Update highbd idct x86 optimizations.

BUG=webm:1412

Change-Id: Ia275940af7d7d8637e9a851a9e39d655bfbe4069

diff --git a/vpx_dsp/x86/highbd_idct16x16_add_sse2.c b/vpx_dsp/x86/highbd_idct16x16_add_sse2.c
index 3d1a43dc86e6cf77740fb9d0e16b8c870a657809..c5d82d9ed0d0ffcdeefc682c434dd3684fa290df 100644 (file)
--- a/vpx_dsp/x86/highbd_idct16x16_add_sse2.c
+++ b/vpx_dsp/x86/highbd_idct16x16_add_sse2.c
@@ -106,20 +106,20 @@ static INLINE void highbd_idct16x16_38_4col(__m128i *const io /*io[16]*/) {
   __m128i temp1[2], sign[2];
 
   // stage 2
-  highbd_multiplication_sse2(io[1], (int)cospi_30_64, (int)cospi_2_64,
-                             &step2[8], &step2[15]);
-  highbd_multiplication_neg_sse2(io[7], (int)cospi_14_64, (int)cospi_18_64,
-                                 &step2[9], &step2[14]);
-  highbd_multiplication_sse2(io[5], (int)cospi_22_64, (int)cospi_10_64,
-                             &step2[10], &step2[13]);
-  highbd_multiplication_neg_sse2(io[3], (int)cospi_6_64, (int)cospi_26_64,
-                                 &step2[11], &step2[12]);
+  highbd_partial_butterfly_sse2(io[1], (int)cospi_30_64, (int)cospi_2_64,
+                                &step2[8], &step2[15]);
+  highbd_partial_butterfly_neg_sse2(io[7], (int)cospi_14_64, (int)cospi_18_64,
+                                    &step2[9], &step2[14]);
+  highbd_partial_butterfly_sse2(io[5], (int)cospi_22_64, (int)cospi_10_64,
+                                &step2[10], &step2[13]);
+  highbd_partial_butterfly_neg_sse2(io[3], (int)cospi_6_64, (int)cospi_26_64,
+                                    &step2[11], &step2[12]);
 
   // stage 3
-  highbd_multiplication_sse2(io[2], (int)cospi_28_64, (int)cospi_4_64,
-                             &step1[4], &step1[7]);
-  highbd_multiplication_neg_sse2(io[6], (int)cospi_12_64, (int)cospi_20_64,
-                                 &step1[5], &step1[6]);
+  highbd_partial_butterfly_sse2(io[2], (int)cospi_28_64, (int)cospi_4_64,
+                                &step1[4], &step1[7]);
+  highbd_partial_butterfly_neg_sse2(io[6], (int)cospi_12_64, (int)cospi_20_64,
+                                    &step1[5], &step1[6]);
   step1[8] = _mm_add_epi32(step2[8], step2[9]);
   step1[9] = _mm_sub_epi32(step2[8], step2[9]);
   step1[10] = _mm_sub_epi32(step2[10], step2[11]);  // step1[10] = -step1[10]
@@ -133,8 +133,8 @@ static INLINE void highbd_idct16x16_38_4col(__m128i *const io /*io[16]*/) {
   abs_extend_64bit_sse2(io[0], temp1, sign);
   step2[0] = multiplication_round_shift_sse2(temp1, sign, (int)cospi_16_64);
   step2[1] = step2[0];
-  highbd_multiplication_sse2(io[4], (int)cospi_24_64, (int)cospi_8_64,
-                             &step2[2], &step2[3]);
+  highbd_partial_butterfly_sse2(io[4], (int)cospi_24_64, (int)cospi_8_64,
+                                &step2[2], &step2[3]);
   highbd_butterfly_sse2(step1[14], step1[9], (int)cospi_24_64, (int)cospi_8_64,
                         &step2[9], &step2[14]);
   highbd_butterfly_sse2(step1[10], step1[13], (int)cospi_8_64, (int)cospi_24_64,
@@ -158,14 +158,14 @@ static INLINE void highbd_idct16x16_10_4col(__m128i *const io /*io[16]*/) {
   __m128i temp[2], sign[2];
 
   // stage 2
-  highbd_multiplication_sse2(io[1], (int)cospi_30_64, (int)cospi_2_64,
-                             &step2[8], &step2[15]);
-  highbd_multiplication_neg_sse2(io[3], (int)cospi_6_64, (int)cospi_26_64,
-                                 &step2[11], &step2[12]);
+  highbd_partial_butterfly_sse2(io[1], (int)cospi_30_64, (int)cospi_2_64,
+                                &step2[8], &step2[15]);
+  highbd_partial_butterfly_neg_sse2(io[3], (int)cospi_6_64, (int)cospi_26_64,
+                                    &step2[11], &step2[12]);
 
