Add base.x86_sse42_utility

This commit is contained in:
Nigel Tao committed 2021-02-18 20:02:58 +11:00
1 parent 492068f3dc
commit 8c8f534a5c
5 files changed
+90 -70

No files matched your search

+17 -17
View File
@@ -23,6 +23,7 @@ pri func hasher.up_x86_sse42!(x: slice base.u8),
// The remaining variables are specific to the SIMD version.
var util : base.x86_sse42_utility
var zeroes : base.x86_m128i
var ones : base.x86_m128i
var weights__left : base.x86_m128i
@@ -38,21 +39,20 @@ pri func hasher.up_x86_sse42!(x: slice base.u8),
var tail_index : base.u64
// zeroes and ones are uniform u16×8 vectors.
zeroes = zeroes.create_mm_set1_epi16(a: 0)
ones = ones.create_mm_set1_epi16(a: 1)
zeroes = util.make_m128i_repeat_u16(a: 0)
ones = util.make_m128i_repeat_u16(a: 1)
// weights__left and weights_right, little-endian, form the sequence 32,
// 31, 30, ..., 1.
weights__left = weights__left.create_mm_set_epi8(
e15: 0x11, e14: 0x12, e13: 0x13, e12: 0x14,
e11: 0x15, e10: 0x16, e9: 0x17, e8: 0x18,
e7: 0x19, e6: 0x1A, e5: 0x1B, e4: 0x1C,
e3: 0x1D, e2: 0x1E, e1: 0x1F, e0: 0x20)
weights_right = weights_right.create_mm_set_epi8(
e15: 0x01, e14: 0x02, e13: 0x03, e12: 0x04,
e11: 0x05, e10: 0x06, e9: 0x07, e8: 0x08,
e7: 0x09, e6: 0x0A, e5: 0x0B, e4: 0x0C,
e3: 0x0D, e2: 0x0E, e1: 0x0F, e0: 0x10)
// weights__left and weights_right form the sequence 32, 31, 30, ..., 1.
weights__left = util.make_m128i_multiple_u8(
a00: 0x20, a01: 0x1F, a02: 0x1E, a03: 0x1D,
a04: 0x1C, a05: 0x1B, a06: 0x1A, a07: 0x19,
a08: 0x18, a09: 0x17, a10: 0x16, a11: 0x15,
a12: 0x14, a13: 0x13, a14: 0x12, a15: 0x11)
weights_right = util.make_m128i_multiple_u8(
a00: 0x10, a01: 0x0F, a02: 0x0E, a03: 0x0D,
a04: 0x0C, a05: 0x0B, a06: 0x0A, a07: 0x09,
a08: 0x08, a09: 0x07, a10: 0x06, a11: 0x05,
a12: 0x04, a13: 0x03, a14: 0x02, a15: 0x01)
// Decompose this.state.
s1 = this.state.low_bits(n: 16)
@@ -86,9 +86,9 @@ pri func hasher.up_x86_sse42!(x: slice base.u8),
// variables s1 and s2. The iterate loop accumulates four parallel u32
// sums in each vector. A post-iterate step merges the four u32 sums
// into a single u32 sum.
v1 = v1.create_mm_setzero_si128()
v2j = v2j.create_mm_setzero_si128()
v2k = v2k.create_mm_setzero_si128()
v1 = util.make_m128i_zeroes()
v2j = util.make_m128i_zeroes()
v2k = util.make_m128i_zeroes()
// The inner loop.
iterate (p = args.x)(length: 32, advance: 32, unroll: 1) {
+3 -2
View File
@@ -78,6 +78,7 @@ pri func decoder.filter_3_distance_4_x86_sse42!(curr: slice base.u8, prev: slice
var curr : slice base.u8
var prev : slice base.u8
var util : base.x86_sse42_utility
var x128 : base.x86_m128i
var a128 : base.x86_m128i
var b128 : base.x86_m128i
@@ -85,7 +86,7 @@ pri func decoder.filter_3_distance_4_x86_sse42!(curr: slice base.u8, prev: slice
var k128 : base.x86_m128i
if args.prev.length() == 0 {
k128 = k128.create_mm_set1_epi8(a: 0xFE)
k128 = util.make_m128i_repeat_u8(a: 0xFE)
iterate (curr = args.curr)(length: 4, advance: 4, unroll: 2) {
// The predictor, p128, is just half (rounded down) of the previous
// pixel, a128. In this branch, b128 stays zero so the average of
@@ -102,7 +103,7 @@ pri func decoder.filter_3_distance_4_x86_sse42!(curr: slice base.u8, prev: slice
}
} else {
k128 = k128.create_mm_set1_epi8(a: 0x01)
k128 = util.make_m128i_repeat_u8(a: 0x01)
iterate (curr = args.curr, prev = args.prev)(length: 4, advance: 4, unroll: 2) {
// Load the pixel from the row above.
b128.load_u32!(a: prev.peek_u32le())