Skip to content
GitLab
Commit 82d504b5 authored by Jingning Han's avatar Jingning Han
Browse files

Use aligned buffer operations in 8x8/16x16 2D-DCT 

This reduces 16x16 2D-DCT runtime from 865 cycles to 837 cycles.

Change-Id: I137758b81cd127b936175284310e81378db64552
parent a32a086d
v1.14.0-linphone 1.4.X experimental feature/update_to_v1.9.0-linphone feature/uwp_nuget forest frame_parallel highbitdepth indianrunnerduck javanwhistlingduck khakicampbell linphone linphone-android linphone-old longtailedduck m31-baseline m49-2623 m52-2743 m54-2840 m56-2924 m66-3359 m68-3440 mandarinduck mcw mcw2 nextgen nextgenv2 pcs-2013 playground sandbox/Jingning/experimental sandbox/Jingning/transcode sandbox/Jingning/vpx sandbox/aconverse@google.com/ansbench sandbox/debargha/playground sandbox/hkuang/frame_parallel sandbox/hkuang@google.com/decode sandbox/jimbankoski@google.com/proposed-aom sandbox/jingning@google.com/decoder_test_suite sandbox/jingning@google.com/experimental sandbox/jzern@google.com/test sandbox/wangch@google.com/vp9 sandbox/yaowu@google.com/mergeaom stable-vp9-decoder v1.12.0-linphone v1.6.1_linphone v1.7.0-linphone v1.9.0-linphone v1.9.0 v1.9.0-rc1 v1.8.2 v1.8.1 v1.8.0 v1.7.0 v1.6.1 v1.6.0 v1.5.0 v1.4.0 v1.3.0
No related merge requests found
Hide whitespace changes
Inline Side-by-side
Showing
with 57 additions and 57 deletions
vp9/encoder/x86/vp9_dct_sse2.c
+ 57
− 57
View file @ 82d504b5
...@@ -134,14 +134,14 @@ void vp9_short_fdct8x8_sse2(int16_t *input, int16_t *output, int pitch) { ...@@ -134,14 +134,14 @@ void vp9_short_fdct8x8_sse2(int16_t *input, int16_t *output, int pitch) {
const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64); const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64);
const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING); const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
// Load input // Load input
__m128i in0 = _mm_loadu_si128((const __m128i *)(input + 0 * stride)); __m128i in0 = _mm_load_si128((const __m128i *)(input + 0 * stride));
__m128i in1 = _mm_loadu_si128((const __m128i *)(input + 1 * stride)); __m128i in1 = _mm_load_si128((const __m128i *)(input + 1 * stride));
__m128i in2 = _mm_loadu_si128((const __m128i *)(input + 2 * stride)); __m128i in2 = _mm_load_si128((const __m128i *)(input + 2 * stride));
__m128i in3 = _mm_loadu_si128((const __m128i *)(input + 3 * stride)); __m128i in3 = _mm_load_si128((const __m128i *)(input + 3 * stride));
__m128i in4 = _mm_loadu_si128((const __m128i *)(input + 4 * stride)); __m128i in4 = _mm_load_si128((const __m128i *)(input + 4 * stride));
__m128i in5 = _mm_loadu_si128((const __m128i *)(input + 5 * stride)); __m128i in5 = _mm_load_si128((const __m128i *)(input + 5 * stride));
__m128i in6 = _mm_loadu_si128((const __m128i *)(input + 6 * stride)); __m128i in6 = _mm_load_si128((const __m128i *)(input + 6 * stride));
__m128i in7 = _mm_loadu_si128((const __m128i *)(input + 7 * stride)); __m128i in7 = _mm_load_si128((const __m128i *)(input + 7 * stride));
// Pre-condition input (shift by two) // Pre-condition input (shift by two)
in0 = _mm_slli_epi16(in0, 2); in0 = _mm_slli_epi16(in0, 2);
in1 = _mm_slli_epi16(in1, 2); in1 = _mm_slli_epi16(in1, 2);
...@@ -363,14 +363,14 @@ void vp9_short_fdct8x8_sse2(int16_t *input, int16_t *output, int pitch) { ...@@ -363,14 +363,14 @@ void vp9_short_fdct8x8_sse2(int16_t *input, int16_t *output, int pitch) {
in6 = _mm_srai_epi16(in6, 1); in6 = _mm_srai_epi16(in6, 1);
in7 = _mm_srai_epi16(in7, 1); in7 = _mm_srai_epi16(in7, 1);
// store results // store results
_mm_storeu_si128((__m128i *)(output + 0 * 8), in0); _mm_store_si128((__m128i *)(output + 0 * 8), in0);
_mm_storeu_si128((__m128i *)(output + 1 * 8), in1); _mm_store_si128((__m128i *)(output + 1 * 8), in1);
_mm_storeu_si128((__m128i *)(output + 2 * 8), in2); _mm_store_si128((__m128i *)(output + 2 * 8), in2);
_mm_storeu_si128((__m128i *)(output + 3 * 8), in3); _mm_store_si128((__m128i *)(output + 3 * 8), in3);
_mm_storeu_si128((__m128i *)(output + 4 * 8), in4); _mm_store_si128((__m128i *)(output + 4 * 8), in4);
_mm_storeu_si128((__m128i *)(output + 5 * 8), in5); _mm_store_si128((__m128i *)(output + 5 * 8), in5);
_mm_storeu_si128((__m128i *)(output + 6 * 8), in6); _mm_store_si128((__m128i *)(output + 6 * 8), in6);
_mm_storeu_si128((__m128i *)(output + 7 * 8), in7); _mm_store_si128((__m128i *)(output + 7 * 8), in7);
} }
} }
...@@ -876,7 +876,7 @@ void vp9_short_fdct16x16_sse2(int16_t *input, int16_t *output, int pitch) { ...@@ -876,7 +876,7 @@ void vp9_short_fdct16x16_sse2(int16_t *input, int16_t *output, int pitch) {
const int stride = pitch >> 1; const int stride = pitch >> 1;
int pass; int pass;
// We need an intermediate buffer between passes. // We need an intermediate buffer between passes.
