Skip to content

GitLab

Commit 6c4af6b7 authored by David Barker's avatar David Barker Committed by Debargha Mukherjee
Browse files
Options

Fix some irregularities in SSE2 variance code 

Change-Id: I1dcf3bd33645aed3347301149808c157eeb44cad
parent 85a8f70c
Hide whitespace changes
Inline Side-by-side
Showing with 61 additions and 61 deletions
aom_dsp/x86/highbd_variance_sse2.c
View file @ 6c4af6b7
......@@ -416,9 +416,9 @@ FNS(sse2);
const uint16_t *dst, ptrdiff_t dst_stride, const uint16_t *sec, \
ptrdiff_t sec_stride, int height, unsigned int *sse, void *unused0, \
void *unused);
#define DECLS(opt1) \
DECL(16, opt1) \
DECL(8, opt1)
#define DECLS(opt) \
DECL(16, opt) \
DECL(8, opt)
DECLS(sse2);
#undef DECL
......@@ -550,18 +550,18 @@ DECLS(sse2);
return (var >= 0) ? (uint32_t)var : 0; \
}
#define FNS(opt1) \
FN(64, 64, 16, 6, 6, opt1, (int64_t)); \
FN(64, 32, 16, 6, 5, opt1, (int64_t)); \
FN(32, 64, 16, 5, 6, opt1, (int64_t)); \
FN(32, 32, 16, 5, 5, opt1, (int64_t)); \
FN(32, 16, 16, 5, 4, opt1, (int64_t)); \
FN(16, 32, 16, 4, 5, opt1, (int64_t)); \
FN(16, 16, 16, 4, 4, opt1, (int64_t)); \
FN(16, 8, 16, 4, 3, opt1, (int64_t)); \
FN(8, 16, 8, 4, 3, opt1, (int64_t)); \
FN(8, 8, 8, 3, 3, opt1, (int64_t)); \
FN(8, 4, 8, 3, 2, opt1, (int64_t));
#define FNS(opt) \
FN(64, 64, 16, 6, 6, opt, (int64_t)); \
FN(64, 32, 16, 6, 5, opt, (int64_t)); \
FN(32, 64, 16, 5, 6, opt, (int64_t)); \
FN(32, 32, 16, 5, 5, opt, (int64_t)); \
FN(32, 16, 16, 5, 4, opt, (int64_t)); \
FN(16, 32, 16, 4, 5, opt, (int64_t)); \
FN(16, 16, 16, 4, 4, opt, (int64_t)); \
FN(16, 8, 16, 4, 3, opt, (int64_t)); \
FN(8, 16, 8, 3, 4, opt, (int64_t)); \
FN(8, 8, 8, 3, 3, opt, (int64_t)); \
FN(8, 4, 8, 3, 2, opt, (int64_t));
FNS(sse2);
......
aom_dsp/x86/variance_sse2.c
View file @ 6c4af6b7
......@@ -345,13 +345,13 @@ unsigned int aom_mse16x16_sse2(const uint8_t *src, int src_stride,
const uint8_t *src, ptrdiff_t src_stride, int x_offset, int y_offset, \
const uint8_t *dst, ptrdiff_t dst_stride, int height, unsigned int *sse, \
void *unused0, void *unused)
#define DECLS(opt1, opt2) \
DECL(4, opt1); \
DECL(8, opt1); \
DECL(16, opt1)
#define DECLS(opt) \
DECL(4, opt); \
DECL(8, opt); \
DECL(16, opt)
DECLS(sse2, sse2);
DECLS(ssse3, ssse3);
DECLS(sse2);
DECLS(ssse3);
#undef DECLS
#undef DECL
......@@ -387,23 +387,23 @@ DECLS(ssse3, ssse3);
return sse - (unsigned int)(cast_prod(cast se * se) >> (wlog2 + hlog2)); \
}
#define FNS(opt1, opt2) \
FN(64, 64, 16, 6, 6, opt1, (int64_t), (int64_t)); \
FN(64, 32, 16, 6, 5, opt1, (int64_t), (int64_t)); \
FN(32, 64, 16, 5, 6, opt1, (int64_t), (int64_t)); \
FN(32, 32, 16, 5, 5, opt1, (int64_t), (int64_t)); \
FN(32, 16, 16, 5, 4, opt1, (int64_t), (int64_t)); \
FN(16, 32, 16, 4, 5, opt1, (int64_t), (int64_t)); \
FN(16, 16, 16, 4, 4, opt1, (uint32_t), (int64_t)); \
FN(16, 8, 16, 4, 3, opt1, (int32_t), (int32_t)); \
FN(8, 16, 8, 3, 4, opt1, (int32_t), (int32_t)); \
FN(8, 8, 8, 3, 3, opt1, (int32_t), (int32_t)); \
