Parameterized API for unary operators

PiperOrigin-RevId: 678574731
google · Oct 24, 2024 · 5de58db · 5de58db
1 parent 79a0d9c
commit 5de58db
Show file tree

Hide file tree

Showing 226 changed files with 6,204 additions and 25,500 deletions.
diff --git a/BUILD.bazel b/BUILD.bazel
@@ -487,6 +487,7 @@ xnnpack_cc_library(
         ":logging",
         ":math",
         ":microparams",
+        ":requantization",
         ":unaligned",
         ":xnnpack_h",
     ],
@@ -973,6 +974,7 @@ xnnpack_cc_library(
         ":microkernel_utils",
         ":microkernels_h",
         ":microparams_init",
+        ":node_type",
         ":normalization",
         ":operator_type",
         ":operator_utils",
@@ -981,9 +983,7 @@ xnnpack_cc_library(
         ":quantization",
         ":xnnpack_h",
         "@pthreadpool",
-    ] + select({
-        "//conditions:default": [],
-    }),
+    ],
 )
 
 xnnpack_cc_library(

diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -425,7 +425,6 @@ SET(OPERATOR_SRCS
   src/operators/deconvolution-nhwc.c
   src/operators/dynamic-fully-connected-nc.c
   src/operators/fully-connected-nc.c
-  src/operators/lut-elementwise-nc.c
   src/operators/max-pooling-nhwc.c
   src/operators/pack-lh.c
   src/operators/reduce-nd.c
@@ -443,50 +442,32 @@ SET(SUBGRAPH_SRCS
   src/memory-planner.c
   src/runtime.c
   src/subgraph.c
-  src/subgraph/abs.c
   src/subgraph/argmax-pooling-2d.c
   src/subgraph/average-pooling-2d.c
-  src/subgraph/bankers-rounding.c
   src/subgraph/batch-matrix-multiply.c
   src/subgraph/binary.c
-  src/subgraph/ceiling.c
-  src/subgraph/clamp.c
   src/subgraph/concatenate.c
-  src/subgraph/convert.c
   src/subgraph/convolution-2d.c
   src/subgraph/copy.c
   src/subgraph/deconvolution-2d.c
   src/subgraph/deprecated.c
   src/subgraph/depth-to-space-2d.c
   src/subgraph/depthwise-convolution-2d.c
-  src/subgraph/elu.c
   src/subgraph/even-split.c
-  src/subgraph/exp.c
-  src/subgraph/floor.c
   src/subgraph/fully-connected-sparse.c
   src/subgraph/fully-connected.c
-  src/subgraph/gelu.c
-  src/subgraph/hardswish.c
-  src/subgraph/leaky-relu.c
-  src/subgraph/log.c
   src/subgraph/max-pooling-2d.c
-  src/subgraph/negate.c
   src/subgraph/pack-lh.c
-  src/subgraph/reciprocal-square-root.c
   src/subgraph/reshape-helpers.c
   src/subgraph/scaled-dot-product-attention.c
-  src/subgraph/sigmoid.c
-  src/subgraph/softmax.c
   src/subgraph/space-to-depth-2d.c
-  src/subgraph/square-root.c
-  src/subgraph/square.c
   src/subgraph/static-constant-pad.c
   src/subgraph/static-reduce.c
   src/subgraph/static-resize-bilinear-2d.c
   src/subgraph/static-slice.c
   src/subgraph/static-transpose.c
-  src/subgraph/tanh.c
   src/subgraph/unpooling-2d.c
+  src/subgraph/unary.c
   src/subgraph/validation.c
   src/tensor.c)
 
@@ -1231,10 +1212,6 @@ IF(XNNPACK_BUILD_TESTS)
   ENDIF()
   TARGET_LINK_LIBRARIES(gemm-microkernel-tester PUBLIC next-prime)
 