   // stage 3
-  highbd_multiplication_sse2(io[2], (int)cospi_28_64, (int)cospi_4_64,
-                             &step1[4], &step1[7]);
+  highbd_partial_butterfly_sse2(io[2], (int)cospi_28_64, (int)cospi_4_64,
+                                &step1[4], &step1[7]);
   step1[8] = step2[8];
   step1[9] = step2[8];
   step1[10] =
@@ -209,25 +209,8 @@ void vpx_highbd_idct16x16_256_add_sse2(const tran_low_t *input, uint16_t *dest,
 
     in = l;
     for (i = 0; i < 2; i++) {
-      in[0] = load_pack_8_32bit(input + 0 * 16);
-      in[1] = load_pack_8_32bit(input + 1 * 16);
-      in[2] = load_pack_8_32bit(input + 2 * 16);
-      in[3] = load_pack_8_32bit(input + 3 * 16);
-      in[4] = load_pack_8_32bit(input + 4 * 16);
-      in[5] = load_pack_8_32bit(input + 5 * 16);
-      in[6] = load_pack_8_32bit(input + 6 * 16);
-      in[7] = load_pack_8_32bit(input + 7 * 16);
-      transpose_16bit_8x8(in, in);
-
-      in[8] = load_pack_8_32bit(input + 0 * 16 + 8);
-      in[9] = load_pack_8_32bit(input + 1 * 16 + 8);
-      in[10] = load_pack_8_32bit(input + 2 * 16 + 8);
-      in[11] = load_pack_8_32bit(input + 3 * 16 + 8);
-      in[12] = load_pack_8_32bit(input + 4 * 16 + 8);
-      in[13] = load_pack_8_32bit(input + 5 * 16 + 8);
-      in[14] = load_pack_8_32bit(input + 6 * 16 + 8);
-      in[15] = load_pack_8_32bit(input + 7 * 16 + 8);
-      transpose_16bit_8x8(in + 8, in + 8);
+      highbd_load_pack_transpose_32bit_8x8(&input[0], 16, &in[0]);
+      highbd_load_pack_transpose_32bit_8x8(&input[8], 16, &in[8]);
       idct16_8col(in, in);
       in = r;
       input += 128;
@@ -249,52 +232,18 @@ void vpx_highbd_idct16x16_256_add_sse2(const tran_low_t *input, uint16_t *dest,
 
     for (i = 0; i < 4; i++) {
       in = all[i];
-      in[0] = _mm_load_si128((const __m128i *)(input + 0 * 16 + 0));
-      in[1] = _mm_load_si128((const __m128i *)(input + 0 * 16 + 4));
-      in[2] = _mm_load_si128((const __m128i *)(input + 1 * 16 + 0));
-      in[3] = _mm_load_si128((const __m128i *)(input + 1 * 16 + 4));
-      in[4] = _mm_load_si128((const __m128i *)(input + 2 * 16 + 0));
-      in[5] = _mm_load_si128((const __m128i *)(input + 2 * 16 + 4));
-      in[6] = _mm_load_si128((const __m128i *)(input + 3 * 16 + 0));
-      in[7] = _mm_load_si128((const __m128i *)(input + 3 * 16 + 4));
-      transpose_32bit_8x4(in, in);
-
-      in[8] = _mm_load_si128((const __m128i *)(input + 0 * 16 + 8));
-      in[9] = _mm_load_si128((const __m128i *)(input + 0 * 16 + 12));
-      in[10] = _mm_load_si128((const __m128i *)(input + 1 * 16 + 8));
-      in[11] = _mm_load_si128((const __m128i *)(input + 1 * 16 + 12));
-      in[12] = _mm_load_si128((const __m128i *)(input + 2 * 16 + 8));
-      in[13] = _mm_load_si128((const __m128i *)(input + 2 * 16 + 12));
-      in[14] = _mm_load_si128((const __m128i *)(input + 3 * 16 + 8));
-      in[15] = _mm_load_si128((const __m128i *)(input + 3 * 16 + 12));
-      transpose_32bit_8x4(in + 8, in + 8);
-
+      highbd_load_transpose_32bit_8x4(&input[0], 16, &in[0]);
+      highbd_load_transpose_32bit_8x4(&input[8], 16, &in[8]);
       highbd_idct16_4col(in);
       input += 4 * 16;
     }
 