FN(8, 4, 8, 3, 2, opt1, (int32_t), (int32_t)); \
FN(4, 8, 4, 2, 3, opt1, (int32_t), (int32_t)); \
FN(4, 4, 4, 2, 2, opt1, (int32_t), (int32_t))
FNS(sse2, sse2);
FNS(ssse3, ssse3);
#define FNS(opt) \
FN(64, 64, 16, 6, 6, opt, (int64_t), (int64_t)); \
FN(64, 32, 16, 6, 5, opt, (int64_t), (int64_t)); \
FN(32, 64, 16, 5, 6, opt, (int64_t), (int64_t)); \
FN(32, 32, 16, 5, 5, opt, (int64_t), (int64_t)); \
FN(32, 16, 16, 5, 4, opt, (int64_t), (int64_t)); \
FN(16, 32, 16, 4, 5, opt, (int64_t), (int64_t)); \
FN(16, 16, 16, 4, 4, opt, (uint32_t), (int64_t)); \
FN(16, 8, 16, 4, 3, opt, (int32_t), (int32_t)); \
FN(8, 16, 8, 3, 4, opt, (int32_t), (int32_t)); \
FN(8, 8, 8, 3, 3, opt, (int32_t), (int32_t)); \
FN(8, 4, 8, 3, 2, opt, (int32_t), (int32_t)); \
FN(4, 8, 4, 2, 3, opt, (int32_t), (int32_t)); \
FN(4, 4, 4, 2, 2, opt, (int32_t), (int32_t))
FNS(sse2);
FNS(ssse3);
#undef FNS
#undef FN
......@@ -415,13 +415,13 @@ FNS(ssse3, ssse3);
const uint8_t *dst, ptrdiff_t dst_stride, const uint8_t *sec, \
ptrdiff_t sec_stride, int height, unsigned int *sse, void *unused0, \
void *unused)
#define DECLS(opt1, opt2) \
DECL(4, opt1); \
DECL(8, opt1); \
DECL(16, opt1)
#define DECLS(opt) \
DECL(4, opt); \
DECL(8, opt); \
DECL(16, opt)
DECLS(sse2, sse2);
DECLS(ssse3, ssse3);
DECLS(sse2);
DECLS(ssse3);
#undef DECL
#undef DECLS
......@@ -458,23 +458,23 @@ DECLS(ssse3, ssse3);
return sse - (unsigned int)(cast_prod(cast se * se) >> (wlog2 + hlog2)); \
}
#define FNS(opt1, opt2) \
FN(64, 64, 16, 6, 6, opt1, (int64_t), (int64_t)); \
FN(64, 32, 16, 6, 5, opt1, (int64_t), (int64_t)); \
FN(32, 64, 16, 5, 6, opt1, (int64_t), (int64_t)); \
FN(32, 32, 16, 5, 5, opt1, (int64_t), (int64_t)); \
FN(32, 16, 16, 5, 4, opt1, (int64_t), (int64_t)); \
FN(16, 32, 16, 4, 5, opt1, (int64_t), (int64_t)); \
FN(16, 16, 16, 4, 4, opt1, (uint32_t), (int64_t)); \
FN(16, 8, 16, 4, 3, opt1, (uint32_t), (int32_t)); \
FN(8, 16, 8, 3, 4, opt1, (uint32_t), (int32_t)); \
FN(8, 8, 8, 3, 3, opt1, (uint32_t), (int32_t)); \
FN(8, 4, 8, 3, 2, opt1, (uint32_t), (int32_t)); \
FN(4, 8, 4, 2, 3, opt1, (uint32_t), (int32_t)); \
FN(4, 4, 4, 2, 2, opt1, (uint32_t), (int32_t))
FNS(sse2, sse);
FNS(ssse3, ssse3);
#define FNS(opt) \
FN(64, 64, 16, 6, 6, opt, (int64_t), (int64_t)); \
FN(64, 32, 16, 6, 5, opt, (int64_t), (int64_t)); \
FN(32, 64, 16, 5, 6, opt, (int64_t), (int64_t)); \
FN(32, 32, 16, 5, 5, opt, (int64_t), (int64_t)); \
FN(32, 16, 16, 5, 4, opt, (int64_t), (int64_t)); \
FN(16, 32, 16, 4, 5, opt, (int64_t), (int64_t)); \
FN(16, 16, 16, 4, 4, opt, (uint32_t), (int64_t)); \
FN(16, 8, 16, 4, 3, opt, (uint32_t), (int32_t)); \
FN(8, 16, 8, 3, 4, opt, (uint32_t), (int32_t)); \
FN(8, 8, 8, 3, 3, opt, (uint32_t), (int32_t)); \
FN(8, 4, 8, 3, 2, opt, (uint32_t), (int32_t)); \
FN(4, 8, 4, 2, 3, opt, (uint32_t), (int32_t)); \
FN(4, 4, 4, 2, 2, opt, (uint32_t), (int32_t))
FNS(sse2);
FNS(ssse3);
#undef FNS
#undef FN
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment