Roll back 'png filter_1_distance_? use more SIMD'

See the previous commit for the rationale.
This commit is contained in:
Nigel Tao committed 2021-01-25 10:08:33 +11:00
1 parent 1660f92686
commit 7de3742121
3 files changed
+35 -157

No files matched your search

+22 -81
View File
@@ -16,95 +16,36 @@
// Filter 1: Sub.
pri func decoder.filter_1_distance_3_sse42!(curr: slice base.u8),
choose cpu_arch >= x86_sse42,
{
var c : slice base.u8
var x128 : base.x86_m128i
var i128 : base.x86_m128i
var j128 : base.x86_m128i
var k128 : base.x86_m128i
var a128 : base.x86_m128i
iterate (c = args.curr)(length: 16, advance: 12, unroll: 1) {
// For distance_3, we only use the first 12 bytes (96 bits) of the
// 128-bit registers. The final 4 bytes (32 bits) are usually junk:
// - [R0 G0 B0 R1 G1 B1 R2 G2 B2 R3 G3 B3 ; junkC junkD junkE junkF]
// For x128, the final bytes are the upcoming pixels:
// - [R0 G0 B0 R1 G1 B1 R2 G2 B2 R3 G3 B3 ; R4 G4 B4 R5]
// For a128, the final bytes are zeroes.
// x128 = [x0, x1, x2, x3; upcoming]
x128.load_slice128!(a: c)
// i128 = [a+x0, x1, x2, x3; junk]
i128 = x128._mm_add_epi8(b: a128)
// j128 = [0, a+x0, x1, x2; junk]
j128 = i128._mm_slli_si128(imm8: 3)
// j128 = [a+x0, a+x0+x1, x1+x2, x2+x3; junk]
j128 = j128._mm_add_epi8(b: i128)
// k128 = [0, 0, a+x0, a+x0+x1; junk]
k128 = j128._mm_slli_si128(imm8: 6)
// k128 = [a+x0, a+x0+x1, a+x0+x1+x2, a+x0+x1+x2+x3; junk]
k128 = k128._mm_add_epi8(b: j128)
// k128 = [a+x0, a+x0+x1, a+x0+x1+x2, a+x0+x1+x2+x3; upcoming]
k128 = k128._mm_blend_epi16(b: x128, imm8: 0xC0)
// Store.
k128.store_slice128!(a: c)
// a128 = [a+x0+x1+x2+x3, 0, 0, 0; zeroes]
a128 = k128._mm_slli_si128(imm8: 4)._mm_srli_si128(imm8: 13)
} else (length: 3, advance: 3, unroll: 1) {
x128.load_u32!(a: c.peek_u24le_as_u32())
x128 = x128._mm_add_epi8(b: a128)
a128 = x128
c.poke_u24le!(a: x128.truncate_u32())
}
}
// This (filter = 1, distance = 3) implementation doesn't actually bench faster
// than the non-SIMD one.
//
// pri func decoder.filter_1_distance_3_sse42!(curr: slice base.u8),
// choose cpu_arch >= x86_sse42,
// {
// var c : slice base.u8
// var x128 : base.x86_m128i
// var a128 : base.x86_m128i
//
// iterate (c = args.curr)(length: 4, advance: 3, unroll: 2) {
// x128.load_u32!(a: c.peek_u32le())
// x128 = x128._mm_add_epi8(b: a128)
// a128 = x128
// c.poke_u24le!(a: x128.truncate_u32())
// } else (length: 3, advance: 3, unroll: 1) {
// x128.load_u32!(a: c.peek_u24le_as_u32())
// x128 = x128._mm_add_epi8(b: a128)
// c.poke_u24le!(a: x128.truncate_u32())
// }
// }
pri func decoder.filter_1_distance_4_sse42!(curr: slice base.u8),
choose cpu_arch >= x86_sse42,
{
var c : slice base.u8
var x128 : base.x86_m128i
var i128 : base.x86_m128i
var j128 : base.x86_m128i
var k128 : base.x86_m128i
var a128 : base.x86_m128i
iterate (c = args.curr)(length: 16, advance: 16, unroll: 1) {
// x128 = [x0, x1, x2, x3]
x128.load_slice128!(a: c)
// i128 = [a+x0, x1, x2, x3]
i128 = x128._mm_add_epi8(b: a128)
// j128 = [0, a+x0, x1, x2]
j128 = i128._mm_slli_si128(imm8: 4)
// j128 = [a+x0, a+x0+x1, x1+x2, x2+x3]
j128 = j128._mm_add_epi8(b: i128)
// k128 = [0, 0, a+x0, a+x0+x1]
k128 = j128._mm_slli_si128(imm8: 8)
// k128 = [a+x0, a+x0+x1, a+x0+x1+x2, a+x0+x1+x2+x3]
k128 = k128._mm_add_epi8(b: j128)
// Store.
k128.store_slice128!(a: c)
// a128 = [a+x0+x1+x2+x3, 0, 0, 0]
a128 = k128._mm_srli_si128(imm8: 12)
} else (length: 4, advance: 4, unroll: 1) {
iterate (c = args.curr)(length: 4, advance: 4, unroll: 2) {
x128.load_u32!(a: c.peek_u32le())
x128 = x128._mm_add_epi8(b: a128)
a128 = x128
+1 -3
View File
@@ -310,9 +310,7 @@ pri func decoder.choose_filter_implementations!() {
// Filter 0 is a no-op. Filter 2, the up filter, should already vectorize
// easily by a good optimizing C compiler.
if this.filter_distance == 3 {
choose filter_1 = [
filter_1_distance_3_sse42,
filter_1_distance_3_fallback]
choose filter_1 = [filter_1_distance_3_fallback]
choose filter_3 = [filter_3_distance_3_fallback]
choose filter_4 = [
filter_4_distance_3_sse42,