int main(int argc, char **argv) {
Var x("x");
Func f("f");
Param<float> u;
f(x) = u;
Target target = get_target_from_environment();
if (target.features & Target::CUDA) {
f.cuda_tile(x, 256);
}
u.set(17.0f);
Image<float> out_17 = f.realize(1024);
u.set(123.0f);
Image<float> out_123 = f.realize(1024);
for (int i = 0; i < 1024; i++) {
if (out_17(i) != 17.0f || out_123(i) != 123.0f) {
printf("Failed!\n");
for (int i = 0; i < 1024; i++) {
printf("%f %f\n", out_17(i), out_123(i));
}
return -1;
}
}
printf("Success!\n");
return 0;
}
int main(int argc, char **argv) {
Var x("x");
Func f("f");
Param<float> u;
f(x) = u;
std::string target = get_target();
if (target == "ptx" || target == "ptx-debug") {
f.cuda_tile(x, 256);
}
u.set(17.0f);
Image<float> out_17 = f.realize(1024);
u.set(123.0f);
Image<float> out_123 = f.realize(1024);
for (int i = 0; i < 1024; i++) {
if (out_17(i) != 17.0f || out_123(i) != 123.0f) {
printf("Failed!\n");
for (int i = 0; i < 1024; i++) {
printf("%f %f\n", out_17(i), out_123(i));
}
return -1;
}
}
printf("Success!\n");
return 0;
}
int simple_rfactor_with_specialize_test(bool compile_module) {
Func f("f"), g("g");
Var x("x"), y("y");
f(x, y) = x + y;
f.compute_root();
g(x, y) = 40;
RDom r(10, 20, 30, 40);
g(r.x, r.y) = min(f(r.x, r.y) + 2, g(r.x, r.y));
Param<int> p;
Var u("u");
Func intm = g.update(0).specialize(p >= 10).rfactor(r.y, u);
intm.compute_root();
intm.vectorize(u, 8);
intm.update(0).vectorize(r.x, 2);
if (compile_module) {
p.set(20);
// Check the call graphs.
Module m = g.compile_to_module({g.infer_arguments()});
CheckCalls checker;
m.functions().front().body.accept(&checker);
CallGraphs expected = {
{g.name(), {}},
{g.update(0).name(), {f.name(), intm.name(), g.name()}},
{intm.name(), {}},
{intm.update(0).name(), {f.name(), intm.name()}},
{f.name(), {}},
};
if (check_call_graphs(checker.calls, expected) != 0) {
return -1;
}
} else {
{
p.set(0);
Image<int> im = g.realize(80, 80);
auto func = [](int x, int y, int z) {
return (10 <= x && x <= 29) && (30 <= y && y <= 69) ? std::min(x + y + 2, 40) : 40;
};
if (check_image(im, func)) {
return -1;
}
}
{
p.set(20);
Image<int> im = g.realize(80, 80);
auto func = [](int x, int y, int z) {
return (10 <= x && x <= 29) && (30 <= y && y <= 69) ? std::min(x + y + 2, 40) : 40;
};
if (check_image(im, func)) {
return -1;
}
}
}
return 0;
}
int tuple_memoize_test(bool toggle_val, int index) {
buffer_index = index;
Param<bool> toggle;
Func f1("f1_" + std::to_string(index)), f2("f2_" + std::to_string(index));
Var x;
f1(x) = Tuple(2*x, 2*x);
f2(x) = Tuple(select(toggle, f1(x)[0], 1),
select(toggle, f1(x)[1], 1));
f1.compute_root().memoize();
f2.set_custom_trace(&single_toggle_trace);
f1.trace_stores();
f2.compile_jit();
set_toggle1 = toggle_val;
toggle.set(set_toggle1);
Realization out = f2.realize(128);
Image<int> out0 = out[0];
Image<int> out1 = out[1];
if (check_correctness_single(out0, set_toggle1) != 0) {
return -1;
}
if (check_correctness_single(out1, set_toggle1) != 0) {
return -1;
}
return 0;
}
int main(int argc, char **argv) {
ImageParam im1(UInt(8), 1);
Buffer<uint8_t> im2(10), im3(20);
Param<int> j;
assert(im1.dimensions() == 1);
assert(im2.dimensions() == 1);
assert(im3.dimensions() == 1);
Func f;
Var x;
f(x) = x + im1.width();
RDom r(0, clamp(im2(j), 0, 99));
f(r) = 37;
im2(3) = 10;
j.set(3);
im1.set(im3);
Buffer<int> result = f.realize(100);
for (int i = 0; i < 100; i++) {
int correct = i < im2(3) ? 37 : (i+20);
if (result(i) != correct) {
printf("result(%d) = %d instead of %d\n", i, result(i), correct);
return -1;
}
}
printf("Success!\n");
return 0;
}
int non_trivial_allocate_predicate_test(bool toggle_val, int index) {
buffer_index = index;
Param<bool> toggle;
Func f1("f1_" + std::to_string(index)), f2("f2_" + std::to_string(index));
Func f3("f3_" + std::to_string(index));
Var x;
// Generate allocate f1[...] if toggle
f1(x) = 2*x;
f2(x) = select(toggle, f1(x), 1);
f3(x) = select(toggle, f2(x), 1);
f1.compute_root().memoize();
f2.compute_root().memoize();
f3.set_custom_trace(&double_toggle_trace);
f1.trace_stores();
f2.trace_stores();
f3.compile_jit();
set_toggle1 = toggle_val;
set_toggle2 = toggle_val;
toggle.set(set_toggle1);
Image<int> out = f3.realize(10);
if (check_correctness_single(out, set_toggle1) != 0) {
return -1;
}
return 0;
}
int single_memoize_test(bool toggle_val, int index) {
buffer_index = index;
Param<bool> toggle;
Func f1("f1_" + std::to_string(index)), f2("f2_" + std::to_string(index));
Var x;
f1(x) = 2*x;
f2(x) = select(toggle, f1(x), 1);
f1.compute_root().memoize();
f2.set_custom_trace(&single_toggle_trace);
f1.trace_stores();
f2.compile_jit();
set_toggle1 = toggle_val;
toggle.set(set_toggle1);
Image<int> out = f2.realize(10);
if (check_correctness_single(out, set_toggle1) != 0) {
return -1;
}
return 0;
}
void check(MemoryType t1, MemoryType t2, MemoryType t3) {
Var x;
// By default, small constant-sized allocations, or
// allocations that can be bounded with a small constant size,
// go on the stack. Other allocations go on the heap.
Func f1, f2, f3;
f1(x) = x;
f1.compute_root().store_in(t1);
f2(x) = x;
f2.compute_root().store_in(t2);
f3(x) = x;
f3.compute_root().store_in(t3);
Func f;
Param<bool> p;
f(x) = (f1(0) + f1(1)) + f2(select(p, 0, 2)) + f2(0) + f3(x % 1000);
p.set(true);
int expected_mallocs = ((t1 == MemoryType::Heap ? 1 : 0) +
(t2 == MemoryType::Heap ? 1 : 0) +
(t3 == MemoryType::Heap ? 1 : 0));
mallocs = 0;
f.set_custom_allocator(my_malloc, my_free);
f.realize(1024);
if (mallocs != expected_mallocs) {
std::cerr << "Wrong number of mallocs for " << t1 << ", " << t2 << ", " << t3 << "\n"
<< "Expected " << expected_mallocs << " got " << mallocs << "\n";
exit(-1);
}
}
int main(int argc, char **argv) {
Var x, y, z;
Func f;
Param<int> k;
k.set(3);
f(x, y, z) = x*y+z*k+1;
f.parallel(x);
f.parallel(y);
f.parallel(z);
Image<int> im = f.realize(64, 64, 64);
for (int x = 0; x < 64; x++) {
for (int y = 0; y < 64; y++) {
for (int z = 0; z < 64; z++) {
if (im(x, y, z) != x*y+z*3+1) {
printf("im(%d, %d, %d) = %d\n", x, y, z, im(x, y, z));
return -1;
}
}
}
}
printf("Success!\n");
return 0;
}
int main(int argc, char **argv) {
// Define a pipeline that dumps some squares to a file using an
// external consumer stage.
Func source;
Var x;
source(x) = x*x;
Param<int> min, extent;
Param<const char *> filename;
Func sink;
std::vector<ExternFuncArgument> args;
args.push_back(source);
args.push_back(filename);
args.push_back(min);
args.push_back(extent);
sink.define_extern("dump_to_file", args, Int(32), 0);
source.compute_root();
sink.compile_jit();
// Dump the first 10 squares to a file
filename.set("halide_test_extern_consumer.txt");
min.set(0);
extent.set(10);
sink.realize();
if (!check_result())
return -1;
// Test ImageParam ExternFuncArgument via passed in image.