     for (i = 0; i < 16; i += 4) {
       int j;
-      out[0] = all[0][i + 0];
-      out[1] = all[1][i + 0];
-      out[2] = all[0][i + 1];
-      out[3] = all[1][i + 1];
-      out[4] = all[0][i + 2];
-      out[5] = all[1][i + 2];
-      out[6] = all[0][i + 3];
-      out[7] = all[1][i + 3];
-      transpose_32bit_8x4(out, out);
-
-      out[8] = all[2][i + 0];
-      out[9] = all[3][i + 0];
-      out[10] = all[2][i + 1];
-      out[11] = all[3][i + 1];
-      out[12] = all[2][i + 2];
-      out[13] = all[3][i + 2];
-      out[14] = all[2][i + 3];
-      out[15] = all[3][i + 3];
-      transpose_32bit_8x4(out + 8, out + 8);
-
+      transpose_32bit_4x4(all[0] + i, out + 0);
+      transpose_32bit_4x4(all[1] + i, out + 4);
+      transpose_32bit_4x4(all[2] + i, out + 8);
+      transpose_32bit_4x4(all[3] + i, out + 12);
       highbd_idct16_4col(out);
 
       for (j = 0; j < 16; ++j) {
@@ -313,16 +262,7 @@ void vpx_highbd_idct16x16_38_add_sse2(const tran_low_t *input, uint16_t *dest,
   if (bd == 8) {
     __m128i in[16], temp[16];
 
-    in[0] = load_pack_8_32bit(input + 0 * 16);
-    in[1] = load_pack_8_32bit(input + 1 * 16);
-    in[2] = load_pack_8_32bit(input + 2 * 16);
-    in[3] = load_pack_8_32bit(input + 3 * 16);
-    in[4] = load_pack_8_32bit(input + 4 * 16);
-    in[5] = load_pack_8_32bit(input + 5 * 16);
-    in[6] = load_pack_8_32bit(input + 6 * 16);
-    in[7] = load_pack_8_32bit(input + 7 * 16);
-    transpose_16bit_8x8(in, in);
-
+    highbd_load_pack_transpose_32bit_8x8(input, 16, in);
     for (i = 8; i < 16; i++) {
       in[i] = _mm_setzero_si128();
     }
@@ -343,30 +283,15 @@ void vpx_highbd_idct16x16_38_add_sse2(const tran_low_t *input, uint16_t *dest,
 
     for (i = 0; i < 2; i++) {
       in = all[i];
-      in[0] = _mm_load_si128((const __m128i *)(input + 0 * 16 + 0));
-      in[1] = _mm_load_si128((const __m128i *)(input + 0 * 16 + 4));
-      in[2] = _mm_load_si128((const __m128i *)(input + 1 * 16 + 0));
-      in[3] = _mm_load_si128((const __m128i *)(input + 1 * 16 + 4));
-      in[4] = _mm_load_si128((const __m128i *)(input + 2 * 16 + 0));
-      in[5] = _mm_load_si128((const __m128i *)(input + 2 * 16 + 4));
-      in[6] = _mm_load_si128((const __m128i *)(input + 3 * 16 + 0));
-      in[7] = _mm_load_si128((const __m128i *)(input + 3 * 16 + 4));
-      transpose_32bit_8x4(in, in);
+      highbd_load_transpose_32bit_8x4(input, 16, in);
       highbd_idct16x16_38_4col(in);
       input += 4 * 16;
     }
 
     for (i = 0; i < 16; i += 4) {
       int j;
-      out[0] = all[0][i + 0];
-      out[1] = all[1][i + 0];
-      out[2] = all[0][i + 1];
-      out[3] = all[1][i + 1];
-      out[4] = all[0][i + 2];
-      out[5] = all[1][i + 2];
-      out[6] = all[0][i + 3];
-      out[7] = all[1][i + 3];
-      transpose_32bit_8x4(out, out);
+      transpose_32bit_4x4(all[0] + i, out + 0);
+      transpose_32bit_4x4(all[1] + i, out + 4);
       highbd_idct16x16_38_4col(out);
 
       for (j = 0; j < 16; ++j) {
@@ -406,11 +331,7 @@ void vpx_highbd_idct16x16_10_add_sse2(const tran_low_t *input, uint16_t *dest,
 
     for (i = 0; i < 2; i++) {
       in = all[i];
-      in[0] = _mm_load_si128((const __m128i *)(input + 0 * 16));
-      in[1] = _mm_load_si128((const __m128i *)(input + 1 * 16));
-      in[2] = _mm_load_si128((const __m128i *)(input + 2 * 16));
-      in[3] = _mm_load_si128((const __m128i *)(input + 3 * 16));
-      transpose_32bit_4x4(in, in);
+      highbd_load_transpose_32bit_4x4(input, 16, in);
       highbd_idct16x16_10_4col(in);
       input += 4 * 16;
     }

diff --git a/vpx_dsp/x86/highbd_idct16x16_add_sse4.c b/vpx_dsp/x86/highbd_idct16x16_add_sse4.c
index 6f5f663946250137c86285379c284ad68d96076c..4c5452cc49dc29836072cd4202a3c9c3b350562d 100644 (file)
--- a/vpx_dsp/x86/highbd_idct16x16_add_sse4.c
+++ b/vpx_dsp/x86/highbd_idct16x16_add_sse4.c
@@ -107,20 +107,20 @@ static INLINE void highbd_idct16x16_38_4col(__m128i *const io /*io[16]*/) {
   __m128i temp1[2];
 
