diff options
| author | Johann <johannkoenig@google.com> | 2017-02-15 17:17:45 -0800 |
|---|---|---|
| committer | Johann Koenig <johannkoenig@google.com> | 2017-02-16 15:02:48 +0000 |
| commit | 44600442dca48a5586105b01403ac03fad54d05b (patch) | |
| tree | 839959ff4937f9527f1e641159890bf512669bdc /vp9/encoder/x86 | |
| parent | 60a10116d13c4394bed943e0213631735d03bb74 (diff) | |
| download | libvpx-44600442dca48a5586105b01403ac03fad54d05b.tar libvpx-44600442dca48a5586105b01403ac03fad54d05b.tar.gz libvpx-44600442dca48a5586105b01403ac03fad54d05b.tar.bz2 libvpx-44600442dca48a5586105b01403ac03fad54d05b.zip | |
bitdepth conversion: really use num elements
The previous implementation confused bit/bytes/elements. It was using
'32' as the multiplier but that was mistakenly adopted because a 32x32
transform embedded the stride.
Change-Id: Ieeb867a332416b9a40580b5e7c9b20088e9e691a
Diffstat (limited to 'vp9/encoder/x86')
| -rw-r--r-- | vp9/encoder/x86/vp9_dct_sse2.asm | 2 | ||||
| -rw-r--r-- | vp9/encoder/x86/vp9_error_sse2.asm | 4 |
2 files changed, 3 insertions, 3 deletions
diff --git a/vp9/encoder/x86/vp9_dct_sse2.asm b/vp9/encoder/x86/vp9_dct_sse2.asm index e24cabba0..8152dce86 100644 --- a/vp9/encoder/x86/vp9_dct_sse2.asm +++ b/vp9/encoder/x86/vp9_dct_sse2.asm @@ -64,6 +64,6 @@ cglobal fwht4x4, 3, 4, 8, input, output, stride psllw m1, 2 STORE_TRAN_LOW 0, outputq, 0, 2, 3 - STORE_TRAN_LOW 1, outputq, 1, 2, 3 + STORE_TRAN_LOW 1, outputq, 8, 2, 3 RET diff --git a/vp9/encoder/x86/vp9_error_sse2.asm b/vp9/encoder/x86/vp9_error_sse2.asm index dcedf913d..5186d3087 100644 --- a/vp9/encoder/x86/vp9_error_sse2.asm +++ b/vp9/encoder/x86/vp9_error_sse2.asm @@ -91,8 +91,8 @@ cglobal block_error_fp, 3, 3, 6, uqc, dqc, size .loop: LOAD_TRAN_LOW 2, uqcq, 0 LOAD_TRAN_LOW 0, dqcq, 0 - LOAD_TRAN_LOW 3, uqcq, 1 - LOAD_TRAN_LOW 1, dqcq, 1 + LOAD_TRAN_LOW 3, uqcq, 8 + LOAD_TRAN_LOW 1, dqcq, 8 INCREMENT_ELEMENTS_TRAN_LOW uqcq, 16 INCREMENT_ELEMENTS_TRAN_LOW dqcq, 16 sub sizeq, 16 |