Image<int32_t> buf = source.realize(10);
ImageParam passed_in(Int(32), 1);
passed_in.set(buf);
Func sink2;
std::vector<ExternFuncArgument> args2;
args2.push_back(passed_in);
args2.push_back(filename);
args2.push_back(min);
args2.push_back(extent);
sink2.define_extern("dump_to_file", args2, Int(32), 0);
sink2.realize();
if (!check_result())
return -1;
printf("Success!\n");
return 0;
}
CoordXform() : m0("m0"), m1("m1"), m2("m2"), m3("m3"), m4("m4"), m5("m5") {
m0.set(m[0]);
m1.set(m[1]);
m2.set(m[2]);
m3.set(m[3]);
m4.set(m[4]);
m5.set(m[5]);
}
int main(int argc, char **argv) {
Param<float> val;
Func f, g;
Var x, y;
f(x, y) = val + cast<uint8_t>(x);
g(x, y) = f(x, y) + f(x - 1, y) + f(x + 1, y);
g.split(y, y, _, 16);
f.store_root();
f.compute_at(g, y).memoize();
val.set(23.0f);
Image<uint8_t> out = g.realize(128, 128);
for (int32_t i = 0; i < 128; i++) {
for (int32_t j = 0; j < 128; j++) {
assert(out(i, j) == (uint8_t)(3 * 23 + i + (i - 1) + (i + 1)));
}
}
}
int main(int argc, char **argv) {
Var x, y, z;
Func f, g;
Param<int> k;
k.set(3);
f(x, y, z) = x*y+z*k+1;
g(x, y, z) = f(x, y, z) + 2;
f.parallel(x);
f.parallel(y);
g.parallel(z);
f.compute_at(g, z);
auto target = get_jit_target_from_environment();
if (target.features_any_of({Target::HVX_64, Target::HVX_128})) {
g.hexagon().vectorize(x, 32);
f.vectorize(x, 32);
}
Buffer<int> im = g.realize(64, 64, 64);
for (int x = 0; x < 64; x++) {
for (int y = 0; y < 64; y++) {
for (int z = 0; z < 64; z++) {
if (im(x, y, z) != x*y+z*3+3) {
printf("im(%d, %d, %d) = %d\n", x, y, z, im(x, y, z));
return -1;
}
}
}
}
printf("Success!\n");
return 0;
}
int main(int argc, char **argv) {
Var x;
Func f;
Param<int> k;
k.set(3);
f(x) = x*k;
f.parallel(x);
Buffer<int> im = f.realize(16);
for (int i = 0; i < 16; i++) {
if (im(i) != i*3) {
printf("im(%d) = %d\n", i, im(i));
return -1;
}
}
printf("Success!\n");
return 0;
}
static int run(int argc, char **argv) {
static const MeCab::Option long_options[] = {
{ "dicdir", 'd', ".", "DIR", "set DIR as dicdir(default \".\" )" },
{ "outdir", 'o', ".", "DIR", "set DIR as output dir" },
{ "model", 'm', 0, "FILE", "use FILE as model file" },
{ "version", 'v', 0, 0, "show the version and exit" },
{ "training-algorithm", 'a', "crf", "(crf|hmm)",
"set training algorithm" },
{ "default-emission-cost", 'E', "4000", "INT",
"set default emission cost for HMM" },
{ "default-transition-cost", 'T', "4000", "INT",
"set default transition cost for HMM" },
{ "help", 'h', 0, 0, "show this help and exit." },
{ 0, 0, 0, 0 }
};
Param param;
if (!param.open(argc, argv, long_options)) {
std::cout << param.what() << "\n\n" << COPYRIGHT
<< "\ntry '--help' for more information." << std::endl;
return -1;
}
if (!param.help_version()) return 0;
ContextID cid;
DecoderFeatureIndex fi;
DictionaryRewriter rewrite;
const std::string dicdir = param.get<std::string>("dicdir");
const std::string outdir = param.get<std::string>("outdir");
const std::string model = param.get<std::string>("model");
#define DCONF(file) create_filename(dicdir, std::string(file)).c_str()
#define OCONF(file) create_filename(outdir, std::string(file)).c_str()
CHECK_DIE(param.load(DCONF(DICRC)))
<< "no such file or directory: " << DCONF(DICRC);
std::string charset;
{
Dictionary dic;
CHECK_DIE(dic.open(DCONF(SYS_DIC_FILE), "r"));
charset = dic.charset();
CHECK_DIE(!charset.empty());
}
int default_emission_cost = 0;
int default_transition_cost = 0;
std::string type = param.get<std::string>("training-algorithm");
toLower(&type);
if (type == "hmm") {
default_emission_cost =
param.get<int>("default-emission-cost");
default_transition_cost =
param.get<int>("default-transition-cost");
CHECK_DIE(default_transition_cost > 0)
<< "default transition cost must be > 0";
CHECK_DIE(default_emission_cost > 0)
<< "default transition cost must be > 0";
param.set("identity-template", 1);
}
CharProperty property;
CHECK_DIE(property.open(param));
property.set_charset(charset.c_str());
const std::string bos = param.get<std::string>("bos-feature");
const int factor = param.get<int>("cost-factor");
std::vector<std::string> dic;
enum_csv_dictionaries(dicdir.c_str(), &dic);
{
CHECK_DIE(dicdir != outdir) <<
"output directory = dictionary directory! "
"Please specify different directory.";
CHECK_DIE(!outdir.empty()) << "output directory is empty";
CHECK_DIE(!model.empty()) << "model file is empty";
CHECK_DIE(fi.open(param)) << fi.what();
CHECK_DIE(factor > 0) << "cost factor needs to be positive value";
CHECK_DIE(!bos.empty()) << "bos-feature is empty";
CHECK_DIE(dic.size()) << "no dictionary is found in " << dicdir;
CHECK_DIE(rewrite.open(DCONF(REWRITE_FILE)));
}
gencid_bos(bos, &rewrite, &cid);
gencid(DCONF(UNK_DEF_FILE), &rewrite, &cid);
for (std::vector<std::string>::const_iterator it = dic.begin();
it != dic.end();
++it) {
gencid(it->c_str(), &rewrite, &cid);
}
std::cout << "emitting "
<< OCONF(LEFT_ID_FILE) << "/ "
<< OCONF(RIGHT_ID_FILE) << std::endl;
cid.build();
cid.save(OCONF(LEFT_ID_FILE), OCONF(RIGHT_ID_FILE));
gendic(DCONF(UNK_DEF_FILE), OCONF(UNK_DEF_FILE), property,
&rewrite, cid, &fi, true, factor, default_emission_cost);
for (std::vector<std::string>::const_iterator it = dic.begin();
it != dic.end();
++it) {
std::string file = *it;
remove_pathname(&file);
gendic(it->c_str(), OCONF(file.c_str()), property,
&rewrite, cid, &fi, false, factor, default_emission_cost);
}
genmatrix(OCONF(MATRIX_DEF_FILE), cid, &fi,
factor, default_transition_cost);
copy(DCONF(CHAR_PROPERTY_DEF_FILE), OCONF(CHAR_PROPERTY_DEF_FILE));
copy(DCONF(REWRITE_FILE), OCONF(REWRITE_FILE));
copy(DCONF(DICRC), OCONF(DICRC));
if (type == "crf")
copy(DCONF(FEATURE_FILE), OCONF(FEATURE_FILE));
#undef OCONF
#undef DCONF
std::cout << "\ndone!\n";
return 0;
}
int tuple_specialize_rdom_predicate_rfactor_test(bool compile_module) {
Func f("f"), g("g");
Var x("x"), y("y"), z("z");
f(x, y, z) = Tuple(x + y + z, x - y + z);
f.compute_root();
RDom r(5, 20, 5, 20, 5, 20);
r.where(r.x*r.x + r.z*r.z <= 200);
r.where(r.y*r.z + r.z*r.z > 100);
Func ref("ref");
ref(x, y) = Tuple(1, 3);
ref(x, y) = Tuple(ref(x, y)[0]*f(r.x, r.y, r.z)[0], ref(x, y)[1] + 2*f(r.x, r.y, r.z)[1]);
Realization ref_rn = ref.realize(10, 10);
g(x, y) = Tuple(1, 3);
g(x, y) = Tuple(g(x, y)[0]*f(r.x, r.y, r.z)[0], g(x, y)[1] + 2*f(r.x, r.y, r.z)[1]);
Param<int> p;
Param<bool> q;
Var u("u"), v("v"), w("w");
Func intm1 = g.update(0).specialize(p >= 5).rfactor({{r.y, v}, {r.z, w}});
intm1.update(0).parallel(v, 2);
RVar rxi("rxi"), rxo("rxo");
intm1.update(0).split(r.x, rxo, rxi, 2);
Var t("t");
Func intm2 = intm1.update(0).specialize(q).rfactor(rxi, t);
Func intm3 = intm1.update(0).specialize(!q).rfactor(rxo, t);
Func intm4 = g.update(0).rfactor({{r.x, u}, {r.z, w}});
intm4.update(0).vectorize(u);
if (compile_module) {
// Check the call graphs.