-  ADD_LIBRARY(unary-operator-tester STATIC test/unary-operator-tester.cc)
-  TARGET_INCLUDE_DIRECTORIES(unary-operator-tester PRIVATE include src test)
-  TARGET_LINK_LIBRARIES(unary-operator-tester PRIVATE XNNPACK pthreadpool GTest::gtest)
-
   ADD_LIBRARY(dwconv-microkernel-tester STATIC test/dwconv-microkernel-tester.cc)
   TARGET_INCLUDE_DIRECTORIES(dwconv-microkernel-tester PRIVATE include src test)
   TARGET_LINK_LIBRARIES(dwconv-microkernel-tester PRIVATE XNNPACK pthreadpool GTest::gtest)
@@ -1245,15 +1222,6 @@ IF(XNNPACK_BUILD_TESTS)
   TARGET_INCLUDE_DIRECTORIES(vbinary-microkernel-tester PRIVATE include src test)
   TARGET_LINK_LIBRARIES(vbinary-microkernel-tester PRIVATE XNNPACK pthreadpool GTest::gtest)
 
-  ADD_LIBRARY(vcvt-microkernel-tester STATIC test/vcvt-microkernel-tester.cc)
-  TARGET_INCLUDE_DIRECTORIES(vcvt-microkernel-tester PRIVATE include src test)
-  TARGET_LINK_LIBRARIES(vcvt-microkernel-tester PRIVATE XNNPACK pthreadpool GTest::gtest)
-
-  ADD_LIBRARY(vunary-microkernel-tester STATIC test/vunary-microkernel-tester.cc)
-  TARGET_INCLUDE_DIRECTORIES(vunary-microkernel-tester PRIVATE include src test)
-  TARGET_LINK_LIBRARIES(vunary-microkernel-tester PRIVATE XNNPACK pthreadpool GTest::gtest)
-  TARGET_LINK_LIBRARIES(vunary-microkernel-tester PUBLIC next-prime)
-
   ADD_LIBRARY(convolution-test-helpers OBJECT test/convolution-test-helpers.cc)
   TARGET_INCLUDE_DIRECTORIES(convolution-test-helpers PRIVATE include src)
   TARGET_LINK_LIBRARIES(convolution-test-helpers PRIVATE xnnpack-base)
@@ -1288,7 +1256,6 @@ IF(XNNPACK_BUILD_TESTS)
       microparams-init
       next-prime
       pthreadpool
-      vunary-microkernel-tester
       XNNPACK)
     ADD_SHARDED_TEST(${TEST}-test 10)
   ENDFOREACH()
@@ -1318,34 +1285,13 @@ IF(XNNPACK_BUILD_TESTS)
     TARGET_LINK_LIBRARIES(subgraph-size-test PRIVATE XNNPACK)
 
     # ---[ Build operator-level unit tests
-    SET(LIBRARY_UNARY_UNIT_TESTS
-        abs-nc
-        bankers-rounding-nc
-        ceiling-nc
-        clamp-nc
-        elu-nc
-        exp-nc
-        floor-nc
-        hardswish-nc
-        leaky-relu-nc
-        log-nc
-        negate-nc
-        reciprocal-square-root-nc
-        sigmoid-nc
-        square-nc
-        square-root-nc
-        tanh-nc
-        truncation-nc)
-    FOREACH(TEST ${LIBRARY_UNARY_UNIT_TESTS})
-      ADD_EXECUTABLE(${TEST}-test test/${TEST}.cc)
-      TARGET_INCLUDE_DIRECTORIES(${TEST}-test PRIVATE src test)
-      TARGET_LINK_LIBRARIES(${TEST}-test PRIVATE
-        GTest::gtest
-        GTest::gtest_main
-        unary-operator-tester
-        XNNPACK)
-      ADD_TEST(NAME ${TEST}-test COMMAND ${TEST}-test)
-    ENDFOREACH()
+    ADD_EXECUTABLE(unary-elementwise-nc-test test/unary-elementwise-nc.cc)
+    TARGET_INCLUDE_DIRECTORIES(unary-elementwise-nc-test PRIVATE src test)
+    TARGET_LINK_LIBRARIES(unary-elementwise-nc-test PRIVATE
+      GTest::gtest
+      GTest::gtest_main
+      XNNPACK)
+    ADD_TEST(NAME unary-elementwise-nc-test COMMAND unary-elementwise-nc-test)
 
     ADD_EXECUTABLE(binary-elementwise-nd-test test/binary-elementwise-nd.cc)
     TARGET_INCLUDE_DIRECTORIES(binary-elementwise-nd-test PRIVATE src test)
@@ -1376,53 +1322,35 @@ IF(XNNPACK_BUILD_TESTS)
 