   // stage 2
-  highbd_multiplication_sse4_1(io[1], (int)cospi_30_64, (int)cospi_2_64,
-                               &step2[8], &step2[15]);
-  highbd_multiplication_sse4_1(io[7], -(int)cospi_18_64, (int)cospi_14_64,
-                               &step2[9], &step2[14]);
-  highbd_multiplication_sse4_1(io[5], (int)cospi_22_64, (int)cospi_10_64,
-                               &step2[10], &step2[13]);
-  highbd_multiplication_sse4_1(io[3], -(int)cospi_26_64, (int)cospi_6_64,
-                               &step2[11], &step2[12]);
+  highbd_partial_butterfly_sse4_1(io[1], (int)cospi_30_64, (int)cospi_2_64,
+                                  &step2[8], &step2[15]);
+  highbd_partial_butterfly_sse4_1(io[7], -(int)cospi_18_64, (int)cospi_14_64,
+                                  &step2[9], &step2[14]);
+  highbd_partial_butterfly_sse4_1(io[5], (int)cospi_22_64, (int)cospi_10_64,
+                                  &step2[10], &step2[13]);
+  highbd_partial_butterfly_sse4_1(io[3], -(int)cospi_26_64, (int)cospi_6_64,
+                                  &step2[11], &step2[12]);
 
   // stage 3
-  highbd_multiplication_sse4_1(io[2], (int)cospi_28_64, (int)cospi_4_64,
-                               &step1[4], &step1[7]);
-  highbd_multiplication_sse4_1(io[6], -(int)cospi_20_64, (int)cospi_12_64,
-                               &step1[5], &step1[6]);
+  highbd_partial_butterfly_sse4_1(io[2], (int)cospi_28_64, (int)cospi_4_64,
+                                  &step1[4], &step1[7]);
+  highbd_partial_butterfly_sse4_1(io[6], -(int)cospi_20_64, (int)cospi_12_64,
+                                  &step1[5], &step1[6]);
   step1[8] = _mm_add_epi32(step2[8], step2[9]);
   step1[9] = _mm_sub_epi32(step2[8], step2[9]);
   step1[10] = _mm_sub_epi32(step2[11], step2[10]);
@@ -134,8 +134,8 @@ static INLINE void highbd_idct16x16_38_4col(__m128i *const io /*io[16]*/) {
   extend_64bit(io[0], temp1);
   step2[0] = multiplication_round_shift_sse4_1(temp1, (int)cospi_16_64);
   step2[1] = step2[0];
-  highbd_multiplication_sse4_1(io[4], (int)cospi_24_64, (int)cospi_8_64,
-                               &step2[2], &step2[3]);
+  highbd_partial_butterfly_sse4_1(io[4], (int)cospi_24_64, (int)cospi_8_64,
+                                  &step2[2], &step2[3]);
   highbd_butterfly_sse4_1(step1[14], step1[9], (int)cospi_24_64,
                           (int)cospi_8_64, &step2[9], &step2[14]);
   highbd_butterfly_sse4_1(step1[10], step1[13], -(int)cospi_8_64,
@@ -159,14 +159,14 @@ static INLINE void highbd_idct16x16_10_4col(__m128i *const io /*io[16]*/) {
   __m128i temp[2];
 
   // stage 2
-  highbd_multiplication_sse4_1(io[1], (int)cospi_30_64, (int)cospi_2_64,
-                               &step2[8], &step2[15]);
-  highbd_multiplication_sse4_1(io[3], -(int)cospi_26_64, (int)cospi_6_64,
-                               &step2[11], &step2[12]);
+  highbd_partial_butterfly_sse4_1(io[1], (int)cospi_30_64, (int)cospi_2_64,
+                                  &step2[8], &step2[15]);
+  highbd_partial_butterfly_sse4_1(io[3], -(int)cospi_26_64, (int)cospi_6_64,
+                                  &step2[11], &step2[12]);
 
