[FFmpeg-cvslog] libswscale: Enable hscale_avx2 for all input sizes.
Alan Kelly
git at videolan.org
Thu Aug 18 17:26:45 EEST 2022
ffmpeg | branch: master | Alan Kelly <alankelly-at-google.com at ffmpeg.org> | Fri Jul 15 16:59:43 2022 +0200| [a38293e4448c9389e604af9858984361a5677a20] | committer: Anton Khirnov
libswscale: Enable hscale_avx2 for all input sizes.
ff_shuffle_filter_coefficients shuffles the tail as required.
Signed-off-by: Anton Khirnov <anton at khirnov.net>
> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=a38293e4448c9389e604af9858984361a5677a20
---
libswscale/utils.c | 19 ++++++++++++++++---
libswscale/x86/swscale.c | 6 ++----
tests/checkasm/sw_scale.c | 2 +-
3 files changed, 19 insertions(+), 8 deletions(-)
diff --git a/libswscale/utils.c b/libswscale/utils.c
index 34503e57f4..baa1791ebe 100644
--- a/libswscale/utils.c
+++ b/libswscale/utils.c
@@ -268,8 +268,7 @@ int ff_shuffle_filter_coefficients(SwsContext *c, int *filterPos,
#if ARCH_X86_64
int i, j, k;
int cpu_flags = av_get_cpu_flags();
- // avx2 hscale filter processes 16 pixel blocks.
- if (!filter || dstW % 16 != 0)
+ if (!filter)
return 0;
if (EXTERNAL_AVX2_FAST(cpu_flags) && !(cpu_flags & AV_CPU_FLAG_SLOW_GATHER)) {
if ((c->srcBpc == 8) && (c->dstBpc <= 14)) {
@@ -281,9 +280,11 @@ int ff_shuffle_filter_coefficients(SwsContext *c, int *filterPos,
}
// Do not swap filterPos for pixels which won't be processed by
// the main loop.
- for (i = 0; i + 8 <= dstW; i += 8) {
+ for (i = 0; i + 16 <= dstW; i += 16) {
FFSWAP(int, filterPos[i + 2], filterPos[i + 4]);
FFSWAP(int, filterPos[i + 3], filterPos[i + 5]);
+ FFSWAP(int, filterPos[i + 10], filterPos[i + 12]);
+ FFSWAP(int, filterPos[i + 11], filterPos[i + 13]);
}
if (filterSize > 4) {
// 16 pixels are processed at a time.
@@ -297,6 +298,18 @@ int ff_shuffle_filter_coefficients(SwsContext *c, int *filterPos,
}
}
}
+ // 4 pixels are processed at a time in the tail.
+ for (; i < dstW; i += 4) {
+ // 4 filter coeffs are processed at a time.
+ int rem = dstW - i >= 4 ? 4 : dstW - i;
+ for (k = 0; k + 4 <= filterSize; k += 4) {
+ for (j = 0; j < rem; ++j) {
+ int from = (i + j) * filterSize + k;
+ int to = i * filterSize + j * 4 + k * 4;
+ memcpy(&filter[to], &filterCopy[from], 4 * sizeof(int16_t));
+ }
+ }
+ }
}
av_free(filterCopy);
}
diff --git a/libswscale/x86/swscale.c b/libswscale/x86/swscale.c
index 89ef9f5d2b..ec1ca0e01c 100644
--- a/libswscale/x86/swscale.c
+++ b/libswscale/x86/swscale.c
@@ -625,10 +625,8 @@ switch(c->dstBpc){ \
if (EXTERNAL_AVX2_FAST(cpu_flags) && !(cpu_flags & AV_CPU_FLAG_SLOW_GATHER)) {
if ((c->srcBpc == 8) && (c->dstBpc <= 14)) {
- if (c->chrDstW % 16 == 0)
- ASSIGN_AVX2_SCALE_FUNC(c->hcScale, c->hChrFilterSize);
- if (c->dstW % 16 == 0)
- ASSIGN_AVX2_SCALE_FUNC(c->hyScale, c->hLumFilterSize);
+ ASSIGN_AVX2_SCALE_FUNC(c->hcScale, c->hChrFilterSize);
+ ASSIGN_AVX2_SCALE_FUNC(c->hyScale, c->hLumFilterSize);
}
}
diff --git a/tests/checkasm/sw_scale.c b/tests/checkasm/sw_scale.c
index cbe4460a99..9c07dd0421 100644
--- a/tests/checkasm/sw_scale.c
+++ b/tests/checkasm/sw_scale.c
@@ -329,7 +329,7 @@ static void check_hscale(void)
ff_sws_init_scale(ctx);
memcpy(filterAvx2, filter, sizeof(uint16_t) * (SRC_PIXELS * MAX_FILTER_WIDTH + MAX_FILTER_WIDTH));
if ((cpu_flags & AV_CPU_FLAG_AVX2) && !(cpu_flags & AV_CPU_FLAG_SLOW_GATHER))
- ff_shuffle_filter_coefficients(ctx, filterPosAvx, width, filterAvx2, SRC_PIXELS);
+ ff_shuffle_filter_coefficients(ctx, filterPosAvx, width, filterAvx2, ctx->dstW);
if (check_func(ctx->hcScale, "hscale_%d_to_%d__fs_%d_dstW_%d", ctx->srcBpc, ctx->dstBpc + 1, width, ctx->dstW)) {
memset(dst0, 0, SRC_PIXELS * sizeof(dst0[0]));
More information about the ffmpeg-cvslog
mailing list