm::CallInst>(inst)) { if (llvm::CallInst *ci = llvm::dyn_cast<llvm::CallInst>(inst)) {
llvm::Function *func = ci->getCalledFunction(); llvm::Function *func = ci->getCalledFunction();
for (int i = 0; i < 6; i++) { for (int i = 0; i < 6; i++) {
if (shifts[i] == func) { if (shifts[i] && (shifts[i] == func)) {
// we matched a call // we matched a call
llvm::Value *shiftedVec = ci->getArgOperand(0); llvm::Value *shiftedVec = ci->getArgOperand(0);
llvm::Value *shiftAmt = ci->getArgOperand(1); llvm::Value *shiftAmt = ci->getArgOperand(1);
if (llvm::isa<llvm::Constant>(shiftAmt)) { if (llvm::isa<llvm::Constant>(shiftAmt)) {
int vectorWidth = g->target->getVectorWidth(); int vectorWidth = g->target->getVectorWidth();
int shuffleVals[vectorWidth]; int * shuffleVals = new int[vectorWidth];
int shiftInt = lGetIntValue(shiftAmt); int shiftInt = lGetIntValue(shiftAmt);
for (int i = 0; i < vectorWidth; i++) { for (int i = 0; i < vectorWidth; i++) {
int s = i + shiftInt; int s = i + shiftInt;
@@ -4998,6 +4996,7 @@ ReplaceStdlibShiftPass::runOnBasicBlock(llvm::BasicBlock &bb) {
shuffleIdxs, "vecShift", ci); shuffleIdxs, "vecShift", ci);
ci->replaceAllUsesWith(shuffle); ci->replaceAllUsesWith(shuffle);
modifiedAny = true; modifiedAny = true;
delete [] shuffleVals;
} else { } else {
PerformanceWarning(SourcePos(), "Stdlib shift() called without constant shift amount."); PerformanceWarning(SourcePos(), "Stdlib shift() called without constant shift amount.");
} }

14
perf.py
View File

@@ -42,6 +42,9 @@ def print_file(line):
def build_test(commands): def build_test(commands):
os.system(commands[4]) os.system(commands[4])
test = os.system(commands[1]) test = os.system(commands[1])
if is_windows:
common.remove_if_exists(".\\X64\\Release1")
os.rename(".\\X64\\Release", ".\\X64\\Release1")
if options.ref: if options.ref:
ref = os.system(commands[3]) ref = os.system(commands[3])
return (options.ref and ref) or test return (options.ref and ref) or test
@@ -345,9 +348,14 @@ def perf(options1, args):
if options.ref != "": if options.ref != "":
options.ref = True options.ref = True
if os.environ.get("ISPC_HOME") != None: if os.environ.get("ISPC_HOME") != None:
if is_windows == False:
if os.path.exists(os.environ["ISPC_HOME"] + os.sep + ispc_test): if os.path.exists(os.environ["ISPC_HOME"] + os.sep + ispc_test):
ispc_test_exists = True ispc_test_exists = True
ispc_test = os.environ["ISPC_HOME"] + os.sep + ispc_test ispc_test = os.environ["ISPC_HOME"] + os.sep + ispc_test
else:
if os.path.exists(os.environ["ISPC_HOME"] + "\\Release\\" + ispc_test):
ispc_test_exists = True
ispc_test = os.environ["ISPC_HOME"] + "\\Release\\" + ispc_test
for counter in PATH_dir: for counter in PATH_dir:
if ispc_test_exists == False: if ispc_test_exists == False:
if os.path.exists(counter + os.sep + ispc_test): if os.path.exists(counter + os.sep + ispc_test):
@@ -419,9 +427,9 @@ def perf(options1, args):
re_command = "make clean >> "+build_log re_command = "make clean >> "+build_log
else: else:
ex_command_ref = "x64\\Release\\ref.exe " + command + " >> " + perf_temp + "_ref" ex_command_ref = "x64\\Release\\ref.exe " + command + " >> " + perf_temp + "_ref"
ex_command = "x64\\Release\\test.exe " + command + " >> " + perf_temp + "_test" ex_command = "x64\\Release1\\test.exe " + command + " >> " + perf_temp + "_test"
bu_command_ref = "msbuild /V:m /p:Platform=x64 /p:Configuration=Release /p:TargetDir=.\ /p:TargetName=ref /t:rebuild >> " + build_log bu_command_ref = "msbuild /V:m /p:Platform=x64 /p:Configuration=Release /p:TargetDir=.\ /p:TargetName=ref /p:ISPC_compiler=ispc_ref /t:rebuild >> " + build_log
bu_command = "msbuild /V:m /p:Platform=x64 /p:Configuration=Release /p:TargetDir=.\ /p:TargetName=test /t:rebuild >> " + build_log bu_command = "msbuild /V:m /p:Platform=x64 /p:Configuration=Release /p:TargetDir=.\ /p:TargetName=test /p:ISPC_compiler=ispc /t:rebuild >> " + build_log
re_command = "msbuild /t:clean >> " + build_log re_command = "msbuild /t:clean >> " + build_log
commands = [ex_command, bu_command, ex_command_ref, bu_command_ref, re_command] commands = [ex_command, bu_command, ex_command_ref, bu_command_ref, re_command]
# parsing config parameters # parsing config parameters

View File

@@ -0,0 +1,16 @@
export uniform int width() { return programCount; }
int foo(uniform float * uniform base, uniform int uOfs, varying int vOfs) {
return (base+uOfs)[vOfs];
}
export void f_f(uniform float RET[], uniform float aFOO[]) {
uniform float * uniform ptr = &aFOO[0];
int val = foo(ptr, programCount, programIndex);
RET[programIndex] = val;
}
export void result(uniform float RET[]) {
RET[programIndex] = 1+programCount+programIndex;
}