[x265] [PATCH v2 3/7] AArch64: Add Neon saoCuStats primitives for low bitdepth
Hari Limaye
hari.limaye at arm.com
Wed May 22 19:06:47 UTC 2024
Add Neon optimised implementations of saoCuStats primitives for low
bitdepth, as none currently exist.
Results of speedtests, compared to the existing C code, when compiled
with LLVM 17 (higher is better):
Neoverse N1:
saoCuStatsBO | 1.25x
saoCuStatsE0 | 2.14x
saoCuStatsE1 | 2.87x
saoCuStatsE2 | 2.65x
saoCuStatsE3 | 2.70x
Neoverse V1:
saoCuStatsBO | 1.09x
saoCuStatsE0 | 2.67x
saoCuStatsE1 | 2.82x
saoCuStatsE2 | 2.93x
saoCuStatsE3 | 3.26x
---
source/common/CMakeLists.txt | 5 +
source/common/aarch64/asm-primitives.cpp | 4 +
source/common/aarch64/sao-prim.cpp | 383 +++++++++++++++++++++++
source/common/aarch64/sao-prim.h | 33 ++
4 files changed, 425 insertions(+)
create mode 100644 source/common/aarch64/sao-prim.cpp
create mode 100644 source/common/aarch64/sao-prim.h
diff --git a/source/common/CMakeLists.txt b/source/common/CMakeLists.txt
index 184b94127..4b00a521c 100644
--- a/source/common/CMakeLists.txt
+++ b/source/common/CMakeLists.txt
@@ -104,6 +104,11 @@ if(ENABLE_ASSEMBLY AND (ARM64 OR CROSS_COMPILE_ARM64))
endif()
set(C_SRCS asm-primitives.cpp pixel-prim.h pixel-prim.cpp filter-prim.h filter-prim.cpp dct-prim.h dct-prim.cpp loopfilter-prim.cpp loopfilter-prim.h intrapred-prim.cpp arm64-utils.cpp arm64-utils.h fun-decls.h)
+
+ if(NOT HIGH_BIT_DEPTH)
+ list(APPEND C_SRCS sao-prim.cpp)
+ endif()
+
enable_language(ASM)
# add ARM assembly/intrinsic files here
diff --git a/source/common/aarch64/asm-primitives.cpp b/source/common/aarch64/asm-primitives.cpp
index f20d1e57d..907ad3a16 100644
--- a/source/common/aarch64/asm-primitives.cpp
+++ b/source/common/aarch64/asm-primitives.cpp
@@ -689,6 +689,7 @@ extern "C" {
#include "dct-prim.h"
#include "loopfilter-prim.h"
#include "intrapred-prim.h"
+#include "sao-prim.h"
namespace X265_NS
{
@@ -1949,6 +1950,9 @@ void setupIntrinsicPrimitives(EncoderPrimitives &p, int cpuMask)
setupDCTPrimitives_neon(p);
setupLoopFilterPrimitives_neon(p);
setupIntraPrimitives_neon(p);
+#if !HIGH_BIT_DEPTH
+ setupSaoPrimitives_neon(p);
+#endif
}
}
diff --git a/source/common/aarch64/sao-prim.cpp b/source/common/aarch64/sao-prim.cpp
new file mode 100644
index 000000000..4f70d76f7
--- /dev/null
+++ b/source/common/aarch64/sao-prim.cpp
@@ -0,0 +1,383 @@
+/*****************************************************************************
+ * Copyright (C) 2024 MulticoreWare, Inc
+ *
+ * Authors: Hari Limaye <hari.limaye at arm.com>
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02111, USA.
+ *
+ * This program is also available under a commercial proprietary license.
+ * For more information, contact us at license @ x265.com.
