@@ -285,7 +285,7 @@ int ff_shuffle_filter_coefficients(SwsContext *c, int *filterPos,
#if ARCH_X86_64
int i = 0, j = 0, k = 0;
int cpu_flags = av_get_cpu_flags();
- if (!filter || dstW % 16 != 0) return 0;
+ if (!filter || (dstW % 4 != 0)) return 0;
if (EXTERNAL_AVX2_FAST(cpu_flags) && !(cpu_flags & AV_CPU_FLAG_SLOW_GATHER)) {
if ((c->srcBpc == 8) && (c->dstBpc <= 14)) {
int16_t *filterCopy = NULL;
@@ -296,9 +296,11 @@ 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.
@@ -312,6 +314,17 @@ int ff_shuffle_filter_coefficients(SwsContext *c, int *filterPos,
}
}
}
+ // 4 pixels are processed at a time in the tail.
+ for (; i + 4 <= dstW; i += 4) {
+ // 4 filter coeffs are processed at a time.
+ for (k = 0; k + 4 <= filterSize; k += 4) {
+ for (j = 0; j < 4; ++j) {
+ int from = (i + j) * filterSize + k;
+ int to = i * filterSize + j * 4 + k * 4;
+ memcpy(&filter[to], &filterCopy[from], 4 * sizeof(int16_t));
+ }
+ }
+ }
}
if (filterCopy)
av_free(filterCopy);
@@ -580,9 +580,9 @@ 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)
+ if (c->chrDstW % 4 == 0)
ASSIGN_AVX2_SCALE_FUNC(c->hcScale, c->hChrFilterSize);
- if (c->dstW % 16 == 0)
+ if (c->dstW % 4 == 0)
ASSIGN_AVX2_SCALE_FUNC(c->hyScale, c->hLumFilterSize);
}
}