ARROW-9131: [C++] Faster ascii_lower and ascii_upper.

maartenbreddels · pitrou · commit d98b9c578065 · 2020-06-15T14:31:20.000+02:00
Following up on #7418 I tried and benchmarked a different way for * ascii_lower * ascii_upper Before (lower is similar): ``` -------------------------------------------------- Benchmark Time CPU Iterations -------------------------------------------------- AsciiUpper_median 4922843 ns 4918961 ns 10 bytes_per_second=3.1457G/s items_per_second=213.17M/s ``` After: ``` -------------------------------------------------- Benchmark Time CPU Iterations -------------------------------------------------- AsciiUpper_median 1391272 ns 1390014 ns 10 bytes_per_second=11.132G/s items_per_second=754.363M/s ``` This is a 3.7x speedup (on a AMD machine). Using http://quick-bench.com/JaDErmVCY23Z1tu6YZns_KBt0qU I found 4.6x speedup for clang 9, 6.4x for GCC 9.2. Also, the test is expanded a bit to include a non-ascii codepoint, to make explicit it is fine to upper or lower case a utf8 string. The non-overlap encoding of utf8 make this ok (see section 2.5 of Unicode Standard Core Specification v13.0). Closes #7434 from maartenbreddels/ARROW-9131 Authored-by: Maarten A. Breddels <maartenbreddels@gmail.com> Signed-off-by: Antoine Pitrou <antoine@python.org>
diff --git a/cpp/src/arrow/compute/kernels/scalar_string.cc b/cpp/src/arrow/compute/kernels/scalar_string.cc
@@ -64,77 +64,30 @@ void StringDataTransform(KernelContext* ctx, const ExecBatch& batch,
   }
 }
 
-// Generated with
-//
-// print("static constexpr uint8_t kAsciiUpperTable[] = {")
-// for i in range(256):
-//     if i > 0: print(', ', end='')
-//     if i >= ord('a') and i <= ord('z'):
-//         print(i - 32, end='')
-//     else:
-//         print(i, end='')
-// print("};")
-
-static constexpr uint8_t kAsciiUpperTable[] = {
-    0,   1,   2,   3,   4,   5,   6,   7,   8,   9,   10,  11,  12,  13,  14,  15,
-    16,  17,  18,  19,  20,  21,  22,  23,  24,  25,  26,  27,  28,  29,  30,  31,
-    32,  33,  34,  35,  36,  37,  38,  39,  40,  41,  42,  43,  44,  45,  46,  47,
-    48,  49,  50,  51,  52,  53,  54,  55,  56,  57,  58,  59,  60,  61,  62,  63,
-    64,  65,  66,  67,  68,  69,  70,  71,  72,  73,  74,  75,  76,  77,  78,  79,
-    80,  81,  82,  83,  84,  85,  86,  87,  88,  89,  90,  91,  92,  93,  94,  95,
-    96,  65,  66,  67,  68,  69,  70,  71,  72,  73,  74,  75,  76,  77,  78,  79,
-    80,  81,  82,  83,  84,  85,  86,  87,  88,  89,  90,  123, 124, 125, 126, 127,
-    128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143,
-    144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159,
-    160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175,
-    176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191,
-    192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207,
-    208, 209, 210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223,
-    224, 225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238, 239,
-    240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253, 254, 255};
-
 void TransformAsciiUpper(const uint8_t* input, int64_t length, uint8_t* output) {
   for (int64_t i = 0; i < length; ++i) {
-    *output++ = kAsciiUpperTable[*input++];
+    const uint8_t utf8_code_unit = *input++;
+    // Code units in the range [a-z] can only be an encoding of an ascii
+    // character/codepoint, not the 2nd, 3rd or 4th code unit (byte) of an different
+    // codepoint. This guaranteed by non-overlap design of the unicode standard. (see
+    // section 2.5 of Unicode Standard Core Specification v13.0)
+    *output++ = ((utf8_code_unit >= 'a') && (utf8_code_unit <= 'z'))
+                    ? (utf8_code_unit - 32)
+                    : utf8_code_unit;
   }
 }
 
 void AsciiUpperExec(KernelContext* ctx, const ExecBatch& batch, Datum* out) {
   StringDataTransform(ctx, batch, TransformAsciiUpper, out);
 }
 
