ref: d9fedf7832292b4f8f83d60444064dc5066cb0e1
parent: fa53008fb739f2ce64a391552cc50804321683c1
author: Parag Salasakar <img.mips1@gmail.com>
date: Wed Jun 17 03:53:06 EDT 2015
mips msa vp9 fdct 32x32 optimization average improvement ~4x-6x Change-Id: Ibcac3ef8ed5e207cf8c121e696570e6b63d3c0f4
--- a/test/dct32x32_test.cc
+++ b/test/dct32x32_test.cc
@@ -386,7 +386,9 @@
INSTANTIATE_TEST_CASE_P(
MSA, Trans32x32Test,
::testing::Values(
- make_tuple(&vp9_fdct32x32_c,
- &vp9_idct32x32_1024_add_msa, 0, VPX_BITS_8)));
+ make_tuple(&vp9_fdct32x32_msa,
+ &vp9_idct32x32_1024_add_msa, 0, VPX_BITS_8),
+ make_tuple(&vp9_fdct32x32_rd_msa,
+ &vp9_idct32x32_1024_add_msa, 1, VPX_BITS_8)));
#endif // HAVE_MSA && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
} // namespace
--- a/vp9/common/mips/msa/vp9_macros_msa.h
+++ b/vp9/common/mips/msa/vp9_macros_msa.h
@@ -720,6 +720,24 @@
}
#define DOTP_SH4_SW(...) DOTP_SH4(v4i32, __VA_ARGS__)
+/* Description : Dot product of word vector elements
+ Arguments : Inputs - mult0, mult1
+ cnst0, cnst1
+ Outputs - out0, out1
+ Return Type - signed word
+ Details : Signed word elements from mult0 are multiplied with
+ signed word elements from cnst0 producing a result
+ twice the size of input i.e. signed double word.
+ Then this multiplication results of adjacent odd-even elements
+ are added together and stored to the out vector
+ (2 signed double word results)
+*/
+#define DOTP_SW2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1) { \+ out0 = (RTYPE)__msa_dotp_s_d((v4i32)mult0, (v4i32)cnst0); \
+ out1 = (RTYPE)__msa_dotp_s_d((v4i32)mult1, (v4i32)cnst1); \
+}
+#define DOTP_SW2_SD(...) DOTP_SW2(v2i64, __VA_ARGS__)
+
/* Description : Dot product & addition of byte vector elements
Arguments : Inputs - mult0, mult1
cnst0, cnst1
@@ -1103,7 +1121,7 @@
Return Type - unsigned halfword
Details : Each unsigned halfword element from 'in0' is saturated to the
value generated with (sat_val+1) bit range.
- The results are in placed to original vectors
+ The results are stored in place
*/
#define SAT_UH2(RTYPE, in0, in1, sat_val) { \in0 = (RTYPE)__msa_sat_u_h((v8u16)in0, sat_val); \
@@ -1125,7 +1143,7 @@
Return Type - unsigned halfword
Details : Each unsigned halfword element from 'in0' is saturated to the
value generated with (sat_val+1) bit range
- The results are in placed to original vectors
+ The results are stored in place
*/
#define SAT_SH2(RTYPE, in0, in1, sat_val) { \in0 = (RTYPE)__msa_sat_s_h((v8i16)in0, sat_val); \
@@ -1438,6 +1456,24 @@
v16i8 zero_m = { 0 }; \\
ILVRL_B2_SH(zero_m, in, out0, out1); \
+}
+
+/* Description : Sign extend halfword elements from input vector and return
