ref: fa829e0e5aa3e43d14ecea5f327b6a2e27818cf8
parent: c9fb719ee15b3c591bdf67618d41779dc0731da3
author: Linfeng Zhang <linfengz@google.com>
date: Thu Aug 3 13:46:21 EDT 2017
Replace multiplication_and_add() with butterfly() in idct x86 code Change-Id: I266e45a3d75a5357c7d6e6f20ab5c6fdbfe4982e
--- a/vpx_dsp/x86/inv_txfm_sse2.c
+++ b/vpx_dsp/x86/inv_txfm_sse2.c
@@ -1001,20 +1001,20 @@
#define IDCT32_34 \
/* Stage1 */ \
- multiplication_and_add(in[1], zero, (int)cospi_31_64, (int)cospi_1_64, \
- &stp1_16, &stp1_31); \
- multiplication_and_add(zero, in[7], (int)cospi_7_64, (int)cospi_25_64, \
- &stp1_19, &stp1_28); \
- multiplication_and_add(in[5], zero, (int)cospi_27_64, (int)cospi_5_64, \
- &stp1_20, &stp1_27); \
- multiplication_and_add(zero, in[3], (int)cospi_3_64, (int)cospi_29_64, \
- &stp1_23, &stp1_24); \
+ butterfly(in[1], zero, (int)cospi_31_64, (int)cospi_1_64, &stp1_16, \
+ &stp1_31); \
+ butterfly(zero, in[7], (int)cospi_7_64, (int)cospi_25_64, &stp1_19, \
+ &stp1_28); \
+ butterfly(in[5], zero, (int)cospi_27_64, (int)cospi_5_64, &stp1_20, \
+ &stp1_27); \
+ butterfly(zero, in[3], (int)cospi_3_64, (int)cospi_29_64, &stp1_23, \
+ &stp1_24); \
\
/* Stage2 */ \
- multiplication_and_add(in[2], zero, (int)cospi_30_64, (int)cospi_2_64, \
- &stp2_8, &stp2_15); \
- multiplication_and_add(zero, in[6], (int)cospi_6_64, (int)cospi_26_64, \
- &stp2_11, &stp2_12); \
+ butterfly(in[2], zero, (int)cospi_30_64, (int)cospi_2_64, &stp2_8, \
+ &stp2_15); \
+ butterfly(zero, in[6], (int)cospi_6_64, (int)cospi_26_64, &stp2_11, \
+ &stp2_12); \
\
stp2_16 = stp1_16; \
stp2_19 = stp1_19; \
@@ -1029,8 +1029,7 @@
stp2_31 = stp1_31; \
\
/* Stage3 */ \
- multiplication_and_add(in[4], zero, (int)cospi_28_64, (int)cospi_4_64, \
- &stp1_4, &stp1_7); \
+ butterfly(in[4], zero, (int)cospi_28_64, (int)cospi_4_64, &stp1_4, &stp1_7); \
\
stp1_8 = stp2_8; \
stp1_11 = stp2_11; \
@@ -1037,15 +1036,15 @@
stp1_12 = stp2_12; \
stp1_15 = stp2_15; \
\
- multiplication_and_add(stp1_31, stp1_16, (int)cospi_28_64, (int)cospi_4_64, \
- &stp1_17, &stp1_30); \
- multiplication_and_add(stp1_19, stp1_28, -(int)cospi_4_64, \
- -(int)cospi_28_64, &stp1_29, &stp1_18); \
+ butterfly(stp1_31, stp1_16, (int)cospi_28_64, (int)cospi_4_64, &stp1_17, \
+ &stp1_30); \
+ butterfly(stp1_19, stp1_28, -(int)cospi_4_64, -(int)cospi_28_64, &stp1_29, \
+ &stp1_18); \
\
- multiplication_and_add(stp1_27, stp1_20, (int)cospi_12_64, (int)cospi_20_64, \
- &stp1_21, &stp1_26); \
- multiplication_and_add(stp1_23, stp1_24, -(int)cospi_20_64, \
- -(int)cospi_12_64, &stp1_25, &stp1_22); \
+ butterfly(stp1_27, stp1_20, (int)cospi_12_64, (int)cospi_20_64, &stp1_21, \
+ &stp1_26); \
+ butterfly(stp1_23, stp1_24, -(int)cospi_20_64, -(int)cospi_12_64, &stp1_25, \
+ &stp1_22); \
\
stp1_16 = stp2_16; \
stp1_31 = stp2_31; \
@@ -1057,8 +1056,8 @@
stp1_28 = stp2_28; \
\
/* Stage4 */ \
- multiplication_and_add(in[0], zero, (int)cospi_16_64, (int)cospi_16_64, \
