OSSForks/pytorch
1
0
Fork 0
mirror of https://github.com/zebrajr/pytorch.git synced 2025-12-06 12:20:52 +01:00
Code Issues Actions 38 Packages Projects Releases Wiki Activity

Revert D25445815: [te] Add fast log approximation based on sleef

Browse Source 
Test Plan: revert-hammer

Differential Revision:
D25445815 (1329066b69)

Original commit changeset: 20696eacd12a

fbshipit-source-id: 38830a6abd16260d60e5dd9a5594e65736a9c782
This commit is contained in:
Edward Yang 2020-12-17 14:53:17 -08:00 committed by Facebook GitHub Bot
parent 6db5e85726
commit ea4ccc730e
6 changed files with 2 additions and 249 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

145
benchmarks/cpp/tensorexpr/bench_approx.cpp
View File

@ -1,145 +0,0 @@
#include <benchmark/benchmark.h>
#include <torch/csrc/jit/tensorexpr/ir_simplifier.h>
#include <torch/csrc/jit/tensorexpr/loopnest.h>
#include <torch/csrc/jit/tensorexpr/tensor.h>
#include <torch/csrc/jit/tensorexpr/llvm_codegen.h>
#include <torch/torch.h>
using namespace torch::jit::tensorexpr;
static void log_sleef(benchmark::State& state) {
KernelScope ks;
auto N = VarHandle("N", kInt);
Placeholder A("A", kFloat, {N});
torch::jit::tensorexpr::Tensor* B =
Compute("B", {N}, [&](const VarHandle& i) {
return log(A.load(i));
});
LoopNest ln({B});
ln.prepareForCodegen();
ln.vectorizeInnerLoops();
Stmt* s = ln.root_stmt();
s = torch::jit::tensorexpr::IRSimplifier::simplify(s);
std::vector<CodeGen::BufferArg> args;
args.emplace_back(B);
args.emplace_back(A);
args.emplace_back(N);
LLVMCodeGen cg(s, args);
at::Tensor A_t = torch::abs(torch::randn({state.range(0)}));
at::Tensor B_t = torch::randn({state.range(0)});
auto B_ref = at::log(A_t);
cg.call({B_t.data_ptr<float>(), A_t.data_ptr<float>(), state.range(0)});
assert(at::allclose(B_t, B_ref));
for (auto _ : state) {
cg.call({B_t.data_ptr<float>(), A_t.data_ptr<float>(), state.range(0)});
}
state.counters["log/s"] = benchmark::Counter(
uint64_t(state.range(0) * state.iterations()), benchmark::Counter::kIsRate);
}
static void log_fast(benchmark::State& state) {
KernelScope ks;
auto N = VarHandle("N", kInt);
Placeholder A("A", kFloat, {N});
torch::jit::tensorexpr::Tensor* B =
Compute("B", {N}, [&](const VarHandle& i) {
return fast_log(A.load(i));
});
LoopNest ln({B});
ln.prepareForCodegen();
ln.vectorizeInnerLoops();
Stmt* s = ln.root_stmt();
s = torch::jit::tensorexpr::IRSimplifier::simplify(s);
std::vector<CodeGen::BufferArg> args;
args.emplace_back(B);
args.emplace_back(A);
args.emplace_back(N);
LLVMCodeGen cg(s, args);
at::Tensor A_t = torch::abs(torch::randn({state.range(0)}));
at::Tensor B_t = torch::randn({state.range(0)});
auto B_ref = at::log(A_t);
cg.call({B_t.data_ptr<float>(), A_t.data_ptr<float>(), state.range(0)});
assert(at::allclose(B_t, B_ref));
for (auto _ : state) {
cg.call({B_t.data_ptr<float>(), A_t.data_ptr<float>(), state.range(0)});
}
state.counters["log/s"] = benchmark::Counter(
uint64_t(state.range(0) * state.iterations()), benchmark::Counter::kIsRate);
}
static void log_aten(benchmark::State& state) {
at::Tensor A_t = torch::abs(torch::randn({state.range(0)}));
at::Tensor B_t = torch::randn({state.range(0)});
for (auto _ : state) {
at::native::log_out(B_t, A_t);
}
state.counters["log/s"] = benchmark::Counter(
uint64_t(state.range(0) * state.iterations()), benchmark::Counter::kIsRate);
}
static void logit_fast(benchmark::State& state) {
KernelScope ks;
auto N = VarHandle("N", kInt);
Placeholder A("A", kFloat, {N});
torch::jit::tensorexpr::Tensor* B =
Compute("B", {N}, [&](const VarHandle& i) {
auto A_elem = A.load(i);
return fast_log(A_elem / (FloatImm::make(1.0f) - A_elem));
});
LoopNest ln({B});
ln.prepareForCodegen();
ln.vectorizeInnerLoops();
Stmt* s = ln.root_stmt();
s = torch::jit::tensorexpr::IRSimplifier::simplify(s);
std::vector<CodeGen::BufferArg> args;
args.emplace_back(B);
args.emplace_back(A);
args.emplace_back(N);
LLVMCodeGen cg(s, args);
at::Tensor A_t = torch::abs(torch::randn({state.range(0)}));
at::Tensor B_t = torch::randn({state.range(0)});
auto B_ref = at::logit(A_t);
cg.call({B_t.data_ptr<float>(), A_t.data_ptr<float>(), state.range(0)});
assert(at::allclose(B_t, B_ref));
for (auto _ : state) {
cg.call({B_t.data_ptr<float>(), A_t.data_ptr<float>(), state.range(0)});
}
state.counters["logit/s"] = benchmark::Counter(
uint64_t(state.range(0) * state.iterations()), benchmark::Counter::kIsRate);
}
static void logit_aten(benchmark::State& state) {
at::Tensor A_t = torch::abs(torch::randn({state.range(0)}));
at::Tensor B_t = torch::randn({state.range(0)});
for (auto _ : state) {
at::native::logit_out(B_t, A_t);
}
state.counters["logit/s"] = benchmark::Counter(
uint64_t(state.range(0) * state.iterations()), benchmark::Counter::kIsRate);
}
BENCHMARK(log_sleef)
->Args({2<<5})
->Args({2<<8})
->Args({2<<12})
->Args({2<<14});
BENCHMARK(log_fast)
->Args({2<<5})
->Args({2<<8})
->Args({2<<12})
->Args({2<<14});
BENCHMARK(log_aten)
->Args({2<<5})
->Args({2<<8})
->Args({2<<12})
->Args({2<<14});
BENCHMARK(logit_fast)
->Args({2<<5})
->Args({2<<8})
->Args({2<<12})
->Args({2<<14});
BENCHMARK(logit_aten)
->Args({2<<5})
->Args({2<<8})
->Args({2<<12})
->Args({2<<14});

