Use widening instructions in Neon z1 predictors
Using vmlal and vsubl allows us to avoid the need for a separate
widening step (in this case by vzip'ing with zero for an unsigned
extension).
While we are here, also take the opportunity to clean up a bunch of
variable declarations.
Change-Id: I32ed1cf5e4f3912c539b84ee513fb9d6ae46e8f5
diff --git a/aom_dsp/arm/intrapred_neon.c b/aom_dsp/arm/intrapred_neon.c
index f66f048..ba17f8a 100644
--- a/aom_dsp/arm/intrapred_neon.c
+++ b/aom_dsp/arm/intrapred_neon.c
@@ -1163,21 +1163,12 @@
// final pixels will be calculated as:
// (above[x] * 32 + 16 + (above[x+1] - above[x]) * shift) >> 5
- uint16x8_t a0, a1;
- uint16x8_t diff, a32;
- uint16x8_t a16;
- uint8x8_t a_mbase_x;
-
- a16 = vdupq_n_u16(16);
- a_mbase_x = vdup_n_u8(above[max_base_x]);
- uint16x8_t v_32 = vdupq_n_u16(32);
+ const uint16x8_t a16 = vdupq_n_u16(16);
+ const uint8x8_t a_mbase_x = vdup_n_u8(above[max_base_x]);
+ const uint8x8_t v_32 = vdup_n_u8(32);
int x = dx;
for (int r = 0; r < W; r++) {
- uint16x8_t res;
- uint16x8_t shift;
- uint8x8x2_t v_tmp_a0_128;
-
int base = x >> frac_bits;
int base_max_diff = (max_base_x - base) >> upsample_above;
if (base_max_diff <= 0) {
@@ -1189,19 +1180,19 @@
if (base_max_diff > H) base_max_diff = H;
+ uint8x8x2_t a01_128;
+ uint16x8_t shift;
if (upsample_above) {
- v_tmp_a0_128 = vld2_u8(above + base);
+ a01_128 = vld2_u8(above + base);
shift = vdupq_n_u16(((x << upsample_above) & 0x3f) >> 1);
} else {
- v_tmp_a0_128.val[0] = vld1_u8(above + base);
- v_tmp_a0_128.val[1] = vld1_u8(above + base + 1);
+ a01_128.val[0] = vld1_u8(above + base);
+ a01_128.val[1] = vld1_u8(above + base + 1);
shift = vdupq_n_u16((x & 0x3f) >> 1);
}
- a0 = vmovl_u8(v_tmp_a0_128.val[0]);
- a1 = vmovl_u8(v_tmp_a0_128.val[1]);
- diff = vsubq_u16(a1, a0); // a[x+1] - a[x]
- a32 = vmlaq_u16(a16, a0, v_32); // a[x] * 32 + 16
- res = vmlaq_u16(a32, diff, shift);
+ uint16x8_t diff = vsubl_u8(a01_128.val[1], a01_128.val[0]);
+ uint16x8_t a32 = vmlal_u8(a16, a01_128.val[0], v_32);
+ uint16x8_t res = vmlaq_u16(a32, diff, shift);
uint8x8_t mask = vld1_u8(BaseMask[base_max_diff]);
dst[r] = vbsl_u8(mask, vshrn_n_u16(res, 5), a_mbase_x);
@@ -1249,13 +1240,10 @@
// final pixels will be calculated as:
// (above[x] * 32 + 16 + (above[x+1] - above[x]) * shift) >> 5
- uint8x16x2_t a0, a1;
- uint16x8x2_t diff, a32;
-
- uint16x8_t a16 = vdupq_n_u16(16);
- uint8x16_t a_mbase_x = vdupq_n_u8(above[max_base_x]);
- uint16x8_t v_32 = vdupq_n_u16(32);
- uint8x16_t v_zero = vdupq_n_u8(0);
+ const uint16x8_t a16 = vdupq_n_u16(16);
