commit 2dbb944760aded60e20ae6dc71f794812c71b67a
author Marat Dukhan <maratek@google.com> Tue May 12 20:43:43 2020 -0700
committer XNNPACK Team <xnnpack-github-robot@google.com> Tue May 12 20:44:15 2020 -0700
tree 0c8c896251e0480fb68a1961047e2651ea572701
parent f1d2da403031416908ff7c734140dccf3705319a

ROUNDZ evaluation stubs

PiperOrigin-RevId: 311256662

BUILD.bazel

diff --git a/BUILD.bazel b/BUILD.bazel
index 96fc30e..13f52ca 100644
--- a/BUILD.bazel
+++ b/BUILD.bazel
@@ -236,6 +236,8 @@
     "src/math/expminus-scalar-lut64-p2.c",
     "src/math/expminus-scalar-p5.c",
     "src/math/roundne-scalar-addsub.c",
+    "src/math/roundz-scalar-addsub.c",
+    "src/math/roundz-scalar-cvt.c",
     "src/math/sigmoid-scalar-lut2048-p1-div.c",
     "src/math/sigmoid-scalar-lut64-p2-div.c",
     "src/math/sigmoid-scalar-p5-div.c",
@@ -498,6 +500,7 @@
     "src/f32-sigmoid/gen/psimd-p5-div-x20.c",
     "src/f32-sigmoid/gen/psimd-p5-div-x24.c",
     "src/math/roundne-psimd-addsub.c",
+    "src/math/roundz-psimd-addsub.c",
     "src/math/sigmoid-psimd-p5-div.c",
 ]
 
@@ -673,6 +676,8 @@
     "src/x8-zip/x4-neon.c",
     "src/x8-zip/xm-neon.c",
     "src/math/roundne-neon-addsub.c",
+    "src/math/roundz-neon-addsub.c",
+    "src/math/roundz-neon-cvt.c",
     "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
     "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
     "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
@@ -917,6 +922,7 @@
 
 NEONV8_UKERNELS = [
     "src/math/roundne-neonv8.c",
+    "src/math/roundz-neonv8.c",
 ]
 
 AARCH64_NEONFP16ARITH_UKERNELS = [
@@ -1073,6 +1079,7 @@
     "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
     "src/x32-packx/x4-sse.c",
     "src/math/roundne-sse-addsub.c",
+    "src/math/roundz-sse-addsub.c",
 ]
 
 SSE2_UKERNELS = [
@@ -1124,6 +1131,7 @@
     "src/math/exp-sse2-p5.c",
     "src/math/expminus-sse2-p5.c",
     "src/math/roundne-sse2-cvt.c",
+    "src/math/roundz-sse2-cvt.c",
     "src/math/sigmoid-sse2-p5-div.c",
     "src/requantization/precise-sse2.c",
     "src/requantization/fp32-sse2.c",
@@ -1147,6 +1155,7 @@
     "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
     "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
     "src/math/roundne-sse41.c",
+    "src/math/roundz-sse41.c",
     "src/requantization/precise-sse4.c",
     "src/requantization/q31-sse4.c",
     "src/requantization/gemmlowp-sse4.c",
@@ -2603,6 +2612,19 @@
     deps = MICROKERNEL_TEST_DEPS,
 )
 
+xnnpack_unit_test(
+    name = "f32_roundz_eval",
+    srcs = [
+        "eval/f32-roundz.cc",
+        "src/xnnpack/AlignedAllocator.h",
+        "src/xnnpack/math-stubs.h",
+    ] + MICROKERNEL_TEST_HDRS,
+    tags = [
+        "notap",
+    ],
+    deps = MICROKERNEL_TEST_DEPS,
+)
+
 xnnpack_benchmark(
     name = "f32_sigmoid_eval",
     srcs = [