- &stp2_1, &stp2_0); \
+ butterfly(in[0], zero, (int)cospi_16_64, (int)cospi_16_64, &stp2_1, \
+ &stp2_0); \
\
stp2_4 = stp1_4; \
stp2_5 = stp1_4; \
@@ -1065,10 +1064,10 @@
stp2_6 = stp1_7; \
stp2_7 = stp1_7; \
\
- multiplication_and_add(stp2_15, stp2_8, (int)cospi_24_64, (int)cospi_8_64, \
- &stp2_9, &stp2_14); \
- multiplication_and_add(stp2_11, stp2_12, -(int)cospi_8_64, \
- -(int)cospi_24_64, &stp2_13, &stp2_10); \
+ butterfly(stp2_15, stp2_8, (int)cospi_24_64, (int)cospi_8_64, &stp2_9, \
+ &stp2_14); \
+ butterfly(stp2_11, stp2_12, -(int)cospi_8_64, -(int)cospi_24_64, &stp2_13, \
+ &stp2_10); \
\
stp2_8 = stp1_8; \
stp2_15 = stp1_15; \
@@ -1098,8 +1097,8 @@
stp1_1 = stp2_1; \
stp1_2 = stp2_1; \
stp1_3 = stp2_0; \
- multiplication_and_add(stp2_6, stp2_5, (int)cospi_16_64, (int)cospi_16_64, \
- &stp1_5, &stp1_6); \
+ butterfly(stp2_6, stp2_5, (int)cospi_16_64, (int)cospi_16_64, &stp1_5, \
+ &stp1_6); \
\
stp1_4 = stp2_4; \
stp1_7 = stp2_7; \
@@ -1116,14 +1115,14 @@
stp1_16 = stp2_16; \
stp1_17 = stp2_17; \
\
- multiplication_and_add(stp2_29, stp2_18, (int)cospi_24_64, (int)cospi_8_64, \
- &stp1_18, &stp1_29); \
- multiplication_and_add(stp2_28, stp2_19, (int)cospi_24_64, (int)cospi_8_64, \
- &stp1_19, &stp1_28); \
- multiplication_and_add(stp2_20, stp2_27, -(int)cospi_8_64, \
- -(int)cospi_24_64, &stp1_27, &stp1_20); \
- multiplication_and_add(stp2_21, stp2_26, -(int)cospi_8_64, \
- -(int)cospi_24_64, &stp1_26, &stp1_21); \
+ butterfly(stp2_29, stp2_18, (int)cospi_24_64, (int)cospi_8_64, &stp1_18, \
+ &stp1_29); \
+ butterfly(stp2_28, stp2_19, (int)cospi_24_64, (int)cospi_8_64, &stp1_19, \
+ &stp1_28); \
+ butterfly(stp2_20, stp2_27, -(int)cospi_8_64, -(int)cospi_24_64, &stp1_27, \
+ &stp1_20); \
+ butterfly(stp2_21, stp2_26, -(int)cospi_8_64, -(int)cospi_24_64, &stp1_26, \
+ &stp1_21); \
\
stp1_22 = stp2_22; \
stp1_23 = stp2_23; \
@@ -1147,10 +1146,10 @@
stp2_14 = stp1_14; \
stp2_15 = stp1_15; \
\
- multiplication_and_add(stp1_13, stp1_10, (int)cospi_16_64, (int)cospi_16_64, \
- &stp2_10, &stp2_13); \
- multiplication_and_add(stp1_12, stp1_11, (int)cospi_16_64, (int)cospi_16_64, \
- &stp2_11, &stp2_12); \
+ butterfly(stp1_13, stp1_10, (int)cospi_16_64, (int)cospi_16_64, &stp2_10, \
+ &stp2_13); \
+ butterfly(stp1_12, stp1_11, (int)cospi_16_64, (int)cospi_16_64, &stp2_11, \
+ &stp2_12); \
\
stp2_16 = _mm_add_epi16(stp1_16, stp1_23); \
stp2_17 = _mm_add_epi16(stp1_17, stp1_22); \
@@ -1193,14 +1192,14 @@
stp1_18 = stp2_18; \
stp1_19 = stp2_19; \
\
- multiplication_and_add(stp2_27, stp2_20, (int)cospi_16_64, (int)cospi_16_64, \
- &stp1_20, &stp1_27); \
- multiplication_and_add(stp2_26, stp2_21, (int)cospi_16_64, (int)cospi_16_64, \
- &stp1_21, &stp1_26); \
- multiplication_and_add(stp2_25, stp2_22, (int)cospi_16_64, (int)cospi_16_64, \
- &stp1_22, &stp1_25); \
- multiplication_and_add(stp2_24, stp2_23, (int)cospi_16_64, (int)cospi_16_64, \
- &stp1_23, &stp1_24); \