+ *****************************************************************************/
+
+#include "sao-prim.h"
+#include "sao.h"
+#include <arm_neon.h>
+
+static inline int8x16_t signOf_neon(const pixel *a, const pixel *b)
+{
+ uint8x16_t s0 = vld1q_u8(a);
+ uint8x16_t s1 = vld1q_u8(b);
+
+ // signOf(a - b) = -(a > b ? -1 : 0) | (a < b ? -1 : 0)
+ int8x16_t cmp0 = vreinterpretq_s8_u8(vcgtq_u8(s0, s1));
+ int8x16_t cmp1 = vreinterpretq_s8_u8(vcgtq_u8(s1, s0));
+ return vorrq_s8(vnegq_s8(cmp0), cmp1);
+}
+
+// Predicate mask indices.
+static const int8_t quad_reg_byte_indices[16] = {
+ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15
+};
+
+static inline int8x16_t mask_inactive_elems(const int rem, int8x16_t edge_type)
+{
+ // Compute a predicate mask where the bits of an element are 0 if the index
+ // is less than the remainder (active), and 1 otherwise.
+ const int8x16_t indices = vld1q_s8(quad_reg_byte_indices);
+ int8x16_t pred = vreinterpretq_s8_u8(vcgeq_s8(indices, vdupq_n_s8(rem)));
+
+ // Use predicate mask to shift "unused lanes" outside of range [-2, 2]
+ pred = vshlq_n_s8(pred, 3);
+ return veorq_s8(edge_type, pred);
+}
+
+/*
+ * Compute Edge Offset statistics (count and stats).
+ * To save some instructions compute count and stats as negative values - since
+ * output of Neon comparison instructions for a matched condition is all 1s (-1).
+ */
+static inline void compute_eo_stats(const int8x16_t edge_type,
+ const int16_t *diff, int16x8_t *count,
+ int32x4_t *stats)
+{
+ // Create a mask for each edge type.
+ int8x16_t mask0 = vreinterpretq_s8_u8(vceqq_s8(edge_type, vdupq_n_s8(-2)));
+ int8x16_t mask1 = vreinterpretq_s8_u8(vceqq_s8(edge_type, vdupq_n_s8(-1)));
+ int8x16_t mask2 = vreinterpretq_s8_u8(vceqq_s8(edge_type, vdupq_n_s8(0)));
+ int8x16_t mask3 = vreinterpretq_s8_u8(vceqq_s8(edge_type, vdupq_n_s8(1)));
+ int8x16_t mask4 = vreinterpretq_s8_u8(vceqq_s8(edge_type, vdupq_n_s8(2)));
+
+ // Compute negative counts for each edge type.
+ count[0] = vpadalq_s8(count[0], mask0);
+ count[1] = vpadalq_s8(count[1], mask1);
+ count[2] = vpadalq_s8(count[2], mask2);
+ count[3] = vpadalq_s8(count[3], mask3);
+ count[4] = vpadalq_s8(count[4], mask4);
+
+ // Widen the masks to 16-bit.
+ int16x8_t mask0_lo = vreinterpretq_s16_s8(vzip1q_s8(mask0, mask0));
+ int16x8_t mask0_hi = vreinterpretq_s16_s8(vzip2q_s8(mask0, mask0));
+ int16x8_t mask1_lo = vreinterpretq_s16_s8(vzip1q_s8(mask1, mask1));
+ int16x8_t mask1_hi = vreinterpretq_s16_s8(vzip2q_s8(mask1, mask1));
+ int16x8_t mask2_lo = vreinterpretq_s16_s8(vzip1q_s8(mask2, mask2));
+ int16x8_t mask2_hi = vreinterpretq_s16_s8(vzip2q_s8(mask2, mask2));
+ int16x8_t mask3_lo = vreinterpretq_s16_s8(vzip1q_s8(mask3, mask3));
+ int16x8_t mask3_hi = vreinterpretq_s16_s8(vzip2q_s8(mask3, mask3));
+ int16x8_t mask4_lo = vreinterpretq_s16_s8(vzip1q_s8(mask4, mask4));
+ int16x8_t mask4_hi = vreinterpretq_s16_s8(vzip2q_s8(mask4, mask4));
+
+ int16x8_t diff_lo = vld1q_s16(diff);
+ int16x8_t diff_hi = vld1q_s16(diff + 8);
+
+ // Compute negative stats for each edge type.
