AuroraMiddleware/skia2
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

somewhat less silly tail loads and stores

Browse Source 
No reason to keep going one at a time when we know there are generally
better ways to handle loading a power-of-two number of low lanes.

This strategy scales up too, with quick answers for 8 (one 8 byte load),
12 (one 8 byte, one 4 byte), etc.

$ ninja -C out monobench; and out/monobench SkRasterPipeline_compile 300

    Before: 46.946ns
    After:  43.341ns

(This happens to be _lowp.  Expect similar small speedups elsewhere.)

Change-Id: I08f87769ea3c9f06ad13d2b1d5326e542b9b63a8
Reviewed-on: https://skia-review.googlesource.com/20903
Reviewed-by: Herb Derby <herb@google.com>
Commit-Queue: Mike Klein <mtklein@chromium.org>
This commit is contained in:
Mike Klein 2017-06-26 16:12:48 -04:00 committed by Skia Commit-Bot
parent 660cc9910f
commit c4fcbed6b2
4 changed files with 5946 additions and 5589 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5909
src/jumper/SkJumper_generated.S
View File

File diff suppressed because it is too large Load Diff

5562
src/jumper/SkJumper_generated_win.S
View File

File diff suppressed because it is too large Load Diff

32
src/jumper/SkJumper_stages.cpp
View File

@ -94,14 +94,14 @@ SI V load(const T* src, size_t tail) {
__builtin_assume(tail < kStride);
if (__builtin_expect(tail, 0)) {
V v{}; // Any inactive lanes are zeroed.
switch (tail-1) {
case 6: v[6] = src[6];
case 5: v[5] = src[5];
case 4: v[4] = src[4];
case 3: v[3] = src[3];
case 2: v[2] = src[2];
case 1: v[1] = src[1];
case 0: v[0] = src[0];
switch (tail) {
case 7: v[6] = src[6];
case 6: v[5] = src[5];
case 5: v[4] = src[4];
case 4: memcpy(&v, src, 4*sizeof(T)); break;
case 3: v[2] = src[2];
case 2: memcpy(&v, src, 2*sizeof(T)); break;
case 1: memcpy(&v, src, 1*sizeof(T)); break;
}
return v;
}
@ -114,14 +114,14 @@ SI void store(T* dst, V v, size_t tail) {
#if defined(JUMPER)
__builtin_assume(tail < kStride);
if (__builtin_expect(tail, 0)) {
switch (tail-1) {
case 6: dst[6] = v[6];
case 5: dst[5] = v[5];
case 4: dst[4] = v[4];
case 3: dst[3] = v[3];
case 2: dst[2] = v[2];
case 1: dst[1] = v[1];
case 0: dst[0] = v[0];
switch (tail) {
case 7: dst[6] = v[6];
case 6: dst[5] = v[5];
case 5: dst[4] = v[4];
case 4: memcpy(dst, &v, 4*sizeof(T)); break;
case 3: dst[2] = v[2];
case 2: memcpy(dst, &v, 2*sizeof(T)); break;
case 1: memcpy(dst, &v, 1*sizeof(T)); break;
}
return;
}

32
src/jumper/SkJumper_stages_lowp.cpp
View File

@ -87,14 +87,14 @@ SI V load(const T* src, size_t tail) {
__builtin_assume(tail < kStride);
if (__builtin_expect(tail, 0)) {
V v{}; // Any inactive lanes are zeroed.
switch (tail-1) {
case 6: v[6] = src[6];
case 5: v[5] = src[5];
case 4: v[4] = src[4];
case 3: v[3] = src[3];
case 2: v[2] = src[2];
case 1: v[1] = src[1];
case 0: v[0] = src[0];
switch (tail) {
case 7: v[6] = src[6];
case 6: v[5] = src[5];
case 5: v[4] = src[4];
case 4: memcpy(&v, src, 4*sizeof(T)); break;
case 3: v[2] = src[2];
case 2: memcpy(&v, src, 2*sizeof(T)); break;
case 1: memcpy(&v, src, 1*sizeof(T)); break;
}
return v;
}
@ -105,14 +105,14 @@ template <typename V, typename T>
SI void store(T* dst, V v, size_t tail) {
__builtin_assume(tail < kStride);
if (__builtin_expect(tail, 0)) {
switch (tail-1) {
case 6: dst[6] = v[6];
case 5: dst[5] = v[5];
case 4: dst[4] = v[4];
case 3: dst[3] = v[3];
case 2: dst[2] = v[2];
case 1: dst[1] = v[1];
case 0: dst[0] = v[0];
switch (tail) {
case 7: dst[6] = v[6];
case 6: dst[5] = v[5];
case 5: dst[4] = v[4];
case 4: memcpy(dst, &v, 4*sizeof(T)); break;
case 3: dst[2] = v[2];
case 2: memcpy(dst, &v, 2*sizeof(T)); break;
case 1: memcpy(dst, &v, 1*sizeof(T)); break;
}
return;
}