Module m = g.compile_to_module({g.infer_arguments()});
CheckCalls checker;
m.functions().front().body.accept(&checker);
CallGraphs expected = {
{g.name(), {}},
{g.update(0).name(), {intm1.name() + ".0", intm1.name() + ".1",
intm4.name() + ".0", intm4.name() + ".1",
g.name() + ".0", g.name() + ".1"}},
{intm1.name(), {}},
{intm1.update(0).name(), {intm2.name() + ".0", intm2.name() + ".1",
intm3.name() + ".0", intm3.name() + ".1",
intm1.name() + ".0", intm1.name() + ".1"}},
{intm2.name(), {}},
{intm2.update(0).name(), {f.name() + ".0", f.name() + ".1",
intm2.name() + ".0", intm2.name() + ".1"}},
{intm3.name(), {}},
{intm3.update(0).name(), {f.name() + ".0", f.name() + ".1",
intm3.name() + ".0", intm3.name() + ".1"}},
{intm4.name(), {}},
{intm4.update(0).name(), {f.name() + ".0", f.name() + ".1",
intm4.name() + ".0", intm4.name() + ".1"}},
{f.name(), {}},
};
if (check_call_graphs(checker.calls, expected) != 0) {
return -1;
}
} else {
{
p.set(10);
q.set(true);
Realization rn = g.realize(10, 10);
Image<int> im1(rn[0]);
Image<int> im2(rn[1]);
Image<int> ref_im1(ref_rn[0]);
Image<int> ref_im2(ref_rn[1]);
auto func1 = [&ref_im1](int x, int y, int z) {
return ref_im1(x, y, z);
};
if (check_image(im1, func1)) {
return -1;
}
auto func2 = [&ref_im2](int x, int y, int z) {
return ref_im2(x, y, z);
};
if (check_image(im2, func2)) {
return -1;
}
}
{
p.set(10);
q.set(false);
Realization rn = g.realize(10, 10);
Image<int> im1(rn[0]);
Image<int> im2(rn[1]);
Image<int> ref_im1(ref_rn[0]);
Image<int> ref_im2(ref_rn[1]);
auto func1 = [&ref_im1](int x, int y, int z) {
return ref_im1(x, y, z);
};
if (check_image(im1, func1)) {
return -1;
}
auto func2 = [&ref_im2](int x, int y, int z) {
return ref_im2(x, y, z);
};
if (check_image(im2, func2)) {
return -1;
}
}
{
p.set(0);
q.set(true);
Realization rn = g.realize(10, 10);
Image<int> im1(rn[0]);
Image<int> im2(rn[1]);
Image<int> ref_im1(ref_rn[0]);
Image<int> ref_im2(ref_rn[1]);
auto func1 = [&ref_im1](int x, int y, int z) {
return ref_im1(x, y, z);
};
if (check_image(im1, func1)) {
return -1;
}
auto func2 = [&ref_im2](int x, int y, int z) {
return ref_im2(x, y, z);
};
if (check_image(im2, func2)) {
return -1;
}
}
{
p.set(0);
q.set(false);
Realization rn = g.realize(10, 10);
Image<int> im1(rn[0]);
Image<int> im2(rn[1]);
Image<int> ref_im1(ref_rn[0]);
Image<int> ref_im2(ref_rn[1]);
auto func1 = [&ref_im1](int x, int y, int z) {
return ref_im1(x, y, z);
};
if (check_image(im1, func1)) {
return -1;
}
auto func2 = [&ref_im2](int x, int y, int z) {
return ref_im2(x, y, z);
};
if (check_image(im2, func2)) {
return -1;
}
}
}
return 0;
}
int intermediate_computed_if_param_test(int index) {
buffer_index = index;
Func f("f_" + std::to_string(index)), g("g_" + std::to_string(index));
Var x("x"), y("y");
Param<int> p;
g(x, y) = x + y;
f(x, y) = x + y;
RDom r(0, 100, 0, 100);
r.where(p > 3);
f(r.x, r.y) += 2*g(r.x, r.y);
// Expect g to be only computed over x=[0,99] and y=[0,99] if param is bigger
// than 3.
g.compute_root();
f.set_custom_trace(&box_bound_trace);
g.trace_stores();
g.trace_realizations();
{
printf("....Set p to 5, expect g to be computed\n");
p.set(5);
run_tracer = false;
niters_expected = 100*100;
niters = 0;
Image<int> im = f.realize(200, 200);
for (int y = 0; y < im.height(); y++) {
for (int x = 0; x < im.width(); x++) {
int correct = x + y;
if ((0 <= x && x <= 99) && (0 <= y && y <= 99)) {
correct = 3*correct;
}
if (im(x, y) != correct) {
printf("im(%d, %d) = %d instead of %d\n",
x, y, im(x, y), correct);
return -1;
}
}
}
if (niters_expected != niters) {
printf("intermediate_computed_if_param_test : Expect niters on g to be %d but got %d instead\n",
niters_expected, niters);
return -1;
}
}
{
printf("....Set p to 0, expect g to be not computed\n");
p.set(0);
run_tracer = false;
niters_expected = 0;
niters = 0;
Image<int> im = f.realize(200, 200);
for (int y = 0; y < im.height(); y++) {
for (int x = 0; x < im.width(); x++) {
int correct = x + y;
if (im(x, y) != correct) {
printf("im(%d, %d) = %d instead of %d\n",
x, y, im(x, y), correct);
return -1;
}
}
}
if (niters_expected != niters) {
printf("intermediate_computed_if_param_test : Expect niters on g to be %d but got %d instead\n",
niters_expected, niters);
return -1;
}
}
return 0;
}
int update_defined_after_wrap_test() {
Func source("source"), g("g");
Var x("x"), y("y");
source(x, y) = x + y;
ImageParam img(Int(32), 2, "img");
Buffer<int> buf = source.realize(200, 200);
img.set(buf);
g(x, y) = img(x, y);
Func wrapper = img.in(g);
// Update of 'g' is defined after img.in(g) is called. g's updates should
// still call img's wrapper.
RDom r(0, 100, 0, 100);
r.where(r.x < r.y);
g(r.x, r.y) += 2*img(r.x, r.y);
Param<bool> param;
Var xi("xi");
RVar rxo("rxo"), rxi("rxi");
g.specialize(param).vectorize(x, 8).unroll(x, 2).split(x, x, xi, 4).parallel(x);
g.update(0).split(r.x, rxo, rxi, 2).unroll(rxi);
Func img_f = img;
img_f.compute_root();
wrapper.compute_root().vectorize(_0, 8).unroll(_0, 2).split(_0, _0, xi, 4).parallel(_0);
{
param.set(true);
// Check the call graphs.
// Expect initialization of 'g' to call 'wrapper' and its update to call
// 'wrapper' and 'g', wrapper' to call 'img_f', 'img_f' to call 'img'
Module m = g.compile_to_module({g.infer_arguments()});
CheckCalls c;
m.functions().front().body.accept(&c);
CallGraphs expected = {
{g.name(), {wrapper.name(), g.name()}},
{wrapper.name(), {img_f.name()}},
{img_f.name(), {img.name()}},
};
if (check_call_graphs(c.calls, expected) != 0) {
return -1;
}
Buffer<int> im = g.realize(200, 200);
auto func = [](int x, int y) {
return ((0 <= x && x <= 99) && (0 <= y && y <= 99) && (x < y)) ? 3*(x + y) : (x + y);
};
if (check_image(im, func)) {
return -1;
}
}
{
param.set(false);
// Check the call graphs.
// Expect initialization of 'g' to call 'wrapper' and its update to call
// 'wrapper' and 'g', wrapper' to call 'img_f', 'img_f' to call 'img'
Module m = g.compile_to_module({g.infer_arguments()});
CheckCalls c;
m.functions().front().body.accept(&c);
CallGraphs expected = {
{g.name(), {wrapper.name(), g.name()}},
{wrapper.name(), {img_f.name()}},
{img_f.name(), {img.name()}},
};
if (check_call_graphs(c.calls, expected) != 0) {
return -1;
}
Buffer<int> im = g.realize(200, 200);
auto func = [](int x, int y) {
return ((0 <= x && x <= 99) && (0 <= y && y <= 99) && (x < y)) ? 3*(x + y) : (x + y);
};
if (check_image(im, func)) {
return -1;
}
}
return 0;
}
int main(int argc, char **argv) {
{
call_count = 0;
Func count_calls;
count_calls.define_extern("count_calls",
std::vector<ExternFuncArgument>(),
UInt(8), 2);
Func f;
f() = count_calls(0, 0);
f.compute_root().memoize();
Image<uint8_t> result1 = f.realize();
Image<uint8_t> result2 = f.realize();
assert(result1(0) == 42);
assert(result2(0) == 42);
assert(call_count == 1);
}
{
call_count = 0;
Param<int32_t> coord;
Func count_calls;
count_calls.define_extern("count_calls",
std::vector<ExternFuncArgument>(),
UInt(8), 2);
Func f, g;
Var x, y;
f() = count_calls(coord, coord);
f.compute_root().memoize();
g(x, y) = f();
coord.set(0);
Image<uint8_t> out1 = g.realize(256, 256);
Image<uint8_t> out2 = g.realize(256, 256);
for (int32_t i = 0; i < 256; i++) {
for (int32_t j = 0; j < 256; j++) {
assert(out1(i, j) == 42);
assert(out2(i, j) == 42);
}
}
assert(call_count == 1);
coord.set(1);
Image<uint8_t> out3 = g.realize(256, 256);
Image<uint8_t> out4 = g.realize(256, 256);
for (int32_t i = 0; i < 256; i++) {
for (int32_t j = 0; j < 256; j++) {
assert(out3(i, j) == 42);
assert(out4(i, j) == 42);
}
}
assert(call_count == 2);
}
{
call_count = 0;
Func count_calls;
count_calls.define_extern("count_calls",
std::vector<ExternFuncArgument>(),
UInt(8), 2);
Func f;
Var x, y;
f(x, y) = count_calls(x, y) + count_calls(x, y);
count_calls.compute_root().memoize();
Image<uint8_t> out1 = f.realize(256, 256);
Image<uint8_t> out2 = f.realize(256, 256);
for (int32_t i = 0; i < 256; i++) {
for (int32_t j = 0; j < 256; j++) {
assert(out1(i, j) == (42 + 42));
assert(out2(i, j) == (42 + 42));
}
}
assert(call_count == 1);
}
call_count = 0;
{
Func count_calls_23;
count_calls_23.define_extern("count_calls_with_arg",
