Commit ebbb3f0a46 for aom
commit ebbb3f0a46cbe277dc01c509f14599a6be9c0502
Author: Frank Galligan <fgalligan@google.com>
Date: Sat Nov 14 11:45:28 2026 -0800
aom_dsp,neon: Optimize highbd_quantize_b pre-scan
Replace the post-loop memset in aom_highbd_quantize_b_neon and
aom_highbd_quantize_b_adaptive_neon with in-loop NEON zero stores and
vmaxvq_u32(vorrq_u32(a, b)) zero checks, and add TX_4X4/TX_8X8 NEON
test coverage in FullPrecisionQuantizeTest.
Speedup: 1.25x-4.36x (-20.7% to -77.1%, aom_highbd_quantize_b_neon &
adaptive TX_4X4..TX_16X16) in test_libaom on Neoverse N3; -1.03% (1T,
-1.15% min) on Neoverse N2 and -0.75% (1T Cortex-A720) / -3.80% (4T)
10b avifenc on Pixel 9 (bit-exact).
TAG=agy
CONV=490488d4-7005-402a-8cc4-c245227f80fe
Change-Id: Idb4f8f2c7be54cf244e501188e2c61e825999d66
diff --git a/aom_dsp/arm/highbd_quantize_neon.c b/aom_dsp/arm/highbd_quantize_neon.c
index 7fe236a86e..4c1e3efd2e 100644
--- a/aom_dsp/arm/highbd_quantize_neon.c
+++ b/aom_dsp/arm/highbd_quantize_neon.c
@@ -11,20 +11,19 @@
#include <arm_neon.h>
#include <assert.h>
-#include <string.h>
#include "config/aom_config.h"
#include "config/aom_dsp_rtcd.h"
#include "aom_dsp/quantize.h"
-static inline uint32_t sum_abs_coeff(const uint32x4_t a) {
+static inline int all_zero_u32x8(const uint32x4_t a, const uint32x4_t b) {
+ const uint32x4_t ab = vorrq_u32(a, b);
#if AOM_ARCH_AARCH64
- return vaddvq_u32(a);
+ return vmaxvq_u32(ab) == 0;
#else
- const uint64x2_t b = vpaddlq_u32(a);
- const uint64x1_t c = vadd_u64(vget_low_u64(b), vget_high_u64(b));
- return (uint32_t)vget_lane_u64(c, 0);
+ const uint32x2_t d = vorr_u32(vget_low_u32(ab), vget_high_u32(ab));
+ return vget_lane_u64(vreinterpret_u64_u32(d), 0) == 0;
#endif
}
@@ -169,8 +168,9 @@ static void highbd_quantize_b_neon(
intptr_t non_zero_count = n_coeffs;
- assert(n_coeffs > 8);
+ assert(n_coeffs >= 16 && (n_coeffs & 7) == 0);
// Pre-scan pass
+ const int32x4_t v_zero_s32 = vdupq_n_s32(0);
const int32x4_t v_zbin_s32x = vdupq_lane_s32(vget_low_s32(v_zbin_s32), 1);
intptr_t i = n_coeffs;
do {
@@ -181,19 +181,18 @@ static void highbd_quantize_b_neon(
const uint32x4_t v_mask_a = vcgeq_s32(v_abs_coeff_a, v_zbin_s32x);
const uint32x4_t v_mask_b = vcgeq_s32(v_abs_coeff_b, v_zbin_s32x);
// If the coefficient is in the base ZBIN range, then discard.
- if (sum_abs_coeff(v_mask_a) + sum_abs_coeff(v_mask_b) == 0) {
+ if (all_zero_u32x8(v_mask_a, v_mask_b)) {
+ vst1q_s32(qcoeff_ptr + i - 8, v_zero_s32);
+ vst1q_s32(qcoeff_ptr + i - 4, v_zero_s32);
+ vst1q_s32(dqcoeff_ptr + i - 8, v_zero_s32);
+ vst1q_s32(dqcoeff_ptr + i - 4, v_zero_s32);
non_zero_count -= 8;
} else {
break;
}
i -= 8;
- } while (i > 0);
-
- const intptr_t remaining_zcoeffs = n_coeffs - non_zero_count;
- memset(qcoeff_ptr + non_zero_count, 0,
- remaining_zcoeffs * sizeof(*qcoeff_ptr));
- memset(dqcoeff_ptr + non_zero_count, 0,
- remaining_zcoeffs * sizeof(*dqcoeff_ptr));
+ // The first 8 coefficients are always quantized below.