+ result in pair of vectors
+ Arguments : Inputs - in (1 input halfword vector)
+ Outputs - out0, out1 (sign extended 2 word vectors)
+ Return Type - signed word
+ Details : Sign bit of halfword elements from input vector 'in' is
+ extracted and interleaved right with same vector 'in0' to
+ generate 4 signed word elements in 'out0'
+ Then interleaved left with same vector 'in0' to
+ generate 4 signed word elements in 'out1'
+*/
+#define UNPCK_SH_SW(in, out0, out1) { \+ v8i16 tmp_m; \
+ \
+ tmp_m = __msa_clti_s_h((v8i16)in, 0); \
+ ILVRL_H2_SW(tmp_m, in, out0, out1); \
}
/* Description : Butterfly of 4 input vectors
--- a/vp9/common/vp9_rtcd_defs.pl
+++ b/vp9/common/vp9_rtcd_defs.pl
@@ -1053,13 +1053,13 @@
specialize qw/vp9_fdct16x16 sse2 msa/;
add_proto qw/void vp9_fdct32x32_1/, "const int16_t *input, tran_low_t *output, int stride";
- specialize qw/vp9_fdct32x32_1 sse2/;
+ specialize qw/vp9_fdct32x32_1 sse2 msa/;
add_proto qw/void vp9_fdct32x32/, "const int16_t *input, tran_low_t *output, int stride";
- specialize qw/vp9_fdct32x32 sse2 avx2/;
+ specialize qw/vp9_fdct32x32 sse2 avx2 msa/;
add_proto qw/void vp9_fdct32x32_rd/, "const int16_t *input, tran_low_t *output, int stride";
- specialize qw/vp9_fdct32x32_rd sse2 avx2/;
+ specialize qw/vp9_fdct32x32_rd sse2 avx2 msa/;
}
#
--- /dev/null
+++ b/vp9/encoder/mips/msa/vp9_fdct32x32_msa.c
@@ -1,0 +1,956 @@
+/*
+ * Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ * Use of this source code is governed by a BSD-style license
+ * that can be found in the LICENSE file in the root of the source
+ * tree. An additional intellectual property rights grant can be found
+ * in the file PATENTS. All contributing project authors may
+ * be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vp9_rtcd.h"
+#include "vp9/encoder/mips/msa/vp9_fdct_msa.h"
+
+static void fdct8x32_1d_column_load_butterfly(const int16_t *input,
+ int32_t src_stride,
+ int16_t *temp_buff) {+ v8i16 in0, in1, in2, in3, in4, in5, in6, in7;
+ v8i16 step0, step1, step2, step3;
+ v8i16 in0_1, in1_1, in2_1, in3_1, in4_1, in5_1, in6_1, in7_1;
+ v8i16 step0_1, step1_1, step2_1, step3_1;
+
+ /* 1st and 2nd set */
+ LD_SH4(input, src_stride, in0, in1, in2, in3);
+ LD_SH4(input + (28 * src_stride), src_stride, in4, in5, in6, in7);
+ LD_SH4(input + (4 * src_stride), src_stride, in0_1, in1_1, in2_1, in3_1);
+ LD_SH4(input + (24 * src_stride), src_stride, in4_1, in5_1, in6_1, in7_1);
+ SLLI_4V(in0, in1, in2, in3, 2);
+ SLLI_4V(in4, in5, in6, in7, 2);
+ SLLI_4V(in0_1, in1_1, in2_1, in3_1, 2);
+ SLLI_4V(in4_1, in5_1, in6_1, in7_1, 2);
+ BUTTERFLY_8(in0, in1, in2, in3, in4, in5, in6, in7,
+ step0, step1, step2, step3, in4, in5, in6, in7);
