aboutsummaryrefslogtreecommitdiff
path: root/src/dsp/arm
diff options
context:
space:
mode:
Diffstat (limited to 'src/dsp/arm')
-rw-r--r--src/dsp/arm/convolve_10bit_neon.cc62
-rw-r--r--src/dsp/arm/convolve_neon.cc12
-rw-r--r--src/dsp/arm/film_grain_neon.cc18
-rw-r--r--src/dsp/arm/intrapred_directional_neon.cc2
-rw-r--r--src/dsp/arm/inverse_transform_neon.cc9
-rw-r--r--src/dsp/arm/loop_filter_10bit_neon.cc2
-rw-r--r--src/dsp/arm/loop_restoration_10bit_neon.cc14
-rw-r--r--src/dsp/arm/loop_restoration_neon.cc6
8 files changed, 59 insertions, 66 deletions
diff --git a/src/dsp/arm/convolve_10bit_neon.cc b/src/dsp/arm/convolve_10bit_neon.cc
index 389f029..1aa0cc7 100644
--- a/src/dsp/arm/convolve_10bit_neon.cc
+++ b/src/dsp/arm/convolve_10bit_neon.cc
@@ -412,30 +412,21 @@ void FilterHorizontal(const uint16_t* LIBGAV1_RESTRICT const src,
void* LIBGAV1_RESTRICT const dest,
const ptrdiff_t pred_stride, const int width,
const int height, const int16x4_t* const v_tap) {
- assert(width < 8 || num_taps != 4);
- // Don't simplify the redundant if conditions with the template parameters,
- // which helps the compiler generate compact code.
- if (width >= 8 && num_taps != 4) {
- FilterHorizontalWidth8AndUp<num_taps, is_compound, is_2d>(
- src, src_stride, dest, pred_stride, width, height, v_tap);
- return;
- }
-
// Horizontal passes only needs to account for number of taps 2 and 4 when
// |width| <= 4.
assert(width <= 4);
assert(num_taps == 2 || num_taps == 4);
if (num_taps == 2 || num_taps == 4) {
- if (width == 4) {
- FilterHorizontalWidth4<num_taps, is_compound, is_2d>(
- src, src_stride, dest, pred_stride, height, v_tap);
- return;
- }
- assert(width == 2);
- if (!is_compound) {
+ if (width == 2 && !is_compound) {
FilterHorizontalWidth2<num_taps, is_2d>(src, src_stride, dest,
pred_stride, height, v_tap);
+ return;
}
+ assert(width == 4);
+ FilterHorizontalWidth4<num_taps, is_compound, is_2d>(
+ src, src_stride, dest, pred_stride, height, v_tap);
+ } else {
+ assert(false);
}
}
@@ -454,19 +445,32 @@ LIBGAV1_ALWAYS_INLINE void DoHorizontalPass(
v_tap[k] = vdup_n_s16(kHalfSubPixelFilters[filter_index][filter_id][k]);
}
- if (filter_index == 2) { // 8 tap.
- FilterHorizontal<8, is_compound, is_2d>(src, src_stride, dst, dst_stride,
- width, height, v_tap);
- } else if (filter_index < 2) { // 6 tap.
- FilterHorizontal<6, is_compound, is_2d>(src + 1, src_stride, dst,
- dst_stride, width, height, v_tap);
- } else if ((filter_index & 0x4) != 0) { // 4 tap.
- // ((filter_index == 4) | (filter_index == 5))
- FilterHorizontal<4, is_compound, is_2d>(src + 2, src_stride, dst,
- dst_stride, width, height, v_tap);
- } else { // 2 tap.
- FilterHorizontal<2, is_compound, is_2d>(src + 3, src_stride, dst,
- dst_stride, width, height, v_tap);
+ // Horizontal filter.
+ // Filter types used for width <= 4 are different from those for width > 4.
+ // When width > 4, the valid filter index range is always [0, 3].
+ // When width <= 4, the valid filter index range is always [4, 5].
