Electron
/
electron
зеркало из https://github.com/electron/electron.git
1
0
Форкнуть 0
Код Задачи Пакеты Проекты Релизы Вики Активность

3108669: arm,dsp: Fix 8bpp Dct64_NEON(). 

https://chromium-review.googlesource.com/c/codecs/libgav1/+/3108669
This commit is contained in:
John Kleinschmidt 2021-08-23 10:53:23 -04:00
Родитель 4cfc7299a5
Коммит 7ed3132312
Не найден ключ, соответствующий данной подписи
Идентификатор ключа GPG: F141756BCD08BB4D
3 изменённых файлов: 102 добавлений и 1 удалений
Показать статистику Скачать Patch файл Скачать Diff файл Показать все файлы Свернуть все файлы

4
patches/config.json
Просмотреть файл

@ -13,5 +13,7 @@
"src/electron/patches/Mantle": "src/third_party/squirrel.mac/vendor/Mantle",
"src/electron/patches/ReactiveObjC": "src/third_party/squirrel.mac/vendor/ReactiveObjC"
"src/electron/patches/ReactiveObjC": "src/third_party/squirrel.mac/vendor/ReactiveObjC",
"src/electron/patches/libgav1": "src/third_party/libgav1/src"
}

1
patches/libgav1/.patches Normal file
Просмотреть файл

@ -0,0 +1 @@
arm_dsp_fix_8bpp_dct64_neon.patch

98
patches/libgav1/arm_dsp_fix_8bpp_dct64_neon.patch Normal file
Просмотреть файл

@ -0,0 +1,98 @@
From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
From: Scott LaVarnway <slavarnway@google.com>
Date: Sat, 14 Aug 2021 09:35:39 -0700
Subject: arm,dsp: Fix 8bpp Dct64_NEON().
When building with clang (-march=armv8.2), an "off by one" error was detected. For now,
disable the fast butterfly.
PiperOrigin-RevId: 390805568
Change-Id: I9749a665b8235ec0f0dcbca46b284223c95af2a9
diff --git a/src/dsp/arm/inverse_transform_neon.cc b/src/dsp/arm/inverse_transform_neon.cc
index 4af8cf7ff68810016cc777719f84888fe00abb1f..0371cd54b3d0f64040a67d8c776270637137134b 100644
--- a/src/dsp/arm/inverse_transform_neon.cc
+++ b/src/dsp/arm/inverse_transform_neon.cc
@@ -388,6 +388,33 @@ LIBGAV1_ALWAYS_INLINE void ButterflyRotation_FirstIsZero(int16x8_t* a,
int16x8_t* b,
const int angle,
const bool flip) {
+#if defined(__ARM_FEATURE_QRDMX) && defined(__aarch64__) && \
+ defined(__clang__) // ARM v8.1-A
+ // Clang optimizes vqrdmulhq_n_s16 and vqsubq_s16 (in HadamardRotation) into
+ // vqrdmlshq_s16 resulting in an "off by one" error. For now, do not use
+ // vqrdmulhq_n_s16().
+ const int16_t cos128 = Cos128(angle);
+ const int16_t sin128 = Sin128(angle);
+ const int32x4_t x0 = vmull_n_s16(vget_low_s16(*b), -sin128);
+ const int32x4_t y0 = vmull_n_s16(vget_low_s16(*b), cos128);
+ const int16x4_t x1 = vqrshrn_n_s32(x0, 12);
+ const int16x4_t y1 = vqrshrn_n_s32(y0, 12);
+
+ const int32x4_t x0_hi = vmull_n_s16(vget_high_s16(*b), -sin128);
+ const int32x4_t y0_hi = vmull_n_s16(vget_high_s16(*b), cos128);
+ const int16x4_t x1_hi = vqrshrn_n_s32(x0_hi, 12);
+ const int16x4_t y1_hi = vqrshrn_n_s32(y0_hi, 12);
+
+ const int16x8_t x = vcombine_s16(x1, x1_hi);
+ const int16x8_t y = vcombine_s16(y1, y1_hi);
+ if (flip) {
+ *a = y;
+ *b = x;
+ } else {
+ *a = x;
+ *b = y;
+ }
+#else
const int16_t cos128 = Cos128(angle);
const int16_t sin128 = Sin128(angle);
// For this function, the max value returned by Sin128() is 4091, which fits
@@ -403,12 +430,40 @@ LIBGAV1_ALWAYS_INLINE void ButterflyRotation_FirstIsZero(int16x8_t* a,
*a = x;
*b = y;
}
+#endif
}
LIBGAV1_ALWAYS_INLINE void ButterflyRotation_SecondIsZero(int16x8_t* a,
int16x8_t* b,
const int angle,
const bool flip) {
+#if defined(__ARM_FEATURE_QRDMX) && defined(__aarch64__) && \
+ defined(__clang__) // ARM v8.1-A
+ // Clang optimizes vqrdmulhq_n_s16 and vqsubq_s16 (in HadamardRotation) into
+ // vqrdmlshq_s16 resulting in an "off by one" error. For now, do not use
+ // vqrdmulhq_n_s16().
+ const int16_t cos128 = Cos128(angle);
+ const int16_t sin128 = Sin128(angle);
+ const int32x4_t x0 = vmull_n_s16(vget_low_s16(*a), cos128);
+ const int32x4_t y0 = vmull_n_s16(vget_low_s16(*a), sin128);
+ const int16x4_t x1 = vqrshrn_n_s32(x0, 12);
+ const int16x4_t y1 = vqrshrn_n_s32(y0, 12);
+
+ const int32x4_t x0_hi = vmull_n_s16(vget_high_s16(*a), cos128);
+ const int32x4_t y0_hi = vmull_n_s16(vget_high_s16(*a), sin128);
+ const int16x4_t x1_hi = vqrshrn_n_s32(x0_hi, 12);
+ const int16x4_t y1_hi = vqrshrn_n_s32(y0_hi, 12);
+
+ const int16x8_t x = vcombine_s16(x1, x1_hi);
+ const int16x8_t y = vcombine_s16(y1, y1_hi);
+ if (flip) {
+ *a = y;
+ *b = x;
+ } else {
+ *a = x;
+ *b = y;
+ }
+#else
const int16_t cos128 = Cos128(angle);
const int16_t sin128 = Sin128(angle);
const int16x8_t x = vqrdmulhq_n_s16(*a, cos128 << 3);
@@ -420,6 +475,7 @@ LIBGAV1_ALWAYS_INLINE void ButterflyRotation_SecondIsZero(int16x8_t* a,
*a = x;
*b = y;
}
+#endif
}
LIBGAV1_ALWAYS_INLINE void HadamardRotation(int16x8_t* a, int16x8_t* b,