+ const uint8x16_t a_mbase_x = vdupq_n_u8(above[max_base_x]);
+ const uint8x8_t v_32 = vdup_n_u8(32);
+ const uint8x16_t v_zero = vdupq_n_u8(0);
int x = dx;
for (int r = 0; r < W; r++) {
@@ -1284,16 +1272,11 @@
a1_128 = vld1q_u8(above + base + 1);
shift = vdupq_n_u16((x & 0x3f) >> 1);
}
- a0 = vzipq_u8(a0_128, v_zero);
- a1 = vzipq_u8(a1_128, v_zero);
- diff.val[0] = vsubq_u16(vreinterpretq_u16_u8(a1.val[0]),
- vreinterpretq_u16_u8(a0.val[0])); // a[x+1] - a[x]
- diff.val[1] = vsubq_u16(vreinterpretq_u16_u8(a1.val[1]),
- vreinterpretq_u16_u8(a0.val[1])); // a[x+1] - a[x]
- a32.val[0] = vmlaq_u16(a16, vreinterpretq_u16_u8(a0.val[0]),
- v_32); // a[x] * 32 + 16
- a32.val[1] = vmlaq_u16(a16, vreinterpretq_u16_u8(a0.val[1]),
- v_32); // a[x] * 32 + 16
+ uint16x8x2_t diff, a32;
+ diff.val[0] = vsubl_u8(vget_low_u8(a1_128), vget_low_u8(a0_128));
+ diff.val[1] = vsubl_u8(vget_high_u8(a1_128), vget_high_u8(a0_128));
+ a32.val[0] = vmlal_u8(a16, vget_low_u8(a0_128), v_32);
+ a32.val[1] = vmlal_u8(a16, vget_high_u8(a0_128), v_32);
res.val[0] = vmlaq_u16(a32.val[0], diff.val[0], shift);
res.val[1] = vmlaq_u16(a32.val[1], diff.val[1], shift);
uint8x16_t v_temp =
@@ -1332,20 +1315,13 @@
// final pixels will be calculated as:
// (above[x] * 32 + 16 + (above[x+1] - above[x]) * shift) >> 5
- uint8x16_t a_mbase_x;
- uint8x16x2_t a0, a1;
- uint16x8x2_t diff, a32;
-
- a_mbase_x = vdupq_n_u8(above[max_base_x]);
- uint16x8_t a16 = vdupq_n_u16(16);
- uint16x8_t v_32 = vdupq_n_u16(32);
- uint8x16_t v_zero = vdupq_n_u8(0);
+ const uint8x16_t a_mbase_x = vdupq_n_u8(above[max_base_x]);
+ const uint16x8_t a16 = vdupq_n_u16(16);
+ const uint8x8_t v_32 = vdup_n_u8(32);
int x = dx;
for (int r = 0; r < N; r++) {
- uint16x8x2_t res;
uint8x16_t res16[2];
- uint8x16_t a0_128, a1_128;
int base = x >> frac_bits;
int base_max_diff = (max_base_x - base);
@@ -1365,20 +1341,14 @@
if (mdiff <= 0) {
res16[jj] = a_mbase_x;
} else {
+ uint16x8x2_t a32, diff, res;
+ uint8x16_t a0_128, a1_128;
a0_128 = vld1q_u8(above + base + j);
a1_128 = vld1q_u8(above + base + j + 1);
- a0 = vzipq_u8(a0_128, v_zero);
- a1 = vzipq_u8(a1_128, v_zero);
- diff.val[0] =
- vsubq_u16(vreinterpretq_u16_u8(a1.val[0]),
- vreinterpretq_u16_u8(a0.val[0])); // a[x+1] - a[x]
- diff.val[1] =
- vsubq_u16(vreinterpretq_u16_u8(a1.val[1]),
- vreinterpretq_u16_u8(a0.val[1])); // a[x+1] - a[x]
- a32.val[0] = vmlaq_u16(a16, vreinterpretq_u16_u8(a0.val[0]),
- v_32); // a[x] * 32 + 16
- a32.val[1] = vmlaq_u16(a16, vreinterpretq_u16_u8(a0.val[1]),
- v_32); // a[x] * 32 + 16