+ BUTTERFLY_8(in0_1, in1_1, in2_1, in3_1, in4_1, in5_1, in6_1, in7_1,
+ step0_1, step1_1, step2_1, step3_1, in4_1, in5_1, in6_1, in7_1);
+ ST_SH4(step0, step1, step2, step3, temp_buff, 8);
+ ST_SH4(in4, in5, in6, in7, temp_buff + (28 * 8), 8);
+ ST_SH4(step0_1, step1_1, step2_1, step3_1, temp_buff + (4 * 8), 8);
+ ST_SH4(in4_1, in5_1, in6_1, in7_1, temp_buff + (24 * 8), 8);
+
+ /* 3rd and 4th set */
+ LD_SH4(input + (8 * src_stride), src_stride, in0, in1, in2, in3);
+ LD_SH4(input + (20 * src_stride), src_stride, in4, in5, in6, in7);
+ LD_SH4(input + (12 * src_stride), src_stride, in0_1, in1_1, in2_1, in3_1);
+ LD_SH4(input + (16 * src_stride), src_stride, in4_1, in5_1, in6_1, in7_1);
+ SLLI_4V(in0, in1, in2, in3, 2);
+ SLLI_4V(in4, in5, in6, in7, 2);
+ SLLI_4V(in0_1, in1_1, in2_1, in3_1, 2);
+ SLLI_4V(in4_1, in5_1, in6_1, in7_1, 2);
+ BUTTERFLY_8(in0, in1, in2, in3, in4, in5, in6, in7,
+ step0, step1, step2, step3, in4, in5, in6, in7);
+ BUTTERFLY_8(in0_1, in1_1, in2_1, in3_1, in4_1, in5_1, in6_1, in7_1,
+ step0_1, step1_1, step2_1, step3_1, in4_1, in5_1, in6_1, in7_1);
+ ST_SH4(step0, step1, step2, step3, temp_buff + (8 * 8), 8);
+ ST_SH4(in4, in5, in6, in7, temp_buff + (20 * 8), 8);
+ ST_SH4(step0_1, step1_1, step2_1, step3_1, temp_buff + (12 * 8), 8);
+ ST_SH4(in4_1, in5_1, in6_1, in7_1, temp_buff + (15 * 8) + 8, 8);
+}
+
+static void fdct8x32_1d_column_even_store(int16_t *input, int16_t *temp) {+ v8i16 in0, in1, in2, in3, in4, in5, in6, in7;
+ v8i16 in8, in9, in10, in11, in12, in13, in14, in15;
+ v8i16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+ v8i16 temp0, temp1;
+
+ /* fdct even */
+ LD_SH4(input, 8, in0, in1, in2, in3);
+ LD_SH4(input + 96, 8, in12, in13, in14, in15);
+ BUTTERFLY_8(in0, in1, in2, in3, in12, in13, in14, in15,
+ vec0, vec1, vec2, vec3, in12, in13, in14, in15);
+ LD_SH4(input + 32, 8, in4, in5, in6, in7);
+ LD_SH4(input + 64, 8, in8, in9, in10, in11);
+ BUTTERFLY_8(in4, in5, in6, in7, in8, in9, in10, in11,
+ vec4, vec5, vec6, vec7, in8, in9, in10, in11);
+
+ /* Stage 3 */
+ ADD4(vec0, vec7, vec1, vec6, vec2, vec5, vec3, vec4, in0, in1, in2, in3);
+ BUTTERFLY_4(in0, in1, in2, in3, temp0, in4, in1, in0);
+ VP9_DOTP_CONST_PAIR(temp0, in4, cospi_16_64, cospi_16_64, temp1, temp0);
+ VP9_FDCT32_POSTPROC_2V_POS_H(temp0, temp1);
+ ST_SH(temp0, temp);
+ ST_SH(temp1, temp + 512);
+
+ VP9_DOTP_CONST_PAIR(in0, in1, cospi_24_64, cospi_8_64, temp1, temp0);
+ VP9_FDCT32_POSTPROC_2V_POS_H(temp0, temp1);
+ ST_SH(temp0, temp + 256);
+ ST_SH(temp1, temp + 768);
+
+ SUB4(vec0, vec7, vec1, vec6, vec2, vec5, vec3, vec4, vec7, vec6, vec5, vec4);
+ VP9_DOTP_CONST_PAIR(vec6, vec5, cospi_16_64, cospi_16_64, vec5, vec6);
+ ADD2(vec4, vec5, vec7, vec6, vec0, vec1);
+ VP9_DOTP_CONST_PAIR(vec1, vec0, cospi_28_64, cospi_4_64, temp1, temp0);
+ VP9_FDCT32_POSTPROC_2V_POS_H(temp0, temp1);
+ ST_SH(temp0, temp + 128);
+ ST_SH(temp1, temp + 896);
+
+ SUB2(vec4, vec5, vec7, vec6, vec4, vec7);
+ VP9_DOTP_CONST_PAIR(vec7, vec4, cospi_12_64, cospi_20_64, temp1, temp0);
+ VP9_FDCT32_POSTPROC_2V_POS_H(temp0, temp1);
+ ST_SH(temp0, temp + 640);
+ ST_SH(temp1, temp + 384);
+
+ VP9_DOTP_CONST_PAIR(in13, in10, cospi_16_64, cospi_16_64, vec2, vec5);
+ VP9_DOTP_CONST_PAIR(in12, in11, cospi_16_64, cospi_16_64, vec3, vec4);
+ ADD4(in8, vec3, in9, vec2, in14, vec5, in15, vec4, in0, vec1, vec6, in2);
+ VP9_DOTP_CONST_PAIR(vec6, vec1, cospi_24_64, cospi_8_64, in1, in3);
+ ADD2(in0, in1, in2, in3, vec0, vec7);
+ VP9_DOTP_CONST_PAIR(vec7, vec0, cospi_30_64, cospi_2_64, temp1, temp0);
+ VP9_FDCT32_POSTPROC_2V_POS_H(temp0, temp1);
+ ST_SH(temp0, temp + 64);
+ ST_SH(temp1, temp + 960);
+
+ SUB2(in0, in1, in2, in3, in0, in2);
+ VP9_DOTP_CONST_PAIR(in2, in0, cospi_14_64, cospi_18_64, temp1, temp0);
+ VP9_FDCT32_POSTPROC_2V_POS_H(temp0, temp1);
+ ST_SH(temp0, temp + 576);
+ ST_SH(temp1, temp + 448);
+
+ SUB2(in9, vec2, in14, vec5, vec2, vec5);
+ VP9_DOTP_CONST_PAIR((-vec2), vec5, cospi_24_64, cospi_8_64, in2, in1);
+ SUB4(in8, vec3, in15, vec4, in3, in2, in0, in1, in3, in0, vec2, vec5);
+ VP9_DOTP_CONST_PAIR(vec5, vec2, cospi_22_64, cospi_10_64, temp1, temp0);
+ VP9_FDCT32_POSTPROC_2V_POS_H(temp0, temp1);
+ ST_SH(temp0, temp + 320);
+ ST_SH(temp1, temp + 704);
+
+ ADD2(in3, in2, in0, in1, vec3, vec4);
+ VP9_DOTP_CONST_PAIR(vec4, vec3, cospi_6_64, cospi_26_64, temp0, temp1);
+ VP9_FDCT32_POSTPROC_2V_POS_H(temp0, temp1);
+ ST_SH(temp0, temp + 192);
+ ST_SH(temp1, temp + 832);
+}
+
+static void fdct8x32_1d_column_odd_store(int16_t *input, int16_t *temp_ptr) {+ v8i16 in16, in17, in18, in19, in20, in21, in22, in23;
+ v8i16 in24, in25, in26, in27, in28, in29, in30, in31, vec4, vec5;
+
+ in20 = LD_SH(input + 32);
+ in21 = LD_SH(input + 40);
+ in26 = LD_SH(input + 80);
+ in27 = LD_SH(input + 88);
+
+ VP9_DOTP_CONST_PAIR(in27, in20, cospi_16_64, cospi_16_64, in20, in27);
+ VP9_DOTP_CONST_PAIR(in26, in21, cospi_16_64, cospi_16_64, in21, in26);
+
+ in18 = LD_SH(input + 16);
+ in19 = LD_SH(input + 24);
+ in28 = LD_SH(input + 96);
+ in29 = LD_SH(input + 104);
+
+ vec4 = in19 - in20;
+ ST_SH(vec4, input + 32);
+ vec4 = in18 - in21;
+ ST_SH(vec4, input + 40);
+ vec4 = in29 - in26;
+ ST_SH(vec4, input + 80);
+ vec4 = in28 - in27;
+ ST_SH(vec4, input + 88);
+
+ in21 = in18 + in21;
+ in20 = in19 + in20;
+ in27 = in28 + in27;
+ in26 = in29 + in26;