+ if (width >= 8) {
+ if (filter_index == 2) { // 8 tap.
+ FilterHorizontalWidth8AndUp<8, is_compound, is_2d>(
+ src, src_stride, dst, dst_stride, width, height, v_tap);
+ } else if (filter_index < 2) { // 6 tap.
+ FilterHorizontalWidth8AndUp<6, is_compound, is_2d>(
+ src + 1, src_stride, dst, dst_stride, width, height, v_tap);
+ } else { // 2 tap.
+ assert(filter_index == 3);
+ FilterHorizontalWidth8AndUp<2, is_compound, is_2d>(
+ src + 3, src_stride, dst, dst_stride, width, height, v_tap);
+ }
+ } else {
+ if ((filter_index & 0x4) != 0) { // 4 tap.
+ // ((filter_index == 4) | (filter_index == 5))
+ FilterHorizontal<4, is_compound, is_2d>(src + 2, src_stride, dst,
+ dst_stride, width, height, v_tap);
+ } else { // 2 tap.
+ assert(filter_index == 3);
+ FilterHorizontal<2, is_compound, is_2d>(src + 3, src_stride, dst,
+ dst_stride, width, height, v_tap);
+ }
}
}
diff --git a/src/dsp/arm/convolve_neon.cc b/src/dsp/arm/convolve_neon.cc
index 5b80da2..97b3f26 100644
--- a/src/dsp/arm/convolve_neon.cc
+++ b/src/dsp/arm/convolve_neon.cc
@@ -371,16 +371,14 @@ void FilterHorizontal(const uint8_t* LIBGAV1_RESTRICT const src,
assert(width <= 4);
assert(filter_index >= 3 && filter_index <= 5);
if (filter_index >= 3 && filter_index <= 5) {
- if (width == 4) {
- FilterHorizontalWidth4<filter_index, is_2d, is_compound>(
- src, src_stride, dest, pred_stride, height, v_tap);
- return;
- }
- assert(width == 2);
- if (!is_compound) {
+ if (width == 2 && !is_compound) {
FilterHorizontalWidth2<filter_index, is_2d>(src, src_stride, dest,
pred_stride, height, v_tap);
+ return;
}
+ assert(width == 4);
+ FilterHorizontalWidth4<filter_index, is_2d, is_compound>(
+ src, src_stride, dest, pred_stride, height, v_tap);
}
}
diff --git a/src/dsp/arm/film_grain_neon.cc b/src/dsp/arm/film_grain_neon.cc
index 76e1151..cde887c 100644
--- a/src/dsp/arm/film_grain_neon.cc
+++ b/src/dsp/arm/film_grain_neon.cc
@@ -682,26 +682,14 @@ inline int16x8_t Clip3(const int16x8_t value, const int16x8_t low,
template <int bitdepth, typename Pixel>
inline int16x8_t GetScalingFactors(const int16_t scaling_lut[],
- const Pixel* source) {
+ const Pixel* source,
+ const int valid_range = 8) {
int16_t start_vals[8];
static_assert(bitdepth <= kBitdepth10,
"NEON Film Grain is not yet implemented for 12bpp.");
#if LIBGAV1_MSAN
- memset(start_vals, 0, sizeof(start_vals));
+ if (valid_range < 8) memset(start_vals, 0, sizeof(start_vals));
#endif
- for (int i = 0; i < 8; ++i) {
- assert(source[i] < (kScalingLookupTableSize << (bitdepth - kBitdepth8)));
- start_vals[i] = scaling_lut[source[i]];
- }
- return vld1q_s16(start_vals);
-}
-
-template <int bitdepth, typename Pixel>
-inline int16x8_t GetScalingFactors(const int16_t scaling_lut[],
- const Pixel* source, const int valid_range) {
- int16_t start_vals[8];
- static_assert(bitdepth <= kBitdepth10,
- "NEON Film Grain is not yet implemented for 12bpp.");
for (int i = 0; i < valid_range; ++i) {
assert(source[i] < (kScalingLookupTableSize << (bitdepth - kBitdepth8)));
start_vals[i] = scaling_lut[source[i]];
diff --git a/src/dsp/arm/intrapred_directional_neon.cc b/src/dsp/arm/intrapred_directional_neon.cc
index e9bdcf0..d36ef5f 100644
--- a/src/dsp/arm/intrapred_directional_neon.cc
+++ b/src/dsp/arm/intrapred_directional_neon.cc
@@ -1752,7 +1752,7 @@ inline void DirectionalZone2FromLeftCol_8x8(
const int index_scale_bits = 6;
// The values in |offset_y| are negative, except for the first element, which
// is zero.
