Use header tables to generate binary microkernel benchmarks

This change replaces the existing quantized binary operator benchmarks. We also now finally have enough automation in place to add binary operator benchmarks easily. Fixes #6396 Also add vunary benchmark to CMake build, and fix some issues found on some targets. PiperOrigin-RevId: 676533418
google · Sep 19, 2024 · 49ce38a · 49ce38a
1 parent a72a8b1
commit 49ce38a
Show file tree

Hide file tree

Showing 12 changed files with 281 additions and 694 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -2083,26 +2083,20 @@ IF(XNNPACK_BUILD_BENCHMARKS)
       qu8-rdsum
       qs8-rsum
       qu8-rsum
-      qs8-vadd
-      qs8-vaddc
       qs8-vcvt
-      qs8-vmul
-      qs8-vmulc
       qu8-f32-vcvt
       qu8-gemm
       qu8-gemm-fp32
       qu8-gemm-rndnu
       qu8-requantization
-      qu8-vadd
-      qu8-vaddc
       qu8-vcvt
-      qu8-vmul
-      qu8-vmulc
       x16-packw
       x32-packw
       x8-lut
       x8-packq
       x8-packw
+      vunary
+      vbinary
       xN-transposec
       xx-transposev)
   FOREACH(BENCH ${MICROKERNEL_BENCHMARKS})

diff --git a/bench/BUILD.bazel b/bench/BUILD.bazel
@@ -241,14 +241,6 @@ xnnpack_benchmark(
     ],
     deps = MICROKERNEL_BENCHMARK_DEPS,
 ) for kernel in [
-    "qs8_vadd",
-    "qs8_vaddc",
-    "qs8_vmul",
-    "qs8_vmulc",
-    "qu8_vadd",
-    "qu8_vaddc",
-    "qu8_vmul",
-    "qu8_vmulc",
     "f16_gavgpool_cw",
     "f16_raddstoreexpminusmax",
     "f16_rmax",
@@ -343,6 +335,15 @@ xnnpack_benchmark(
     deps = MICROKERNEL_BENCHMARK_DEPS,
 )
 
+xnnpack_benchmark(
+    name = "vbinary_bench",
+    srcs = ["vbinary.cc"],
+    # TODO(b/367939259): This is not really that slow, but --config=ios_x86_64 fails to pass
+    # --benchmark_min_time=1x to the benchmark.
+    tags = xnnpack_slow_benchmark_tags(),
+    deps = MICROKERNEL_BENCHMARK_DEPS,
+)
+
 xnnpack_benchmark(
     name = "f32_igemm_bench",
     srcs = [

diff --git a/bench/qs8-vadd.cc b/bench/qs8-vadd.cc
diff --git a/bench/qs8-vaddc.cc b/bench/qs8-vaddc.cc
diff --git a/bench/qs8-vmul.cc b/bench/qs8-vmul.cc
diff --git a/bench/qs8-vmulc.cc b/bench/qs8-vmulc.cc