[arm]: Add aom_highbd_quantize_b{,_32x32,_64x64}_neon().
3.4x to 4.9x faster than "C" depending on the last nonzero coeff position.
Bug: b/217282899
Change-Id: I4b70e1dec8cb37a245fa68515bd384a0cb6e29cf
diff --git a/aom_dsp/aom_dsp.cmake b/aom_dsp/aom_dsp.cmake
index 3f88f11..0f65315 100644
--- a/aom_dsp/aom_dsp.cmake
+++ b/aom_dsp/aom_dsp.cmake
@@ -287,6 +287,7 @@
"${AOM_ROOT}/aom_dsp/x86/highbd_variance_sse4.c")
list(APPEND AOM_DSP_ENCODER_INTRIN_NEON
+ "${AOM_ROOT}/aom_dsp/arm/highbd_quantize_neon.c"
"${AOM_ROOT}/aom_dsp/arm/highbd_variance_neon.c")
endif()
diff --git a/aom_dsp/aom_dsp_rtcd_defs.pl b/aom_dsp/aom_dsp_rtcd_defs.pl
index 0d782f7..0a57d7f 100755
--- a/aom_dsp/aom_dsp_rtcd_defs.pl
+++ b/aom_dsp/aom_dsp_rtcd_defs.pl
@@ -689,19 +689,19 @@
if (aom_config("CONFIG_AV1_ENCODER") eq "yes" && aom_config("CONFIG_AV1_HIGHBITDEPTH") eq "yes") {
add_proto qw/void aom_highbd_quantize_b/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
- specialize qw/aom_highbd_quantize_b sse2 avx2/;
+ specialize qw/aom_highbd_quantize_b sse2 avx2 neon/;
add_proto qw/void aom_highbd_quantize_b_adaptive/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
specialize qw/aom_highbd_quantize_b_adaptive sse2 avx2/;
add_proto qw/void aom_highbd_quantize_b_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
- specialize qw/aom_highbd_quantize_b_32x32 sse2/;
+ specialize qw/aom_highbd_quantize_b_32x32 sse2 neon/;
add_proto qw/void aom_highbd_quantize_b_32x32_adaptive/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
specialize qw/aom_highbd_quantize_b_32x32_adaptive sse2 avx2/;
add_proto qw/void aom_highbd_quantize_b_64x64/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
- specialize qw/aom_highbd_quantize_b_64x64 sse2/;
+ specialize qw/aom_highbd_quantize_b_64x64 sse2 neon/;
add_proto qw/void aom_highbd_quantize_b_64x64_adaptive/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
specialize qw/aom_highbd_quantize_b_64x64_adaptive sse2/;
diff --git a/aom_dsp/arm/highbd_quantize_neon.c b/aom_dsp/arm/highbd_quantize_neon.c
new file mode 100644
index 0000000..3deb80d
--- /dev/null
+++ b/aom_dsp/arm/highbd_quantize_neon.c
@@ -0,0 +1,232 @@
+/*
+ * Copyright (c) 2022, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+#include <assert.h>
+
+#include "aom_dsp/arm/mem_neon.h"
+
+#include "av1/common/quant_common.h"
+#include "av1/encoder/av1_quantize.h"
+
+static INLINE uint32_t sum_abs_coeff(const uint32x4_t a) {
+#if defined(__aarch64__)
+ return vaddvq_u32(a);
+#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);
+#endif
+}
+
+static INLINE uint16x4_t
+quantize_4(const tran_low_t *coeff_ptr, tran_low_t *qcoeff_ptr,
+ tran_low_t *dqcoeff_ptr, int32x4_t v_quant_s32,
+ int32x4_t v_dequant_s32, int32x4_t v_round_s32, int32x4_t v_zbin_s32,
+ int32x4_t v_quant_shift_s32, int log_scale) {
+ const int32x4_t v_coeff = vld1q_s32(coeff_ptr);
+ const int32x4_t v_coeff_sign =
+ vreinterpretq_s32_u32(vcltq_s32(v_coeff, vdupq_n_s32(0)));
+ const int32x4_t v_abs_coeff = vabsq_s32(v_coeff);
+ // if (abs_coeff < zbins[rc != 0]),
