[FFmpeg-devel] [PATCH 2/6] lavc/audiodsp: properly unroll vector_clipf

["Rémi Denis-Courmont" <remi@remlab.net>]

Given that source and destination can alias, the compiler was forced to
perform each read-modify-write sequentially. We cannot use the `restrict`
qualifier to avoid this here because the AC-3 encoder uses the function
in-place. Instead this commit provides an explicit guarantee to the
compiler that batches of 8 elements will not overlap, so that it can
interleave calculations.

In practice contemporary optimising compilers are able to unroll and keep
the temporary array in FPU registers (without spilling).

On SiFive-U74, this speeds the same signs branch by 4x, and the
opposite signs branch 1.5x.
---
 libavcodec/audiodsp.c | 40 +++++++++++++++++-----------------------
 1 file changed, 17 insertions(+), 23 deletions(-)

diff --git a/libavcodec/audiodsp.c b/libavcodec/audiodsp.c
index c5427d3535..9e83f06aaa 100644
--- a/libavcodec/audiodsp.c
+++ b/libavcodec/audiodsp.c
@@ -38,41 +38,35 @@ static inline float clipf_c_one(float a, uint32_t mini,
 static void vector_clipf_c_opposite_sign(float *dst, const float *src,
                                          float min, float max, int len)
 {
-    int i;
     uint32_t mini        = av_float2int(min);
     uint32_t maxi        = av_float2int(max);
     uint32_t maxisign    = maxi ^ (1U << 31);
 
-    for (i = 0; i < len; i += 8) {
-        dst[i + 0] = clipf_c_one(src[i + 0], mini, maxi, maxisign);
-        dst[i + 1] = clipf_c_one(src[i + 1], mini, maxi, maxisign);
-        dst[i + 2] = clipf_c_one(src[i + 2], mini, maxi, maxisign);
-        dst[i + 3] = clipf_c_one(src[i + 3], mini, maxi, maxisign);
-        dst[i + 4] = clipf_c_one(src[i + 4], mini, maxi, maxisign);
-        dst[i + 5] = clipf_c_one(src[i + 5], mini, maxi, maxisign);
-        dst[i + 6] = clipf_c_one(src[i + 6], mini, maxi, maxisign);
-        dst[i + 7] = clipf_c_one(src[i + 7], mini, maxi, maxisign);
+    for (int i = 0; i < len; i += 8) {
+        float tmp[8];
+
+        for (int j = 0; j < 8; j++)
+            tmp[j]= clipf_c_one(src[i + j], mini, maxi, maxisign);
+        for (int j = 0; j < 8; j++)
+            dst[i + j] = tmp[j];
     }
 }
 
 static void vector_clipf_c(float *dst, const float *src, int len,
                            float min, float max)
 {
-    int i;
-
     if (min < 0 && max > 0) {
         vector_clipf_c_opposite_sign(dst, src, min, max, len);
-    } else {
-        for (i = 0; i < len; i += 8) {
-            dst[i]     = av_clipf(src[i], min, max);
-            dst[i + 1] = av_clipf(src[i + 1], min, max);
-            dst[i + 2] = av_clipf(src[i + 2], min, max);
-            dst[i + 3] = av_clipf(src[i + 3], min, max);
-            dst[i + 4] = av_clipf(src[i + 4], min, max);
-            dst[i + 5] = av_clipf(src[i + 5], min, max);
-            dst[i + 6] = av_clipf(src[i + 6], min, max);
-            dst[i + 7] = av_clipf(src[i + 7], min, max);
-        }
+        return;
+    }
+
+    for (int i = 0; i < len; i += 8) {
+        float tmp[8];
+
+        for (int j = 0; j < 8; j++)
+            tmp[j]= av_clipf(src[i + j], min, max);
+        for (int j = 0; j < 8; j++)
+            dst[i + j] = tmp[j];
     }
 }
 
-- 
2.45.2

_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel

To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".