Implement faster mju_isZeroByte.

PiperOrigin-RevId: 814148714
Change-Id: Iaf5ed9d3502eaaaee8d46a986f2db00811ae0707

src/engine/engine_util_misc.c

@@ -1332,26 +1332,8 @@ int mju_isZero(const mjtNum* vec, int n) {
 
 // return 1 if all elements are 0
 int mju_isZeroByte(const unsigned char* vec, int n) {
-  size_t i = 0;
-
-  // unroll using 8-byte chunks
-  const uint64_t* vec64 = (const uint64_t*)vec;
-  size_t n64 = n / sizeof(uint64_t);
-  for (; i < n64; ++i) {
-    if (vec64[i]) {
-      return 0;
-    }
-  }
-
-  // remaining bytes
-  i *= sizeof(uint64_t);
-  for (; i < n; ++i) {
-    if (vec[i]) {
-      return 0;
-    }
-  }
-
-  return 1;
+  if (!n || *vec) return !n;
+  return memcmp(vec, vec + 1, n - 1) == 0;
 }
 
 

src/engine/engine_util_misc.h

@@ -136,7 +136,7 @@ MJAPI int mju_isBad(mjtNum x);
 // return 1 if all elements are numerically 0 (-0.0 treated as zero)
 MJAPI int mju_isZero(const mjtNum* vec, int n);
 
-// return 1 if all elements are 0x00, can be ~2x faster than mju_isZero
+// return 1 if all elements are 0x00, faster than mju_isZero
 MJAPI int mju_isZeroByte(const unsigned char* vec, int n);
 
 // set integer vector to 0

test/benchmark/iszero_benchmark_test.cc

@@ -46,6 +46,11 @@ ABSL_ATTRIBUTE_NO_TAIL_CALL static void IzZeroBenchmark(
 }
 
 // Define benchmarks.
+void BM_isZeroByte_10(benchmark::State& state) {
+  IzZeroBenchmark(state, 10, true);
+}
+BENCHMARK(BM_isZeroByte_10);
+
 void BM_isZeroByte_1e2(benchmark::State& state) {
   IzZeroBenchmark(state, 100, true);
 }
@@ -56,6 +61,11 @@ void BM_isZeroByte_1e5(benchmark::State& state) {
 }
 BENCHMARK(BM_isZeroByte_1e5);
 
+void BM_isZero_10(benchmark::State& state) {
+  IzZeroBenchmark(state, 10, false);
+}
+BENCHMARK(BM_isZero_10);
+
 void BM_isZero_1e2(benchmark::State& state) {
   IzZeroBenchmark(state, 100, false);
 }

test/engine/engine_util_misc_test.cc

@@ -360,6 +360,35 @@ TEST_F(UtilMiscTest, MjuIsZero) {
   EXPECT_EQ(mju_isZeroByte((const unsigned char*)vec, sizeof(mjtNum)), 0);
 }
 
+TEST_F(UtilMiscTest, MjuIsZeroByte) {
+  // Zero length array
+  EXPECT_TRUE(mju_isZeroByte(nullptr, 0));
+
+  // zero length array with non-null pointer
+  unsigned char vec0[1] = {0};
+  EXPECT_TRUE(mju_isZeroByte(vec0, sizeof(vec0)));
+
+  // one zero element array
+  unsigned char vec1[1] = {0};
+  EXPECT_TRUE(mju_isZeroByte(vec1, sizeof(vec1)));
+
+  // one non-zero element array
+  unsigned char vec2[2] = {1};
+  EXPECT_FALSE(mju_isZeroByte(vec2, sizeof(vec2)));
+
+  // Non-zero at start
+  unsigned char vec3[3] = {1, 0, 0};
+  EXPECT_FALSE(mju_isZeroByte(vec3, sizeof(vec3)));
+
+  // Non-zero at end
+  unsigned char vec4[3] = {0, 0, 1};
+  EXPECT_FALSE(mju_isZeroByte(vec4, sizeof(vec4)));
+
+  // Non-zero in middle
+  unsigned char vec5[3] = {0, 1, 0};
+  EXPECT_FALSE(mju_isZeroByte(vec5, sizeof(vec5)));
+}
+
 // --------------------------------- Interpolation -----------------------------
 
 using InterpolationTest = MujocoTest;