Unify naming of EXPMINUS evaluation stubs
Include range reduction specification in names of EXPMINUS evaluation stubs
PiperOrigin-RevId: 343601475
diff --git a/BUILD.bazel b/BUILD.bazel
index e1324a7..33bb87d 100644
--- a/BUILD.bazel
+++ b/BUILD.bazel
@@ -449,9 +449,9 @@
"src/f32-vunary/gen/vsqr-scalar-x1.c",
"src/f32-vunary/gen/vsqr-scalar-x2.c",
"src/f32-vunary/gen/vsqr-scalar-x4.c",
- "src/math/expminus-scalar-lut64-p2.c",
- "src/math/expminus-scalar-lut2048-p1.c",
- "src/math/expminus-scalar-p5.c",
+ "src/math/expminus-scalar-rr2-lut64-p2.c",
+ "src/math/expminus-scalar-rr2-lut2048-p1.c",
+ "src/math/expminus-scalar-rr2-p5.c",
"src/math/roundd-scalar-addsub.c",
"src/math/roundd-scalar-cvt.c",
"src/math/roundd-scalar-floor.c",
@@ -1657,9 +1657,9 @@
"src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
"src/math/exp-neonfma-rr2-lut64-p2.c",
"src/math/exp-neonfma-rr2-p5.c",
- "src/math/expminus-neonfma-lut64-p2.c",
- "src/math/expminus-neonfma-lut2048-p1.c",
- "src/math/expminus-neonfma-p5.c",
+ "src/math/expminus-neonfma-rr2-lut64-p2.c",
+ "src/math/expminus-neonfma-rr2-lut2048-p1.c",
+ "src/math/expminus-neonfma-rr2-p5.c",
"src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
"src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
"src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
@@ -2134,7 +2134,7 @@
"src/f32-vrnd/gen/vrndz-sse2-x8.c",
"src/math/exp-sse2-rr2-lut64-p2.c",
"src/math/exp-sse2-rr2-p5.c",
- "src/math/expminus-sse2-p5.c",
+ "src/math/expminus-sse2-rr2-p5.c",
"src/math/expm1minus-sse2-rr2-p6.c",
"src/math/roundd-sse2-cvt.c",
"src/math/roundne-sse2-cvt.c",
@@ -2681,7 +2681,7 @@
"src/math/exp-avx2-rr2-lut8-p3-perm.c",
"src/math/exp-avx2-rr2-lut8-p4-perm.c",
"src/math/exp-avx2-rr2-p5.c",
- "src/math/expminus-avx2-p5.c",
+ "src/math/expminus-avx2-rr2-p5.c",
"src/math/extexp-avx2-p5.c",
"src/math/sigmoid-avx2-rr1-lut64-p2-gather-div.c",
"src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr1fma.c",
diff --git a/CMakeLists.txt b/CMakeLists.txt
index 93c1c6b..d42e49e 100755
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -572,9 +572,9 @@
src/f32-vunary/gen/vsqr-scalar-x1.c
src/f32-vunary/gen/vsqr-scalar-x2.c
src/f32-vunary/gen/vsqr-scalar-x4.c
- src/math/expminus-scalar-lut64-p2.c
- src/math/expminus-scalar-lut2048-p1.c
- src/math/expminus-scalar-p5.c
+ src/math/expminus-scalar-rr2-lut64-p2.c
+ src/math/expminus-scalar-rr2-lut2048-p1.c
+ src/math/expminus-scalar-rr2-p5.c
src/math/roundd-scalar-addsub.c
src/math/roundd-scalar-cvt.c
src/math/roundd-scalar-floor.c
@@ -1097,9 +1097,9 @@
src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c
src/math/exp-neonfma-rr2-lut64-p2.c
src/math/exp-neonfma-rr2-p5.c
- src/math/expminus-neonfma-lut64-p2.c
- src/math/expminus-neonfma-lut2048-p1.c
- src/math/expminus-neonfma-p5.c
+ src/math/expminus-neonfma-rr2-lut64-p2.c
+ src/math/expminus-neonfma-rr2-lut2048-p1.c
+ src/math/expminus-neonfma-rr2-p5.c
src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c
src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c
src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c
@@ -1568,7 +1568,7 @@
src/f32-vrnd/gen/vrndz-sse2-x8.c
src/math/exp-sse2-rr2-lut64-p2.c
src/math/exp-sse2-rr2-p5.c
- src/math/expminus-sse2-p5.c
+ src/math/expminus-sse2-rr2-p5.c
src/math/expm1minus-sse2-rr2-p6.c
src/math/roundd-sse2-cvt.c
src/math/roundne-sse2-cvt.c
@@ -2109,7 +2109,7 @@
src/math/exp-avx2-rr2-lut8-p3-perm.c
src/math/exp-avx2-rr2-lut8-p4-perm.c
