summaryrefslogtreecommitdiff
path: root/lib/jxl/transpose-inl.h
diff options
context:
space:
mode:
Diffstat (limited to 'lib/jxl/transpose-inl.h')
-rw-r--r--lib/jxl/transpose-inl.h124
1 files changed, 63 insertions, 61 deletions
diff --git a/lib/jxl/transpose-inl.h b/lib/jxl/transpose-inl.h
index d12b129..4674420 100644
--- a/lib/jxl/transpose-inl.h
+++ b/lib/jxl/transpose-inl.h
@@ -74,50 +74,51 @@ JXL_INLINE_TRANSPOSE void GenericTransposeBlock(TransposeSimdTag<true>,
static_assert(COLS_or_0 % 8 == 0, "Invalid number of columns");
for (size_t n = 0; n < ROWS; n += 8) {
for (size_t m = 0; m < COLS; m += 8) {
- auto i0 = from.LoadPart(BlockDesc<8>(), n + 0, m + 0);
- auto i1 = from.LoadPart(BlockDesc<8>(), n + 1, m + 0);
- auto i2 = from.LoadPart(BlockDesc<8>(), n + 2, m + 0);
- auto i3 = from.LoadPart(BlockDesc<8>(), n + 3, m + 0);
- auto i4 = from.LoadPart(BlockDesc<8>(), n + 4, m + 0);
- auto i5 = from.LoadPart(BlockDesc<8>(), n + 5, m + 0);
- auto i6 = from.LoadPart(BlockDesc<8>(), n + 6, m + 0);
- auto i7 = from.LoadPart(BlockDesc<8>(), n + 7, m + 0);
+ const BlockDesc<8> d;
+ auto i0 = from.LoadPart(d, n + 0, m + 0);
+ auto i1 = from.LoadPart(d, n + 1, m + 0);
+ auto i2 = from.LoadPart(d, n + 2, m + 0);
+ auto i3 = from.LoadPart(d, n + 3, m + 0);
+ auto i4 = from.LoadPart(d, n + 4, m + 0);
+ auto i5 = from.LoadPart(d, n + 5, m + 0);
+ auto i6 = from.LoadPart(d, n + 6, m + 0);
+ auto i7 = from.LoadPart(d, n + 7, m + 0);
// Surprisingly, this straightforward implementation (24 cycles on port5)
// is faster than load128+insert and LoadDup128+ConcatUpperLower+blend.
- const auto q0 = InterleaveLower(i0, i2);
- const auto q1 = InterleaveLower(i1, i3);
- const auto q2 = InterleaveUpper(i0, i2);
- const auto q3 = InterleaveUpper(i1, i3);
- const auto q4 = InterleaveLower(i4, i6);
- const auto q5 = InterleaveLower(i5, i7);
- const auto q6 = InterleaveUpper(i4, i6);
- const auto q7 = InterleaveUpper(i5, i7);
-
- const auto r0 = InterleaveLower(q0, q1);
- const auto r1 = InterleaveUpper(q0, q1);
- const auto r2 = InterleaveLower(q2, q3);
- const auto r3 = InterleaveUpper(q2, q3);
- const auto r4 = InterleaveLower(q4, q5);
- const auto r5 = InterleaveUpper(q4, q5);
- const auto r6 = InterleaveLower(q6, q7);
- const auto r7 = InterleaveUpper(q6, q7);
-
- i0 = ConcatLowerLower(r4, r0);
- i1 = ConcatLowerLower(r5, r1);
- i2 = ConcatLowerLower(r6, r2);
- i3 = ConcatLowerLower(r7, r3);
- i4 = ConcatUpperUpper(r4, r0);
- i5 = ConcatUpperUpper(r5, r1);
- i6 = ConcatUpperUpper(r6, r2);
- i7 = ConcatUpperUpper(r7, r3);
- to.StorePart(BlockDesc<8>(), i0, m + 0, n + 0);
- to.StorePart(BlockDesc<8>(), i1, m + 1, n + 0);
- to.StorePart(BlockDesc<8>(), i2, m + 2, n + 0);
- to.StorePart(BlockDesc<8>(), i3, m + 3, n + 0);
- to.StorePart(BlockDesc<8>(), i4, m + 4, n + 0);
- to.StorePart(BlockDesc<8>(), i5, m + 5, n + 0);
- to.StorePart(BlockDesc<8>(), i6, m + 6, n + 0);
- to.StorePart(BlockDesc<8>(), i7, m + 7, n + 0);
+ const auto q0 = InterleaveLower(d, i0, i2);
+ const auto q1 = InterleaveLower(d, i1, i3);
+ const auto q2 = InterleaveUpper(d, i0, i2);
+ const auto q3 = InterleaveUpper(d, i1, i3);
+ const auto q4 = InterleaveLower(d, i4, i6);
+ const auto q5 = InterleaveLower(d, i5, i7);
+ const auto q6 = InterleaveUpper(d, i4, i6);
+ const auto q7 = InterleaveUpper(d, i5, i7);
+
+ const auto r0 = InterleaveLower(d, q0, q1);
+ const auto r1 = InterleaveUpper(d, q0, q1);
+ const auto r2 = InterleaveLower(d, q2, q3);
+ const auto r3 = InterleaveUpper(d, q2, q3);
+ const auto r4 = InterleaveLower(d, q4, q5);
+ const auto r5 = InterleaveUpper(d, q4, q5);
+ const auto r6 = InterleaveLower(d, q6, q7);
+ const auto r7 = InterleaveUpper(d, q6, q7);
+
+ i0 = ConcatLowerLower(d, r4, r0);
+ i1 = ConcatLowerLower(d, r5, r1);
+ i2 = ConcatLowerLower(d, r6, r2);
+ i3 = ConcatLowerLower(d, r7, r3);
+ i4 = ConcatUpperUpper(d, r4, r0);
+ i5 = ConcatUpperUpper(d, r5, r1);
+ i6 = ConcatUpperUpper(d, r6, r2);
+ i7 = ConcatUpperUpper(d, r7, r3);
+ to.StorePart(d, i0, m + 0, n + 0);
+ to.StorePart(d, i1, m + 1, n + 0);
+ to.StorePart(d, i2, m + 2, n + 0);
+ to.StorePart(d, i3, m + 3, n + 0);
+ to.StorePart(d, i4, m + 4, n + 0);
+ to.StorePart(d, i5, m + 5, n + 0);
+ to.StorePart(d, i6, m + 6, n + 0);
+ to.StorePart(d, i7, m + 7, n + 0);
}
}
}
@@ -137,25 +138,26 @@ JXL_INLINE_TRANSPOSE void GenericTransposeBlock(TransposeSimdTag<true>,
static_assert(COLS_or_0 % 4 == 0, "Invalid number of columns");
for (size_t n = 0; n < ROWS; n += 4) {
for (size_t m = 0; m < COLS; m += 4) {
- const auto p0 = from.LoadPart(BlockDesc<4>(), n + 0, m + 0);
- const auto p1 = from.LoadPart(BlockDesc<4>(), n + 1, m + 0);
- const auto p2 = from.LoadPart(BlockDesc<4>(), n + 2, m + 0);
- const auto p3 = from.LoadPart(BlockDesc<4>(), n + 3, m + 0);
-
- const auto q0 = InterleaveLower(p0, p2);
- const auto q1 = InterleaveLower(p1, p3);
- const auto q2 = InterleaveUpper(p0, p2);
- const auto q3 = InterleaveUpper(p1, p3);
-
- const auto r0 = InterleaveLower(q0, q1);
- const auto r1 = InterleaveUpper(q0, q1);
- const auto r2 = InterleaveLower(q2, q3);
- const auto r3 = InterleaveUpper(q2, q3);
-
- to.StorePart(BlockDesc<4>(), r0, m + 0, n + 0);
- to.StorePart(BlockDesc<4>(), r1, m + 1, n + 0);
- to.StorePart(BlockDesc<4>(), r2, m + 2, n + 0);
- to.StorePart(BlockDesc<4>(), r3, m + 3, n + 0);
+ const BlockDesc<4> d;
+ const auto p0 = from.LoadPart(d, n + 0, m + 0);
+ const auto p1 = from.LoadPart(d, n + 1, m + 0);
+ const auto p2 = from.LoadPart(d, n + 2, m + 0);
+ const auto p3 = from.LoadPart(d, n + 3, m + 0);
+
+ const auto q0 = InterleaveLower(d, p0, p2);
+ const auto q1 = InterleaveLower(d, p1, p3);
+ const auto q2 = InterleaveUpper(d, p0, p2);
+ const auto q3 = InterleaveUpper(d, p1, p3);
+
+ const auto r0 = InterleaveLower(d, q0, q1);
+ const auto r1 = InterleaveUpper(d, q0, q1);
+ const auto r2 = InterleaveLower(d, q2, q3);
+ const auto r3 = InterleaveUpper(d, q2, q3);
+
+ to.StorePart(d, r0, m + 0, n + 0);
+ to.StorePart(d, r1, m + 1, n + 0);
+ to.StorePart(d, r2, m + 2, n + 0);
+ to.StorePart(d, r3, m + 3, n + 0);
}
}
}