int16_t intermediate[256]; DECLARE_ALIGNED_ARRAY(16, int16_t, intermediate, 256);
int16_t *in = input; int16_t *in = input;
int16_t *out = intermediate; int16_t *out = intermediate;
// Constants // Constants
...@@ -919,22 +919,22 @@ void vp9_short_fdct16x16_sse2(int16_t *input, int16_t *output, int pitch) { ...@@ -919,22 +919,22 @@ void vp9_short_fdct16x16_sse2(int16_t *input, int16_t *output, int pitch) {
__m128i res08, res09, res10, res11, res12, res13, res14, res15; __m128i res08, res09, res10, res11, res12, res13, res14, res15;
// Load and pre-condition input. // Load and pre-condition input.
if (0 == pass) { if (0 == pass) {
in00 = _mm_loadu_si128((const __m128i *)(in + 0 * stride)); in00 = _mm_load_si128((const __m128i *)(in + 0 * stride));
in01 = _mm_loadu_si128((const __m128i *)(in + 1 * stride)); in01 = _mm_load_si128((const __m128i *)(in + 1 * stride));
in02 = _mm_loadu_si128((const __m128i *)(in + 2 * stride)); in02 = _mm_load_si128((const __m128i *)(in + 2 * stride));
in03 = _mm_loadu_si128((const __m128i *)(in + 3 * stride)); in03 = _mm_load_si128((const __m128i *)(in + 3 * stride));
in04 = _mm_loadu_si128((const __m128i *)(in + 4 * stride)); in04 = _mm_load_si128((const __m128i *)(in + 4 * stride));
in05 = _mm_loadu_si128((const __m128i *)(in + 5 * stride)); in05 = _mm_load_si128((const __m128i *)(in + 5 * stride));
in06 = _mm_loadu_si128((const __m128i *)(in + 6 * stride)); in06 = _mm_load_si128((const __m128i *)(in + 6 * stride));
in07 = _mm_loadu_si128((const __m128i *)(in + 7 * stride)); in07 = _mm_load_si128((const __m128i *)(in + 7 * stride));
in08 = _mm_loadu_si128((const __m128i *)(in + 8 * stride)); in08 = _mm_load_si128((const __m128i *)(in + 8 * stride));
in09 = _mm_loadu_si128((const __m128i *)(in + 9 * stride)); in09 = _mm_load_si128((const __m128i *)(in + 9 * stride));
in10 = _mm_loadu_si128((const __m128i *)(in + 10 * stride)); in10 = _mm_load_si128((const __m128i *)(in + 10 * stride));
in11 = _mm_loadu_si128((const __m128i *)(in + 11 * stride)); in11 = _mm_load_si128((const __m128i *)(in + 11 * stride));
in12 = _mm_loadu_si128((const __m128i *)(in + 12 * stride)); in12 = _mm_load_si128((const __m128i *)(in + 12 * stride));
in13 = _mm_loadu_si128((const __m128i *)(in + 13 * stride)); in13 = _mm_load_si128((const __m128i *)(in + 13 * stride));
in14 = _mm_loadu_si128((const __m128i *)(in + 14 * stride)); in14 = _mm_load_si128((const __m128i *)(in + 14 * stride));
in15 = _mm_loadu_si128((const __m128i *)(in + 15 * stride)); in15 = _mm_load_si128((const __m128i *)(in + 15 * stride));
// x = x << 2 // x = x << 2
in00 = _mm_slli_epi16(in00, 2); in00 = _mm_slli_epi16(in00, 2);
in01 = _mm_slli_epi16(in01, 2); in01 = _mm_slli_epi16(in01, 2);
...@@ -953,22 +953,22 @@ void vp9_short_fdct16x16_sse2(int16_t *input, int16_t *output, int pitch) { ...@@ -953,22 +953,22 @@ void vp9_short_fdct16x16_sse2(int16_t *input, int16_t *output, int pitch) {
in14 = _mm_slli_epi16(in14, 2); in14 = _mm_slli_epi16(in14, 2);
in15 = _mm_slli_epi16(in15, 2); in15 = _mm_slli_epi16(in15, 2);
} else { } else {
in00 = _mm_loadu_si128((const __m128i *)(in + 0 * 16)); in00 = _mm_load_si128((const __m128i *)(in + 0 * 16));
