Merge pull request #4023 from Feoramund/simd-index

Vectorize `index_byte`

tests/benchmark/all.odin

@@ -1,4 +1,5 @@
 package benchmarks
 
+@(require) import "bytes"
 @(require) import "crypto"
 @(require) import "hash"

tests/benchmark/bytes/benchmark_bytes.odin

@@ -0,0 +1,116 @@
+package benchmark_bytes
+
+import "core:bytes"
+import "core:fmt"
+import "core:log"
+import "core:testing"
+import "core:time"
+
+
+// These are the normal, unoptimized algorithms.
+
+plain_index_byte :: proc(s: []u8, c: byte) -> (res: int) #no_bounds_check {
+	for i := 0; i < len(s); i += 1 {
+		if s[i] == c {
+			return i
+		}
+	}
+	return -1
+}
+
+plain_last_index_byte :: proc(s: []u8, c: byte) -> (res: int) #no_bounds_check {
+	for i := len(s)-1; i >= 0; i -= 1 {
+		if s[i] == c {
+			return i
+		}
+	}
+	return -1
+}
+
+sizes := [?]int {
+	15, 16, 17,
+	31, 32, 33,
+	256,
+	512,
+	1024,
+	1024 * 1024,
+	1024 * 1024 * 1024,
+}
+
+run_trial_size :: proc(p: proc([]u8, byte) -> int, size: int, idx: int, warmup: int, runs: int) -> (timing: time.Duration) {
+	data := make([]u8, size)
+	defer delete(data)
+
+	for i in 0..<size {
+		data[i] = u8('0' + i % 10)
+	}
+	data[idx] = 'z'
+
+	accumulator: int
+
+	for _ in 0..<warmup {
+		accumulator += p(data, 'z')
+	}
+
+	for _ in 0..<runs {
+		start := time.now()
+		accumulator += p(data, 'z')
+		done := time.since(start)
+		timing += done
+	}
+
+	timing /= time.Duration(runs)
+
+	log.debug(accumulator)
+	return
+}
+
+HOT :: 3
+
+@test
+benchmark_plain_index_cold :: proc(t: ^testing.T) {
+	report: string
+	for size in sizes {
+		timing := run_trial_size(plain_index_byte, size, size - 1, 0, 1)
+		report = fmt.tprintf("%s\n        +++ % 8M | %v", report, size, timing)
+		timing = run_trial_size(plain_last_index_byte, size, 0, 0, 1)
+		report = fmt.tprintf("%s\n (last) +++ % 8M | %v", report, size, timing)
+	}
+	log.info(report)
+}
+
+@test
+benchmark_plain_index_hot :: proc(t: ^testing.T) {
+	report: string
+	for size in sizes {
+		timing := run_trial_size(plain_index_byte, size, size - 1, HOT, HOT)
+		report = fmt.tprintf("%s\n        +++ % 8M | %v", report, size, timing)
+		timing = run_trial_size(plain_last_index_byte, size, 0, HOT, HOT)
+		report = fmt.tprintf("%s\n (last) +++ % 8M | %v", report, size, timing)
+	}
+	log.info(report)
+}
+
+@test
+benchmark_simd_index_cold :: proc(t: ^testing.T) {
+	report: string
+	for size in sizes {
+		timing := run_trial_size(bytes.index_byte, size, size - 1, 0, 1)
+		report = fmt.tprintf("%s\n        +++ % 8M | %v", report, size, timing)
+		timing = run_trial_size(bytes.last_index_byte, size, 0, 0, 1)
+		report = fmt.tprintf("%s\n (last) +++ % 8M | %v", report, size, timing)
+	}
+	log.info(report)
+}
+
+@test
+benchmark_simd_index_hot :: proc(t: ^testing.T) {
+	report: string
+	for size in sizes {
+		timing := run_trial_size(bytes.index_byte, size, size - 1, HOT, HOT)
+		report = fmt.tprintf("%s\n        +++ % 8M | %v", report, size, timing)
+		timing = run_trial_size(bytes.last_index_byte, size, 0, HOT, HOT)
+		report = fmt.tprintf("%s\n (last) +++ % 8M | %v", report, size, timing)
+	}
+	log.info(report)
+}

tests/core/bytes/test_core_bytes.odin

@@ -0,0 +1,89 @@
+package test_core_bytes
+
+import "core:bytes"
+import "core:slice"
+import "core:testing"
+
+@private SIMD_SCAN_WIDTH :: 8 * size_of(uintptr)
+
+@test
+test_index_byte_sanity :: proc(t: ^testing.T) {
+	// We must be able to find the byte at the correct index.
+	data := make([]u8, 2 * SIMD_SCAN_WIDTH)
+	defer delete(data)
+	slice.fill(data, '-')
+
+	INDEX_MAX :: SIMD_SCAN_WIDTH - 1
+
+	for offset in 0..<INDEX_MAX {
+		for idx in 0..<INDEX_MAX {
+			sub := data[offset:]
+			sub[idx] = 'o'
+			if !testing.expect_value(t, bytes.index_byte(sub, 'o'), idx) {
+				return
+			}
+			if !testing.expect_value(t, bytes.last_index_byte(sub, 'o'), idx) {
+				return
+			}
+			sub[idx] = '-'
+		}
+	}
+}
+
+@test
+test_index_byte_empty :: proc(t: ^testing.T) {
+	a: [1]u8
+	testing.expect_value(t, bytes.index_byte(a[0:0], 'o'), -1)
+	testing.expect_value(t, bytes.last_index_byte(a[0:0], 'o'), -1)
+}
+
+@test
+test_index_byte_multiple_hits :: proc(t: ^testing.T) {
+	for n in 5..<256 {
+		data := make([]u8, n)
+		defer delete(data)
+		slice.fill(data, '-')
+
+		data[n-1] = 'o'
+		data[n-3] = 'o'
+		data[n-5] = 'o'
+
+		// Find the first one.
+		if !testing.expect_value(t, bytes.index_byte(data, 'o'), n-5) {
+			return
+		}
+
+		// Find the last one.
+		if !testing.expect_value(t, bytes.last_index_byte(data, 'o'), n-1) {
+			return
+		}
+	}
+}
+
+@test
+test_index_byte_zero :: proc(t: ^testing.T) {
+	// This test protects against false positives in uninitialized memory.
+	for n in 1..<256 {
+		data := make([]u8, n + 64)
+		defer delete(data)
+		slice.fill(data, '-')
+
+		// Positive hit.
+		data[n-1] = 0
+		if !testing.expect_value(t, bytes.index_byte(data[:n], 0), n-1) {
+			return
+		}
+		if !testing.expect_value(t, bytes.last_index_byte(data[:n], 0), n-1) {
+			return
+		}
+
+		// Test for false positives.
+		data[n-1] = '-'
+		if !testing.expect_value(t, bytes.index_byte(data[:n], 0), -1) {
+			return
+		}
+		if !testing.expect_value(t, bytes.last_index_byte(data[:n], 0), -1) {
+			return
+		}
+	}
+}

tests/core/normal.odin

@@ -9,6 +9,7 @@ download_assets :: proc() {
 	}
 }
 
+@(require) import "bytes"
 @(require) import "c/libc"
 @(require) import "compress"
 @(require) import "container"