src/math/exp-avx2-rr2-p5.c
- src/math/expminus-avx2-p5.c
+ src/math/expminus-avx2-rr2-p5.c
src/math/extexp-avx2-p5.c
src/math/sigmoid-avx2-rr1-lut64-p2-gather-div.c
src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr1fma.c
diff --git a/eval/f32-expminus.cc b/eval/f32-expminus.cc
index 40255aa..cf9c08b 100644
--- a/eval/f32-expminus.cc
+++ b/eval/f32-expminus.cc
@@ -52,46 +52,46 @@
}
#if XNN_ARCH_X86 || XNN_ARCH_X86_64
- static void f32_expminus__sse2_p5(benchmark::State& state) {
- ExpError(state, xnn_math_f32_expminus__sse2_p5, 4);
+ static void f32_expminus__sse2_rr2_p5(benchmark::State& state) {
+ ExpError(state, xnn_math_f32_expminus__sse2_rr2_p5, 4);
}
- static void f32_expminus__avx2_p5(benchmark::State& state) {
- ExpError(state, xnn_math_f32_expminus__avx2_p5, 8);
+ static void f32_expminus__avx2_rr2_p5(benchmark::State& state) {
+ ExpError(state, xnn_math_f32_expminus__avx2_rr2_p5, 8);
}
- BENCHMARK(f32_expminus__sse2_p5)->Unit(benchmark::kMillisecond)->Iterations(1);
- BENCHMARK(f32_expminus__avx2_p5)->Unit(benchmark::kMillisecond)->Iterations(1);
+ BENCHMARK(f32_expminus__sse2_rr2_p5)->Unit(benchmark::kMillisecond)->Iterations(1);
+ BENCHMARK(f32_expminus__avx2_rr2_p5)->Unit(benchmark::kMillisecond)->Iterations(1);
#endif // XNN_ARCH_X86 || XNN_ARCH_X86_64
#if XNN_ARCH_ARM || XNN_ARCH_ARM64
- static void f32_expminus__neonfma_p5(benchmark::State& state) {
- ExpError(state, xnn_math_f32_expminus__neonfma_p5, 4);
+ static void f32_expminus__neonfma_rr2_p5(benchmark::State& state) {
+ ExpError(state, xnn_math_f32_expminus__neonfma_rr2_p5, 4);
}
- static void f32_expminus__neonfma_lut64_p2(benchmark::State& state) {
- ExpError(state, xnn_math_f32_expminus__neonfma_lut64_p2, 4);
+ static void f32_expminus__neonfma_rr2_lut64_p2(benchmark::State& state) {
+ ExpError(state, xnn_math_f32_expminus__neonfma_rr2_lut64_p2, 4);
}
- static void f32_expminus__neonfma_lut2048_p1(benchmark::State& state) {
- ExpError(state, xnn_math_f32_expminus__neonfma_lut2048_p1, 4);
+ static void f32_expminus__neonfma_rr2_lut2048_p1(benchmark::State& state) {
+ ExpError(state, xnn_math_f32_expminus__neonfma_rr2_lut2048_p1, 4);
}
- BENCHMARK(f32_expminus__neonfma_p5)->Unit(benchmark::kMillisecond)->Iterations(1);
- BENCHMARK(f32_expminus__neonfma_lut64_p2)->Unit(benchmark::kMillisecond)->Iterations(1);
- BENCHMARK(f32_expminus__neonfma_lut2048_p1)->Unit(benchmark::kMillisecond)->Iterations(1);
+ BENCHMARK(f32_expminus__neonfma_rr2_p5)->Unit(benchmark::kMillisecond)->Iterations(1);
+ BENCHMARK(f32_expminus__neonfma_rr2_lut64_p2)->Unit(benchmark::kMillisecond)->Iterations(1);
+ BENCHMARK(f32_expminus__neonfma_rr2_lut2048_p1)->Unit(benchmark::kMillisecond)->Iterations(1);
#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
-static void f32_expminus__scalar_p5(benchmark::State& state) {
- ExpError(state, xnn_math_f32_expminus__scalar_p5, 1);
+static void f32_expminus__scalar_rr2_p5(benchmark::State& state) {
+ ExpError(state, xnn_math_f32_expminus__scalar_rr2_p5, 1);
}
-static void f32_expminus__scalar_lut64_p2(benchmark::State& state) {
- ExpError(state, xnn_math_f32_expminus__scalar_lut64_p2, 1);
+static void f32_expminus__scalar_rr2_lut64_p2(benchmark::State& state) {
+ ExpError(state, xnn_math_f32_expminus__scalar_rr2_lut64_p2, 1);
}
-static void f32_expminus__scalar_lut2048_p1(benchmark::State& state) {
- ExpError(state, xnn_math_f32_expminus__scalar_lut2048_p1, 1);
+static void f32_expminus__scalar_rr2_lut2048_p1(benchmark::State& state) {