+ } while (i != 8);
// DC and first 3 AC
v_mask_lo =
@@ -216,7 +215,7 @@ static void highbd_quantize_b_neon(
get_max_lane_eob(iscan, v_eobmax, vcombine_u16(v_mask_lo, v_mask_hi));
intptr_t count = non_zero_count - 8;
- for (; count > 0; count -= 8) {
+ for (; count != 0; count -= 8) {
coeff_ptr += 8;
qcoeff_ptr += 8;
dqcoeff_ptr += 8;
@@ -302,13 +301,14 @@ static void highbd_quantize_b_adaptive_neon(
int16x8_t v_eobmax = vdupq_n_s16(-1);
int16x8_t v_eobmin = vdupq_n_s16((int16_t)n_coeffs);
- assert(n_coeffs > 8);
+ assert(n_coeffs >= 16 && (n_coeffs & 7) == 0);
// Pre-scan pass
const int32x4_t v_zbin_s32x = vdupq_lane_s32(vget_low_s32(v_zbin_s32), 1);
const int prescan_add_1 =
ROUND_POWER_OF_TWO(dequant_ptr[1] * EOB_FACTOR, 7 + AOM_QM_BITS);
const int32x4_t v_zbin_prescan =
vaddq_s32(v_zbin_s32x, vdupq_n_s32(prescan_add_1));
+ const int32x4_t v_zero_s32 = vdupq_n_s32(0);
intptr_t non_zero_count = n_coeffs;
intptr_t i = n_coeffs;
do {
@@ -319,19 +319,18 @@ static void highbd_quantize_b_adaptive_neon(
const uint32x4_t v_mask_a = vcgeq_s32(v_abs_coeff_a, v_zbin_prescan);
const uint32x4_t v_mask_b = vcgeq_s32(v_abs_coeff_b, v_zbin_prescan);
// If the coefficient is in the base ZBIN range, then discard.
- if (sum_abs_coeff(v_mask_a) + sum_abs_coeff(v_mask_b) == 0) {
+ if (all_zero_u32x8(v_mask_a, v_mask_b)) {
+ vst1q_s32(qcoeff_ptr + i - 8, v_zero_s32);
+ vst1q_s32(qcoeff_ptr + i - 4, v_zero_s32);
+ vst1q_s32(dqcoeff_ptr + i - 8, v_zero_s32);
+ vst1q_s32(dqcoeff_ptr + i - 4, v_zero_s32);
non_zero_count -= 8;
} else {
break;
}
i -= 8;
- } while (i > 0);
-
- const intptr_t remaining_zcoeffs = n_coeffs - non_zero_count;
- memset(qcoeff_ptr + non_zero_count, 0,
- remaining_zcoeffs * sizeof(*qcoeff_ptr));
- memset(dqcoeff_ptr + non_zero_count, 0,
- remaining_zcoeffs * sizeof(*dqcoeff_ptr));
+ // The first 8 coefficients are always quantized below.
+ } while (i != 8);
// DC and first 3 AC
v_mask_lo =
@@ -354,7 +353,7 @@ static void highbd_quantize_b_adaptive_neon(
vcombine_u16(v_mask_lo, v_mask_hi), n_coeffs);
intptr_t count = non_zero_count - 8;
- for (; count > 0; count -= 8) {
+ for (; count != 0; count -= 8) {
coeff_ptr += 8;
qcoeff_ptr += 8;
dqcoeff_ptr += 8;
diff --git a/test/quantize_func_test.cc b/test/quantize_func_test.cc
index 7e13313f74..3e35c8dc50 100644
--- a/test/quantize_func_test.cc
+++ b/test/quantize_func_test.cc
@@ -922,6 +922,10 @@ const QuantizeParam<QuantizeFunc> kQParamArrayNEON[] = {
static_cast<TX_SIZE>(TX_32X32), TYPE_B, AOM_BITS_12),
make_tuple(&aom_highbd_quantize_b_64x64_c, &aom_highbd_quantize_b_64x64_neon,
static_cast<TX_SIZE>(TX_64X64), TYPE_B, AOM_BITS_12),
+ make_tuple(&aom_highbd_quantize_b_c, &aom_highbd_quantize_b_neon,
+ static_cast<TX_SIZE>(TX_4X4), TYPE_B, AOM_BITS_12),
+ make_tuple(&aom_highbd_quantize_b_c, &aom_highbd_quantize_b_neon,
+ static_cast<TX_SIZE>(TX_8X8), TYPE_B, AOM_BITS_12),
#if !CONFIG_REALTIME_ONLY
make_tuple(&aom_highbd_quantize_b_adaptive_c,
&aom_highbd_quantize_b_adaptive_neon,
@@ -932,6 +936,12 @@ const QuantizeParam<QuantizeFunc> kQParamArrayNEON[] = {
make_tuple(&aom_highbd_quantize_b_64x64_adaptive_c,
&aom_highbd_quantize_b_64x64_adaptive_neon,
static_cast<TX_SIZE>(TX_64X64), TYPE_B, AOM_BITS_12),
+ make_tuple(&aom_highbd_quantize_b_adaptive_c,
+ &aom_highbd_quantize_b_adaptive_neon, static_cast<TX_SIZE>(TX_4X4),
+ TYPE_B, AOM_BITS_12),
+ make_tuple(&aom_highbd_quantize_b_adaptive_c,
+ &aom_highbd_quantize_b_adaptive_neon, static_cast<TX_SIZE>(TX_8X8),
+ TYPE_B, AOM_BITS_12),
#endif // !CONFIG_REALTIME_ONLY
#endif // CONFIG_AV1_HIGHBITDEPTH
};