+ butterfly(stp2_27, stp2_20, (int)cospi_16_64, (int)cospi_16_64, &stp1_20, \
+ &stp1_27); \
+ butterfly(stp2_26, stp2_21, (int)cospi_16_64, (int)cospi_16_64, &stp1_21, \
+ &stp1_26); \
+ butterfly(stp2_25, stp2_22, (int)cospi_16_64, (int)cospi_16_64, &stp1_22, \
+ &stp1_25); \
+ butterfly(stp2_24, stp2_23, (int)cospi_16_64, (int)cospi_16_64, &stp1_23, \
+ &stp1_24); \
\
stp1_28 = stp2_28; \
stp1_29 = stp2_29; \
--- a/vpx_dsp/x86/inv_txfm_sse2.h
+++ b/vpx_dsp/x86/inv_txfm_sse2.h
@@ -78,10 +78,9 @@
}
// Multiply elements by constants and add them together.
-static INLINE void multiplication_and_add(const __m128i in0, const __m128i in1,
- const int c0, const int c1,
- __m128i *const res0,
- __m128i *const res1) {+static INLINE void butterfly(const __m128i in0, const __m128i in1, const int c0,
+ const int c1, __m128i *const res0,
+ __m128i *const res1) {const __m128i cst0 = pair_set_epi16(c0, -c1);
const __m128i cst1 = pair_set_epi16(c1, c0);
const __m128i lo = _mm_unpacklo_epi16(in0, in1);
@@ -90,8 +89,6 @@
*res1 = idct_calc_wraplow_sse2(lo, hi, cst1);
}
-#define butterfly multiplication_and_add
-
// Functions to allow 8 bit optimisations to be used when profile 0 is used with
// highbitdepth enabled
static INLINE __m128i load_input_data4(const tran_low_t *data) {@@ -315,16 +312,16 @@
__m128i step1[8], step2[8];
// stage 1
- multiplication_and_add(in[1], in[7], (int)cospi_28_64, (int)cospi_4_64,
- &step1[4], &step1[7]);
- multiplication_and_add(in[5], in[3], (int)cospi_12_64, (int)cospi_20_64,
- &step1[5], &step1[6]);
+ butterfly(in[1], in[7], (int)cospi_28_64, (int)cospi_4_64, &step1[4],
+ &step1[7]);
+ butterfly(in[5], in[3], (int)cospi_12_64, (int)cospi_20_64, &step1[5],
+ &step1[6]);
// stage 2
- multiplication_and_add(in[0], in[4], (int)cospi_16_64, (int)cospi_16_64,
- &step2[1], &step2[0]);
- multiplication_and_add(in[2], in[6], (int)cospi_24_64, (int)cospi_8_64,
- &step2[2], &step2[3]);
+ butterfly(in[0], in[4], (int)cospi_16_64, (int)cospi_16_64, &step2[1],
+ &step2[0]);
+ butterfly(in[2], in[6], (int)cospi_24_64, (int)cospi_8_64, &step2[2],
+ &step2[3]);
step2[4] = _mm_add_epi16(step1[4], step1[5]);
step2[5] = _mm_sub_epi16(step1[4], step1[5]);
@@ -336,8 +333,8 @@
step1[1] = _mm_add_epi16(step2[1], step2[2]);
step1[2] = _mm_sub_epi16(step2[1], step2[2]);
step1[3] = _mm_sub_epi16(step2[0], step2[3]);
- multiplication_and_add(step2[6], step2[5], (int)cospi_16_64, (int)cospi_16_64,
- &step1[5], &step1[6]);
+ butterfly(step2[6], step2[5], (int)cospi_16_64, (int)cospi_16_64, &step1[5],
+ &step1[6]);
// stage 4
out[0] = _mm_add_epi16(step1[0], step2[7]);
@@ -412,20 +409,20 @@
__m128i step1[16], step2[16];
// stage 2
- multiplication_and_add(io[1], io[15], (int)cospi_30_64, (int)cospi_2_64,
- &step2[8], &step2[15]);
- multiplication_and_add(io[9], io[7], (int)cospi_14_64, (int)cospi_18_64,
- &step2[9], &step2[14]);
- multiplication_and_add(io[5], io[11], (int)cospi_22_64, (int)cospi_10_64,