   // stage 3
-  highbd_multiplication_sse4_1(io[2], (int)cospi_28_64, (int)cospi_4_64,
-                               &step1[4], &step1[7]);
+  highbd_partial_butterfly_sse4_1(io[2], (int)cospi_28_64, (int)cospi_4_64,
+                                  &step1[4], &step1[7]);
   step1[8] = step2[8];
   step1[9] = step2[8];
   step1[10] = step2[11];
@@ -208,25 +208,8 @@ void vpx_highbd_idct16x16_256_add_sse4_1(const tran_low_t *input,
 
     in = l;
     for (i = 0; i < 2; i++) {
-      in[0] = load_pack_8_32bit(input + 0 * 16);
-      in[1] = load_pack_8_32bit(input + 1 * 16);
-      in[2] = load_pack_8_32bit(input + 2 * 16);
-      in[3] = load_pack_8_32bit(input + 3 * 16);
-      in[4] = load_pack_8_32bit(input + 4 * 16);
-      in[5] = load_pack_8_32bit(input + 5 * 16);
-      in[6] = load_pack_8_32bit(input + 6 * 16);
-      in[7] = load_pack_8_32bit(input + 7 * 16);
-      transpose_16bit_8x8(in, in);
-
-      in[8] = load_pack_8_32bit(input + 0 * 16 + 8);
-      in[9] = load_pack_8_32bit(input + 1 * 16 + 8);
-      in[10] = load_pack_8_32bit(input + 2 * 16 + 8);
-      in[11] = load_pack_8_32bit(input + 3 * 16 + 8);
-      in[12] = load_pack_8_32bit(input + 4 * 16 + 8);
-      in[13] = load_pack_8_32bit(input + 5 * 16 + 8);
-      in[14] = load_pack_8_32bit(input + 6 * 16 + 8);
-      in[15] = load_pack_8_32bit(input + 7 * 16 + 8);
-      transpose_16bit_8x8(in + 8, in + 8);
+      highbd_load_pack_transpose_32bit_8x8(&input[0], 16, &in[0]);
+      highbd_load_pack_transpose_32bit_8x8(&input[8], 16, &in[8]);
       idct16_8col(in, in);
       in = r;
       input += 128;
@@ -248,52 +231,18 @@ void vpx_highbd_idct16x16_256_add_sse4_1(const tran_low_t *input,
 
     for (i = 0; i < 4; i++) {
       in = all[i];
-      in[0] = _mm_load_si128((const __m128i *)(input + 0 * 16 + 0));
-      in[1] = _mm_load_si128((const __m128i *)(input + 0 * 16 + 4));
-      in[2] = _mm_load_si128((const __m128i *)(input + 1 * 16 + 0));
-      in[3] = _mm_load_si128((const __m128i *)(input + 1 * 16 + 4));
-      in[4] = _mm_load_si128((const __m128i *)(input + 2 * 16 + 0));
-      in[5] = _mm_load_si128((const __m128i *)(input + 2 * 16 + 4));
-      in[6] = _mm_load_si128((const __m128i *)(input + 3 * 16 + 0));
-      in[7] = _mm_load_si128((const __m128i *)(input + 3 * 16 + 4));
-      transpose_32bit_8x4(in, in);
-
-      in[8] = _mm_load_si128((const __m128i *)(input + 0 * 16 + 8));
-      in[9] = _mm_load_si128((const __m128i *)(input + 0 * 16 + 12));
-      in[10] = _mm_load_si128((const __m128i *)(input + 1 * 16 + 8));
-      in[11] = _mm_load_si128((const __m128i *)(input + 1 * 16 + 12));
-      in[12] = _mm_load_si128((const __m128i *)(input + 2 * 16 + 8));
-      in[13] = _mm_load_si128((const __m128i *)(input + 2 * 16 + 12));
-      in[14] = _mm_load_si128((const __m128i *)(input + 3 * 16 + 8));
-      in[15] = _mm_load_si128((const __m128i *)(input + 3 * 16 + 12));
-      transpose_32bit_8x4(in + 8, in + 8);
-
+      highbd_load_transpose_32bit_8x4(&input[0], 16, &in[0]);
+      highbd_load_transpose_32bit_8x4(&input[8], 16, &in[8]);
       highbd_idct16_4col(in);
       input += 4 * 16;
     }
 
     for (i = 0; i < 16; i += 4) {
       int j;
-      out[0] = all[0][i + 0];
-      out[1] = all[1][i + 0];
-      out[2] = all[0][i + 1];
-      out[3] = all[1][i + 1];
-      out[4] = all[0][i + 2];
-      out[5] = all[1][i + 2];
-      out[6] = all[0][i + 3];
-      out[7] = all[1][i + 3];
-      transpose_32bit_8x4(out, out);
-
-      out[8] = all[2][i + 0];
-      out[9] = all[3][i + 0];
-      out[10] = all[2][i + 1];
-      out[11] = all[3][i + 1];
-      out[12] = all[2][i + 2];
-      out[13] = all[3][i + 2];
-      out[14] = all[2][i + 3];
-      out[15] = all[3][i + 3];
-      transpose_32bit_8x4(out + 8, out + 8);
-
+      transpose_32bit_4x4(all[0] + i, out + 0);
+      transpose_32bit_4x4(all[1] + i, out + 4);
+      transpose_32bit_4x4(all[2] + i, out + 8);
+      transpose_32bit_4x4(all[3] + i, out + 12);
       highbd_idct16_4col(out);
 
       for (j = 0; j < 16; ++j) {
@@ -312,16 +261,7 @@ void vpx_highbd_idct16x16_38_add_sse4_1(const tran_low_t *input, uint16_t *dest,
   if (bd == 8) {
     __m128i in[16], temp[16];
 