     # ---[ Build subgraph-level unit tests
     SET(LIBRARY_SUBGRAPH_UNIT_TESTS
-        abs
-        abs-reshape
         argmax-pooling-2d
         average-pooling-2d
         average-pooling-2d-reshape
-        bankers-rounding
         binary
-        ceiling
-        clamp
         concatenate2
         concatenate3
         concatenate4
         concatenate5
-        convert
         copy
         depth-to-space-2d
-        elu
-        gelu
-        exp
         even-split2
         even-split3
         even-split4
-        floor
         global-average-pooling-1d
         global-average-pooling-2d
         global-sum-pooling-1d
         global-sum-pooling-2d
-        hardswish
-        leaky-relu
-        log
         max-pooling-2d
-        negate
-        reciprocal-square-root
         reshape-helpers
-        sigmoid
         static-slice
         softmax
         space-to-depth-2d
-        square
-        square-root
         static-constant-pad
         static-reduce
         static-reshape
         static-resize-bilinear-2d
         static-transpose
-        tanh
         transpose-reshape
+        unary
         unpooling-2d)
     FOREACH(TEST ${LIBRARY_SUBGRAPH_UNIT_TESTS})
       ADD_EXECUTABLE(${TEST}-test test/${TEST}.cc)
@@ -1747,7 +1675,6 @@ IF(XNNPACK_BUILD_TESTS)
     ADD_EXECUTABLE(${TEST}-test test/${TEST}.cc)
     TARGET_INCLUDE_DIRECTORIES(${TEST}-test PRIVATE include src test)
     TARGET_LINK_LIBRARIES(${TEST}-test PRIVATE
-        vcvt-microkernel-tester
         GTest::gmock
         GTest::gtest
         GTest::gtest_main
@@ -1800,7 +1727,6 @@ IF(XNNPACK_BUILD_TESTS)
     ADD_EXECUTABLE(${TEST}-test test/${TEST}.cc)
     TARGET_INCLUDE_DIRECTORIES(${TEST}-test PRIVATE include src test)
     TARGET_LINK_LIBRARIES(${TEST}-test PRIVATE
-        vunary-microkernel-tester
         GTest::gmock
         GTest::gtest
         GTest::gtest_main
@@ -1945,7 +1871,6 @@ IF(XNNPACK_BUILD_BENCHMARKS)
         softmax
         square
         square-root
-        truncation
         tanh)
     FOREACH(BENCH ${LIBRARY_OPERATOR_BENCHMARKS})
       ADD_EXECUTABLE(${BENCH}-bench bench/${BENCH}.cc)
@@ -1963,7 +1888,6 @@ IF(XNNPACK_BUILD_BENCHMARKS)
       f16-conv-hwc2chw
       f16-dwconv
       f16-dwconv2d-chw
-      f16-f32-vcvt
       f16-f32acc-gemm
       f16-f32acc-igemm
       f16-f32acc-rdsum
@@ -1982,16 +1906,13 @@ IF(XNNPACK_BUILD_BENCHMARKS)
       f32-conv-hwc2chw
       f32-dwconv
       f32-dwconv2d-chw
-      f32-f16-vcvt
       f32-gemm
       f32-gemm-goi-minmax
       f32-gemm-minmax
       f32-igemm
       f32-im2col-gemm
       f32-qc4w-gemm
       f32-qc8w-gemm
-      f32-qs8-vcvt
-      f32-qu8-vcvt
       f32-raddexpminusmax
       f32-raddextexp
       f32-raddstoreexpminusmax
@@ -2014,21 +1935,15 @@ IF(XNNPACK_BUILD_BENCHMARKS)
       qd8-f32-qc8w-gemm
       qp8-f32-qc4w-gemm
       qp8-f32-qb4w-gemm
-      qs16-qs8-vcvt
       qs8-dwconv
-      qs8-f16-vcvt
-      qs8-f32-vcvt
       qs8-gemm
       qs8-qc8w-gemm-fp32
       qu8-rdsum
       qs8-rsum
       qu8-rsum
-      qs8-vcvt
-      qu8-f32-vcvt
       qu8-gemm
       qu8-gemm-fp32
       qu8-gemm-rndnu
-      qu8-vcvt
       x16-packw
       x32-packw
       x8-lut

diff --git a/bench/BUILD.bazel b/bench/BUILD.bazel
@@ -228,29 +228,6 @@ xnnpack_benchmark(
     ]),
 )
 
