[WIP] implement slice-by-4 with SIMD instructions

src/crc32.ts

@@ -1,12 +1,30 @@
-import { makeUint8Array } from "./utils.ts"
+import { makeUint8Array, parseWasm } from "./utils.ts"
 
-const wasm = "AGFzbQEAAAABCgJgAABgAn9/AXwDAwIAAQUDAQACBw0DAW0CAAF0AAABYwABCpUBAkkBA38DQCABIQBBACECA0AgAEEBdiAAQQFxQaCG4u1+bHMhACACQQFqIgJBCEcNAAsgAUECdCAANgIAIAFBAWoiAUGAAkcNAAsLSQEBfyABQX9zIQFBgIAEIQJBgIAEIABqIQADQCABQf8BcSACLQAAc0ECdCgCACABQQh2cyEBIAJBAWoiAiAASQ0ACyABQX9zuAs"
+const head = "AGFzbQEAAAAB"
+const test = "BQFgAAF7AwIBAAoKAQgAQQD9D/1iCw"
+const defBasic = "CwJgAABgA39/fwF8AwMCAAEFAwEAAgcNAwFtAgABdAAAAWMAAQqWAQI"
+const defSIMD = "EQNgAABgA39/fwF8YAJ/fwF8AwUEAAEAAgUDAQACBxEEAW0CAAF0AAABeAACAWMAAwqjAwQ"
+const implBasic = "SQEDfwNAIAEhAEEAIQIDQCAAQQF2IABBAXFBoIbi7X5scyEAIAJBAWoiAkEIRw0ACyABQQJ0IAA2AgAgAUEBaiIBQYACRw0ACwtKACABQX9zIQFBgIAEIAJyIQJBgIAEIABqIQADQCABQf8BcSACLQAAc0ECdCgCACABQQh2cyEBIAJBAWoiAiAASQ0ACyABQX9zuAs"
+const implSIMD = "TwEDfwNAQQAhAgNAIAIgAXIoAgAiAEH/AXFBAnQoAgAgAEEIdnMhACACQYAIaiICIAFyIAA2AgAgAkGAGEcNAAsgAUEEaiIBQYAIRw0ACwu7AQICfwF7IAFBf3MhAUGAgAQhAkGAgAQgAGoiAEECdUECdCIDQYSABE8EQANAIAEgAigCAHP9Ef0MA////wL///8B////AP////0OQQL9qwH9DAAAAAAABAAAAAgAAAAMAAD9UCIE/RsAKAIAIAT9GwEoAgBzIAT9GwIoAgBzIAT9GwMoAgBzIQEgAkEEaiICIANHDQALCyACIABJBHwgAEGAgARrIAFBf3MgAkGAgARrEAEFIAFBf3O4Cws"
+
+let enableSIMD = true
+try {
+  parseWasm([head, test])
+} catch(_) {
+  enableSIMD = false
+}
+
+const instance = new WebAssembly.Instance(parseWasm(
+  [head, enableSIMD ? defSIMD : defBasic, implBasic, enableSIMD ? implSIMD : ""]
+))
+const { t, c, x, m } = instance.exports as {
+  t(): void, x(): void, m: WebAssembly.Memory
+  c(length: number, init: number): number
+}
+
+t() // initialize the table of precomputed CRCs ; this takes 1 kB in the first page of Memory
+enableSIMD && x() // extend the table of precomputed CRCs ; this takes another 3 kB
 
-const instance = new WebAssembly.Instance(
-  new WebAssembly.Module(Uint8Array.from(atob(wasm), c => c.charCodeAt(0)))
-)
-const { t, c, m } = instance.exports as { t(): void, c(length: number, init: number): number, m: WebAssembly.Memory }
-t() // initialize the table of precomputed CRCs ; this takes 8 kB in the second page of Memory
 export const memory = m // for testing
 
 // Someday we'll have BYOB stream readers and encodeInto etc.