-    in[0] = load_pack_8_32bit(input + 0 * 16);
-    in[1] = load_pack_8_32bit(input + 1 * 16);
-    in[2] = load_pack_8_32bit(input + 2 * 16);
-    in[3] = load_pack_8_32bit(input + 3 * 16);
-    in[4] = load_pack_8_32bit(input + 4 * 16);
-    in[5] = load_pack_8_32bit(input + 5 * 16);
-    in[6] = load_pack_8_32bit(input + 6 * 16);
-    in[7] = load_pack_8_32bit(input + 7 * 16);
-    transpose_16bit_8x8(in, in);
-
+    highbd_load_pack_transpose_32bit_8x8(&input[0], 16, &in[0]);
     for (i = 8; i < 16; i++) {
       in[i] = _mm_setzero_si128();
     }
@@ -342,30 +282,15 @@ void vpx_highbd_idct16x16_38_add_sse4_1(const tran_low_t *input, uint16_t *dest,
 
     for (i = 0; i < 2; i++) {
       in = all[i];
-      in[0] = _mm_load_si128((const __m128i *)(input + 0 * 16 + 0));
-      in[1] = _mm_load_si128((const __m128i *)(input + 0 * 16 + 4));
-      in[2] = _mm_load_si128((const __m128i *)(input + 1 * 16 + 0));
-      in[3] = _mm_load_si128((const __m128i *)(input + 1 * 16 + 4));
-      in[4] = _mm_load_si128((const __m128i *)(input + 2 * 16 + 0));
-      in[5] = _mm_load_si128((const __m128i *)(input + 2 * 16 + 4));
-      in[6] = _mm_load_si128((const __m128i *)(input + 3 * 16 + 0));
-      in[7] = _mm_load_si128((const __m128i *)(input + 3 * 16 + 4));
-      transpose_32bit_8x4(in, in);
+      highbd_load_transpose_32bit_8x4(input, 16, in);
       highbd_idct16x16_38_4col(in);
       input += 4 * 16;
     }
 
     for (i = 0; i < 16; i += 4) {
       int j;
-      out[0] = all[0][i + 0];
-      out[1] = all[1][i + 0];
-      out[2] = all[0][i + 1];
-      out[3] = all[1][i + 1];
-      out[4] = all[0][i + 2];
-      out[5] = all[1][i + 2];
-      out[6] = all[0][i + 3];
-      out[7] = all[1][i + 3];
-      transpose_32bit_8x4(out, out);
+      transpose_32bit_4x4(all[0] + i, out + 0);
+      transpose_32bit_4x4(all[1] + i, out + 4);
       highbd_idct16x16_38_4col(out);
 
       for (j = 0; j < 16; ++j) {
@@ -405,11 +330,7 @@ void vpx_highbd_idct16x16_10_add_sse4_1(const tran_low_t *input, uint16_t *dest,
 
     for (i = 0; i < 2; i++) {
       in = all[i];
-      in[0] = _mm_load_si128((const __m128i *)(input + 0 * 16));
-      in[1] = _mm_load_si128((const __m128i *)(input + 1 * 16));
-      in[2] = _mm_load_si128((const __m128i *)(input + 2 * 16));
-      in[3] = _mm_load_si128((const __m128i *)(input + 3 * 16));
-      transpose_32bit_4x4(in, in);
+      highbd_load_transpose_32bit_4x4(input, 16, in);
       highbd_idct16x16_10_4col(in);
       input += 4 * 16;
     }

diff --git a/vpx_dsp/x86/highbd_inv_txfm_sse2.h b/vpx_dsp/x86/highbd_inv_txfm_sse2.h
index 81187e0fdd757008e680cfb68fe3b5dc90faf3e2..c58e49c0252e8c87718d5341f7b6565e67f6f454 100644 (file)
--- a/vpx_dsp/x86/highbd_inv_txfm_sse2.h
+++ b/vpx_dsp/x86/highbd_inv_txfm_sse2.h
@@ -16,6 +16,7 @@
 #include "./vpx_config.h"
 #include "vpx/vpx_integer.h"
 #include "vpx_dsp/inv_txfm.h"
+#include "vpx_dsp/x86/transpose_sse2.h"
 #include "vpx_dsp/x86/txfm_common_sse2.h"
 
 static INLINE void extend_64bit(const __m128i in,
@@ -84,6 +85,7 @@ static INLINE __m128i multiplication_round_shift_sse2(
   const __m128i pair_c = pair_set_epi32(c << 2, 0);
   __m128i t0, t1;
 