-[xnnpack_benchmark(
-    name = "%s_bench" % kernel,
-    srcs = [
-        "%s.cc" % kernel.replace("_", "-"),
-        "vcvt-benchmark.h",
-    ],
-    deps = MICROKERNEL_BENCHMARK_DEPS,
-) for kernel in [
-    "qs8_f16_vcvt",
-    "qs8_f32_vcvt",
-    "qs8_vcvt",
-    "qs16_qs8_vcvt",
-    "qu8_f32_vcvt",
-    "qu8_vcvt",
-    "f16_f32_vcvt",
-    "f16_qs8_vcvt",
-    "f32_f16_vcvt",
-    "f32_qs8_vcvt",
-    "f32_qu8_vcvt",
-    "s32_f32_vcvt",
-    "u32_f32_vcvt",
-]]
-
 [xnnpack_benchmark(
     name = "%s_bench" % kernel,
     srcs = [
@@ -590,7 +567,6 @@ xnnpack_benchmark(
     "square",
     "square_root",
     "tanh",
-    "truncation",
 ]]
 
 xnnpack_benchmark(

diff --git a/bench/abs.cc b/bench/abs.cc
@@ -14,15 +14,11 @@
 
 
 static void xnnpack_abs_f16(benchmark::State& state) {
-  benchmark_unary_operator<xnn_float16, xnn_float16>(
-      xnn_create_abs_nc_f16, xnn_reshape_abs_nc_f16, xnn_setup_abs_nc_f16,
-      state);
+  benchmark_unary_operator<xnn_float16, xnn_float16>(state, xnn_unary_abs);
 }
 
 static void xnnpack_abs_f32(benchmark::State& state) {
-  benchmark_unary_operator<float, float>(xnn_create_abs_nc_f32,
-                                         xnn_reshape_abs_nc_f32,
-                                         xnn_setup_abs_nc_f32, state);
+  benchmark_unary_operator<float, float>(state, xnn_unary_abs);
 }
 
 BENCHMARK(xnnpack_abs_f16)

diff --git a/bench/bankers-rounding.cc b/bench/bankers-rounding.cc
@@ -14,14 +14,12 @@
 
 static void xnnpack_bankers_rounding_f16(benchmark::State& state) {
   benchmark_unary_operator<xnn_float16, xnn_float16>(
-      xnn_create_bankers_rounding_nc_f16, xnn_reshape_bankers_rounding_nc_f16,
-      xnn_setup_bankers_rounding_nc_f16, state);
+      state, xnn_unary_bankers_rounding);
 }
 
 static void xnnpack_bankers_rounding_f32(benchmark::State& state) {
   benchmark_unary_operator<float, float>(
-      xnn_create_bankers_rounding_nc_f32, xnn_reshape_bankers_rounding_nc_f32,
-      xnn_setup_bankers_rounding_nc_f32, state);
+      state, xnn_unary_bankers_rounding);
 }
 
 BENCHMARK(xnnpack_bankers_rounding_f16)

diff --git a/bench/ceiling.cc b/bench/ceiling.cc
@@ -14,15 +14,11 @@
 
 
 static void xnnpack_ceiling_f16(benchmark::State& state) {
-  benchmark_unary_operator<xnn_float16, xnn_float16>(xnn_create_ceiling_nc_f16,
-                                             xnn_reshape_ceiling_nc_f16,
-                                             xnn_setup_ceiling_nc_f16, state);
+  benchmark_unary_operator<xnn_float16, xnn_float16>(state, xnn_unary_ceiling);
 }
 
 static void xnnpack_ceiling_f32(benchmark::State& state) {
-  benchmark_unary_operator<float, float>(xnn_create_ceiling_nc_f32,
-                                         xnn_reshape_ceiling_nc_f32,
-                                         xnn_setup_ceiling_nc_f32, state);
+  benchmark_unary_operator<float, float>(state, xnn_unary_ceiling);
 }
 
 BENCHMARK(xnnpack_ceiling_f16)