+ ExpError(state, xnn_math_f32_expminus__scalar_rr2_lut2048_p1, 1);
}
-BENCHMARK(f32_expminus__scalar_p5)->Unit(benchmark::kMillisecond)->Iterations(1);
-BENCHMARK(f32_expminus__scalar_lut64_p2)->Unit(benchmark::kMillisecond)->Iterations(1);
-BENCHMARK(f32_expminus__scalar_lut2048_p1)->Unit(benchmark::kMillisecond)->Iterations(1);
+BENCHMARK(f32_expminus__scalar_rr2_p5)->Unit(benchmark::kMillisecond)->Iterations(1);
+BENCHMARK(f32_expminus__scalar_rr2_lut64_p2)->Unit(benchmark::kMillisecond)->Iterations(1);
+BENCHMARK(f32_expminus__scalar_rr2_lut2048_p1)->Unit(benchmark::kMillisecond)->Iterations(1);
#ifndef XNNPACK_BENCHMARK_NO_MAIN
BENCHMARK_MAIN();
diff --git a/src/math/expminus-avx2-p5.c b/src/math/expminus-avx2-rr2-p5.c
similarity index 98%
rename from src/math/expminus-avx2-p5.c
rename to src/math/expminus-avx2-rr2-p5.c
index ad287c2..407129e 100644
--- a/src/math/expminus-avx2-p5.c
+++ b/src/math/expminus-avx2-rr2-p5.c
@@ -11,7 +11,7 @@
#include <xnnpack/math-stubs.h>
-void xnn_math_f32_expminus__avx2_p5(
+void xnn_math_f32_expminus__avx2_rr2_p5(
size_t n,
const float* input,
float* output)
diff --git a/src/math/expminus-neonfma-lut2048-p1.c b/src/math/expminus-neonfma-rr2-lut2048-p1.c
similarity index 98%
rename from src/math/expminus-neonfma-lut2048-p1.c
rename to src/math/expminus-neonfma-rr2-lut2048-p1.c
index 3eba912..8bc3b56 100644
--- a/src/math/expminus-neonfma-lut2048-p1.c
+++ b/src/math/expminus-neonfma-rr2-lut2048-p1.c
@@ -15,7 +15,7 @@
// Table of exp2(k / 2048) values decremented (as integer) by (k << 12), k = 0..2048
extern XNN_INTERNAL const float xnn_table_exp2minus_k_over_2048[2048];
-void xnn_math_f32_expminus__neonfma_lut2048_p1(
+void xnn_math_f32_expminus__neonfma_rr2_lut2048_p1(
size_t n,
const float* input,
float* output)
diff --git a/src/math/expminus-neonfma-lut64-p2.c b/src/math/expminus-neonfma-rr2-lut64-p2.c
similarity index 98%
rename from src/math/expminus-neonfma-lut64-p2.c
rename to src/math/expminus-neonfma-rr2-lut64-p2.c
index 5443778..3cbeaa8 100644
--- a/src/math/expminus-neonfma-lut64-p2.c
+++ b/src/math/expminus-neonfma-rr2-lut64-p2.c
@@ -15,7 +15,7 @@
// Table of exp2(k / 64) values decremented (as integer) by (k << 17), k = 0..63
extern XNN_INTERNAL const float xnn_table_exp2minus_k_over_64[64];
-void xnn_math_f32_expminus__neonfma_lut64_p2(
+void xnn_math_f32_expminus__neonfma_rr2_lut64_p2(
size_t n,
const float* input,
float* output)
diff --git a/src/math/expminus-neonfma-p5.c b/src/math/expminus-neonfma-rr2-p5.c
similarity index 98%
rename from src/math/expminus-neonfma-p5.c
rename to src/math/expminus-neonfma-rr2-p5.c
index 9b25775..748b561 100644
--- a/src/math/expminus-neonfma-p5.c
+++ b/src/math/expminus-neonfma-rr2-p5.c
@@ -11,7 +11,7 @@
#include <xnnpack/math-stubs.h>
-void xnn_math_f32_expminus__neonfma_p5(
+void xnn_math_f32_expminus__neonfma_rr2_p5(
size_t n,
const float* input,
float* output)
diff --git a/src/math/expminus-scalar-lut2048-p1.c b/src/math/expminus-scalar-rr2-lut2048-p1.c
similarity index 98%
rename from src/math/expminus-scalar-lut2048-p1.c
rename to src/math/expminus-scalar-rr2-lut2048-p1.c
index acf5a44..16fcf1c 100644
--- a/src/math/expminus-scalar-lut2048-p1.c
+++ b/src/math/expminus-scalar-rr2-lut2048-p1.c
@@ -15,7 +15,7 @@
// Table of exp2(k / 2048) values decremented (as integer) by (k << 12), k = 0..2048
extern XNN_INTERNAL const uint32_t xnn_table_exp2minus_k_over_2048[2048];
