diff options
author | Roland Scheidegger <sroland@vmware.com> | 2016-01-02 04:59:09 +0100 |
---|---|---|
committer | Roland Scheidegger <sroland@vmware.com> | 2016-01-08 00:34:14 +0100 |
commit | 2923c7a0ed92a29da029183356e81ad55e615cf7 (patch) | |
tree | bd4f83e3bd33c2109472a7c4fc76dcdc6638cc01 /src/gallium/auxiliary/util/u_sse.h | |
parent | fad283ba9e691d0d5d170f388e75542f2c39e559 (diff) | |
download | external_mesa3d-2923c7a0ed92a29da029183356e81ad55e615cf7.zip external_mesa3d-2923c7a0ed92a29da029183356e81ad55e615cf7.tar.gz external_mesa3d-2923c7a0ed92a29da029183356e81ad55e615cf7.tar.bz2 |
llvmpipe: do 64bit plane calculations in the sse path
The sse path was pretty much disabled for practical purposes because the
largest allowed fb size was 128x128. So, adapt it for 64bit plane calculations.
This is actually not that difficult, though a problem is that we can't do
a signed 32x32->64bit mul, only unsigned, so need to fix that up. Overall,
the code still looks reasonable, though it's not like changes there in
setup really make much of a difference in the end...
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Diffstat (limited to 'src/gallium/auxiliary/util/u_sse.h')
-rw-r--r-- | src/gallium/auxiliary/util/u_sse.h | 92 |
1 files changed, 80 insertions, 12 deletions
diff --git a/src/gallium/auxiliary/util/u_sse.h b/src/gallium/auxiliary/util/u_sse.h index 7f8e5a1..cae4138 100644 --- a/src/gallium/auxiliary/util/u_sse.h +++ b/src/gallium/auxiliary/util/u_sse.h @@ -166,14 +166,49 @@ _mm_shuffle_epi8(__m128i a, __m128i mask) #endif /* !PIPE_ARCH_SSSE3 */ +/* + * Provide an SSE implementation of _mm_mul_epi32() in terms of + * _mm_mul_epu32(). + * + * Basically, albeit surprising at first (and second, and third...) look + * if a * b is done signed instead of unsigned, can just + * subtract b from the high bits of the result if a is negative + * (and the same for a if b is negative). Modular arithmetic at its best! + * + * So for int32 a,b in crude pseudo-code ("*" here denoting a widening mul) + * fixupb = (signmask(b) & a) << 32ULL + * fixupa = (signmask(a) & b) << 32ULL + * a * b = (unsigned)a * (unsigned)b - fixupb - fixupa + * = (unsigned)a * (unsigned)b -(fixupb + fixupa) + * + * This does both lo (dwords 0/2) and hi parts (1/3) at the same time due + * to some optimization potential. + */ +static inline __m128i +mm_mullohi_epi32(const __m128i a, const __m128i b, __m128i *res13) +{ + __m128i a13, b13, mul02, mul13; + __m128i anegmask, bnegmask, fixup, fixup02, fixup13; + a13 = _mm_shuffle_epi32(a, _MM_SHUFFLE(2,3,0,1)); + b13 = _mm_shuffle_epi32(b, _MM_SHUFFLE(2,3,0,1)); + anegmask = _mm_srai_epi32(a, 31); + bnegmask = _mm_srai_epi32(b, 31); + fixup = _mm_add_epi32(_mm_and_si128(anegmask, b), + _mm_and_si128(bnegmask, a)); + mul02 = _mm_mul_epu32(a, b); + mul13 = _mm_mul_epu32(a13, b13); + fixup02 = _mm_slli_epi64(fixup, 32); + fixup13 = _mm_and_si128(fixup, _mm_set_epi32(-1,0,-1,0)); + *res13 = _mm_sub_epi64(mul13, fixup13); + return _mm_sub_epi64(mul02, fixup02); +} /* Provide an SSE2 implementation of _mm_mullo_epi32() in terms of * _mm_mul_epu32(). * - * I suspect this works fine for us because one of our operands is - * always positive, but not sure that this can be used for general - * signed integer multiplication. + * This always works regardless the signs of the operands, since + * the high bits (which would be different) aren't used. * * This seems close enough to the speed of SSE4 and the real * _mm_mullo_epi32() intrinsic as to not justify adding an sse4 @@ -188,6 +223,12 @@ static inline __m128i mm_mullo_epi32(const __m128i a, const __m128i b) /* Interleave the results, either with shuffles or (slightly * faster) direct bit operations: + * XXX: might be only true for some cpus (in particular 65nm + * Core 2). On most cpus (including that Core 2, but not Nehalem...) + * using _mm_shuffle_ps/_mm_shuffle_epi32 might also be faster + * than using the 3 instructions below. But logic should be fine + * as well, we can't have optimal solution for all cpus (if anything, + * should just use _mm_mullo_epi32() if sse41 is available...). */ #if 0 __m128i ba8 = _mm_shuffle_epi32(ba, 8); @@ -214,17 +255,44 @@ transpose4_epi32(const __m128i * restrict a, __m128i * restrict q, __m128i * restrict r) { - __m128i t0 = _mm_unpacklo_epi32(*a, *b); - __m128i t1 = _mm_unpacklo_epi32(*c, *d); - __m128i t2 = _mm_unpackhi_epi32(*a, *b); - __m128i t3 = _mm_unpackhi_epi32(*c, *d); - - *o = _mm_unpacklo_epi64(t0, t1); - *p = _mm_unpackhi_epi64(t0, t1); - *q = _mm_unpacklo_epi64(t2, t3); - *r = _mm_unpackhi_epi64(t2, t3); + __m128i t0 = _mm_unpacklo_epi32(*a, *b); + __m128i t1 = _mm_unpacklo_epi32(*c, *d); + __m128i t2 = _mm_unpackhi_epi32(*a, *b); + __m128i t3 = _mm_unpackhi_epi32(*c, *d); + + *o = _mm_unpacklo_epi64(t0, t1); + *p = _mm_unpackhi_epi64(t0, t1); + *q = _mm_unpacklo_epi64(t2, t3); + *r = _mm_unpackhi_epi64(t2, t3); } + +/* + * Same as above, except the first two values are already interleaved + * (i.e. contain 64bit values). + */ +static inline void +transpose2_64_2_32(const __m128i * restrict a01, + const __m128i * restrict a23, + const __m128i * restrict c, + const __m128i * restrict d, + __m128i * restrict o, + __m128i * restrict p, + __m128i * restrict q, + __m128i * restrict r) +{ + __m128i t0 = *a01; + __m128i t1 = _mm_unpacklo_epi32(*c, *d); + __m128i t2 = *a23; + __m128i t3 = _mm_unpackhi_epi32(*c, *d); + + *o = _mm_unpacklo_epi64(t0, t1); + *p = _mm_unpackhi_epi64(t0, t1); + *q = _mm_unpacklo_epi64(t2, t3); + *r = _mm_unpackhi_epi64(t2, t3); +} + + #define SCALAR_EPI32(m, i) _mm_shuffle_epi32((m), _MM_SHUFFLE(i,i,i,i)) |