- int16x8_t offset_y = left_y;
+ int16x8_t offset_y;
int16x8_t shift_upsampled = left_y;
// The shift argument must be a constant, otherwise use upsample_shift
// directly.
diff --git a/src/dsp/arm/inverse_transform_neon.cc b/src/dsp/arm/inverse_transform_neon.cc
index 452f14a..cc4e4a4 100644
--- a/src/dsp/arm/inverse_transform_neon.cc
+++ b/src/dsp/arm/inverse_transform_neon.cc
@@ -345,11 +345,12 @@ LIBGAV1_ALWAYS_INLINE void ButterflyRotation_FirstIsZero(int16x8_t* a,
int16x8_t* b,
const int angle,
const bool flip) {
+ // Clang < 14 targeting armv8.1-a+ optimizes vqrdmulhq_n_s16 and vqsubq_s16
+ // (in HadamardRotation) into vqrdmlshq_s16 resulting in an "off by one"
+ // error. This behavior was fixed in 14.0.0:
+ // https://github.com/llvm/llvm-project/commit/82973edfb72a95b442fa6d2bb404e15a4031855e
#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().
+ defined(__clang__) && __clang_major__ < 14
const int16_t cos128 = Cos128(angle);
const int16_t sin128 = Sin128(angle);
const int32x4_t x0 = vmull_n_s16(vget_low_s16(*b), -sin128);
diff --git a/src/dsp/arm/loop_filter_10bit_neon.cc b/src/dsp/arm/loop_filter_10bit_neon.cc
index a9dd98f..abdc074 100644
--- a/src/dsp/arm/loop_filter_10bit_neon.cc
+++ b/src/dsp/arm/loop_filter_10bit_neon.cc
@@ -444,7 +444,6 @@ void Horizontal6_NEON(void* const dest, const ptrdiff_t stride,
const uint64x1_t need_filter6 = vreinterpret_u64_u16(is_flat3_mask);
if (vget_lane_u64(need_filter6, 0) == 0) {
// Filter6() does not apply, but Filter4() applies to one or more values.
- p0q0_output = p0q0;
p1q1_output = vbslq_u16(needs_filter_mask_8, f4_p1q1, p1q1);
p0q0_output = vbslq_u16(needs_filter_mask_8, f4_p0q0, p0q0);
} else {
@@ -526,7 +525,6 @@ void Vertical6_NEON(void* const dest, const ptrdiff_t stride, int outer_thresh,
const uint64x1_t need_filter6 = vreinterpret_u64_u16(is_flat3_mask);
if (vget_lane_u64(need_filter6, 0) == 0) {
// Filter6() does not apply, but Filter4() applies to one or more values.
- p0q0_output = p0q0;
p1q1_output = vbslq_u16(needs_filter_mask_8, f4_p1q1, p1q1);
p0q0_output = vbslq_u16(needs_filter_mask_8, f4_p0q0, p0q0);
} else {
diff --git a/src/dsp/arm/loop_restoration_10bit_neon.cc b/src/dsp/arm/loop_restoration_10bit_neon.cc
index 410bc20..9191080 100644
--- a/src/dsp/arm/loop_restoration_10bit_neon.cc
+++ b/src/dsp/arm/loop_restoration_10bit_neon.cc
@@ -1130,7 +1130,13 @@ inline void LookupIntermediate(const uint16x8_t sum, const uint16x8_t index,
const uint8x8_t idx = vqmovn_u16(index);
uint8_t temp[8];
vst1_u8(temp, idx);
- *ma = vsetq_lane_u8(kSgrMaLookup[temp[0]], *ma, offset + 0);
+ // offset == 0 is assumed to be the first call to this function. The value is
+ // duplicated to avoid -Wuninitialized warnings under gcc.