+ diff.val[0] = vsubl_u8(vget_low_u8(a1_128), vget_low_u8(a0_128));
+ diff.val[1] = vsubl_u8(vget_high_u8(a1_128), vget_high_u8(a0_128));
+ a32.val[0] = vmlal_u8(a16, vget_low_u8(a0_128), v_32);
+ a32.val[1] = vmlal_u8(a16, vget_high_u8(a0_128), v_32);
res.val[0] = vmlaq_u16(a32.val[0], diff.val[0], shift);
res.val[1] = vmlaq_u16(a32.val[1], diff.val[1], shift);
@@ -1424,20 +1394,15 @@
// final pixels will be calculated as:
// (above[x] * 32 + 16 + (above[x+1] - above[x]) * shift) >> 5
- uint8x16x2_t a0, a1;
- uint16x8x2_t a32, diff;
-
- uint16x8_t a16 = vdupq_n_u16(16);
- uint8x16_t a_mbase_x = vdupq_n_u8(above[max_base_x]);
- uint8x16_t max_base_x128 = vdupq_n_u8(max_base_x);
- uint16x8_t v_32 = vdupq_n_u16(32);
- uint8x16_t v_zero = vdupq_n_u8(0);
- uint8x16_t step = vdupq_n_u8(16);
+ const uint16x8_t a16 = vdupq_n_u16(16);
+ const uint8x16_t a_mbase_x = vdupq_n_u8(above[max_base_x]);
+ const uint8x16_t max_base_x128 = vdupq_n_u8(max_base_x);
+ const uint8x8_t v_32 = vdup_n_u8(32);
+ const uint8x16_t v_zero = vdupq_n_u8(0);
+ const uint8x16_t step = vdupq_n_u8(16);
int x = dx;
for (int r = 0; r < N; r++, dst += stride) {
- uint16x8x2_t res;
-
int base = x >> frac_bits;
if (base >= max_base_x) {
for (int i = r; i < N; ++i) {
@@ -1451,7 +1416,6 @@
}
uint16x8_t shift = vdupq_n_u16((x & 0x3f) >> 1);
- uint8x16_t a0_128, a1_128, res128, mask128;
uint8x16_t base_inc128 =
vaddq_u8(vdupq_n_u8(base), vcombine_u8(vcreate_u8(0x0706050403020100),
vcreate_u8(0x0F0E0D0C0B0A0908)));
@@ -1461,20 +1425,14 @@
if (mdif <= 0) {
vst1q_u8(dst + j, a_mbase_x);
} else {
+ uint16x8x2_t a32, diff, res;
+ uint8x16_t a0_128, a1_128, mask128, res128;
a0_128 = vld1q_u8(above + base + j);
a1_128 = vld1q_u8(above + base + 1 + j);
- a0 = vzipq_u8(a0_128, v_zero);
- a1 = vzipq_u8(a1_128, v_zero);
- diff.val[0] =
- vsubq_u16(vreinterpretq_u16_u8(a1.val[0]),
- vreinterpretq_u16_u8(a0.val[0])); // a[x+1] - a[x]
- diff.val[1] =
- vsubq_u16(vreinterpretq_u16_u8(a1.val[1]),
- vreinterpretq_u16_u8(a0.val[1])); // a[x+1] - a[x]
- a32.val[0] = vmlaq_u16(a16, vreinterpretq_u16_u8(a0.val[0]),
- v_32); // a[x] * 32 + 16
- a32.val[1] = vmlaq_u16(a16, vreinterpretq_u16_u8(a0.val[1]),
- v_32); // a[x] * 32 + 16
+ diff.val[0] = vsubl_u8(vget_low_u8(a1_128), vget_low_u8(a0_128));
+ diff.val[1] = vsubl_u8(vget_high_u8(a1_128), vget_high_u8(a0_128));
+ a32.val[0] = vmlal_u8(a16, vget_low_u8(a0_128), v_32);
+ a32.val[1] = vmlal_u8(a16, vget_high_u8(a0_128), v_32);
res.val[0] = vmlaq_u16(a32.val[0], diff.val[0], shift);
res.val[1] = vmlaq_u16(a32.val[1], diff.val[1], shift);
uint8x16_t v_temp =