Internal::vec(ExternFuncArgument(cast<uint8_t>(23))),
UInt(8), 2);
Func count_calls_42;
count_calls_42.define_extern("count_calls_with_arg",
Internal::vec(ExternFuncArgument(cast<uint8_t>(42))),
UInt(8), 2);
Func f;
Var x, y;
f(x, y) = count_calls_23(x, y) + count_calls_42(x, y);
count_calls_23.compute_root().memoize();
count_calls_42.compute_root().memoize();
Image<uint8_t> out1 = f.realize(256, 256);
Image<uint8_t> out2 = f.realize(256, 256);
for (int32_t i = 0; i < 256; i++) {
for (int32_t j = 0; j < 256; j++) {
assert(out1(i, j) == (23 + 42));
assert(out2(i, j) == (23 + 42));
}
}
assert(call_count_with_arg == 2);
}
{
Param<uint8_t> val1;
Param<uint8_t> val2;
call_count_with_arg = 0;
Func count_calls_val1;
count_calls_val1.define_extern("count_calls_with_arg",
Internal::vec(ExternFuncArgument(Expr(val1))),
UInt(8), 2);
Func count_calls_val2;
count_calls_val2.define_extern("count_calls_with_arg",
Internal::vec(ExternFuncArgument(Expr(val2))),
UInt(8), 2);
Func f;
Var x, y;
f(x, y) = count_calls_val1(x, y) + count_calls_val2(x, y);
count_calls_val1.compute_root().memoize();
count_calls_val2.compute_root().memoize();
val1.set(23);
val2.set(42);
Image<uint8_t> out1 = f.realize(256, 256);
Image<uint8_t> out2 = f.realize(256, 256);
val1.set(42);
Image<uint8_t> out3 = f.realize(256, 256);
val1.set(23);
Image<uint8_t> out4 = f.realize(256, 256);
val1.set(42);
Image<uint8_t> out5 = f.realize(256, 256);
val2.set(57);
Image<uint8_t> out6 = f.realize(256, 256);
for (int32_t i = 0; i < 256; i++) {
for (int32_t j = 0; j < 256; j++) {
assert(out1(i, j) == (23 + 42));
assert(out2(i, j) == (23 + 42));
assert(out3(i, j) == (42 + 42));
assert(out4(i, j) == (23 + 42));
assert(out5(i, j) == (42 + 42));
assert(out6(i, j) == (42 + 57));
}
}
assert(call_count_with_arg == 4);
}
{
Param<float> val;
call_count_with_arg = 0;
Func count_calls;
count_calls.define_extern("count_calls_with_arg",
Internal::vec(ExternFuncArgument(cast<uint8_t>(val))),
UInt(8), 2);
Func f;
Var x, y;
f(x, y) = count_calls(x, y) + count_calls(x, y);
count_calls.compute_root().memoize();
val.set(23.0f);
Image<uint8_t> out1 = f.realize(256, 256);
val.set(23.4f);
Image<uint8_t> out2 = f.realize(256, 256);
for (int32_t i = 0; i < 256; i++) {
for (int32_t j = 0; j < 256; j++) {
assert(out1(i, j) == (23 + 23));
assert(out2(i, j) == (23 + 23));
}
}
assert(call_count_with_arg == 2);
}
{
Param<float> val;
call_count_with_arg = 0;
Func count_calls;
count_calls.define_extern("count_calls_with_arg",
Internal::vec(ExternFuncArgument(memoize_tag(cast<uint8_t>(val)))),
UInt(8), 2);
Func f;
Var x, y;
f(x, y) = count_calls(x, y) + count_calls(x, y);
count_calls.compute_root().memoize();
val.set(23.0f);
Image<uint8_t> out1 = f.realize(256, 256);
val.set(23.4f);
Image<uint8_t> out2 = f.realize(256, 256);
for (int32_t i = 0; i < 256; i++) {
for (int32_t j = 0; j < 256; j++) {
assert(out1(i, j) == (23 + 23));
assert(out2(i, j) == (23 + 23));
}
}
assert(call_count_with_arg == 1);
}
{
// Case with bounds computed not equal to bounds realized.
Param<float> val;
Param<int32_t> index;
call_count_with_arg = 0;
Func count_calls;
count_calls.define_extern("count_calls_with_arg",
Internal::vec(ExternFuncArgument(cast<uint8_t>(val))),
UInt(8), 2);
Func f, g, h;
Var x;
f(x) = count_calls(x, 0) + cast<uint8_t>(x);
g(x) = f(x);
h(x) = g(4) + g(index);
f.compute_root().memoize();
g.vectorize(x, 8).compute_at(h, x);
val.set(23.0f);
index.set(2);
Image<uint8_t> out1 = h.realize(1);
assert(out1(0) == (uint8_t)(2 * 23 + 4 + 2));
assert(call_count_with_arg == 3);
index.set(4);
out1 = h.realize(1);
assert(out1(0) == (uint8_t)(2 * 23 + 4 + 4));
assert(call_count_with_arg == 4);
}
{
// Test Tuple case
Param<float> val;
call_count_with_arg = 0;
Func count_calls;
count_calls.define_extern("count_calls_with_arg",
Internal::vec(ExternFuncArgument(cast<uint8_t>(val))),
UInt(8), 2);
Func f;
Var x, y, xi, yi;
f(x, y) = Tuple(count_calls(x, y) + cast<uint8_t>(x), x);
count_calls.compute_root().memoize();
f.compute_root().memoize();
Func g;
g(x, y) = Tuple(f(x, y)[0] + f(x - 1, y)[0] + f(x + 1, y)[0], f(x, y)[1]);
val.set(23.0f);
Realization out = g.realize(128, 128);
Image<uint8_t> out0 = out[0];
Image<int32_t> out1 = out[1];
for (int32_t i = 0; i < 100; i++) {
for (int32_t j = 0; j < 100; j++) {
assert(out0(i, j) == (uint8_t)(3 * 23 + i + (i - 1) + (i + 1)));
assert(out1(i, j) == i);
}
}
out = g.realize(128, 128);
out0 = out[0];
out1 = out[1];
for (int32_t i = 0; i < 100; i++) {
for (int32_t j = 0; j < 100; j++) {
assert(out0(i, j) == (uint8_t)(3 * 23 + i + (i - 1) + (i + 1)));
assert(out1(i, j) == i);
}
}
assert(call_count_with_arg == 1);
}
{
// Test cache eviction
Param<float> val;
call_count_with_arg = 0;
Func count_calls;
count_calls.define_extern("count_calls_with_arg",
Internal::vec(ExternFuncArgument(cast<uint8_t>(val))),
UInt(8), 2);
Func f;
Var x, y, xi, yi;
f(x, y) = count_calls(x, y) + cast<uint8_t>(x);
count_calls.compute_root().memoize();
Func g;
g(x, y) = f(x, y) + f(x - 1, y) + f(x + 1, y);
Internal::JITSharedRuntime::memoization_cache_set_size(1000000);
for (int v = 0; v < 1000; v++) {
int r = rand() % 256;
val.set((float)r);
Image<uint8_t> out1 = g.realize(128, 128);
for (int32_t i = 0; i < 100; i++) {
for (int32_t j = 0; j < 100; j++) {
assert(out1(i, j) == (uint8_t)(3 * r + i + (i - 1) + (i + 1)));
}
}
}
// TODO work out an assertion on call count here.
fprintf(stderr, "Call count is %d.\n", call_count_with_arg);
// Return cache size to default.
Internal::JITSharedRuntime::memoization_cache_set_size(0);
}
{
// Test flushing entire cache with a single element larger than the cache
Param<float> val;
call_count_with_arg = 0;
Func count_calls;
count_calls.define_extern("count_calls_with_arg",
Internal::vec(ExternFuncArgument(cast<uint8_t>(val))),
UInt(8), 2);
Func f;
Var x, y, xi, yi;
f(x, y) = count_calls(x, y) + cast<uint8_t>(x);
count_calls.compute_root().memoize();
Func g;
g(x, y) = f(x, y) + f(x - 1, y) + f(x + 1, y);
Internal::JITSharedRuntime::memoization_cache_set_size(1000000);
for (int v = 0; v < 1000; v++) {
int r = rand() % 256;
val.set((float)r);
Image<uint8_t> out1 = g.realize(128, 128);
for (int32_t i = 0; i < 100; i++) {
for (int32_t j = 0; j < 100; j++) {
assert(out1(i, j) == (uint8_t)(3 * r + i + (i - 1) + (i + 1)));
}
}
}
// TODO work out an assertion on call count here.
fprintf(stderr, "Call count before oversize realize is %d.\n", call_count_with_arg);
call_count_with_arg = 0;
Image<uint8_t> big = g.realize(1024, 1024);
Image<uint8_t> big2 = g.realize(1024, 1024);
// TODO work out an assertion on call count here.
fprintf(stderr, "Call count after oversize realize is %d.\n", call_count_with_arg);
call_count_with_arg = 0;
for (int v = 0; v < 1000; v++) {
int r = rand() % 256;
val.set((float)r);
Image<uint8_t> out1 = g.realize(128, 128);
for (int32_t i = 0; i < 100; i++) {
for (int32_t j = 0; j < 100; j++) {
assert(out1(i, j) == (uint8_t)(3 * r + i + (i - 1) + (i + 1)));
}
}
}
fprintf(stderr, "Call count is %d.\n", call_count_with_arg);
// Return cache size to default.
Internal::JITSharedRuntime::memoization_cache_set_size(0);
}
{
// Test parallel cache access
Param<float> val;
Func count_calls;
count_calls.define_extern("count_calls_with_arg_parallel",
Internal::vec(ExternFuncArgument(cast<uint8_t>(val))),
UInt(8), 3);
Func f;
Var x, y;
// Ensure that all calls map to the same cache key, but pass a thread ID
// through to avoid having to do locking or an atomic add
f(x, y) = count_calls(x, y % 4, memoize_tag(y / 16, 0)) + cast<uint8_t>(x);
Func g;
g(x, y) = f(x, y) + f(x - 1, y) + f(x + 1, y);
count_calls.compute_at(f, y).memoize();
f.compute_at(g, y).memoize();
g.parallel(y, 16);
val.set(23.0f);
Internal::JITSharedRuntime::memoization_cache_set_size(1000000);
Image<uint8_t> out = g.realize(128, 128);
for (int32_t i = 0; i < 128; i++) {
for (int32_t j = 0; j < 128; j++) {
assert(out(i, j) == (uint8_t)(3 * 23 + i + (i - 1) + (i + 1)));
}
}
// TODO work out an assertion on call counts here.
for (int i = 0; i < 8; i++) {
fprintf(stderr, "Call count for thread %d is %d.\n", i, call_count_with_arg_parallel[i]);
}
// Return cache size to default.