+ int16x8_t stats0 = vmulq_s16(diff_lo, mask0_lo);
+ int16x8_t stats1 = vmulq_s16(diff_lo, mask1_lo);
+ int16x8_t stats2 = vmulq_s16(diff_lo, mask2_lo);
+ int16x8_t stats3 = vmulq_s16(diff_lo, mask3_lo);
+ int16x8_t stats4 = vmulq_s16(diff_lo, mask4_lo);
+ stats0 = vmlaq_s16(stats0, diff_hi, mask0_hi);
+ stats1 = vmlaq_s16(stats1, diff_hi, mask1_hi);
+ stats2 = vmlaq_s16(stats2, diff_hi, mask2_hi);
+ stats3 = vmlaq_s16(stats3, diff_hi, mask3_hi);
+ stats4 = vmlaq_s16(stats4, diff_hi, mask4_hi);
+
+ stats[0] = vpadalq_s16(stats[0], stats0);
+ stats[1] = vpadalq_s16(stats[1], stats1);
+ stats[2] = vpadalq_s16(stats[2], stats2);
+ stats[3] = vpadalq_s16(stats[3], stats3);
+ stats[4] = vpadalq_s16(stats[4], stats4);
+}
+
+/*
+ * Reduce and store Edge Offset statistics (count and stats).
+ */
+static inline void reduce_eo_stats(int32x4_t *vstats, int16x8_t *vcount,
+ int32_t *stats, int32_t *count)
+{
+ // s_eoTable maps edge types to memory in order: {2, 0, 1, 3, 4}.
+ int16x8_t c01 = vpaddq_s16(vcount[2], vcount[0]);
+ int16x8_t c23 = vpaddq_s16(vcount[1], vcount[3]);
+ int16x8_t c0123 = vpaddq_s16(c01, c23);
+
+ // Subtract from current count, as we calculate the negation.
+ vst1q_s32(count, vsubq_s32(vld1q_s32(count), vpaddlq_s16(c0123)));
+ count[4] -= vaddvq_s16(vcount[4]);
+
+ int32x4_t s01 = vpaddq_s32(vstats[2], vstats[0]);
+ int32x4_t s23 = vpaddq_s32(vstats[1], vstats[3]);
+ int32x4_t s0123 = vpaddq_s32(s01, s23);
+
+ // Subtract from current stats, as we calculate the negation.
+ vst1q_s32(stats, vsubq_s32(vld1q_s32(stats), s0123));
+ stats[4] -= vaddvq_s32(vstats[4]);
+}
+
+namespace X265_NS {
+void saoCuStatsBO_neon(const int16_t *diff, const pixel *rec, intptr_t stride,
+ int endX, int endY, int32_t *stats, int32_t *count)
+{
+ // Additional temporary buffer for accumulation.
+ int32_t stats_tmp[32] = { 0 };
+ int32_t count_tmp[32] = { 0 };
+
+ // Byte-addressable pointers to buffers, to optimise address calculation.
+ uint8_t *stats_b[2] = {
+ reinterpret_cast<uint8_t *>(stats),
+ reinterpret_cast<uint8_t *>(stats_tmp),
+ };
+ uint8_t *count_b[2] = {
+ reinterpret_cast<uint8_t *>(count),
+ reinterpret_cast<uint8_t *>(count_tmp),
+ };
+
+ // Combine shift for index calculation with shift for address calculation.
+ const int right_shift = X265_DEPTH - X265_NS::SAO::SAO_BO_BITS;
+ const int left_shift = 2;
+ const int shift = right_shift - left_shift;
+ // Mask out bits 7, 1 & 0 to account for combination of shifts.