+  assert(c >= 0);
   t0 = multiply_apply_sign_sse2(in[0], sign[0], pair_c);
   t1 = multiply_apply_sign_sse2(in[1], sign[1], pair_c);
   t0 = dct_const_round_shift_64bit(t0);
@@ -99,6 +101,7 @@ static INLINE __m128i multiplication_neg_round_shift_sse2(
   const __m128i pair_c = pair_set_epi32(c << 2, 0);
   __m128i t0, t1;
 
+  assert(c >= 0);
   t0 = multiply_apply_sign_sse2(in[0], sign[0], pair_c);
   t1 = multiply_apply_sign_sse2(in[1], sign[1], pair_c);
   t0 = _mm_sub_epi64(_mm_setzero_si128(), t0);
@@ -118,6 +121,8 @@ static INLINE void highbd_butterfly_sse2(const __m128i in0, const __m128i in1,
   const __m128i pair_c1 = pair_set_epi32(c1 << 2, 0);
   __m128i temp1[4], temp2[4], sign1[2], sign2[2];
 
+  assert(c0 >= 0);
+  assert(c1 >= 0);
   abs_extend_64bit_sse2(in0, temp1, sign1);
   abs_extend_64bit_sse2(in1, temp2, sign2);
   temp1[2] = multiply_apply_sign_sse2(temp1[0], sign1[0], pair_c1);
@@ -140,43 +145,61 @@ static INLINE void highbd_butterfly_sse2(const __m128i in0, const __m128i in1,
   *out1 = pack_4(temp2[0], temp2[1]);
 }
 
-static INLINE void highbd_butterfly_cospi16_sse2(const __m128i in0,
-                                                 const __m128i in1,
+// Note: c0 and c1 must be non negative.
+static INLINE void highbd_partial_butterfly_sse2(const __m128i in, const int c0,
+                                                 const int c1,
                                                  __m128i *const out0,
                                                  __m128i *const out1) {
-  __m128i temp1[2], temp2, sign[2];
-
-  temp2 = _mm_add_epi32(in0, in1);
-  abs_extend_64bit_sse2(temp2, temp1, sign);
-  *out0 = multiplication_round_shift_sse2(temp1, sign, (int)cospi_16_64);
-  temp2 = _mm_sub_epi32(in0, in1);
-  abs_extend_64bit_sse2(temp2, temp1, sign);
-  *out1 = multiplication_round_shift_sse2(temp1, sign, (int)cospi_16_64);
-}
-
-// Note: c0 and c1 must be non negative.
-static INLINE void highbd_multiplication_sse2(const __m128i in, const int c0,
-                                              const int c1, __m128i *const out0,
-                                              __m128i *const out1) {
   __m128i temp[2], sign[2];
 
+  assert(c0 >= 0);
+  assert(c1 >= 0);
   abs_extend_64bit_sse2(in, temp, sign);
   *out0 = multiplication_round_shift_sse2(temp, sign, c0);
   *out1 = multiplication_round_shift_sse2(temp, sign, c1);
 }
 
 // Note: c0 and c1 must be non negative.
-static INLINE void highbd_multiplication_neg_sse2(const __m128i in,
-                                                  const int c0, const int c1,
-                                                  __m128i *const out0,
-                                                  __m128i *const out1) {
+static INLINE void highbd_partial_butterfly_neg_sse2(const __m128i in,
+                                                     const int c0, const int c1,
+                                                     __m128i *const out0,
+                                                     __m128i *const out1) {
   __m128i temp[2], sign[2];
 
+  assert(c0 >= 0);
+  assert(c1 >= 0);
   abs_extend_64bit_sse2(in, temp, sign);
   *out0 = multiplication_neg_round_shift_sse2(temp, sign, c1);
   *out1 = multiplication_round_shift_sse2(temp, sign, c0);
 }
 