Internal::JITSharedRuntime::memoization_cache_set_size(0);
}
{
Param<float> val;
Func f;
Var x, y;
f(x, y) = cast<uint8_t>((x << 8) + y);
Func prev_func = f;
Func stage[4];
for (int i = 0; i < 4; i++) {
std::vector<ExternFuncArgument> args(3);
args[0] = cast<int32_t>(i);
args[1] = cast<int32_t>(val);
args[2] = prev_func;
stage[i].define_extern("count_calls_staged",
args,
UInt(8), 2);
prev_func = stage[i];
}
f.compute_root();
for (int i = 0; i < 3; i++) {
stage[i].compute_root();
}
stage[3].compute_root().memoize();
val.set(23.0f);
Image<uint8_t> result = stage[3].realize(128, 128);
for (int32_t i = 0; i < 128; i++) {
for (int32_t j = 0; j < 128; j++) {
assert(result(i, j) == (uint8_t)((i << 8) + j + 4 * 23));
}
}
for (int i = 0; i < 4; i++) {
fprintf(stderr, "Call count for stage %d is %d.\n", i, call_count_staged[i]);
}
result = stage[3].realize(128, 128);
for (int32_t i = 0; i < 128; i++) {
for (int32_t j = 0; j < 128; j++) {
assert(result(i, j) == (uint8_t)((i << 8) + j + 4 * 23));
}
}
for (int i = 0; i < 4; i++) {
fprintf(stderr, "Call count for stage %d is %d.\n", i, call_count_staged[i]);
}
}
fprintf(stderr, "Success!\n");
return 0;
}
int main(int argc, char **argv) {
// Test bool
check_range<bool, uint8_t>(0, 2, 0, 1,
0, 2, 0, 1,
0, 256, 0, 1,
"<bool, uint8_t> exhaustive");
// Exhaustively test 8-bit cases
check_range<uint8_t, uint8_t>(0, 256, 0, 1,
0, 256, 0, 1,
0, 256, 0, 1,
"<uint8_t, uint8_t> exhaustive");
check_range<int8_t, uint8_t>(0, 256, -128, 1,
0, 256, -128, 1,
0, 256, 0, 1,
"<int8_t, uint8_t> exhaustive");
check_range<uint8_t, float>(0, 256, 0, 1,
0, 256, 0, 1,
0, 256, 0, 1/255.0f,
"<uint8_t, float> exhaustive");
check_range<int8_t, float>(0, 256, -128, 1,
0, 256, -128, 1,
0, 256, 0, 1/255.0f,
"<int8_t, float> exhaustive");
// Check all delta values for 16-bit, verify swapping arguments doesn't break
check_range<uint16_t, uint16_t>(0, 65536, 0, 1,
65535, 1, 0, 1,
0, 257, 255, 1,
"<uint16_t, uint16_t> all zero starts");
check_range<uint16_t, uint16_t>(65535, 1, 0, 1,
0, 65536, 0, 1,
0, 257, 255, 1,
"<uint16_t, uint16_t> all one starts");
// Verify different bit sizes for value and weight types
check_range<uint16_t, uint8_t>(0, 1, 0, 1,
65535, 1, 0, 1,
0, 255, 1, 1,
"<uint16_t, uint8_t> zero, one uint8_t weight test");
check_range<uint16_t, uint32_t>(0, 1, 0, 1,
65535, 1, 0, 1,
std::numeric_limits<int32_t>::min(), 257, 255 * 65535, 1,
"<uint16_t, uint8_t> zero, one uint32_t weight test");
check_range<uint32_t, uint8_t>(0, 1, 0, 1,
1 << 31, 1, 0, 1,
0, 255, 0, 1,
"<uint32_t, uint8_t> weight test");
check_range<uint32_t, uint16_t>(0, 1, 0, 1,
1 << 31, 1, 0, 1,
0, 65535, 0, 1,
"<uint32_t, uint16_t> weight test");
// Verify float weights with integer values
check_range<uint16_t, float>(0, 1, 0, 1,
65535, 1, 0, 1,
0, 257, 0, 255.0f/65535.0f,
"<uint16_t, float> zero, one float weight test");
check_range<int16_t, uint16_t>(0, 65536, -32768, 1,
0, 1, 0, 1,
0, 257, 0, 255,
"<int16_t, uint16_t> all zero starts");
#if 0 // takes too long, difficult to test with uint32_t
// Check all delta values for 32-bit, do it in signed arithmetic
check_range<int32_t, uint32_t>(std::numeric_limits<int32_t>::min(), std::numeric_limits<int32_t>::max(), 0, 1,
1 << 31, 1, 0, 1,
0, 1, 1 << 31, 1,
"<uint32_t, uint32_t> all zero starts");
#endif
check_range<float, float>(0, 100, 0, .01,
0, 100, 0, .01,
0, 100, 0, .01,
"<float, float> float values 0 to 1 by 1/100ths");
check_range<float, float>(0, 100, -5, .1,
0, 100, 0, .1,
0, 100, 0, .1,
"<float, float> float values -5 to 5 by 1/100ths");
// Verify float values with integer weights
check_range<float, uint8_t>(0, 100, -5, .1,
0, 100, 0, .1,
0, 255, 0, 1,
"<float, uint8_t> float values -5 to 5 by 1/100ths");
check_range<float, uint16_t>(0, 100, -5, .1,
0, 100, 0, .1,
0, 255, 0, 257,
"<float, uint16_t> float values -5 to 5 by 1/100ths");
check_range<float, uint32_t>(0, 100, -5, .1,
0, 100, 0, .1,
std::numeric_limits<int32_t>::min(), 257, 255 * 65535, 1,
"<float, uint32_t> float values -5 to 5 by 1/100ths");
// Check constant and constant case:
Func lerp_constants("lerp_constants");
lerp_constants() = lerp(0, cast<uint32_t>(1023), .5f);
Image<uint32_t> result = lerp_constants.realize();
uint32_t expected = evaluate<uint32_t>(cast<uint32_t>(lerp(0, cast<uint16_t>(1023), .5f)));
if (result(0) != expected)
std::cerr << "Expected " << expected << " got " << result(0) << std::endl;
assert(result(0) == expected);
// Add a little more coverage for uint32_t as this was failing
// without being detected for a long time.
Image<uint8_t> input_a_img(16, 16);
Image<uint8_t> input_b_img(16, 16);
for (int i = 0; i < 16; i ++) {
for (int j = 0; j < 16; j ++) {
input_a_img(i, j) = (i << 4) + j;
input_b_img(i, j) = ((15 - i) << 4) + (15 - j);
}
}
ImageParam input_a(UInt(8), 2);
ImageParam input_b(UInt(8), 2);
Var x, y;
Func lerp_with_casts;
Param<float> w;
lerp_with_casts(x, y) = lerp(cast<int32_t>(input_a(x, y)), cast<int32_t>(input_b(x, y)), w);
lerp_with_casts.vectorize(x, 4);
input_a.set(input_a_img);
input_b.set(input_b_img);
w.set(0.0f);
Image<int32_t> result_should_be_a = lerp_with_casts.realize(16, 16);
w.set(1.0f);
Image<int32_t> result_should_be_b = lerp_with_casts.realize(16, 16);
for (int i = 0; i < 16; i ++) {
for (int j = 0; j < 16; j ++) {
assert(input_a_img(i, j) == result_should_be_a(i, j));
assert(input_b_img(i, j) == result_should_be_b(i, j));
}
}
std::cout << "Success!" << std::endl;
}
int gpu_intermediate_computed_if_param_test(int index) {
buffer_index = index;
Func f("f_" + std::to_string(index)), g("g_" + std::to_string(index)), h("h_" + std::to_string(index));
Var x("x"), y("y");
Param<int> p;
g(x, y) = x + y;
h(x, y) = 10;
f(x, y) = x + y;
RDom r1(0, 100, 0, 100);
r1.where(p > 3);
f(r1.x, r1.y) += 2*g(r1.x, r1.y);
RDom r2(0, 100, 0, 100);
r2.where(p <= 3);
f(r2.x, r2.y) += h(r2.x, r2.y) + g(r2.x, r2.y);
f.update(0).specialize(p >= 2).gpu_tile(r1.x, r1.y, 4, 4);
g.compute_root();
h.compute_root();
h.gpu_tile(x, y, 8, 8);
{
printf("....Set p to 5, expect g to be computed\n");
p.set(5);
run_tracer = false;
niters_expected = 100*100;
niters = 0;
Image<int> im = f.realize(200, 200);
for (int y = 0; y < im.height(); y++) {
for (int x = 0; x < im.width(); x++) {
int correct = x + y;
if ((0 <= x && x <= 99) && (0 <= y && y <= 99)) {
correct = 3*correct;
}
if (im(x, y) != correct) {
printf("im(%d, %d) = %d instead of %d\n",
x, y, im(x, y), correct);
return -1;
}
}
}
}
{
printf("....Set p to 0, expect g to be not computed\n");
p.set(0);
run_tracer = false;
niters_expected = 0;
niters = 0;
Image<int> im = f.realize(200, 200);
for (int y = 0; y < im.height(); y++) {
for (int x = 0; x < im.width(); x++) {
int correct = x + y;
if ((0 <= x && x <= 99) && (0 <= y && y <= 99)) {
correct += 10 + correct;
}
if (im(x, y) != correct) {
printf("im(%d, %d) = %d instead of %d\n",
x, y, im(x, y), correct);
return -1;
}
}
}
}
return 0;
}
int main(int argc, char **argv) {
{
Param<bool> param;
Func f;
Var x;
f(x) = select(param, x*3, x*17);
// Vectorize when the output is large enough
Expr cond = (f.output_buffer().width() >= 4);
f.specialize(cond).vectorize(x, 4);
// This has created a specialization of f that is
// vectorized. Now we want to further specialize both the
// default case and the special case based on param. We can
// retrieve a reference to the specialization using the same
// condition again:
f.specialize(cond).specialize(param);
// Now specialize the narrow case on param as well
f.specialize(param);
f.set_custom_trace(&my_trace);
f.trace_stores();
Image<int> out(100);
// Just check that all the specialization didn't change the output.