- &step2[10], &step2[13]);
- multiplication_and_add(io[13], io[3], (int)cospi_6_64, (int)cospi_26_64,
- &step2[11], &step2[12]);
+ butterfly(io[1], io[15], (int)cospi_30_64, (int)cospi_2_64, &step2[8],
+ &step2[15]);
+ butterfly(io[9], io[7], (int)cospi_14_64, (int)cospi_18_64, &step2[9],
+ &step2[14]);
+ butterfly(io[5], io[11], (int)cospi_22_64, (int)cospi_10_64, &step2[10],
+ &step2[13]);
+ butterfly(io[13], io[3], (int)cospi_6_64, (int)cospi_26_64, &step2[11],
+ &step2[12]);
// stage 3
- multiplication_and_add(io[2], io[14], (int)cospi_28_64, (int)cospi_4_64,
- &step1[4], &step1[7]);
- multiplication_and_add(io[10], io[6], (int)cospi_12_64, (int)cospi_20_64,
- &step1[5], &step1[6]);
+ butterfly(io[2], io[14], (int)cospi_28_64, (int)cospi_4_64, &step1[4],
+ &step1[7]);
+ butterfly(io[10], io[6], (int)cospi_12_64, (int)cospi_20_64, &step1[5],
+ &step1[6]);
step1[8] = _mm_add_epi16(step2[8], step2[9]);
step1[9] = _mm_sub_epi16(step2[8], step2[9]);
step1[10] = _mm_sub_epi16(step2[11], step2[10]);
@@ -436,14 +433,14 @@
step1[15] = _mm_add_epi16(step2[14], step2[15]);
// stage 4
- multiplication_and_add(io[0], io[8], (int)cospi_16_64, (int)cospi_16_64,
- &step2[1], &step2[0]);
- multiplication_and_add(io[4], io[12], (int)cospi_24_64, (int)cospi_8_64,
- &step2[2], &step2[3]);
- multiplication_and_add(step1[14], step1[9], (int)cospi_24_64, (int)cospi_8_64,
- &step2[9], &step2[14]);
- multiplication_and_add(step1[10], step1[13], -(int)cospi_8_64,
- -(int)cospi_24_64, &step2[13], &step2[10]);
+ butterfly(io[0], io[8], (int)cospi_16_64, (int)cospi_16_64, &step2[1],
+ &step2[0]);
+ butterfly(io[4], io[12], (int)cospi_24_64, (int)cospi_8_64, &step2[2],
+ &step2[3]);
+ butterfly(step1[14], step1[9], (int)cospi_24_64, (int)cospi_8_64, &step2[9],
+ &step2[14]);
+ butterfly(step1[10], step1[13], -(int)cospi_8_64, -(int)cospi_24_64,
+ &step2[13], &step2[10]);
step2[5] = _mm_sub_epi16(step1[4], step1[5]);
step1[4] = _mm_add_epi16(step1[4], step1[5]);
step2[6] = _mm_sub_epi16(step1[7], step1[6]);
@@ -458,8 +455,8 @@
step1[1] = _mm_add_epi16(step2[1], step2[2]);
step1[2] = _mm_sub_epi16(step2[1], step2[2]);
step1[3] = _mm_sub_epi16(step2[0], step2[3]);
- multiplication_and_add(step2[6], step2[5], (int)cospi_16_64, (int)cospi_16_64,
- &step1[5], &step1[6]);
+ butterfly(step2[6], step2[5], (int)cospi_16_64, (int)cospi_16_64, &step1[5],
+ &step1[6]);
step1[8] = _mm_add_epi16(step2[8], step2[11]);
step1[9] = _mm_add_epi16(step2[9], step2[10]);
step1[10] = _mm_sub_epi16(step2[9], step2[10]);
@@ -478,10 +475,10 @@
step2[5] = _mm_sub_epi16(step1[2], step1[5]);
step2[6] = _mm_sub_epi16(step1[1], step1[6]);
step2[7] = _mm_sub_epi16(step1[0], step1[7]);
- multiplication_and_add(step1[13], step1[10], (int)cospi_16_64,
- (int)cospi_16_64, &step2[10], &step2[13]);
- multiplication_and_add(step1[12], step1[11], (int)cospi_16_64,
- (int)cospi_16_64, &step2[11], &step2[12]);