+ const uint32x4_t v_zbin_mask = vcgeq_s32(v_abs_coeff, v_zbin_s32);
+ const int32x4_t v_log_scale = vdupq_n_s32(log_scale);
+ // const int64_t tmp = (int64_t)abs_coeff + log_scaled_round;
+ const int32x4_t v_tmp = vaddq_s32(v_abs_coeff, v_round_s32);
+ // const int32_t tmpw32 = tmp * wt;
+ const int32x4_t v_tmpw32 = vmulq_s32(v_tmp, vdupq_n_s32((1 << AOM_QM_BITS)));
+ // const int32_t tmp2 = (int32_t)((tmpw32 * quant64) >> 16);
+ const int32x4_t v_tmp2 = vqdmulhq_s32(v_tmpw32, v_quant_s32);
+ // const int32_t tmp3 =
+ // ((((tmp2 + tmpw32)<< log_scale) * (int64_t)(quant_shift << 15)) >> 32);
+ const int32x4_t v_tmp3 = vqdmulhq_s32(
+ vshlq_s32(vaddq_s32(v_tmp2, v_tmpw32), v_log_scale), v_quant_shift_s32);
+ // const int abs_qcoeff = vmask ? (int)tmp3 >> AOM_QM_BITS : 0;
+ const int32x4_t v_abs_qcoeff = vandq_s32(vreinterpretq_s32_u32(v_zbin_mask),
+ vshrq_n_s32(v_tmp3, AOM_QM_BITS));
+ // const tran_low_t abs_dqcoeff = (abs_qcoeff * dequant_iwt) >> log_scale;
+ // vshlq_s32 will shift right if shift value is negative.
+ const int32x4_t v_abs_dqcoeff =
+ vshlq_s32(vmulq_s32(v_abs_qcoeff, v_dequant_s32), vnegq_s32(v_log_scale));
+ // qcoeff_ptr[rc] = (tran_low_t)((abs_qcoeff ^ coeff_sign) - coeff_sign);
+ const int32x4_t v_qcoeff =
+ vsubq_s32(veorq_s32(v_abs_qcoeff, v_coeff_sign), v_coeff_sign);
+ // dqcoeff_ptr[rc] = (tran_low_t)((abs_dqcoeff ^ coeff_sign) - coeff_sign);
+ const int32x4_t v_dqcoeff =
+ vsubq_s32(veorq_s32(v_abs_dqcoeff, v_coeff_sign), v_coeff_sign);
+
+ vst1q_s32(qcoeff_ptr, v_qcoeff);
+ vst1q_s32(dqcoeff_ptr, v_dqcoeff);
+
+ // Used to find eob.
+ const uint32x4_t nz_qcoeff_mask = vcgtq_s32(v_abs_qcoeff, vdupq_n_s32(0));
+ return vmovn_u32(nz_qcoeff_mask);
+}
+
+static INLINE int16x8_t get_max_lane_eob(const int16_t *iscan,
+ int16x8_t v_eobmax,
+ uint16x8_t v_mask) {
+ const int16x8_t v_iscan = vld1q_s16(&iscan[0]);
+ const int16x8_t v_iscan_plus1 = vaddq_s16(v_iscan, vdupq_n_s16(1));
+ const int16x8_t v_nz_iscan = vbslq_s16(v_mask, v_iscan_plus1, vdupq_n_s16(0));
+ return vmaxq_s16(v_eobmax, v_nz_iscan);
+}
+
+static INLINE uint16_t get_max_eob(int16x8_t v_eobmax) {
+#ifdef __aarch64__
+ return (uint16_t)vmaxvq_s16(v_eobmax);
+#else
+ const int16x4_t v_eobmax_3210 =
+ vmax_s16(vget_low_s16(v_eobmax), vget_high_s16(v_eobmax));
+ const int64x1_t v_eobmax_xx32 =
+ vshr_n_s64(vreinterpret_s64_s16(v_eobmax_3210), 32);
+ const int16x4_t v_eobmax_tmp =
+ vmax_s16(v_eobmax_3210, vreinterpret_s16_s64(v_eobmax_xx32));
+ const int64x1_t v_eobmax_xxx3 =
+ vshr_n_s64(vreinterpret_s64_s16(v_eobmax_tmp), 16);
+ const int16x4_t v_eobmax_final =
+ vmax_s16(v_eobmax_tmp, vreinterpret_s16_s64(v_eobmax_xxx3));
+ return (uint16_t)vget_lane_s16(v_eobmax_final, 0);
+#endif
+}
+
+static void highbd_quantize_b_neon(
+ const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr,
+ const int16_t *round_ptr, const int16_t *quant_ptr,
+ const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
+ tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
+ const int16_t *scan, const int16_t *iscan, const int log_scale) {
+ (void)scan;