param.set(true);
reset_trace();
f.realize(out);
for (int i = 0; i < out.width(); i++) {
int correct = i*3;
if (out(i) != correct) {
printf("out(%d) was %d instead of %d\n",
i, out(i), correct);
}
}
param.set(false);
f.realize(out);
for (int i = 0; i < out.width(); i++) {
int correct = i*17;
if (out(i) != correct) {
printf("out(%d) was %d instead of %d\n",
i, out(i), correct);
}
}
// Should have used vector stores
if (!vector_store || scalar_store) {
printf("This was supposed to use vector stores\n");
return -1;
}
// Now try a smaller input
out = Image<int>(3);
param.set(true);
reset_trace();
f.realize(out);
for (int i = 0; i < out.width(); i++) {
int correct = i*3;
if (out(i) != correct) {
printf("out(%d) was %d instead of %d\n",
i, out(i), correct);
}
}
param.set(false);
f.realize(out);
for (int i = 0; i < out.width(); i++) {
int correct = i*17;
if (out(i) != correct) {
printf("out(%d) was %d instead of %d\n",
i, out(i), correct);
}
}
// Should have used scalar stores
if (vector_store || !scalar_store) {
printf("This was supposed to use scalar stores\n");
return -1;
}
}
{
Func f1, f2, g1, g2;
Var x;
// Define pipeline A
f1(x) = x + 7;
g1(x) = f1(x) + f1(x + 1);
// Define pipeline B
f2(x) = x * 34;
g2(x) = f2(x) + f2(x - 1);
// Switch between them based on a boolean param
Param<bool> param;
Func out;
out(x) = select(param, g1(x), g2(x));
// These will be outside the condition that specializes out,
// but skip stages will nuke their allocation and computation
// for us.
f1.compute_root();
g1.compute_root();
f2.compute_root();
out.specialize(param);
// Count allocations.
out.set_custom_allocator(&my_malloc, &my_free);
reset_alloc_counts();
param.set(true);
out.realize(100);
if (empty_allocs != 1 || nonempty_allocs != 2 || frees != 3) {
printf("There were supposed to be 1 empty alloc, 2 nonempty allocs, and 3 frees.\n"
"Instead we got %d empty allocs, %d nonempty allocs, and %d frees.\n",
empty_allocs, nonempty_allocs, frees);
return -1;
}
reset_alloc_counts();
param.set(false);
out.realize(100);
if (empty_allocs != 2 || nonempty_allocs != 1 || frees != 3) {
printf("There were supposed to be 2 empty allocs, 1 nonempty alloc, and 3 frees.\n"
"Instead we got %d empty allocs, %d nonempty allocs, and %d frees.\n",
empty_allocs, nonempty_allocs, frees);
return -1;
}
}
{
// Specialize for interleaved vs planar inputs
ImageParam im(Float(32), 1);
im.set_stride(0, Expr()); // unconstrain the stride
Func f;
Var x;
f(x) = im(x);
// If we have a stride of 1 it's worth vectorizing, but only if the width is also > 8.
f.specialize(im.stride(0) == 1 && im.width() >= 8).vectorize(x, 8);
f.trace_stores();
f.set_custom_trace(&my_trace);
// Check bounds inference is still cool with widths < 8
f.infer_input_bounds(5);
int m = im.get().min(0), e = im.get().extent(0);
if (m != 0 || e != 5) {
printf("min, extent = %d, %d instead of 0, 5\n", m, e);
return -1;
}
// Check we don't crash with the small input, and that it uses scalar stores
reset_trace();
f.realize(5);
if (!scalar_store || vector_store) {
printf("These stores were supposed to be scalar.\n");
return -1;
}
// Check we don't crash with a larger input, and that it uses vector stores
Image<float> image(100);
im.set(image);
reset_trace();
f.realize(100);
if (scalar_store || !vector_store) {
printf("These stores were supposed to be vector.\n");
return -1;
}
}
{
// Bounds required of the input change depending on the param
ImageParam im(Float(32), 1);
Param<bool> param;
Func f;
Var x;
f(x) = select(param, im(x + 10), im(x - 10));
f.specialize(param);
param.set(true);
f.infer_input_bounds(100);
int m = im.get().min(0);
if (m != 10) {
printf("min %d instead of 10\n", m);
return -1;
}
param.set(false);
im.set(Buffer());
f.infer_input_bounds(100);
m = im.get().min(0);
if (m != -10) {
printf("min %d instead of -10\n", m);
return -1;
}
}
{
// Specialize an update definition
Func f;
Var x;
Param<int> start, size;
RDom r(start, size);
f(x) = x;
f(r) = 10 - r;
// Special-case for when we only update one element of f
f.update().specialize(size == 1);
// Also special-case updating no elements of f
f.update().specialize(size == 0);
start.set(0);
size.set(1);
// Not crashing is enough
f.realize(100);
}
{
// What happens to bounds inference if an input is not used at
// all for a given specialization?
ImageParam im(Float(32), 1);
Param<bool> param;
Func f;
Var x;
f(x) = select(param, im(x), 0.0f);
f.specialize(param);
param.set(false);
Image<float> image(10);
im.set(image);
// The image is too small, but that should be OK, because the
// param is false so the image will never be used.
f.realize(100);
}
{
// Specialization inherits the scheduling directives done so far:
ImageParam im(Int(32), 2);
Func f;
Var x, y;
f(x, y) = im(x, y);
Expr cond = f.output_buffer().width() >= 4;
// Unroll y by two innermost.
f.reorder(y, x).unroll(y, 2).reorder(x, y);
// Vectorize if the output is at least 4-wide. Inherits the
// unrolling already done.
f.specialize(cond).vectorize(x, 4);
// Confirm that the unrolling applies to both cases using bounds inference:
f.infer_input_bounds(3, 1);
if (im.get().extent(0) != 3) {
printf("extent(0) was supposed to be 3.\n");
return -1;
}
if (im.get().extent(1) != 2) {
// Height is 2, because the unrolling also happens in the
// specialized case.
printf("extent(1) was supposed to be 2.\n");
return -1;
}
}
{
// Check we don't need to specialize intermediate stages.
ImageParam im(Int(32), 1);
Func f, g, h, out;
Var x;
f(x) = im(x);
g(x) = f(x);
h(x) = g(x);
out(x) = h(x);
Expr w = out.output_buffer().extent(0);
out.output_buffer().set_min(0, 0);
f.compute_root().specialize(w >= 4).vectorize(x, 4);
g.compute_root().vectorize(x, 4);
h.compute_root().vectorize(x, 4);
out.specialize(w >= 4).vectorize(x, 4);
Image<int> input(3), output(3);
// Shouldn't throw a bounds error:
im.set(input);
out.realize(output);
}
{
// Check specializations of stages nested in other stages simplify appropriately.
ImageParam im(Int(32), 2);
Param<bool> cond1, cond2;
Func f, out;
Var x, y;
f(x, y) = im(x, y);
out(x, y) = f(x, y);
f.compute_at(out, x).specialize(cond1 && cond2).vectorize(x, 4);
out.compute_root().specialize(cond1 && cond2).vectorize(x, 4);
if_then_else_count = 0;
CountIfThenElse pass1;
for (auto ff : out.compile_to_module(out.infer_arguments()).functions()) {
pass1.mutate(ff.body);
}
Image<int> input(3, 3), output(3, 3);
// Shouldn't throw a bounds error:
im.set(input);
out.realize(output);
if (if_then_else_count != 1) {
printf("Expected 1 IfThenElse stmts. Found %d.\n", if_then_else_count);
return -1;
}
}
{
// Check specializations of stages nested in other stages simplify appropriately.
ImageParam im(Int(32), 2);
Param<bool> cond1, cond2;
Func f, out;
Var x, y;
f(x, y) = im(x, y);
out(x, y) = f(x, y);
f.compute_at(out, x).specialize(cond1).vectorize(x, 4);
out.compute_root().specialize(cond1 && cond2).vectorize(x, 4);
if_then_else_count = 0;
CountIfThenElse pass2;
for (auto ff : out.compile_to_module(out.infer_arguments()).functions()) {
pass2.mutate(ff.body);
}
Image<int> input(3, 3), output(3, 3);
// Shouldn't throw a bounds error:
im.set(input);
out.realize(output);
// There should have been 2 Ifs total: They are the
// outer cond1 && cond2, and the condition in the true case
// should have been simplified away. The If in the false
// branch cannot be simplified.
if (if_then_else_count != 2) {
printf("Expected 2 IfThenElse stmts. Found %d.\n", if_then_else_count);
return -1;
}
}
printf("Success!\n");
return 0;
}
int main(int argc, char **argv) {
if (get_jit_target_from_environment().has_feature(Target::Profile)) {
// The profiler adds lots of extra prints, so counting the
// number of prints is not useful.