+ const int mask = 0x7c;
+
+ // Compute statistics into temporary buffers.
+ for (int y = 0; y < endY; y++)
+ {
+ int x = 0;
+ for (; x + 8 < endX; x += 8)
+ {
+ uint64_t class_idx_64 =
+ *reinterpret_cast<const uint64_t *>(rec + x) >> shift;
+
+ for (int i = 0; i < 8; ++i)
+ {
+ const int idx = i & 1;
+ const int off = (class_idx_64 >> (i * 8)) & mask;
+ *reinterpret_cast<uint32_t*>(stats_b[idx] + off) += diff[x + i];
+ *reinterpret_cast<uint32_t*>(count_b[idx] + off) += 1;
+ }
+ }
+
+ if (x < endX)
+ {
+ uint64_t class_idx_64 =
+ *reinterpret_cast<const uint64_t *>(rec + x) >> shift;
+
+ for (int i = 0; (i + x) < endX; ++i)
+ {
+ const int idx = i & 1;
+ const int off = (class_idx_64 >> (i * 8)) & mask;
+ *reinterpret_cast<uint32_t*>(stats_b[idx] + off) += diff[x + i];
+ *reinterpret_cast<uint32_t*>(count_b[idx] + off) += 1;
+ }
+ }
+
+ diff += MAX_CU_SIZE;
+ rec += stride;
+ }
+
+ // Reduce temporary buffers to destination using Neon.
+ for (int i = 0; i < 32; i += 4)
+ {
+ int32x4_t s0 = vld1q_s32(stats_tmp + i);
+ int32x4_t s1 = vld1q_s32(stats + i);
+ vst1q_s32(stats + i, vaddq_s32(s0, s1));
+
+ int32x4_t c0 = vld1q_s32(count_tmp + i);
+ int32x4_t c1 = vld1q_s32(count + i);
+ vst1q_s32(count + i, vaddq_s32(c0, c1));
+ }
+}
+
+void saoCuStatsE0_neon(const int16_t *diff, const pixel *rec, intptr_t stride,
+ int endX, int endY, int32_t *stats, int32_t *count)
+{
+ // Separate buffers for each edge type, so that we can vectorise.
+ int16x8_t tmp_count[5] = { vdupq_n_s16(0), vdupq_n_s16(0), vdupq_n_s16(0),
+ vdupq_n_s16(0), vdupq_n_s16(0) };
+ int32x4_t tmp_stats[5] = { vdupq_n_s32(0), vdupq_n_s32(0), vdupq_n_s32(0),
+ vdupq_n_s32(0), vdupq_n_s32(0) };
+
+ for (int y = 0; y < endY; y++)
+ {
+ // Calculate negated sign_left(x) directly, to save negation when
+ // reusing sign_right(x) as sign_left(x + 1).
+ int8x16_t neg_sign_left = vdupq_n_s8(x265_signOf(rec[-1] - rec[0]));
+ for (int x = 0; x < endX; x += 16)
+ {
+ int8x16_t sign_right = signOf_neon(rec + x, rec + x + 1);
+
+ // neg_sign_left(x) = sign_right(x + 1), reusing one from previous
+ // iteration.
+ neg_sign_left = vextq_s8(neg_sign_left, sign_right, 15);
+
+ // Subtract instead of add, as sign_left is negated.
+ int8x16_t edge_type = vsubq_s8(sign_right, neg_sign_left);
+
+ // For reuse in the next iteration.
+ neg_sign_left = sign_right;
+
+ edge_type = mask_inactive_elems(endX - x, edge_type);
+ compute_eo_stats(edge_type, diff + x, tmp_count, tmp_stats);
+ }
+
+ diff += MAX_CU_SIZE;
+ rec += stride;
+ }
+
+ reduce_eo_stats(tmp_stats, tmp_count, stats, count);
+}
+
+void saoCuStatsE1_neon(const int16_t *diff, const pixel *rec, intptr_t stride,
+ int8_t *upBuff1, int endX, int endY, int32_t *stats,
+ int32_t *count)
+{
+ // Separate buffers for each edge type, so that we can vectorise.
