From 51d07554a039004ff278009852f9d33c0f76bf91 Mon Sep 17 00:00:00 2001
From: George Steed <george.steed@arm.com>
Date: Tue, 27 Aug 2024 14:33:25 +0100
Subject: [PATCH] [AArch64] Add SME implementation of ScaleRowDown2Linear

There is no benefit from an SVE version of this kernel for devices with
an SVE vector length of 128-bits, so skip directly to SME instead.  We
do not use the ZA tile here, so this is a purely streaming-SVE (SSVE)
implementation.

Change-Id: Ie6b91bd4407130ba2653838088e81e72e4460f68
Reviewed-on: https://chromium-review.googlesource.com/c/libyuv/libyuv/+/5913884
Reviewed-by: Justin Green <greenjustin@google.com>
Reviewed-by: Frank Barchard <fbarchard@chromium.org>
---
 include/libyuv/scale_row.h |  4 ++++
 source/scale.cc            |  5 +++--
 source/scale_sme.cc        | 40 ++++++++++++++++++++++++++++++++++++++
 3 files changed, 47 insertions(+), 2 deletions(-)

diff --git a/include/libyuv/scale_row.h b/include/libyuv/scale_row.h
index 6e488a4cd..37907f377 100644
--- a/include/libyuv/scale_row.h
+++ b/include/libyuv/scale_row.h
@@ -1416,6 +1416,10 @@ void ScaleRowDown2Linear_NEON(const uint8_t* src_ptr,
                               ptrdiff_t src_stride,
                               uint8_t* dst,
                               int dst_width);
+void ScaleRowDown2Linear_SME(const uint8_t* src_ptr,
+                             ptrdiff_t src_stride,
+                             uint8_t* dst,
+                             int dst_width);
 void ScaleRowDown2Box_NEON(const uint8_t* src_ptr,
                            ptrdiff_t src_stride,
                            uint8_t* dst,
diff --git a/source/scale.cc b/source/scale.cc
index 2f9bb4990..610be07cd 100644
--- a/source/scale.cc
+++ b/source/scale.cc
@@ -76,8 +76,9 @@ static void ScalePlaneDown2(int src_width,
 #endif
 #if defined(HAS_SCALEROWDOWN2_SME)
   if (TestCpuFlag(kCpuHasSME)) {
-    if (filtering == kFilterNone) {
-      ScaleRowDown2 = ScaleRowDown2_SME;
+    if (filtering == kFilterNone || filtering == kFilterLinear) {
+      ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_SME
+                                               : ScaleRowDown2Linear_SME;
     }
   }
 #endif
diff --git a/source/scale_sme.cc b/source/scale_sme.cc
index d4f0a827d..34c6884a5 100644
--- a/source/scale_sme.cc
+++ b/source/scale_sme.cc
@@ -56,6 +56,46 @@ __arm_locally_streaming void ScaleRowDown2_SME(const uint8_t* src_ptr,
       : "memory", "cc", "z0", "z1", "p0");
 }
 
+__arm_locally_streaming void ScaleRowDown2Linear_SME(const uint8_t* src_ptr,
+                                                     ptrdiff_t src_stride,
+                                                     uint8_t* dst,
+                                                     int dst_width) {
+  // Streaming-SVE only, no use of ZA tile.
+  (void)src_stride;
+  int vl;
+  asm volatile(
+      "cntb     %x[vl]                                       \n"
+      "subs     %w[dst_width], %w[dst_width], %w[vl]         \n"
+      "b.lt     2f                                           \n"
+
+      "1:                                                    \n"
+      "ptrue    p0.b                                         \n"
+      "ld2b     {z0.b, z1.b}, p0/z, [%[src_ptr]]             \n"
+      "incb     %[src_ptr], all, mul #2                      \n"
+      "urhadd   z0.b, p0/m, z0.b, z1.b                       \n"
+      "subs     %w[dst_width], %w[dst_width], %w[vl]         \n"
+      "st1b     {z0.b}, p0, [%[dst_ptr]]                     \n"
+      "incb     %[dst_ptr]                                   \n"
+      "b.ge     1b                                           \n"
+
+      "2:                                                    \n"
+      "adds     %w[dst_width], %w[dst_width], %w[vl]         \n"
+      "b.eq     99f                                          \n"
+
+      "whilelt  p0.b, wzr, %w[dst_width]                     \n"
+      "ld2b     {z0.b, z1.b}, p0/z, [%[src_ptr]]             \n"
+      "urhadd   z0.b, p0/m, z0.b, z1.b                       \n"
+      "st1b     {z0.b}, p0, [%[dst_ptr]]                     \n"
+
+      "99:                                                   \n"
+      : [src_ptr] "+r"(src_ptr),      // %[src_ptr]
+        [dst_ptr] "+r"(dst),          // %[dst_ptr]
+        [dst_width] "+r"(dst_width),  // %[dst_width]
+        [vl] "=r"(vl)                 // %[vl]
+      :
+      : "memory", "cc", "z0", "z1", "p0");
+}
+
 #endif  // !defined(LIBYUV_DISABLE_SME) && defined(CLANG_HAS_SME) &&
         // defined(__aarch64__)