printf("Skipping test because profiler is active\n");
return 0;
}
Var x;
{
Func f;
f(x) = print(x * x, "the answer is", 42.0f, "unsigned", cast<uint32_t>(145));
f.set_custom_print(halide_print);
Buffer<int32_t> result = f.realize(10);
for (int32_t i = 0; i < 10; i++) {
if (result(i) != i * i) {
return -1;
}
}
assert(messages.size() == 10);
for (size_t i = 0; i < messages.size(); i++) {
long square;
float forty_two;
unsigned long one_forty_five;
int scan_count = sscanf(messages[i].c_str(), "%ld the answer is %f unsigned %lu",
&square, &forty_two, &one_forty_five);
assert(scan_count == 3);
assert(square == static_cast<long long>(i * i));
assert(forty_two == 42.0f);
assert(one_forty_five == 145);
}
}
messages.clear();
{
Func f;
Param<int> param;
param.set(127);
// Test a string containing a printf format specifier (It should print it as-is).
f(x) = print_when(x == 3, x * x, "g", 42.0f, "%s", param);
f.set_custom_print(halide_print);
Buffer<int32_t> result = f.realize(10);
for (int32_t i = 0; i < 10; i++) {
if (result(i) != i * i) {
return -1;
}
}
assert(messages.size() == 1);
long nine;
float forty_two;
long p;
int scan_count = sscanf(messages[0].c_str(), "%ld g %f %%s %ld",
&nine, &forty_two, &p);
assert(scan_count == 3);
assert(nine == 9);
assert(forty_two == 42.0f);
assert(p == 127);
}
messages.clear();
{
Func f;
// Test a single message longer than 8K.
std::vector<Expr> args;
for (int i = 0; i < 500; i++) {
uint64_t n = i;
n *= n;
n *= n;
n *= n;
n *= n;
n += 100;
int32_t hi = n >> 32;
int32_t lo = n & 0xffffffff;
args.push_back((cast<uint64_t>(hi) << 32) | lo);
Expr dn = cast<double>((float)(n));
args.push_back(dn);
}
f(x) = print(args);
f.set_custom_print(halide_print);
Buffer<uint64_t> result = f.realize(1);
if (result(0) != 100) {
return -1;
}
assert(messages.back().size() == 8191);
}
messages.clear();
// Check that Halide's stringification of floats and doubles
// matches %f and %e respectively.
#ifndef _WIN32
// msvc's library has different ideas about how %f and %e should come out.
{
Func f, g;
const int N = 1000000;
Expr e = reinterpret(Float(32), random_uint());
// Make sure we cover some special values.
e = select(x == 0, 0.0f,
x == 1, -0.0f,
x == 2, std::numeric_limits<float>::infinity(),
x == 3, -std::numeric_limits<float>::infinity(),
x == 4, std::numeric_limits<float>::quiet_NaN(),
x == 5, -std::numeric_limits<float>::quiet_NaN(),
e);
e = select(x == 5, std::numeric_limits<float>::denorm_min(),
x == 6, -std::numeric_limits<float>::denorm_min(),
x == 7, std::numeric_limits<float>::min(),
x == 8, -std::numeric_limits<float>::min(),
x == 9, std::numeric_limits<float>::max(),
x == 10, -std::numeric_limits<float>::max(),
x == 11, 1.0f - 1.0f / (1 << 22),
e);
f(x) = print(e);
f.set_custom_print(halide_print);
Buffer<float> imf = f.realize(N);
assert(messages.size() == (size_t)N);
char correct[1024];
for (int i = 0; i < N; i++) {
snprintf(correct, sizeof(correct), "%f\n", imf(i));
// OS X prints -nan as nan
#ifdef __APPLE__
if (messages[i] == "-nan\n") messages[i] = "nan\n";
#endif
if (messages[i] != correct) {
printf("float %d: %s vs %s for %10.20e\n", i, messages[i].c_str(), correct, imf(i));
return -1;
}
}
messages.clear();
g(x) = print(reinterpret(Float(64), (cast<uint64_t>(random_uint()) << 32) | random_uint()));
g.set_custom_print(halide_print);
Buffer<double> img = g.realize(N);
assert(messages.size() == (size_t)N);
for (int i = 0; i < N; i++) {
snprintf(correct, sizeof(correct), "%e\n", img(i));
#ifdef __APPLE__
if (messages[i] == "-nan\n") messages[i] = "nan\n";
#endif
if (messages[i] != correct) {
printf("double %d: %s vs %s for %10.20e\n", i, messages[i].c_str(), correct, img(i));
return -1;
}
}
}
#endif
printf("Success!\n");
return 0;
}
int main(int argc, char **argv) {
Func f1, f2, f3, f4, f5;
Func g1, g2, g3, g4, g5;
Var x, xi;
Expr v = x*1.34f + 1.0142f;
Param<float> p;
p.set(1.0f);
// Test accuracy of reciprocals.
// First prevent any optimizations by hiding 1.0 in a param.
f1(x) = p / v;
// Now test various vectorization widths with an explicit 1.0. On
// arm 2 and 4 trigger optimizations. On x86 4 and 8 do.
f2(x) = fast_inverse(v);
f2.vectorize(x, 2);
f3(x) = fast_inverse(v);
f3.vectorize(x, 4);
f4(x) = fast_inverse(v);
f4.vectorize(x, 8);
// Same thing for reciprocal square root.
g1(x) = p / sqrt(v);
g2(x) = fast_inverse_sqrt(v);
g2.vectorize(x, 2);
g3(x) = fast_inverse_sqrt(v);
g3.vectorize(x, 4);
g4(x) = fast_inverse_sqrt(v);
g4.vectorize(x, 8);
// Also test both on the GPU.
f5(x) = fast_inverse(v);
g5(x) = fast_inverse_sqrt(v);
Target t = get_jit_target_from_environment();
if (t.has_gpu_feature()) {
f5.gpu_tile(x, xi, 16);
g5.gpu_tile(x, xi, 16);
}
Buffer<float> imf1 = f1.realize(10000);
Buffer<float> imf2 = f2.realize(10000);
Buffer<float> imf3 = f3.realize(10000);
Buffer<float> imf4 = f4.realize(10000);
Buffer<float> imf5 = f5.realize(10000);
Buffer<float> img1 = g1.realize(10000);
Buffer<float> img2 = g2.realize(10000);
Buffer<float> img3 = g3.realize(10000);
Buffer<float> img4 = g4.realize(10000);
Buffer<float> img5 = g5.realize(10000);
printf("Testing accuracy of inverse\n");
check(imf1, imf2);
check(imf1, imf3);
check(imf1, imf4);
check(imf1, imf5);
printf("Pass.\n");
printf("Testing accuracy of inverse sqrt\n");
check(img1, img2);
check(img1, img3);
check(img1, img4);
check(img1, img5);
printf("Pass.\n");
printf("Success!\n");
return 0;
}
int main(int argc, char **argv) {
Func f1, f2, f3, f4;
Func g1, g2, g3, g4;
Var x;
Expr v = x*1.34f + 1.0142f;
Param<float> p;
p.set(1.0f);
// Test accuracy of reciprocals.
// First prevent any optimizations by hiding 1.0 in a param.
f1(x) = p / v;
// Now test various vectorization widths with an explicit 1.0. On
// arm 2 and 4 trigger optimizations. On x86 4 and 8 do.
f2(x) = 1.0f / v;
f2.vectorize(x, 2);
f3(x) = 1.0f / v;
f3.vectorize(x, 4);
f4(x) = 1.0f / v;
f4.vectorize(x, 8);
// Same thing for reciprocal square root.
g1(x) = p / sqrt(v);
g2(x) = 1.0f / sqrt(v);
g2.vectorize(x, 2);
g3(x) = 1.0f / sqrt(v);
g3.vectorize(x, 4);
g4(x) = 1.0f / sqrt(v);
g4.vectorize(x, 8);
Image<float> imf1 = f1.realize(10000);
Image<float> imf2 = f2.realize(10000);
Image<float> imf3 = f3.realize(10000);
Image<float> imf4 = f4.realize(10000);
Image<float> img1 = g1.realize(10000);
Image<float> img2 = g2.realize(10000);
Image<float> img3 = g3.realize(10000);
Image<float> img4 = g4.realize(10000);
printf("Testing accuracy of inverse\n");
check(imf1, imf2);
check(imf1, imf3);
check(imf1, imf4);
printf("Pass.\n");
printf("Testing accuracy of inverse sqrt\n");
check(img1, img2);
check(img1, img3);
check(img1, img4);
printf("Pass.\n");
printf("Success!\n");
return 0;
}
int main(int argc, char **argv) {
Target target = get_jit_target_from_environment();
if (target.has_feature(Target::Profile)) {
// The profiler adds lots of extra prints, so counting the
// number of prints is not useful.
printf("Skipping test because profiler is active\n");
return 0;
}
if (target.has_feature(Target::Debug)) {
// Same thing here: the runtime debug adds lots of extra prints,
// so counting the number of prints is not useful.
printf("Skipping test because runtime debug is active\n");
return 0;
}
Var x;
{
Func f;
f(x) = print(x * x, "the answer is", 42.0f, "unsigned", cast<uint32_t>(145));
f.set_custom_print(halide_print);
Buffer<int32_t> result = f.realize(10);
for (int32_t i = 0; i < 10; i++) {
if (result(i) != i * i) {
return -1;
}
}
assert(messages.size() == 10);
for (size_t i = 0; i < messages.size(); i++) {
long square;
float forty_two;
unsigned long one_forty_five;
int scan_count = sscanf(messages[i].c_str(), "%ld the answer is %f unsigned %lu",
&square, &forty_two, &one_forty_five);
assert(scan_count == 3);
assert(square == static_cast<long long>(i * i));
assert(forty_two == 42.0f);
assert(one_forty_five == 145);
}
}
messages.clear();
{
Func f;
Param<int> param;
param.set(127);
// Test a string containing a printf format specifier (It should print it as-is).
f(x) = print_when(x == 3, x * x, "g", 42.0f, "%s", param);
f.set_custom_print(halide_print);
Buffer<int32_t> result = f.realize(10);
for (int32_t i = 0; i < 10; i++) {
if (result(i) != i * i) {
return -1;
}
}
assert(messages.size() == 1);
long nine;
float forty_two;
long p;
int scan_count = sscanf(messages[0].c_str(), "%ld g %f %%s %ld",
&nine, &forty_two, &p);
assert(scan_count == 3);
assert(nine == 9);
assert(forty_two == 42.0f);
assert(p == 127);
}
messages.clear();
{
Func f;
// Test a single message longer than 8K.
std::vector<Expr> args;
for (int i = 0; i < 500; i++) {
uint64_t n = i;
n *= n;
n *= n;
n *= n;
n *= n;
n += 100;
int32_t hi = n >> 32;
int32_t lo = n & 0xffffffff;
args.push_back((cast<uint64_t>(hi) << 32) | lo);
Expr dn = cast<double>((float)(n));
args.push_back(dn);
}
f(x) = print(args);
f.set_custom_print(halide_print);
Buffer<uint64_t> result = f.realize(1);
if (result(0) != 100) {
return -1;
}
assert(messages.back().size() == 8191);
}
messages.clear();
// Check that Halide's stringification of floats and doubles
// matches %f and %e respectively.
