[x265] [PATCH 3/3] AArch64: Add Armv8.4 Neon DotProd implementations of sse_pp

Hari Limaye hari.limaye at arm.com
Tue Jun 25 12:52:48 UTC 2024


Add implementations of sse_pp primitives using Neon DotProd
instructions, which are mandatory from Armv8.4. These implementations
use a single UDOT instruction in place of multiple UMULL, UADALP
sequences.
---
 source/common/CMakeLists.txt             |   2 +-
 source/common/aarch64/asm-primitives.cpp |  11 ++
 source/common/aarch64/fun-decls.h        |   1 +
 source/common/aarch64/ssd-neon-dotprod.S | 165 +++++++++++++++++++++++
 4 files changed, 178 insertions(+), 1 deletion(-)
 create mode 100644 source/common/aarch64/ssd-neon-dotprod.S

diff --git a/source/common/CMakeLists.txt b/source/common/CMakeLists.txt
index 568c1a50c..d94ea8795 100644
--- a/source/common/CMakeLists.txt
+++ b/source/common/CMakeLists.txt
@@ -112,7 +112,7 @@ if(ENABLE_ASSEMBLY AND (ARM64 OR CROSS_COMPILE_ARM64))
     set(A_SRCS asm.S mc-a.S mc-a-common.S sad-a.S pixel-util.S pixel-util-common.S p2s.S p2s-common.S ipfilter.S ipfilter-common.S blockcopy8.S blockcopy8-common.S ssd-a.S ssd-a-common.S)
     set(A_SRCS_SVE asm-sve.S blockcopy8-sve.S p2s-sve.S pixel-util-sve.S)
     set(A_SRCS_SVE2 mc-a-sve2.S pixel-util-sve2.S ipfilter-sve2.S ssd-a-sve2.S)
-    set(A_SRCS_NEON_DOTPROD sad-neon-dotprod.S)
+    set(A_SRCS_NEON_DOTPROD sad-neon-dotprod.S ssd-neon-dotprod.S)
     set(VEC_PRIMITIVES)
 
     set(ARM_ASMS "${A_SRCS}" CACHE INTERNAL "ARM Assembly Sources")
diff --git a/source/common/aarch64/asm-primitives.cpp b/source/common/aarch64/asm-primitives.cpp
index fc7d205db..32d75ee35 100644
--- a/source/common/aarch64/asm-primitives.cpp
+++ b/source/common/aarch64/asm-primitives.cpp
@@ -1172,6 +1172,17 @@ void setupNeonDotProdPrimitives(EncoderPrimitives &p)
     LUMA_PU_MULTIPLE_16(sad, pixel_sad, neon_dotprod);
     LUMA_PU_MULTIPLE_16(sad_x3, sad_x3, neon_dotprod);
     LUMA_PU_MULTIPLE_16(sad_x4, sad_x4, neon_dotprod);
+
+    // sse_pp
+    ALL_LUMA_TU(sse_pp, pixel_sse_pp, neon_dotprod);
+    p.chroma[X265_CSP_I420].cu[BLOCK_420_4x4].sse_pp   = PFX(pixel_sse_pp_4x4_neon_dotprod);
+    p.chroma[X265_CSP_I420].cu[BLOCK_420_8x8].sse_pp   = PFX(pixel_sse_pp_8x8_neon_dotprod);
+    p.chroma[X265_CSP_I420].cu[BLOCK_420_16x16].sse_pp = PFX(pixel_sse_pp_16x16_neon_dotprod);
+    p.chroma[X265_CSP_I420].cu[BLOCK_420_32x32].sse_pp = PFX(pixel_sse_pp_32x32_neon_dotprod);
+    p.chroma[X265_CSP_I422].cu[BLOCK_422_4x8].sse_pp   = PFX(pixel_sse_pp_4x8_neon_dotprod);
+    p.chroma[X265_CSP_I422].cu[BLOCK_422_8x16].sse_pp  = PFX(pixel_sse_pp_8x16_neon_dotprod);
+    p.chroma[X265_CSP_I422].cu[BLOCK_422_16x32].sse_pp = PFX(pixel_sse_pp_16x32_neon_dotprod);
+    p.chroma[X265_CSP_I422].cu[BLOCK_422_32x64].sse_pp = PFX(pixel_sse_pp_32x64_neon_dotprod);
 }
 #else // !HIGH_BIT_DEPTH
 void setupNeonDotProdPrimitives(EncoderPrimitives &)
diff --git a/source/common/aarch64/fun-decls.h b/source/common/aarch64/fun-decls.h
index ba5496032..363725fef 100644
--- a/source/common/aarch64/fun-decls.h
+++ b/source/common/aarch64/fun-decls.h
@@ -175,6 +175,7 @@ DECLS(sve2);
 FUNCDEF_PU_MULT_16(int, pixel_sad, neon_dotprod, const pixel*, intptr_t, const pixel*, intptr_t);
 FUNCDEF_PU_MULT_16(void, sad_x3, neon_dotprod, const pixel*, const pixel*, const pixel*, const pixel*, intptr_t, int32_t*);
 FUNCDEF_PU_MULT_16(void, sad_x4, neon_dotprod, const pixel*, const pixel*, const pixel*, const pixel*, const pixel*, intptr_t, int32_t*);
+FUNCDEF_PU(sse_t, pixel_sse_pp, neon_dotprod, const pixel*, intptr_t, const pixel*, intptr_t);
 
 void PFX(pixel_planecopy_cp_neon(const uint8_t* src, intptr_t srcStride, pixel* dst, intptr_t dstStride, int width, int height, int shift));
 