+static INLINE void highbd_butterfly_cospi16_sse2(const __m128i in0,
+                                                 const __m128i in1,
+                                                 __m128i *const out0,
+                                                 __m128i *const out1) {
+  __m128i temp1[2], temp2, sign[2];
+
+  temp2 = _mm_add_epi32(in0, in1);
+  abs_extend_64bit_sse2(temp2, temp1, sign);
+  *out0 = multiplication_round_shift_sse2(temp1, sign, (int)cospi_16_64);
+  temp2 = _mm_sub_epi32(in0, in1);
+  abs_extend_64bit_sse2(temp2, temp1, sign);
+  *out1 = multiplication_round_shift_sse2(temp1, sign, (int)cospi_16_64);
+}
+
+// Only do addition and subtraction butterfly, size = 16, 32
+static INLINE void highbd_add_sub_butterfly(const __m128i *in, __m128i *out,
+                                            int size) {
+  int i = 0;
+  const int num = size >> 1;
+  const int bound = size - 1;
+  while (i < num) {
+    out[i] = _mm_add_epi32(in[i], in[bound - i]);
+    out[bound - i] = _mm_sub_epi32(in[i], in[bound - i]);
+    i++;
+  }
+}
+
 static INLINE void highbd_idct8_stage4(const __m128i *const in,
                                        __m128i *const out) {
   out[0] = _mm_add_epi32(in[0], in[7]);
@@ -313,6 +336,44 @@ static INLINE __m128i load_pack_8_32bit(const tran_low_t *const input) {
   return _mm_packs_epi32(t0, t1);
 }
 
+static INLINE void highbd_load_pack_transpose_32bit_8x8(const tran_low_t *input,
+                                                        const int stride,
+                                                        __m128i *const in) {
+  in[0] = load_pack_8_32bit(input + 0 * stride);
+  in[1] = load_pack_8_32bit(input + 1 * stride);
+  in[2] = load_pack_8_32bit(input + 2 * stride);
+  in[3] = load_pack_8_32bit(input + 3 * stride);
+  in[4] = load_pack_8_32bit(input + 4 * stride);
+  in[5] = load_pack_8_32bit(input + 5 * stride);
+  in[6] = load_pack_8_32bit(input + 6 * stride);
+  in[7] = load_pack_8_32bit(input + 7 * stride);
+  transpose_16bit_8x8(in, in);
+}
+
+static INLINE void highbd_load_transpose_32bit_8x4(const tran_low_t *input,
+                                                   const int stride,
+                                                   __m128i *in) {
+  in[0] = _mm_load_si128((const __m128i *)(input + 0 * stride + 0));
+  in[1] = _mm_load_si128((const __m128i *)(input + 0 * stride + 4));
+  in[2] = _mm_load_si128((const __m128i *)(input + 1 * stride + 0));
+  in[3] = _mm_load_si128((const __m128i *)(input + 1 * stride + 4));
+  in[4] = _mm_load_si128((const __m128i *)(input + 2 * stride + 0));
+  in[5] = _mm_load_si128((const __m128i *)(input + 2 * stride + 4));
+  in[6] = _mm_load_si128((const __m128i *)(input + 3 * stride + 0));
+  in[7] = _mm_load_si128((const __m128i *)(input + 3 * stride + 4));
+  transpose_32bit_8x4(in, in);
+}
+
+static INLINE void highbd_load_transpose_32bit_4x4(const tran_low_t *input,
+                                                   const int stride,
+                                                   __m128i *in) {
+  in[0] = _mm_load_si128((const __m128i *)(input + 0 * stride));
+  in[1] = _mm_load_si128((const __m128i *)(input + 1 * stride));
+  in[2] = _mm_load_si128((const __m128i *)(input + 2 * stride));
+  in[3] = _mm_load_si128((const __m128i *)(input + 3 * stride));
+  transpose_32bit_4x4(in, in);
+}
+
 static INLINE void highbd_write_buffer_8(uint16_t *dest, const __m128i in,
                                          const int bd) {
   const __m128i final_rounding = _mm_set1_epi16(1 << 5);

diff --git a/vpx_dsp/x86/highbd_inv_txfm_sse4.h b/vpx_dsp/x86/highbd_inv_txfm_sse4.h
index 66af7c6996e4fb6dcd49caa2501ecb00b668e7ee..68344ff797e94f11602c3e62c1ce50b176bc1b0e 100644 (file)
--- a/vpx_dsp/x86/highbd_inv_txfm_sse4.h
+++ b/vpx_dsp/x86/highbd_inv_txfm_sse4.h
@@ -73,10 +73,10 @@ static INLINE void highbd_butterfly_cospi16_sse4_1(const __m128i in0,
   *out1 = multiplication_round_shift_sse4_1(temp1, (int)cospi_16_64);
 }
 
-static INLINE void highbd_multiplication_sse4_1(const __m128i in, const int c0,
-                                                const int c1,
-                                                __m128i *const out0,
-                                                __m128i *const out1) {
+static INLINE void highbd_partial_butterfly_sse4_1(const __m128i in,
+                                                   const int c0, const int c1,
+                                                   __m128i *const out0,
+                                                   __m128i *const out1) {
   __m128i temp[2];
 
   extend_64bit(in, temp);