+ const int16x4_t v_quant = vld1_s16(quant_ptr);
+ const int16x4_t v_dequant = vld1_s16(dequant_ptr);
+ const int16x4_t v_zero = vdup_n_s16(0);
+ const uint16x4_t v_round_select = vcgt_s16(vdup_n_s16(log_scale), v_zero);
+ const int16x4_t v_round_no_scale = vld1_s16(round_ptr);
+ const int16x4_t v_round_log_scale =
+ vqrdmulh_n_s16(v_round_no_scale, (int16_t)(1 << (15 - log_scale)));
+ const int16x4_t v_round =
+ vbsl_s16(v_round_select, v_round_log_scale, v_round_no_scale);
+ const int16x4_t v_quant_shift = vld1_s16(quant_shift_ptr);
+ const int16x4_t v_zbin_no_scale = vld1_s16(zbin_ptr);
+ const int16x4_t v_zbin_log_scale =
+ vqrdmulh_n_s16(v_zbin_no_scale, (int16_t)(1 << (15 - log_scale)));
+ const int16x4_t v_zbin =
+ vbsl_s16(v_round_select, v_zbin_log_scale, v_zbin_no_scale);
+ int32x4_t v_round_s32 = vmovl_s16(v_round);
+ int32x4_t v_quant_s32 = vshlq_n_s32(vmovl_s16(v_quant), 15);
+ int32x4_t v_dequant_s32 = vmovl_s16(v_dequant);
+ int32x4_t v_quant_shift_s32 = vshlq_n_s32(vmovl_s16(v_quant_shift), 15);
+ int32x4_t v_zbin_s32 = vmovl_s16(v_zbin);
+ uint16x4_t v_mask_lo, v_mask_hi;
+ int16x8_t v_eobmax = vdupq_n_s16(-1);
+
+ int non_zero_count = n_coeffs;
+
+ assert(n_coeffs > 8);
+ // Pre-scan pass
+ const int32x4_t v_zbin_s32x = vdupq_lane_s32(vget_low_s32(v_zbin_s32), 1);
+ int i = n_coeffs;
+ do {
+ const int32x4_t v_coeff_a = vld1q_s32(coeff_ptr + i - 4 - 1);
+ const int32x4_t v_coeff_b = vld1q_s32(coeff_ptr + i - 8 - 1);
+ const int32x4_t v_abs_coeff_a = vabsq_s32(v_coeff_a);
+ const int32x4_t v_abs_coeff_b = vabsq_s32(v_coeff_b);
+ 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) {
+ non_zero_count -= 8;
+ } else {
+ break;
+ }
+ i -= 8;
+ } while (i > 0);
+
+ const int 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));
+
+ // DC and first 3 AC
+ v_mask_lo =
+ quantize_4(coeff_ptr, qcoeff_ptr, dqcoeff_ptr, v_quant_s32, v_dequant_s32,
+ v_round_s32, v_zbin_s32, v_quant_shift_s32, log_scale);
+
+ // overwrite the DC constants with AC constants
+ v_round_s32 = vdupq_lane_s32(vget_low_s32(v_round_s32), 1);
+ v_quant_s32 = vdupq_lane_s32(vget_low_s32(v_quant_s32), 1);
+ v_dequant_s32 = vdupq_lane_s32(vget_low_s32(v_dequant_s32), 1);
+ v_quant_shift_s32 = vdupq_lane_s32(vget_low_s32(v_quant_shift_s32), 1);
+ v_zbin_s32 = vdupq_lane_s32(vget_low_s32(v_zbin_s32), 1);
+
+ // 4 more AC
+ v_mask_hi = quantize_4(coeff_ptr + 4, qcoeff_ptr + 4, dqcoeff_ptr + 4,
+ v_quant_s32, v_dequant_s32, v_round_s32, v_zbin_s32,
+ v_quant_shift_s32, log_scale);
+
+ v_eobmax =
+ get_max_lane_eob(iscan, v_eobmax, vcombine_u16(v_mask_lo, v_mask_hi));
+
+ int count = non_zero_count - 8;
+ for (; count > 0; count -= 8) {
+ coeff_ptr += 8;
+ qcoeff_ptr += 8;
+ dqcoeff_ptr += 8;
+ iscan += 8;
+ v_mask_lo = quantize_4(coeff_ptr, qcoeff_ptr, dqcoeff_ptr, v_quant_s32,
+ v_dequant_s32, v_round_s32, v_zbin_s32,
+ v_quant_shift_s32, log_scale);
+ v_mask_hi = quantize_4(coeff_ptr + 4, qcoeff_ptr + 4, dqcoeff_ptr + 4,
+ v_quant_s32, v_dequant_s32, v_round_s32, v_zbin_s32,
+ v_quant_shift_s32, log_scale);
+ // Find the max lane eob for 8 coeffs.