+ butterfly(step1[13], step1[10], (int)cospi_16_64, (int)cospi_16_64,
+ &step2[10], &step2[13]);
+ butterfly(step1[12], step1[11], (int)cospi_16_64, (int)cospi_16_64,
+ &step2[11], &step2[12]);
// stage 7
io[0] = _mm_add_epi16(step2[0], step1[15]);
@@ -615,26 +612,26 @@
transpose_16bit_4x8(l, io);
// stage 2
- multiplication_and_add(io[1], zero, (int)cospi_30_64, (int)cospi_2_64,
- &step2[8], &step2[15]);
- multiplication_and_add(zero, io[3], (int)cospi_6_64, (int)cospi_26_64,
- &step2[11], &step2[12]);
+ butterfly(io[1], zero, (int)cospi_30_64, (int)cospi_2_64, &step2[8],
+ &step2[15]);
+ butterfly(zero, io[3], (int)cospi_6_64, (int)cospi_26_64, &step2[11],
+ &step2[12]);
// stage 3
- multiplication_and_add(io[2], zero, (int)cospi_28_64, (int)cospi_4_64,
- &step1[4], &step1[7]);
+ butterfly(io[2], zero, (int)cospi_28_64, (int)cospi_4_64, &step1[4],
+ &step1[7]);
// stage 4
- multiplication_and_add(io[0], zero, (int)cospi_16_64, (int)cospi_16_64,
- &step1[1], &step1[0]);
- multiplication_and_add(step2[15], step2[8], (int)cospi_24_64, (int)cospi_8_64,
- &step2[9], &step2[14]);
- multiplication_and_add(step2[11], step2[12], -(int)cospi_8_64,
- -(int)cospi_24_64, &step2[13], &step2[10]);
+ butterfly(io[0], zero, (int)cospi_16_64, (int)cospi_16_64, &step1[1],
+ &step1[0]);
+ butterfly(step2[15], step2[8], (int)cospi_24_64, (int)cospi_8_64, &step2[9],
+ &step2[14]);
+ butterfly(step2[11], step2[12], -(int)cospi_8_64, -(int)cospi_24_64,
+ &step2[13], &step2[10]);
// stage 5
- multiplication_and_add(step1[7], step1[4], (int)cospi_16_64, (int)cospi_16_64,
- &step1[5], &step1[6]);
+ butterfly(step1[7], step1[4], (int)cospi_16_64, (int)cospi_16_64, &step1[5],
+ &step1[6]);
step1[8] = _mm_add_epi16(step2[8], step2[11]);
step1[9] = _mm_add_epi16(step2[9], step2[10]);
step1[10] = _mm_sub_epi16(step2[9], step2[10]);
@@ -653,10 +650,10 @@
step2[5] = _mm_sub_epi16(step1[1], step1[5]);
step2[6] = _mm_sub_epi16(step1[1], step1[6]);
step2[7] = _mm_sub_epi16(step1[0], step1[7]);
- multiplication_and_add(step1[13], step1[10], (int)cospi_16_64,
- (int)cospi_16_64, &step2[10], &step2[13]);
- multiplication_and_add(step1[12], step1[11], (int)cospi_16_64,
- (int)cospi_16_64, &step2[11], &step2[12]);
+ butterfly(step1[13], step1[10], (int)cospi_16_64, (int)cospi_16_64,
+ &step2[10], &step2[13]);
+ butterfly(step1[12], step1[11], (int)cospi_16_64, (int)cospi_16_64,
+ &step2[11], &step2[12]);
// stage 7
io[0] = _mm_add_epi16(step2[0], step1[15]);
--- a/vpx_dsp/x86/inv_txfm_ssse3.h
+++ b/vpx_dsp/x86/inv_txfm_ssse3.h
@@ -92,8 +92,8 @@
step1[1] = _mm_add_epi16(step2[0], step2[2]);
step1[2] = _mm_sub_epi16(step2[0], step2[2]);
step1[3] = _mm_sub_epi16(step2[0], step2[3]);
- multiplication_and_add(step2[6], step2[5], (int)cospi_16_64, (int)cospi_16_64,
- &step1[5], &step1[6]);
+ butterfly(step2[6], step2[5], (int)cospi_16_64, (int)cospi_16_64, &step1[5],
+ &step1[6]);
// stage 4
io[0] = _mm_add_epi16(step1[0], step2[7]);
--
⑨