+ int16x8_t tmp_count[5] = { vdupq_n_s16(0), vdupq_n_s16(0), vdupq_n_s16(0),
+ vdupq_n_s16(0), vdupq_n_s16(0) };
+ int32x4_t tmp_stats[5] = { vdupq_n_s32(0), vdupq_n_s32(0), vdupq_n_s32(0),
+ vdupq_n_s32(0), vdupq_n_s32(0) };
+
+ // Negate upBuff1 (sign_up), so we can subtract and save repeated negations.
+ for (int x = 0; x < endX; x += 16)
+ {
+ vst1q_s8(upBuff1 + x, vnegq_s8(vld1q_s8(upBuff1 + x)));
+ }
+
+ for (int y = 0; y < endY; y++)
+ {
+ for (int x = 0; x < endX; x += 16)
+ {
+ int8x16_t sign_up = vld1q_s8(upBuff1 + x);
+ int8x16_t sign_down = signOf_neon(rec + x, rec + x + stride);
+
+ // Subtract instead of add, as sign_up is negated.
+ int8x16_t edge_type = vsubq_s8(sign_down, sign_up);
+
+ // For reuse in the next iteration.
+ vst1q_s8(upBuff1 + x, sign_down);
+
+ edge_type = mask_inactive_elems(endX - x, edge_type);
+ compute_eo_stats(edge_type, diff + x, tmp_count, tmp_stats);
+ }
+
+ diff += MAX_CU_SIZE;
+ rec += stride;
+ }
+
+ reduce_eo_stats(tmp_stats, tmp_count, stats, count);
+}
+
+void saoCuStatsE2_neon(const int16_t *diff, const pixel *rec, intptr_t stride,
+ int8_t *upBuff1, int8_t *upBufft, int endX, int endY,
+ int32_t *stats, int32_t *count)
+{
+ // Separate buffers for each edge type, so that we can vectorise.
+ int16x8_t tmp_count[5] = { vdupq_n_s16(0), vdupq_n_s16(0), vdupq_n_s16(0),
+ vdupq_n_s16(0), vdupq_n_s16(0) };
+ int32x4_t tmp_stats[5] = { vdupq_n_s32(0), vdupq_n_s32(0), vdupq_n_s32(0),
+ vdupq_n_s32(0), vdupq_n_s32(0) };
+
+ // Negate upBuff1 (sign_up) so we can subtract and save repeated negations.
+ for (int x = 0; x < endX; x += 16)
+ {
+ vst1q_s8(upBuff1 + x, vnegq_s8(vld1q_s8(upBuff1 + x)));
+ }
+
+ for (int y = 0; y < endY; y++)
+ {
+ upBufft[0] = x265_signOf(rec[-1] - rec[stride]);
+ for (int x = 0; x < endX; x += 16)
+ {
+ int8x16_t sign_up = vld1q_s8(upBuff1 + x);
+ int8x16_t sign_down = signOf_neon(rec + x, rec + x + stride + 1);
+
+ // Subtract instead of add, as sign_up is negated.
+ int8x16_t edge_type = vsubq_s8(sign_down, sign_up);
+
+ // For reuse in the next iteration.
+ vst1q_s8(upBufft + x + 1, sign_down);
+
+ edge_type = mask_inactive_elems(endX - x, edge_type);
+ compute_eo_stats(edge_type, diff + x, tmp_count, tmp_stats);
+ }
+
+ std::swap(upBuff1, upBufft);
+
+ rec += stride;
+ diff += MAX_CU_SIZE;
+ }
+
+ reduce_eo_stats(tmp_stats, tmp_count, stats, count);
+}
+
+void saoCuStatsE3_neon(const int16_t *diff, const pixel *rec, intptr_t stride,
+ int8_t *upBuff1, int endX, int endY, int32_t *stats,
+ int32_t *count)
+{
+ // Separate buffers for each edge type, so that we can vectorise.