diff --git a/source/common/aarch64/ssd-neon-dotprod.S b/source/common/aarch64/ssd-neon-dotprod.S
new file mode 100644
index 000000000..4df4fb35b
--- /dev/null
+++ b/source/common/aarch64/ssd-neon-dotprod.S
@@ -0,0 +1,165 @@
+/*****************************************************************************
+ * 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 "asm.S"
+
+.arch armv8.2-a+dotprod
+
+#ifdef __APPLE__
+.section __RODATA,__rodata
+#else
+.section .rodata
+#endif
+
+.align 4
+
+.text
+
+// Fully unrolled.
+.macro SSE_PP_4xN h
+function PFX(pixel_sse_pp_4x\h\()_neon_dotprod)
+    movi            v0.4s, #0
+.rept \h / 4
+    ldr             s16, [x0]
+    ldr             s17, [x2]
+    add             x0, x0, x1
+    add             x2, x2, x3
+    ld1             {v16.s}[1], [x0], x1
+    ld1             {v16.s}[2], [x0], x1
+    ld1             {v16.s}[3], [x0], x1
+    ld1             {v17.s}[1], [x2], x3
+    ld1             {v17.s}[2], [x2], x3
+    ld1             {v17.s}[3], [x2], x3
+
+    uabd            v1.16b, v16.16b, v17.16b
+    udot            v0.4s, v1.16b, v1.16b
+.endr
+    addv            s0, v0.4s
+    fmov            w0, s0
+    ret
+endfunc
+.endm
+
+SSE_PP_4xN 4
+SSE_PP_4xN 8
+
+// Fully unrolled.
+.macro SSE_PP_8xN h
+function PFX(pixel_sse_pp_8x\h\()_neon_dotprod)
+    movi            v0.4s, #0
+.rept \h
+    ld1             {v16.8b}, [x0], x1
+    ld1             {v17.8b}, [x2], x3
+
+    uabd            v1.8b, v16.8b, v17.8b
+    udot            v0.2s, v1.8b, v1.8b
+.endr
+    addv            s0, v0.4s
+    fmov            w0, s0
+    ret
+endfunc
+.endm
+
+SSE_PP_8xN 8
+SSE_PP_8xN 16
+
+// Fully unrolled.
+.macro SSE_PP_16xN h
+function PFX(pixel_sse_pp_16x\h\()_neon_dotprod)
+    movi            v0.4s, #0
+    movi            v1.4s, #0
+.rept \h / 2
+    ld1             {v16.16b}, [x0], x1
+    ld1             {v17.16b}, [x2], x3
+    ld1             {v18.16b}, [x0], x1
+    ld1             {v19.16b}, [x2], x3
+
+    uabd            v2.16b, v16.16b, v17.16b
+    udot            v0.4s, v2.16b, v2.16b
+    uabd            v3.16b, v18.16b, v19.16b
+    udot            v1.4s, v3.16b, v3.16b
+.endr
+    add             v0.4s, v0.4s, v1.4s
+    addv            s0, v0.4s
+    fmov            w0, s0
+    ret
+endfunc
+.endm
+
+SSE_PP_16xN 16
+SSE_PP_16xN 32
+
+// Loop unrolled to process 4 rows per iteration.
+.macro SSE_PP_32xN h
+function PFX(pixel_sse_pp_32x\h\()_neon_dotprod)
+    mov             w12, #(\h / 4)
+    movi            v0.4s, #0
+    movi            v1.4s, #0
+.Loop_sse_pp_32_x\h:
+    sub             w12, w12, #1
+.rept 4
+    ld1             {v16.16b,v17.16b}, [x0], x1
+    ld1             {v18.16b,v19.16b}, [x2], x3
+
+    uabd            v2.16b, v16.16b, v18.16b
+    udot            v0.4s, v2.16b, v2.16b
+    uabd            v3.16b, v17.16b, v19.16b
+    udot            v1.4s, v3.16b, v3.16b
+.endr
+    cbnz            w12, .Loop_sse_pp_32_x\h
+    add             v0.4s, v0.4s, v1.4s
+    addv            s0, v0.4s
+    fmov            w0, s0
+    ret
+endfunc
+.endm
+
+SSE_PP_32xN 32
+SSE_PP_32xN 64
+
+// Loop unrolled to process 4 rows per iteration.
+function PFX(pixel_sse_pp_64x64_neon_dotprod)
+    mov             w12, #16
+    movi            v0.4s, #0
+    movi            v1.4s, #0
+.Loop_sse_pp_64:
+    sub             w12, w12, #1
+.rept 4
+    ld1             {v16.16b-v19.16b}, [x0], x1
+    ld1             {v20.16b-v23.16b}, [x2], x3
+
+    uabd            v2.16b, v16.16b, v20.16b
+    udot            v0.4s, v2.16b, v2.16b
+    uabd            v3.16b, v17.16b, v21.16b
+    udot            v1.4s, v3.16b, v3.16b
+    uabd            v4.16b, v18.16b, v22.16b
+    udot            v0.4s, v4.16b, v4.16b
+    uabd            v5.16b, v19.16b, v23.16b
+    udot            v1.4s, v5.16b, v5.16b
+.endr
+    cbnz            w12, .Loop_sse_pp_64
+    add             v0.4s, v0.4s, v1.4s
+    addv            s0, v0.4s
+    fmov            w0, s0
+    ret
+endfunc
-- 
2.42.1

-------------- next part --------------
A non-text attachment was scrubbed...
Name: 0003-AArch64-Add-Armv8.4-Neon-DotProd-implementations-of-.patch
Type: text/x-patch
Size: 9000 bytes
Desc: not available
URL: <http://mailman.videolan.org/pipermail/x265-devel/attachments/20240625/174da450/attachment.bin>


More information about the x265-devel mailing list