-// Generated with
-//
-// print("static constexpr uint8_t kAsciiLowerTable[] = {")
-// for i in range(256):
-//     if i > 0: print(', ', end='')
-//     if i >= ord('A') and i <= ord('Z'):
-//         print(i + 32, end='')
-//     else:
-//         print(i, end='')
-// print("};")
-
-static constexpr uint8_t kAsciiLowerTable[] = {
-    0,   1,   2,   3,   4,   5,   6,   7,   8,   9,   10,  11,  12,  13,  14,  15,
-    16,  17,  18,  19,  20,  21,  22,  23,  24,  25,  26,  27,  28,  29,  30,  31,
-    32,  33,  34,  35,  36,  37,  38,  39,  40,  41,  42,  43,  44,  45,  46,  47,
-    48,  49,  50,  51,  52,  53,  54,  55,  56,  57,  58,  59,  60,  61,  62,  63,
-    64,  97,  98,  99,  100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111,
-    112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 91,  92,  93,  94,  95,
-    96,  97,  98,  99,  100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111,
-    112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127,
-    128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143,
-    144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159,
-    160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175,
-    176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191,
-    192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207,
-    208, 209, 210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223,
-    224, 225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238, 239,
-    240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253, 254, 255};
-
 void TransformAsciiLower(const uint8_t* input, int64_t length, uint8_t* output) {
   for (int64_t i = 0; i < length; ++i) {
-    *output++ = kAsciiLowerTable[*input++];
+    // As with TransformAsciiUpper, the same guarantee holds for the range [A-Z]
+    const uint8_t utf8_code_unit = *input++;
+    *output++ = ((utf8_code_unit >= 'A') && (utf8_code_unit <= 'Z'))
+                    ? (utf8_code_unit + 32)
+                    : utf8_code_unit;
   }
 }
 
diff --git a/cpp/src/arrow/compute/kernels/scalar_string_test.cc b/cpp/src/arrow/compute/kernels/scalar_string_test.cc
@@ -57,13 +57,13 @@ TYPED_TEST(TestStringKernels, AsciiLength) {
 }
 
 TYPED_TEST(TestStringKernels, AsciiUpper) {
-  this->CheckUnary("ascii_upper", "[\"aAa&\", null, \"\", \"b\"]", this->string_type(),
-                   "[\"AAA&\", null, \"\", \"B\"]");
+  this->CheckUnary("ascii_upper", "[\"aAazZæÆ&\", null, \"\", \"b\"]",
+                   this->string_type(), "[\"AAAZZæÆ&\", null, \"\", \"B\"]");
 }
 
 TYPED_TEST(TestStringKernels, AsciiLower) {
-  this->CheckUnary("ascii_lower", "[\"aAa&\", null, \"\", \"b\"]", this->string_type(),
-                   "[\"aaa&\", null, \"\", \"b\"]");
+  this->CheckUnary("ascii_lower", "[\"aAazZæÆ&\", null, \"\", \"b\"]",
+                   this->string_type(), "[\"aaazzæÆ&\", null, \"\", \"b\"]");
 }
 
 TYPED_TEST(TestStringKernels, Strptime) {

Original file line number	Diff line number	Diff line change
`@@ -57,13 +57,13 @@ TYPED_TEST(TestStringKernels, AsciiLength) {`
`57`	`57`	`}`
`58`	`58`
`59`	`59`	`TYPED_TEST(TestStringKernels, AsciiUpper) {`
`60`		`- this->CheckUnary("ascii_upper", "[\"aAa&\", null, \"\", \"b\"]", this->string_type(),`
`61`		`- "[\"AAA&\", null, \"\", \"B\"]");`
	`60`	`+ this->CheckUnary("ascii_upper", "[\"aAazZæÆ&\", null, \"\", \"b\"]",`
	`61`	`+ this->string_type(), "[\"AAAZZæÆ&\", null, \"\", \"B\"]");`
`62`	`62`	`}`
`63`	`63`
`64`	`64`	`TYPED_TEST(TestStringKernels, AsciiLower) {`
`65`		`- this->CheckUnary("ascii_lower", "[\"aAa&\", null, \"\", \"b\"]", this->string_type(),`
`66`		`- "[\"aaa&\", null, \"\", \"b\"]");`
	`65`	`+ this->CheckUnary("ascii_lower", "[\"aAazZæÆ&\", null, \"\", \"b\"]",`
	`66`	`+ this->string_type(), "[\"aaazzæÆ&\", null, \"\", \"b\"]");`
`67`	`67`	`}`
`68`	`68`
`69`	`69`	`TYPED_TEST(TestStringKernels, Strptime) {`