src/crc32.wat

@@ -23,10 +23,11 @@
   )
   ;; this computes the CRC32 of what you put in the module's second page of Memory
   ;; (do not overwrite the first page!)
-  (func $crc32 (param $len i32) (param $crc i32) (result f64) (local $i i32)
+  (func $crc32 (param $len i32) (param $crc i32) (param $i i32) (result f64)
     (local.set $crc (i32.xor (local.get $crc) (i32.const -1)))
-    (local.set $i (i32.const 0x10000))
+    (local.set $i (i32.or (i32.const 0x10000) (local.get $i)))
     (local.set $len (i32.add (i32.const 0x10000) (local.get $len)))
+
     (loop
       (i32.and (local.get $crc) (i32.const 0xFF))
       (i32.load8_u (local.get $i))
@@ -42,6 +43,86 @@
     (i32.xor (local.get $crc) (i32.const -1))
     f64.convert_i32_u ;; return a positive Number
   )
+  (func $extendTable (local $crc i32) (local $i i32) (local $j i32)
+    (loop
+      (local.set $j (i32.const 0))
+      (loop
+        (i32.load (i32.or (local.get $j) (local.get $i)))
+        local.tee $crc
+
+        (i32.and (i32.const 0xFF))
+        (i32.shl (i32.const 2))
+        i32.load
+
+        (i32.shr_u  (local.get $crc) (i32.const 8))
+        (local.set $crc (i32.xor))
+
+        (local.tee $j (i32.add (local.get $j) (i32.const 0x400)))
+        (i32.or (local.get $i))
+
+        local.get $crc
+        i32.store
+        (br_if 0 (i32.ne (local.get $j) (i32.const 0xC00)))
+      )
+      (local.tee $i (i32.add (local.get $i) (i32.const 4)))
+      (br_if 0 (i32.ne (i32.const 0x400)))
+    ) ;; in total, the tables occupy the first 4 kB of the first mem page
+  )
+  (func $crc32x4 (param $len i32) (param $crc i32) (result f64) (local $i i32) (local $l8 i32) (local $v v128)
+    (local.set $crc (i32.xor (local.get $crc) (i32.const -1)))
+    (local.set $i (i32.const 0x10000))
+    (local.tee $len (i32.add (i32.const 0x10000) (local.get $len)))
+    i32.const 2
+    i32.shr_s
+    i32.const 2
+    i32.shl
+    local.tee $l8
+
+    (if (i32.ge_u (i32.const 0x10004))
+    (loop
+      (i32.xor (local.get $crc) (i32.load (local.get $i)))
+
+      i32x4.splat
+      v128.const i32x4 0xFFFFFF03 0xFFFFFF02 0xFFFFFF01 0xFFFFFF00
+      i8x16.swizzle ;; this was called 'v8x16.swizzle' if you have an older WABT
+
+      (i32x4.shl (i32.const 2))
+      v128.const i32x4 0x00000000 0x00000400 0x00000800 0x00000C00
+      v128.or
+      local.tee $v
+      (i32.load (i32x4.extract_lane 0))
+
+      local.get $v
+      (i32.load (i32x4.extract_lane 1))
+      i32.xor
+
+      local.get $v
+      (i32.load (i32x4.extract_lane 2))
+      i32.xor
+
+      local.get $v
+      (i32.load (i32x4.extract_lane 3))
+      i32.xor
+
+      local.set $crc ;; store the updated CRC
+      (local.tee $i (i32.add (local.get $i) (i32.const 4)))
+      (br_if 0 (i32.ne (local.get $l8)))
+    ))
+
+    (if (result f64) (i32.lt_u (local.get $i) (local.get $len))
+    (then
+      (call $crc32 
+        (i32.sub (local.get $len) (i32.const 0x10000))
+        (i32.xor (local.get $crc) (i32.const -1))
+        (i32.sub (local.get $i) (i32.const 0x10000))
+      )
+    ) (else
+      (i32.xor (local.get $crc) (i32.const -1))
+      f64.convert_i32_u ;; return a positive Number
+    ))
+  )
   (export "t" (func $genTable))
-  (export "c" (func $crc32))
-)
+  ;; (export "c" (func $crc32))
+  (export "c" (func $crc32x4))
+  (export "x" (func $extendTable))
+)

src/utils.ts

@@ -2,3 +2,8 @@ export const makeBuffer = (size: number) => new DataView(new ArrayBuffer(size))
 export const makeUint8Array = (thing: any) => new Uint8Array(thing.buffer || thing)
 export const clampInt32 = (n: bigint) => Math.min(0xffffffff, Number(n))
 export const clampInt16 = (n: bigint) => Math.min(0xffff, Number(n))
+
+export const parseWasm = (base64: string[]) => {
+  const chunks = base64.flatMap(s => Array.from(atob(s), c => c.charCodeAt(0)))
+  return new WebAssembly.Module(Uint8Array.of(...chunks))
+}

test/crc32.test.ts

@@ -4,8 +4,8 @@ import { crc32, memory } from "../src/crc32.ts"
 const table = await Deno.readFile("./test/table.array")
 
 Deno.test("the CRC32 module precomputes CRCs for each byte using the polynomial 0xEDB88320", () => {
-  const actual = new Uint8Array(memory.buffer.slice(0, 0x0400))
-  const expected = table.slice(0, 0x400)
+  const actual = new Uint8Array(memory.buffer).subarray(0, 0x1000)
+  const expected = table.subarray(0, 0x1000)
   assertEquals(actual, expected)
 })