Tweak some adler32 hasher.up_x86_sse42 var names

This commit is contained in:
Nigel Tao committed 2021-04-28 00:32:32 +10:00
1 parent abd37fa3f8
commit 0a6cce5e0e
2 files changed
+27 -27

No files matched your search

+8 -8
View File
@@ -20886,8 +20886,8 @@ wuffs_adler32__hasher__up_x86_sse42(
__m128i v_ones = {0};
__m128i v_weights__left = {0};
__m128i v_weights_right = {0};
__m128i v_p__left = {0};
__m128i v_p_right = {0};
__m128i v_q__left = {0};
__m128i v_q_right = {0};
__m128i v_v1 = {0};
__m128i v_v2 = {0};
__m128i v_v2j = {0};
@@ -20918,13 +20918,13 @@ wuffs_adler32__hasher__up_x86_sse42(
v_p.len = 32;
uint8_t* i_end0_p = v_p.ptr + (((i_slice_p.len - (size_t)(v_p.ptr - i_slice_p.ptr)) / 32) * 32);
while (v_p.ptr < i_end0_p) {
v_p__left = _mm_lddqu_si128((const __m128i*)(const void*)(v_p.ptr));
v_p_right = _mm_lddqu_si128((const __m128i*)(const void*)(v_p.ptr + 16));
v_q__left = _mm_lddqu_si128((const __m128i*)(const void*)(v_p.ptr));
v_q_right = _mm_lddqu_si128((const __m128i*)(const void*)(v_p.ptr + 16));
v_v2j = _mm_add_epi32(v_v2j, v_v1);
v_v1 = _mm_add_epi32(v_v1, _mm_sad_epu8(v_p__left, v_zeroes));
v_v1 = _mm_add_epi32(v_v1, _mm_sad_epu8(v_p_right, v_zeroes));
v_v2k = _mm_add_epi32(v_v2k, _mm_madd_epi16(v_ones, _mm_maddubs_epi16(v_p__left, v_weights__left)));
v_v2k = _mm_add_epi32(v_v2k, _mm_madd_epi16(v_ones, _mm_maddubs_epi16(v_p_right, v_weights_right)));
v_v1 = _mm_add_epi32(v_v1, _mm_sad_epu8(v_q__left, v_zeroes));
v_v1 = _mm_add_epi32(v_v1, _mm_sad_epu8(v_q_right, v_zeroes));
v_v2k = _mm_add_epi32(v_v2k, _mm_madd_epi16(v_ones, _mm_maddubs_epi16(v_q__left, v_weights__left)));
v_v2k = _mm_add_epi32(v_v2k, _mm_madd_epi16(v_ones, _mm_maddubs_epi16(v_q_right, v_weights_right)));
v_p.ptr += 32;
}
v_p.len = 0;
+19 -19
View File
@@ -28,8 +28,8 @@ pri func hasher.up_x86_sse42!(x: slice base.u8),
var ones : base.x86_m128i
var weights__left : base.x86_m128i
var weights_right : base.x86_m128i
var p__left : base.x86_m128i
var p_right : base.x86_m128i
var q__left : base.x86_m128i
var q_right : base.x86_m128i
var v1 : base.x86_m128i
var v2 : base.x86_m128i
var v2j : base.x86_m128i
@@ -92,13 +92,13 @@ pri func hasher.up_x86_sse42!(x: slice base.u8),
// The inner loop.
iterate (p = args.x)(length: 32, advance: 32, unroll: 1) {
// Split the 32-byte p into left and right halves. SSE4.2 works
// with 16-byte registers.
// SSE4.2 works with 16-byte registers. Split the 32-byte p into
// left and right halves.
//
// Let p__left = [u8×16: p00, p01, p02, ..., p15]
// Let p_right = [u8×16: p16, p17, p18, ..., p31]
p__left = util.make_m128i_slice128(a: p[.. 16])
p_right = util.make_m128i_slice128(a: p[16 .. 32])
// Let q__left = [u8×16: p00, p01, p02, ..., p15]
// Let q_right = [u8×16: p16, p17, p18, ..., p31]
q__left = util.make_m128i_slice128(a: p[.. 16])
q_right = util.make_m128i_slice128(a: p[16 .. 32])
// For v2j, we need to calculate the sums of the s1j terms for each
// of p's 32 elements. This is simply 32 times the same number,
@@ -108,14 +108,14 @@ pri func hasher.up_x86_sse42!(x: slice base.u8),
// For v1, we need to add the elements of p. Computing the sum of
// absolute differences (_mm_sad_epu8) with zero just sums the
// elements. p__left._mm_sad_epu8(b: zeroes) equals
// elements. q__left._mm_sad_epu8(b: zeroes) equals
// [u64×2: p00 + p01 + ... + p07, p08 + p09 + ... + p15]
// This is equivalent (little-endian) to:
// [u32×4: p00 + p01 + ... + p07, 0, p08 + p09 + ... + p15, 0]
// We accumulate those "sum of p__left elements" in v1, and ditto
// for the p_right elements.
v1 = v1._mm_add_epi32(b: p__left._mm_sad_epu8(b: zeroes))
v1 = v1._mm_add_epi32(b: p_right._mm_sad_epu8(b: zeroes))
// We accumulate those "sum of q__left's elements" in v1, and ditto
// for q_right's elements.
v1 = v1._mm_add_epi32(b: q__left._mm_sad_epu8(b: zeroes))
v1 = v1._mm_add_epi32(b: q_right._mm_sad_epu8(b: zeroes))
// For v2k, we need to calculate a weighted sum: ((32 * p00) + (31
// * p01) + (30 * p02) + ... + (1 * p31)), which splits naturally
@@ -128,19 +128,19 @@ pri func hasher.up_x86_sse42!(x: slice base.u8),
// ...
// ((18*p14)+(17*p15))]
//
// The ones._mm_madd_epi16(b: etc) call is likewise a multiply-add
// (note that it's "madd" not "add"). Multiplying by 1 is a no-op,
// so this sums u16 pairs to produce u32 values:
// The ones._mm_madd_epi16(b: etc) call is a multiply-add (note
// that it's "madd" not "add"). Multiplying by 1 is a no-op, so
// this sums u16 pairs to produce u32 values:
// [u32×4: ((32*p00)+(31*p01)+(30*p02)+(29*p03)),
// ((28*p04)+(27*p05)+(26*p06)+(25*p07)),
// ...
// ((20*p12)+(19*p13)+(18*p14)+(17*p15))]
//
// Ditto again for the p_right elements.
// Ditto again for q_right's elements.
v2k = v2k._mm_add_epi32(b: ones._mm_madd_epi16(b:
p__left._mm_maddubs_epi16(b: weights__left)))
q__left._mm_maddubs_epi16(b: weights__left)))
v2k = v2k._mm_add_epi32(b: ones._mm_madd_epi16(b:
p_right._mm_maddubs_epi16(b: weights_right)))
q_right._mm_maddubs_epi16(b: weights_right)))
}
// Merge the four parallel u32 sums (v1) into the single u32 sum (s1).