-void xnn_math_f32_expminus__scalar_lut2048_p1(
+void xnn_math_f32_expminus__scalar_rr2_lut2048_p1(
size_t n,
const float* input,
float* output)
diff --git a/src/math/expminus-scalar-lut64-p2.c b/src/math/expminus-scalar-rr2-lut64-p2.c
similarity index 98%
rename from src/math/expminus-scalar-lut64-p2.c
rename to src/math/expminus-scalar-rr2-lut64-p2.c
index a8ce276..c0a7826 100644
--- a/src/math/expminus-scalar-lut64-p2.c
+++ b/src/math/expminus-scalar-rr2-lut64-p2.c
@@ -15,7 +15,7 @@
// Table of exp2(k / 64) values decremented (as integer) by (k << 17), k = 0..63
extern XNN_INTERNAL const uint32_t xnn_table_exp2minus_k_over_64[64];
-void xnn_math_f32_expminus__scalar_lut64_p2(
+void xnn_math_f32_expminus__scalar_rr2_lut64_p2(
size_t n,
const float* input,
float* output)
diff --git a/src/math/expminus-scalar-p5.c b/src/math/expminus-scalar-rr2-p5.c
similarity index 98%
rename from src/math/expminus-scalar-p5.c
rename to src/math/expminus-scalar-rr2-p5.c
index b9df440..dc876b1 100644
--- a/src/math/expminus-scalar-p5.c
+++ b/src/math/expminus-scalar-rr2-p5.c
@@ -12,7 +12,7 @@
#include <fp16/bitcasts.h>
-void xnn_math_f32_expminus__scalar_p5(
+void xnn_math_f32_expminus__scalar_rr2_p5(
size_t n,
const float* input,
float* output)
diff --git a/src/math/expminus-sse2-p5.c b/src/math/expminus-sse2-rr2-p5.c
similarity index 98%
rename from src/math/expminus-sse2-p5.c
rename to src/math/expminus-sse2-rr2-p5.c
index ac3b408..893781c 100644
--- a/src/math/expminus-sse2-p5.c
+++ b/src/math/expminus-sse2-rr2-p5.c
@@ -11,7 +11,7 @@
#include <xnnpack/math-stubs.h>
-void xnn_math_f32_expminus__sse2_p5(
+void xnn_math_f32_expminus__sse2_rr2_p5(
size_t n,
const float* input,
float* output)
diff --git a/src/xnnpack/math-stubs.h b/src/xnnpack/math-stubs.h
index 683bebc..2ccf5d4 100644
--- a/src/xnnpack/math-stubs.h
+++ b/src/xnnpack/math-stubs.h
@@ -97,14 +97,14 @@
DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_exp__avx512f_rr2_lut32_p2_perm2)
DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_exp__avx512f_rr2_lut32_p2_perm2_scalef)
-DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_expminus__neonfma_p5)
-DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_expminus__neonfma_lut64_p2)
-DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_expminus__neonfma_lut2048_p1)
-DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_expminus__sse2_p5)
-DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_expminus__avx2_p5)
-DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_expminus__scalar_p5)
-DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_expminus__scalar_lut64_p2)
-DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_expminus__scalar_lut2048_p1)
+DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_expminus__neonfma_rr2_p5)
+DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_expminus__neonfma_rr2_lut64_p2)
+DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_expminus__neonfma_rr2_lut2048_p1)
+DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_expminus__sse2_rr2_p5)
+DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_expminus__avx2_rr2_p5)
+DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_expminus__scalar_rr2_p5)
+DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_expminus__scalar_rr2_lut64_p2)
+DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_expminus__scalar_rr2_lut2048_p1)
DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_expm1minus__sse2_rr2_p6)