+ v_eobmax =
+ get_max_lane_eob(iscan, v_eobmax, vcombine_u16(v_mask_lo, v_mask_hi));
+ }
+
+ *eob_ptr = get_max_eob(v_eobmax);
+}
+
+void aom_highbd_quantize_b_neon(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+ const int16_t *zbin_ptr,
+ const int16_t *round_ptr,
+ const int16_t *quant_ptr,
+ const int16_t *quant_shift_ptr,
+ tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+ const int16_t *dequant_ptr, uint16_t *eob_ptr,
+ const int16_t *scan, const int16_t *iscan) {
+ highbd_quantize_b_neon(coeff_ptr, n_coeffs, zbin_ptr, round_ptr, quant_ptr,
+ quant_shift_ptr, qcoeff_ptr, dqcoeff_ptr, dequant_ptr,
+ eob_ptr, scan, iscan, 0);
+}
+
+void aom_highbd_quantize_b_32x32_neon(
+ const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr,
+ const int16_t *round_ptr, const int16_t *quant_ptr,
+ const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
+ tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
+ const int16_t *scan, const int16_t *iscan) {
+ highbd_quantize_b_neon(coeff_ptr, n_coeffs, zbin_ptr, round_ptr, quant_ptr,
+ quant_shift_ptr, qcoeff_ptr, dqcoeff_ptr, dequant_ptr,
+ eob_ptr, scan, iscan, 1);
+}
+
+void aom_highbd_quantize_b_64x64_neon(
+ const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr,
+ const int16_t *round_ptr, const int16_t *quant_ptr,
+ const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
+ tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
+ const int16_t *scan, const int16_t *iscan) {
+ highbd_quantize_b_neon(coeff_ptr, n_coeffs, zbin_ptr, round_ptr, quant_ptr,
+ quant_shift_ptr, qcoeff_ptr, dqcoeff_ptr, dequant_ptr,
+ eob_ptr, scan, iscan, 2);
+}
diff --git a/test/quantize_func_test.cc b/test/quantize_func_test.cc
index fd18ffd..90957d8 100644
--- a/test/quantize_func_test.cc
+++ b/test/quantize_func_test.cc
@@ -709,6 +709,7 @@
static_cast<TX_SIZE>(TX_32X32), TYPE_B, AOM_BITS_8),
make_tuple(&aom_quantize_b_64x64_c, &aom_quantize_b_64x64_neon,
static_cast<TX_SIZE>(TX_64X64), TYPE_B, AOM_BITS_8),
+
#if CONFIG_AV1_HIGHBITDEPTH
make_tuple(&highbd_quan16x16_wrapper<av1_highbd_quantize_fp_c>,
&highbd_quan16x16_wrapper<av1_highbd_quantize_fp_neon>,
@@ -719,6 +720,12 @@
make_tuple(&highbd_quan64x64_wrapper<av1_highbd_quantize_fp_c>,
&highbd_quan64x64_wrapper<av1_highbd_quantize_fp_neon>,
static_cast<TX_SIZE>(TX_64X64), TYPE_FP, AOM_BITS_12),
+ make_tuple(&aom_highbd_quantize_b_c, &aom_highbd_quantize_b_neon,
+ static_cast<TX_SIZE>(TX_16X16), TYPE_B, AOM_BITS_12),
+ make_tuple(&aom_highbd_quantize_b_32x32_c, &aom_highbd_quantize_b_32x32_neon,
+ 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),
#endif
};