+ if (offset == 0) {
+ *ma = vdupq_n_u8(kSgrMaLookup[temp[0]]);
+ } else {
+ *ma = vsetq_lane_u8(kSgrMaLookup[temp[0]], *ma, offset + 0);
+ }
*ma = vsetq_lane_u8(kSgrMaLookup[temp[1]], *ma, offset + 1);
*ma = vsetq_lane_u8(kSgrMaLookup[temp[2]], *ma, offset + 2);
*ma = vsetq_lane_u8(kSgrMaLookup[temp[3]], *ma, offset + 3);
@@ -1712,8 +1718,6 @@ LIBGAV1_ALWAYS_INLINE void BoxSumFilterPreProcess3(
s[0] = Load1QMsanU16(src + 0, overread_in_bytes + 0);
s[1] = Load1QMsanU16(src + 8, overread_in_bytes + 16);
Square(s[0], sq);
- // Quiet "may be used uninitialized" warning.
- mas[0] = mas[1] = vdupq_n_u8(0);
BoxFilterPreProcess3Lo(s, scale, sum3, square_sum3, sq, &mas[0], bs);
int x = 0;
@@ -2067,8 +2071,6 @@ LIBGAV1_ALWAYS_INLINE void BoxFilterPass2(
s[0] = Load1QMsanU16(src0 + 0, overread_in_bytes + 0);
s[1] = Load1QMsanU16(src0 + 8, overread_in_bytes + 16);
Square(s[0], sq);
- // Quiet "may be used uninitialized" warning.
- mas[0] = mas[1] = vdupq_n_u8(0);
BoxFilterPreProcess3Lo(s, scale, sum3, square_sum3, sq, &mas[0], bs);
int x = 0;
@@ -2255,8 +2257,6 @@ inline void BoxFilterLastRow(
s[0] = Load1QMsanU16(src0 + 0, overread_in_bytes + 0);
s[1] = Load1QMsanU16(src0 + 8, overread_in_bytes + 16);
Square(s[0], sq);
- // Quiet "may be used uninitialized" warning.
- ma3[0] = ma3[1] = vdupq_n_u8(0);
BoxFilterPreProcessLastRowLo(s, scales, sum3, sum5, square_sum3, square_sum5,
sq, &ma3[0], &ma5[0], b3, b5);
diff --git a/src/dsp/arm/loop_restoration_neon.cc b/src/dsp/arm/loop_restoration_neon.cc
index cd8552e..adb8f36 100644
--- a/src/dsp/arm/loop_restoration_neon.cc
+++ b/src/dsp/arm/loop_restoration_neon.cc
@@ -1125,7 +1125,11 @@ inline void CalculateIntermediate(const uint16x8_t sum,
val = AdjustValue(val, idx, 101); // 101 is the last index which value is 3.
val = AdjustValue(val, idx, 169); // 169 is the last index which value is 2.
val = AdjustValue(val, idx, 254); // 254 is the last index which value is 1.
- *ma = (offset == 0) ? vcombine_u8(val, vget_high_u8(*ma))
+ // offset == 0 is assumed to be the first call to this function. Note
+ // vget_high_u8(*ma) is not used in this case to avoid a -Wuninitialized
+ // warning with some versions of gcc. vdup_n_u8(0) could work as well, but in
+ // most cases clang and gcc generated better code with this version.
+ *ma = (offset == 0) ? vcombine_u8(val, val)
: vcombine_u8(vget_low_u8(*ma), val);
// b = ma * b * one_over_n
generated by cgit v1.2.3 (git 2.39.1) at 2025-04-07 14:47:27 +0000