in01 = _mm_loadu_si128((const __m128i *)(in + 1 * 16)); in01 = _mm_load_si128((const __m128i *)(in + 1 * 16));
in02 = _mm_loadu_si128((const __m128i *)(in + 2 * 16)); in02 = _mm_load_si128((const __m128i *)(in + 2 * 16));
in03 = _mm_loadu_si128((const __m128i *)(in + 3 * 16)); in03 = _mm_load_si128((const __m128i *)(in + 3 * 16));
in04 = _mm_loadu_si128((const __m128i *)(in + 4 * 16)); in04 = _mm_load_si128((const __m128i *)(in + 4 * 16));
in05 = _mm_loadu_si128((const __m128i *)(in + 5 * 16)); in05 = _mm_load_si128((const __m128i *)(in + 5 * 16));
in06 = _mm_loadu_si128((const __m128i *)(in + 6 * 16)); in06 = _mm_load_si128((const __m128i *)(in + 6 * 16));
in07 = _mm_loadu_si128((const __m128i *)(in + 7 * 16)); in07 = _mm_load_si128((const __m128i *)(in + 7 * 16));
in08 = _mm_loadu_si128((const __m128i *)(in + 8 * 16)); in08 = _mm_load_si128((const __m128i *)(in + 8 * 16));
in09 = _mm_loadu_si128((const __m128i *)(in + 9 * 16)); in09 = _mm_load_si128((const __m128i *)(in + 9 * 16));
in10 = _mm_loadu_si128((const __m128i *)(in + 10 * 16)); in10 = _mm_load_si128((const __m128i *)(in + 10 * 16));
in11 = _mm_loadu_si128((const __m128i *)(in + 11 * 16)); in11 = _mm_load_si128((const __m128i *)(in + 11 * 16));
in12 = _mm_loadu_si128((const __m128i *)(in + 12 * 16)); in12 = _mm_load_si128((const __m128i *)(in + 12 * 16));
in13 = _mm_loadu_si128((const __m128i *)(in + 13 * 16)); in13 = _mm_load_si128((const __m128i *)(in + 13 * 16));
in14 = _mm_loadu_si128((const __m128i *)(in + 14 * 16)); in14 = _mm_load_si128((const __m128i *)(in + 14 * 16));
in15 = _mm_loadu_si128((const __m128i *)(in + 15 * 16)); in15 = _mm_load_si128((const __m128i *)(in + 15 * 16));
// x = (x + 1) >> 2 // x = (x + 1) >> 2
in00 = _mm_add_epi16(in00, kOne); in00 = _mm_add_epi16(in00, kOne);
in01 = _mm_add_epi16(in01, kOne); in01 = _mm_add_epi16(in01, kOne);
...@@ -1475,14 +1475,14 @@ void vp9_short_fdct16x16_sse2(int16_t *input, int16_t *output, int pitch) { ...@@ -1475,14 +1475,14 @@ void vp9_short_fdct16x16_sse2(int16_t *input, int16_t *output, int pitch) {
// 06 16 26 36 46 56 66 76 // 06 16 26 36 46 56 66 76
// 07 17 27 37 47 57 67 77 // 07 17 27 37 47 57 67 77
// Store results // Store results
_mm_storeu_si128((__m128i *)(out + 8 + 0 * 16), tr2_0); _mm_store_si128((__m128i *)(out + 8 + 0 * 16), tr2_0);
_mm_storeu_si128((__m128i *)(out + 8 + 1 * 16), tr2_1); _mm_store_si128((__m128i *)(out + 8 + 1 * 16), tr2_1);
_mm_storeu_si128((__m128i *)(out + 8 + 2 * 16), tr2_2); _mm_store_si128((__m128i *)(out + 8 + 2 * 16), tr2_2);
_mm_storeu_si128((__m128i *)(out + 8 + 3 * 16), tr2_3); _mm_store_si128((__m128i *)(out + 8 + 3 * 16), tr2_3);
_mm_storeu_si128((__m128i *)(out + 8 + 4 * 16), tr2_4); _mm_store_si128((__m128i *)(out + 8 + 4 * 16), tr2_4);
_mm_storeu_si128((__m128i *)(out + 8 + 5 * 16), tr2_5); _mm_store_si128((__m128i *)(out + 8 + 5 * 16), tr2_5);
_mm_storeu_si128((__m128i *)(out + 8 + 6 * 16), tr2_6); _mm_store_si128((__m128i *)(out + 8 + 6 * 16), tr2_6);
_mm_storeu_si128((__m128i *)(out + 8 + 7 * 16), tr2_7); _mm_store_si128((__m128i *)(out + 8 + 7 * 16), tr2_7);
} }
out += 8*16; out += 8*16;
} }
......
Supports Markdown
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