32
test/cpp/tensorexpr/test_aten.cpp
View File

@ -733,38 +733,6 @@ TEST(ATen, logFloat) {
}
}
TEST(ATen, fastLogFloat) {
KernelScope kernel_scope;
const int kTotalSize = 128 * 128;
Placeholder a_buf(BufHandle("A", {ExprHandle(kTotalSize)}, kFloat));
Placeholder b_buf(BufHandle("B", {ExprHandle(kTotalSize)}, kFloat));
VarHandle index = VarHandle("index", kInt);
ExprHandle load_a = a_buf.load(index);
Stmt* store_b = b_buf.store({index}, fast_log(load_a));
Stmt* stmt = For::make(index, 0, kTotalSize, store_b);
PaddedBuffer<float> a_v(kTotalSize);
PaddedBuffer<float> b_v(kTotalSize);
for (int i = 0; i < kTotalSize; ++i) {
a_v(i) = at::randn({1}).item().to<float>();
}
SimpleIREvaluator ir_eval(stmt, a_buf, b_buf);
ir_eval(a_v, b_v);
for (int i = 0; i < kTotalSize; ++i) {
auto test = b_v(i);
auto ref = std::log(a_v(i));
if (std::isnan(ref)) {
ASSERT_EQ(std::isnan(test), true);
} else {
ASSERT_FLOAT_EQ(test, ref);
}
}
}
TEST(ATen, log10Float) {
KernelScope kernel_scope;
const int kTotalSize = 128;