#ifndef _WIN32
// msvc's library has different ideas about how %f and %e should come out.
{
Func f, g;
const int N = 1000000;
Expr e = reinterpret(Float(32), random_uint());
// Make sure we cover some special values.
e = select(x == 0, 0.0f,
x == 1, -0.0f,
x == 2, std::numeric_limits<float>::infinity(),
x == 3, -std::numeric_limits<float>::infinity(),
x == 4, std::numeric_limits<float>::quiet_NaN(),
x == 5, -std::numeric_limits<float>::quiet_NaN(),
e);
e = select(x == 5, std::numeric_limits<float>::denorm_min(),
x == 6, -std::numeric_limits<float>::denorm_min(),
x == 7, std::numeric_limits<float>::min(),
x == 8, -std::numeric_limits<float>::min(),
x == 9, std::numeric_limits<float>::max(),
x == 10, -std::numeric_limits<float>::max(),
x == 11, 1.0f - 1.0f / (1 << 22),
e);
f(x) = print(e);
f.set_custom_print(halide_print);
Buffer<float> imf = f.realize(N);
assert(messages.size() == (size_t)N);
char correct[1024];
for (int i = 0; i < N; i++) {
snprintf(correct, sizeof(correct), "%f\n", imf(i));
// Some versions of the std library can emit some NaN patterns
// as "-nan", due to sloppy conversion (or not) of the sign bit.
// Halide considers all NaN's equivalent, so paper over this
// noise in the test by normalizing all -nan -> nan.
if (messages[i] == "-nan\n") messages[i] = "nan\n";
if (!strcmp(correct, "-nan\n")) strcpy(correct, "nan\n");
if (messages[i] != correct) {
printf("float %d: %s vs %s for %10.20e\n", i, messages[i].c_str(), correct, imf(i));
return -1;
}
}
messages.clear();
g(x) = print(reinterpret(Float(64), (cast<uint64_t>(random_uint()) << 32) | random_uint()));
g.set_custom_print(halide_print);
Buffer<double> img = g.realize(N);
assert(messages.size() == (size_t)N);
for (int i = 0; i < N; i++) {
snprintf(correct, sizeof(correct), "%e\n", img(i));
// Some versions of the std library can emit some NaN patterns
// as "-nan", due to sloppy conversion (or not) of the sign bit.
// Halide considers all NaN's equivalent, so paper over this
// noise in the test by normalizing all -nan -> nan.
if (messages[i] == "-nan\n") messages[i] = "nan\n";
if (!strcmp(correct, "-nan\n")) strcpy(correct, "nan\n");
if (messages[i] != correct) {
printf("double %d: %s vs %s for %10.20e\n", i, messages[i].c_str(), correct, img(i));
return -1;
}
}
}
#endif
messages.clear();
{
Func f;
// Test a vectorized print.
f(x) = print(x * 3);
f.set_custom_print(halide_print);
f.vectorize(x, 32);
if (target.features_any_of({Target::HVX_64, Target::HVX_128})) {
f.hexagon();
}
Buffer<int> result = f.realize(128);
if (!target.features_any_of({Target::HVX_64, Target::HVX_128})) {
assert((int)messages.size() == result.width());
for (size_t i = 0; i < messages.size(); i++) {
assert(messages[i] == std::to_string(i * 3) + "\n");
}
} else {
// The Hexagon simulator prints directly to stderr, so we
// can't read the messages.
}
}
messages.clear();
{
Func f;
// Test a vectorized print_when.
f(x) = print_when(x % 2 == 0, x * 3);
f.set_custom_print(halide_print);
f.vectorize(x, 32);
if (target.features_any_of({Target::HVX_64, Target::HVX_128})) {
f.hexagon();
}
Buffer<int> result = f.realize(128);
if (!target.features_any_of({Target::HVX_64, Target::HVX_128})) {
assert((int)messages.size() == result.width() / 2);
for (size_t i = 0; i < messages.size(); i++) {
assert(messages[i] == std::to_string(i * 2 * 3) + "\n");
}
} else {
// The Hexagon simulator prints directly to stderr, so we
// can't read the messages.
}
}
printf("Success!\n");
return 0;
}
int main(int argc, char **argv) {
Func f;
Var x;
f(x) = x;
f.compute_root();
// Halide will partition a loop into three pieces in a few
// situations. The pieces are 1) a messy prologue, 2) a clean
// steady state, and 3) a messy epilogue. One way to trigger this
// is if you use a boundary condition helper:
{
Func g = BoundaryConditions::repeat_edge(f, 0, 100);
count_partitions(g, 3);
}
// If you vectorize or otherwise split, then the last vector
// (which gets shifted leftwards) is its own partition. This
// removes some clamping logic from the inner loop.
{
Func g;
g(x) = f(x);
g.vectorize(x, 8);
count_partitions(g, 2);
}
// The slicing applies to every loop level starting from the
// outermost one, but only recursively simplifies the clean steady
// state. It either splits things three (start, middle, end). So
// adding a boundary condition to a 2D computation will produce 5
// code paths for the top, bottom, left, right, and center of the
// image.
{
Var y;
Func g;
g(x, y) = x + y;
g.compute_root();
Func h = BoundaryConditions::mirror_image(g, 0, 10, 0, 10);
count_partitions(h, 5);
}
// If you split and also have a boundary condition, or have
// multiple boundary conditions at play (e.g. because you're
// blurring an inlined Func that uses a boundary condition), then
// there are still only three partitions. The steady state is the
// slice of the loop where *all* of the boundary conditions and
// splitting logic simplify away.
{
Func g = BoundaryConditions::mirror_interior(f, 0, 10);
Func h;
Param<int> t1, t2;
h(x) = g(x-1) + g(x+1);
h.vectorize(x, 8);
count_partitions(h, 3);
}
// You can manually control the splitting behavior using the
// 'likely' intrinsic. When used on one side of a select, min,
// max, or clamp, it tags the select, min, max, or clamp as likely
// to simplify to that expression in the steady state case, and
// tries to solve for loop variable values for which this is true.
{
// So this code should produce a prologue that evaluates to sin(x), and
// a steady state that evaluates to 1:
Func g;
g(x) = select(x < 10, sin(x), likely(1.0f));
// There should be two partitions
count_partitions(g, 2);
// But only one should call sin
count_sin_calls(g, 1);
}
{
// This code should produce a prologue and epilogue that
// evaluate sin(x), and a steady state that evaluates to 1:
Func g;
g(x) = select(x < 10 || x > 100, sin(x), likely(1.0f));
// There should be three partitions
count_partitions(g, 3);
// With calls to sin in the prologue and epilogue.
count_sin_calls(g, 2);
}
// As a specialize case, we treat clamped ramps as likely to
// simplify to the clamped expression. This handles the many
// existing cases where people have written their boundary
// condition manually using clamp.
{
Func g;
g(x) = f(clamp(x, 0, 10)); // treated as clamp(likely(x), 0, 10)
g.vectorize(x, 8);
count_partitions(g, 3);
}
// Using the likely intrinsic pulls some IR relating to the
// condition outside of the loop. We'd better check that this
// respects lets and doesn't do any combinatorial expansion. We'll
// do this with a nasty comparison:
{
Func g;
Var y;
// Have an inner reduction loop that the comparisons depend on
// to make things harder.
RDom r(0, 5);
const int N = 25;
// Make some nasty expressions to compare to.
Expr e[N];
e[0] = y;
for (int i = 1; i < N; i++) {
e[i] = e[i-1] * e[i-1] + y + r;
}
// Make a nasty condition that uses all of these.
Expr nasty = cast<bool>(1);
for (int i = 0; i < N; i++) {
nasty = nasty && (x*(i+1) < e[i]);
}
// Have an innermost loop over c to complicate things further.
Var c;
g(c, x, y) = sum(select(nasty, likely(10), c + r));
// Check that it doesn't take the age of the world to compile,
// and that it produces the right number of partitions.
count_partitions(g, 3);
}
// Make sure partitions that occur outside of the actual bounds
// don't mess things up.
{
Func g;
Var x;
Param<int> limit;
g(x) = select(x > limit, likely(3), 2);
// If either of these realize calls iterates from 0 to limit,
// and then from limit to 10, we'll have a nice segfault.
limit.set(10000000);
Buffer<int> result = g.realize(10);
limit.set(-10000000);
result = g.realize(10);
}
// The performance of this behavior is tested in
// test/performance/boundary_conditions.cpp
printf("Success!\n");
return 0;
}