dany/libwebp
1
0
Fork 0
mirror of https://github.com/webmproject/libwebp.git synced 2024-12-25 05:08:21 +01:00
Code Issues Actions Packages Projects Releases Wiki Activity

rework AddVector_SSE2

Browse Source 
Take advantage of the known sizes used by VP8LHistogramAdd() and remove
loop for the remainder. The loop was being auto-vectorized making the
code larger and slower than the vectorized C code.

For larger sizes the new code is ~4-7% faster than the old code with
about the same improvement against the vectorized C code. For the
minimal size (40), the new code is ~30% faster than the C and old SSE2
code.

The LINE_SIZE==8 option is removed with this change. It had been set to
16 for its entire life and clang-16 was unrolling the LINE_SIZE==8 case
by 2 in any case; they both profile similarly.

Change-Id: I2376e2dca3bffa38477b4a432f4c533419e3be0e
This commit is contained in:
James Zern 2024-11-12 14:58:23 -08:00
parent 2ddaaf0aa5
commit 7bda3deb89
1 changed files with 31 additions and 11 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

42
src/dsp/lossless_enc_sse2.c
View File

@ -175,36 +175,56 @@ static void CollectColorRedTransforms_SSE2(const uint32_t* WEBP_RESTRICT argb,
// Note we are adding uint32_t's as *signed* int32's (using _mm_add_epi32). But
// that's ok since the histogram values are less than 1<<28 (max picture size).
#define LINE_SIZE 16 // 8 or 16
static void AddVector_SSE2(const uint32_t* WEBP_RESTRICT a,
const uint32_t* WEBP_RESTRICT b,
uint32_t* WEBP_RESTRICT out, int size) {
int i;
for (i = 0; i + LINE_SIZE <= size; i += LINE_SIZE) {
int i = 0;
int aligned_size = size & ~15;
// Size is, at minimum, NUM_DISTANCE_CODES (40) and may be as large as
// NUM_LITERAL_CODES (256) + NUM_LENGTH_CODES (24) + (0 or a non-zero power of
// 2). See the usage in VP8LHistogramAdd().
assert(size >= 16);
assert(size % 2 == 0);
do {
const __m128i a0 = _mm_loadu_si128((const __m128i*)&a[i + 0]);
const __m128i a1 = _mm_loadu_si128((const __m128i*)&a[i + 4]);
#if (LINE_SIZE == 16)
const __m128i a2 = _mm_loadu_si128((const __m128i*)&a[i + 8]);
const __m128i a3 = _mm_loadu_si128((const __m128i*)&a[i + 12]);
#endif
const __m128i b0 = _mm_loadu_si128((const __m128i*)&b[i + 0]);
const __m128i b1 = _mm_loadu_si128((const __m128i*)&b[i + 4]);
#if (LINE_SIZE == 16)
const __m128i b2 = _mm_loadu_si128((const __m128i*)&b[i + 8]);
const __m128i b3 = _mm_loadu_si128((const __m128i*)&b[i + 12]);
#endif
_mm_storeu_si128((__m128i*)&out[i + 0], _mm_add_epi32(a0, b0));
_mm_storeu_si128((__m128i*)&out[i + 4], _mm_add_epi32(a1, b1));
#if (LINE_SIZE == 16)
_mm_storeu_si128((__m128i*)&out[i + 8], _mm_add_epi32(a2, b2));
_mm_storeu_si128((__m128i*)&out[i + 12], _mm_add_epi32(a3, b3));
#endif
i += 16;
} while (i != aligned_size);
if ((size & 8) != 0) {
const __m128i a0 = _mm_loadu_si128((const __m128i*)&a[i + 0]);
const __m128i a1 = _mm_loadu_si128((const __m128i*)&a[i + 4]);
const __m128i b0 = _mm_loadu_si128((const __m128i*)&b[i + 0]);
const __m128i b1 = _mm_loadu_si128((const __m128i*)&b[i + 4]);
_mm_storeu_si128((__m128i*)&out[i + 0], _mm_add_epi32(a0, b0));
_mm_storeu_si128((__m128i*)&out[i + 4], _mm_add_epi32(a1, b1));
i += 8;
}
for (; i < size; ++i) {
out[i] = a[i] + b[i];
size &= 7;
if (size == 4) {
const __m128i a0 = _mm_loadu_si128((const __m128i*)&a[i]);
const __m128i b0 = _mm_loadu_si128((const __m128i*)&b[i]);
_mm_storeu_si128((__m128i*)&out[i], _mm_add_epi32(a0, b0));
} else if (size == 2) {
const __m128i a0 = _mm_loadl_epi64((const __m128i*)&a[i]);
const __m128i b0 = _mm_loadl_epi64((const __m128i*)&b[i]);
_mm_storel_epi64((__m128i*)&out[i], _mm_add_epi32(a0, b0));
}
}
#define LINE_SIZE 16 // 8 or 16
static void AddVectorEq_SSE2(const uint32_t* WEBP_RESTRICT a,
uint32_t* WEBP_RESTRICT out, int size) {
int i;