32
test/cpp/tensorexpr/test_llvm.cpp
View File

@ -217,38 +217,6 @@ TEST(LLVM, BitCast) {
}
}
TEST(LLVM, fastLogFloat) {
KernelScope kernel_scope;
const int kTotalSize = 128 * 128;
Placeholder a_buf(BufHandle("A", {ExprHandle(kTotalSize)}, kFloat));
Placeholder b_buf(BufHandle("B", {ExprHandle(kTotalSize)}, kFloat));
VarHandle index = VarHandle("index", kInt);
ExprHandle load_a = a_buf.load(index);
Stmt* store_b = b_buf.store({index}, fast_log(load_a));
Stmt* stmt = For::make(index, 0, kTotalSize, store_b);
PaddedBuffer<float> a_v(kTotalSize);
PaddedBuffer<float> b_v(kTotalSize);
for (int i = 0; i < kTotalSize; ++i) {
a_v(i) = at::randn({1}).item().to<float>();
}
LLVMCodeGen ir_eval(stmt, {a_buf, b_buf});
ir_eval.call({a_v, b_v});
for (int i = 0; i < kTotalSize; ++i) {
auto test = b_v(i);
auto ref = std::log(a_v(i));
if (std::isnan(ref)) {
ASSERT_EQ(std::isnan(test), true);
} else {
ASSERT_FLOAT_EQ(test, ref);
}
}
}
TEST(LLVM, LetTest01) {
KernelScope kernel_scope;

7
torch/csrc/jit/tensorexpr/eval.h
View File

@ -337,12 +337,9 @@ class SimpleIREvaluator : public CodeGen, public IRVisitor {
std::vector<T> result_v(lhs_v.size());
for (size_t i = 0; i < lhs_v.size(); i++) {
switch (op_type) {
case IRNodeType::kLshift: {
typename std::make_unsigned<T>::type a =
static_cast<typename std::make_unsigned<T>::type>(lhs_v[i]);
result_v[i] = a << rhs_v[i];
case IRNodeType::kLshift:
result_v[i] = lhs_v[i] << rhs_v[i];
break;
}
case IRNodeType::kRshift:
result_v[i] = lhs_v[i] >> rhs_v[i];
break;

34
torch/csrc/jit/tensorexpr/expr.cpp
View File

@ -128,40 +128,6 @@ ExprHandle fabs(const ExprHandle& v) {
return Intrinsics::make(kFabs, v);
}
ExprHandle fast_log(const ExprHandle& v) {
// this implementation is taken from sleef:
// https://github.com/shibatch/sleef/blob/master/src/libm/sleefsp.c#L1131
// to generate coefficients, this tool is provided
// https://github.com/shibatch/sleef/blob/master/src/gencoef/gencoef.txt
auto ilogb2kf = [](ExprHandle x) {
auto y = (bitcast<int32_t>(x) >> IntImm::make(23)) & IntImm::make(0xff);
return y - IntImm::make(0x7f);
};
auto ldexp3kf = [](ExprHandle x, ExprHandle e) {
return bitcast<float>(bitcast<int32_t>(x) + (e << IntImm::make(23)));
};
auto e = ilogb2kf(v * FloatImm::make(1.0 / 0.75));
auto m = ldexp3kf(v, IntImm::make(-1) * e);
auto one = FloatImm::make(1.0f);
auto x = (m - one) / (m + one);
auto x2 = x * x;
auto mlaf = [](ExprHandle x, ExprHandle y, float z) {
return x * y + FloatImm::make(z);
};
auto t = FloatImm::make(0.2392828464508056640625);
t = mlaf(t, x2, 0.28518211841583251953125);
t = mlaf(t, x2, 0.400005877017974853515625);
t = mlaf(t, x2, 0.666666686534881591796875);
t = mlaf(t, x2, 2.0);
x = x * t + FloatImm::make(0.693147180559945286226764) * e;
x = IfThenElse::make(v < FloatImm::make(0), FloatImm::make(std::numeric_limits<float>::quiet_NaN()), x);
x = IfThenElse::make(v == FloatImm::make(0), FloatImm::make(-std::numeric_limits<float>::infinity()), x);
return x;
}
ExprHandle log(const ExprHandle& v) {
return Intrinsics::make(kLog, v);
}

1
torch/csrc/jit/tensorexpr/expr.h
View File

@ -290,7 +290,6 @@ TORCH_API ExprHandle exp(const ExprHandle& v);
TORCH_API ExprHandle expm1(const ExprHandle& v);
TORCH_API ExprHandle fabs(const ExprHandle& v);
TORCH_API ExprHandle log(const ExprHandle& v);
TORCH_API ExprHandle fast_log(const ExprHandle& v);
TORCH_API ExprHandle log2(const ExprHandle& v);
TORCH_API ExprHandle log10(const ExprHandle& v);
TORCH_API ExprHandle log1p(const ExprHandle& v);