Unroll some std/png sse128 loops

wuffs_png_decode_filt_1_dist_3/clang9                     1.57GB/s ± 0%  1.57GB/s ± 0%     ~     (p=0.151 n=5+5)
wuffs_png_decode_filt_1_dist_4/clang9                     4.62GB/s ± 0%  5.87GB/s ± 1%  +26.93%  (p=0.008 n=5+5)
wuffs_png_decode_filt_2_dist_3/clang9                     13.3GB/s ± 0%  13.1GB/s ± 1%   -1.52%  (p=0.008 n=5+5)
wuffs_png_decode_filt_2_dist_4/clang9                     13.3GB/s ± 0%  13.1GB/s ± 1%   -1.50%  (p=0.008 n=5+5)
wuffs_png_decode_filt_3_dist_3/clang9                      910MB/s ± 0%   909MB/s ± 0%     ~     (p=0.095 n=5+5)
wuffs_png_decode_filt_3_dist_4/clang9                     1.78GB/s ± 0%  1.84GB/s ± 0%   +3.65%  (p=0.008 n=5+5)
wuffs_png_decode_filt_4_dist_3/clang9                      542MB/s ± 0%   571MB/s ± 0%   +5.31%  (p=0.016 n=4+5)
wuffs_png_decode_filt_4_dist_4/clang9                      844MB/s ± 0%   819MB/s ± 0%   -3.00%  (p=0.008 n=5+5)

wuffs_png_decode_image_40k_24bpp/clang9                    127MB/s ± 0%   127MB/s ± 0%     ~     (p=0.310 n=5+5)
wuffs_png_decode_image_552k_32bpp_ignore_checksum/clang9   326MB/s ± 0%   322MB/s ± 0%   -1.19%  (p=0.008 n=5+5)
wuffs_png_decode_image_552k_32bpp_verify_checksum/clang9   259MB/s ± 0%   256MB/s ± 0%   -1.31%  (p=0.008 n=5+5)
wuffs_png_decode_image_4002k_24bpp/clang9                  128MB/s ± 0%   128MB/s ± 0%     ~     (p=0.278 n=5+5)

wuffs_png_decode_filt_1_dist_3/gcc10                      1.84GB/s ± 0%  1.84GB/s ± 0%     ~     (p=0.190 n=4+5)
wuffs_png_decode_filt_1_dist_4/gcc10                      3.38GB/s ± 1%  5.41GB/s ± 1%  +59.98%  (p=0.008 n=5+5)
wuffs_png_decode_filt_2_dist_3/gcc10                      8.29GB/s ± 0%  8.30GB/s ± 0%     ~     (p=0.222 n=5+5)
wuffs_png_decode_filt_2_dist_4/gcc10                      8.28GB/s ± 0%  8.30GB/s ± 0%   +0.24%  (p=0.008 n=5+5)
wuffs_png_decode_filt_3_dist_3/gcc10                      1.07GB/s ± 0%  1.08GB/s ± 0%     ~     (p=0.421 n=5+5)
wuffs_png_decode_filt_3_dist_4/gcc10                      2.13GB/s ± 0%  2.36GB/s ± 0%  +10.74%  (p=0.008 n=5+5)
wuffs_png_decode_filt_4_dist_3/gcc10                       587MB/s ± 0%   616MB/s ± 0%   +4.95%  (p=0.016 n=4+5)
wuffs_png_decode_filt_4_dist_4/gcc10                       796MB/s ± 0%   871MB/s ± 0%   +9.44%  (p=0.008 n=5+5)

wuffs_png_decode_image_40k_24bpp/gcc10                     130MB/s ± 0%   131MB/s ± 0%   +0.78%  (p=0.008 n=5+5)
wuffs_png_decode_image_552k_32bpp_ignore_checksum/gcc10    320MB/s ± 0%   325MB/s ± 0%   +1.69%  (p=0.008 n=5+5)
wuffs_png_decode_image_552k_32bpp_verify_checksum/gcc10    270MB/s ± 0%   273MB/s ± 0%   +1.30%  (p=0.008 n=5+5)
wuffs_png_decode_image_4002k_24bpp/gcc10                   131MB/s ± 0%   131MB/s ± 0%   +0.40%  (p=0.008 n=5+5)
This commit is contained in:
Nigel Tao committed 2021-01-22 23:25:19 +11:00
1 parent 5d88f3cb37
commit d09788938b
2 files changed
+152 -13

No files matched your search

+6 -6
View File
@@ -26,7 +26,7 @@
// var x128 : base.sse128_i
// var a128 : base.sse128_i
//
// iterate (c = args.curr)(length: 4, advance: 3, unroll: 1) {
// 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
@@ -45,7 +45,7 @@ pri func decoder.filter_1_distance_4_sse128!(curr: slice base.u8),
var x128 : base.sse128_i
var a128 : base.sse128_i
iterate (c = args.curr)(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
@@ -79,7 +79,7 @@ pri func decoder.filter_3_distance_4_sse128!(curr: slice base.u8, prev: slice ba
if args.prev.length() == 0 {
k128 = k128.create_mm_set1_epi8(a: 0xFE)
iterate (c = args.curr)(length: 4, advance: 4, unroll: 1) {
iterate (c = 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
// a128 and b128 is just half of a128. _mm_avg_epu8 rounds up, but
@@ -96,7 +96,7 @@ pri func decoder.filter_3_distance_4_sse128!(curr: slice base.u8, prev: slice ba
} else {
k128 = k128.create_mm_set1_epi8(a: 0x01)
iterate (c = args.curr, p = args.prev)(length: 4, advance: 4, unroll: 1) {
iterate (c = args.curr, p = args.prev)(length: 4, advance: 4, unroll: 2) {
// Load the pixel from the row above.
b128.load_u32!(a: p.peek_u32le())
@@ -147,7 +147,7 @@ pri func decoder.filter_4_distance_3_sse128!(curr: slice base.u8, prev: slice ba
var z128 : base.sse128_i
// § The advance is 3, not 4.
iterate (c = args.curr, p = args.prev)(length: 4, advance: 3, unroll: 1) {
iterate (c = args.curr, p = args.prev)(length: 4, advance: 3, unroll: 2) {
b128.load_u32!(a: p.peek_u32le())
b128 = b128._mm_unpacklo_epi8(b: z128)
pa128 = b128._mm_sub_epi16(b: c128)
@@ -217,7 +217,7 @@ pri func decoder.filter_4_distance_4_sse128!(curr: slice base.u8, prev: slice ba
var smallest128 : base.sse128_i
var z128 : base.sse128_i
iterate (c = args.curr, p = args.prev)(length: 4, advance: 4, unroll: 1) {
iterate (c = args.curr, p = args.prev)(length: 4, advance: 4, unroll: 2) {
// Load the pixel from the row above.
b128.load_u32!(a: p.peek_u32le())