+ int16x8_t tmp_count[5] = { vdupq_n_s16(0), vdupq_n_s16(0), vdupq_n_s16(0),
+ vdupq_n_s16(0), vdupq_n_s16(0) };
+ int32x4_t tmp_stats[5] = { vdupq_n_s32(0), vdupq_n_s32(0), vdupq_n_s32(0),
+ vdupq_n_s32(0), vdupq_n_s32(0) };
+
+ // Negate upBuff1 (sign_up) so we can subtract and save repeated negations.
+ for (int x = 0; x < endX; x += 16)
+ {
+ vst1q_s8(upBuff1 + x, vnegq_s8(vld1q_s8(upBuff1 + x)));
+ }
+
+ for (int y = 0; y < endY; y++)
+ {
+ for (int x = 0; x < endX; x += 16)
+ {
+ int8x16_t sign_up = vld1q_s8(upBuff1 + x);
+ int8x16_t sign_down = signOf_neon(rec + x, rec + x + stride - 1);
+
+ // subtraction instead of addition, as sign_up is negated.
+ int8x16_t edge_type = vsubq_s8(sign_down, sign_up);
+
+ // For reuse in the next iteration.
+ vst1q_s8(upBuff1 + x - 1, sign_down);
+
+ edge_type = mask_inactive_elems(endX - x, edge_type);
+ compute_eo_stats(edge_type, diff + x, tmp_count, tmp_stats);
+ }
+
+ upBuff1[endX - 1] = x265_signOf(rec[endX] - rec[endX - 1 + stride]);
+
+ rec += stride;
+ diff += MAX_CU_SIZE;
+ }
+
+ reduce_eo_stats(tmp_stats, tmp_count, stats, count);
+}
+
+void setupSaoPrimitives_neon(EncoderPrimitives &p)
+{
+ p.saoCuStatsBO = saoCuStatsBO_neon;
+ p.saoCuStatsE0 = saoCuStatsE0_neon;
+ p.saoCuStatsE1 = saoCuStatsE1_neon;
+ p.saoCuStatsE2 = saoCuStatsE2_neon;
+ p.saoCuStatsE3 = saoCuStatsE3_neon;
+}
+} // namespace X265_NS
diff --git a/source/common/aarch64/sao-prim.h b/source/common/aarch64/sao-prim.h
new file mode 100644
index 000000000..1437ccd75
--- /dev/null
+++ b/source/common/aarch64/sao-prim.h
@@ -0,0 +1,33 @@
+/*****************************************************************************
+ * Copyright (C) 2024 MulticoreWare, Inc
+ *
+ * Authors: Hari Limaye <hari.limaye at arm.com>
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02111, USA.
+ *
+ * This program is also available under a commercial proprietary license.
+ * For more information, contact us at license @ x265.com.
+ *****************************************************************************/
+
+#ifndef X265_COMMON_AARCH64_SAO_PRIM_H
+#define X265_COMMON_AARCH64_SAO_PRIM_H
+
+#include "primitives.h"
+
+namespace X265_NS {
+void setupSaoPrimitives_neon(EncoderPrimitives &p);
+}
+
+#endif // X265_COMMON_AARCH64_SAO_PRIM_H
--
2.42.1
-------------- next part --------------
A non-text attachment was scrubbed...
Name: v2-0003-AArch64-Add-Neon-saoCuStats-primitives-for-low-bi.patch
Type: text/x-patch
Size: 19520 bytes
Desc: not available
URL: <http://mailman.videolan.org/pipermail/x265-devel/attachments/20240522/8b941219/attachment-0001.bin>
More information about the x265-devel
mailing list