ref: 43a30d3a1a6b627fa05ba63f4c51414ced781ccb
parent: 4a8c248744500f9caf00588ca312efce5659e45e
author: Johann <johann.koenig@duck.com>
date: Mon Nov 12 06:30:03 EST 2018
quantize: use aarch64 vmaxv Simplify max value calculation on aarch64 by using vmaxv. Much faster for 4x4 but diminishing returns as the block size grows. Only the vp9 quantize has a speed test hooked up. Anticipate similar results for the other quantize versions. Before: [ RUN ] NEON/VP9QuantizeTest.DISABLED_Speed/2 [ BENCH ] Bypass calculations 4x4 31.6 ms ( ±0.0 ms ) [ BENCH ] Full calculations 4x4 31.6 ms ( ±0.0 ms ) [ BENCH ] Bypass calculations 8x8 17.7 ms ( ±0.0 ms ) [ BENCH ] Full calculations 8x8 17.7 ms ( ±0.0 ms ) [ BENCH ] Bypass calculations 16x16 14.2 ms ( ±0.0 ms ) [ BENCH ] Full calculations 16x16 14.2 ms ( ±0.0 ms ) [ OK ] NEON/VP9QuantizeTest.DISABLED_Speed/2 (1906 ms) [ RUN ] NEON/VP9QuantizeTest.DISABLED_Speed/3 [ BENCH ] Bypass calculations 32x32 18.6 ms ( ±0.0 ms ) [ BENCH ] Full calculations 32x32 18.6 ms ( ±0.0 ms ) After: [ RUN ] NEON/VP9QuantizeTest.DISABLED_Speed/2 [ BENCH ] Bypass calculations 4x4 29.1 ms ( ±0.0 ms ) [ BENCH ] Full calculations 4x4 29.1 ms ( ±0.0 ms ) [ BENCH ] Bypass calculations 8x8 16.9 ms ( ±0.0 ms ) [ BENCH ] Full calculations 8x8 16.9 ms ( ±0.0 ms ) [ BENCH ] Bypass calculations 16x16 14.1 ms ( ±0.0 ms ) [ BENCH ] Full calculations 16x16 14.1 ms ( ±0.0 ms ) [ OK ] NEON/VP9QuantizeTest.DISABLED_Speed/2 (1803 ms) [ RUN ] NEON/VP9QuantizeTest.DISABLED_Speed/3 [ BENCH ] Bypass calculations 32x32 18.6 ms ( ±0.0 ms ) [ BENCH ] Full calculations 32x32 18.6 ms ( ±0.0 ms ) Change-Id: Ic95812b3fdbd4e47b4dcb8ed46c68a9617de38d2
--- a/vp8/encoder/arm/neon/fastquantizeb_neon.c
+++ b/vp8/encoder/arm/neon/fastquantizeb_neon.c
@@ -26,9 +26,11 @@
zig_zag1 = vld1q_u16(inv_zig_zag + 8);
int16x8_t x0, x1, sz0, sz1, y0, y1;
uint16x8_t eob0, eob1;
+#ifndef __aarch64__
uint16x4_t eob_d16;
uint32x2_t eob_d32;
uint32x4_t eob_q32;
+#endif // __arch64__
/* sign of z: z >> 15 */
sz0 = vshrq_n_s16(z0, 15);
@@ -66,11 +68,17 @@
/* select the largest value */
eob0 = vmaxq_u16(eob0, eob1);
+#ifdef __aarch64__
+ *d->eob = (int8_t)vmaxvq_u16(eob0);
+#else
eob_d16 = vmax_u16(vget_low_u16(eob0), vget_high_u16(eob0));
eob_q32 = vmovl_u16(eob_d16);
eob_d32 = vmax_u32(vget_low_u32(eob_q32), vget_high_u32(eob_q32));
eob_d32 = vpmax_u32(eob_d32, eob_d32);
+ vst1_lane_s8((int8_t *)d->eob, vreinterpret_s8_u32(eob_d32), 0);
+#endif // __aarch64__
+
/* qcoeff = x */
vst1q_s16(d->qcoeff, x0);
vst1q_s16(d->qcoeff + 8, x1);
@@ -78,6 +86,4 @@
/* dqcoeff = x * dequant */
vst1q_s16(d->dqcoeff, vmulq_s16(dequant0, x0));
vst1q_s16(d->dqcoeff + 8, vmulq_s16(dequant1, x1));
-
- vst1_lane_s8((int8_t *)d->eob, vreinterpret_s8_u32(eob_d32), 0);
}
--- a/vp9/encoder/arm/neon/vp9_quantize_neon.c
+++ b/vp9/encoder/arm/neon/vp9_quantize_neon.c
@@ -97,6 +97,9 @@
store_s16q_to_tran_low(qcoeff_ptr + i, v_qcoeff);
store_s16q_to_tran_low(dqcoeff_ptr + i, v_dqcoeff);
}
+#ifdef __aarch64__
+ *eob_ptr = vmaxvq_s16(v_eobmax_76543210);
+#else
{const int16x4_t v_eobmax_3210 = vmax_s16(vget_low_s16(v_eobmax_76543210),
vget_high_s16(v_eobmax_76543210));
@@ -111,6 +114,7 @@
*eob_ptr = (uint16_t)vget_lane_s16(v_eobmax_final, 0);
}
+#endif // __aarch64__
}
static INLINE int32x4_t extract_sign_bit(int32x4_t a) {@@ -226,6 +230,9 @@
dqcoeff_ptr += 8;
}
+#ifdef __aarch64__
+ *eob_ptr = vmaxvq_u16(eob_max);
+#else
{const uint16x4_t eob_max_0 =
vmax_u16(vget_low_u16(eob_max), vget_high_u16(eob_max));
@@ -233,5 +240,6 @@
const uint16x4_t eob_max_2 = vpmax_u16(eob_max_1, eob_max_1);
vst1_lane_u16(eob_ptr, eob_max_2, 0);
}
+#endif // __aarch64__
}
}
--- a/vpx_dsp/arm/quantize_neon.c
+++ b/vpx_dsp/arm/quantize_neon.c
@@ -135,6 +135,9 @@
} while (n_coeffs > 0);
}
+#ifdef __aarch64__
+ *eob_ptr = vmaxvq_u16(eob_max);
+#else
{const uint16x4_t eob_max_0 =
vmax_u16(vget_low_u16(eob_max), vget_high_u16(eob_max));
@@ -142,6 +145,7 @@
const uint16x4_t eob_max_2 = vpmax_u16(eob_max_1, eob_max_1);
vst1_lane_u16(eob_ptr, eob_max_2, 0);
}
+#endif // __aarch64__
}
static INLINE int32x4_t extract_sign_bit(int32x4_t a) {@@ -288,6 +292,9 @@
}
}
+#ifdef __aarch64__
+ *eob_ptr = vmaxvq_u16(eob_max);
+#else
{const uint16x4_t eob_max_0 =
vmax_u16(vget_low_u16(eob_max), vget_high_u16(eob_max));
@@ -295,4 +302,5 @@
const uint16x4_t eob_max_2 = vpmax_u16(eob_max_1, eob_max_1);
vst1_lane_u16(eob_ptr, eob_max_2, 0);
}
+#endif // __aarch64__
}