+
+ LD_SH4(input + 48, 8, in22, in23, in24, in25);
+ VP9_DOTP_CONST_PAIR(in25, in22, cospi_16_64, cospi_16_64, in22, in25);
+ VP9_DOTP_CONST_PAIR(in24, in23, cospi_16_64, cospi_16_64, in23, in24);
+
+ in16 = LD_SH(input);
+ in17 = LD_SH(input + 8);
+ in30 = LD_SH(input + 112);
+ in31 = LD_SH(input + 120);
+
+ vec4 = in17 - in22;
+ ST_SH(vec4, input + 16);
+ vec4 = in16 - in23;
+ ST_SH(vec4, input + 24);
+ vec4 = in31 - in24;
+ ST_SH(vec4, input + 96);
+ vec4 = in30 - in25;
+ ST_SH(vec4, input + 104);
+
+ ADD4(in16, in23, in17, in22, in30, in25, in31, in24, in16, in17, in30, in31);
+ VP9_DOTP_CONST_PAIR(in26, in21, cospi_24_64, cospi_8_64, in18, in29);
+ VP9_DOTP_CONST_PAIR(in27, in20, cospi_24_64, cospi_8_64, in19, in28);
+ ADD4(in16, in19, in17, in18, in30, in29, in31, in28, in27, in22, in21, in25);
+ VP9_DOTP_CONST_PAIR(in21, in22, cospi_28_64, cospi_4_64, in26, in24);
+ ADD2(in27, in26, in25, in24, in23, in20);
+ VP9_DOTP_CONST_PAIR(in20, in23, cospi_31_64, cospi_1_64, vec4, vec5);
+ VP9_FDCT32_POSTPROC_2V_POS_H(vec5, vec4);
+ ST_SH(vec5, temp_ptr);
+ ST_SH(vec4, temp_ptr + 960);
+
+ SUB2(in27, in26, in25, in24, in22, in21);
+ VP9_DOTP_CONST_PAIR(in21, in22, cospi_15_64, cospi_17_64, vec5, vec4);
+ VP9_FDCT32_POSTPROC_2V_POS_H(vec5, vec4);
+ ST_SH(vec5, temp_ptr + 448);
+ ST_SH(vec4, temp_ptr + 512);
+
+ SUB4(in17, in18, in16, in19, in31, in28, in30, in29, in23, in26, in24, in20);
+ VP9_DOTP_CONST_PAIR((-in23), in20, cospi_28_64, cospi_4_64, in27, in25);
+ SUB2(in26, in27, in24, in25, in23, in20);
+ VP9_DOTP_CONST_PAIR(in20, in23, cospi_23_64, cospi_9_64, vec4, vec5);
+ VP9_FDCT32_POSTPROC_2V_POS_H(vec5, vec4);
+ ST_SH(vec4, temp_ptr + 704);
+ ST_SH(vec5, temp_ptr + 256);
+
+ ADD2(in26, in27, in24, in25, in22, in21);
+ VP9_DOTP_CONST_PAIR(in21, in22, cospi_7_64, cospi_25_64, vec4, vec5);
+ VP9_FDCT32_POSTPROC_2V_POS_H(vec5, vec4);
+ ST_SH(vec4, temp_ptr + 192);
+ ST_SH(vec5, temp_ptr + 768);
+
+ LD_SH4(input + 16, 8, in22, in23, in20, in21);
+ LD_SH4(input + 80, 8, in26, in27, in24, in25);
+ in16 = in20;
+ in17 = in21;
+ VP9_DOTP_CONST_PAIR(-in16, in27, cospi_24_64, cospi_8_64, in20, in27);
+ VP9_DOTP_CONST_PAIR(-in17, in26, cospi_24_64, cospi_8_64, in21, in26);
+ SUB4(in23, in20, in22, in21, in25, in26, in24, in27, in28, in17, in18, in31);
+ VP9_DOTP_CONST_PAIR(in18, in17, cospi_12_64, cospi_20_64, in29, in30);
+ ADD2(in28, in29, in31, in30, in16, in19);
+ VP9_DOTP_CONST_PAIR(in19, in16, cospi_27_64, cospi_5_64, vec5, vec4);
+ VP9_FDCT32_POSTPROC_2V_POS_H(vec5, vec4);
+ ST_SH(vec5, temp_ptr + 832);
+ ST_SH(vec4, temp_ptr + 128);
+
+ SUB2(in28, in29, in31, in30, in17, in18);
+ VP9_DOTP_CONST_PAIR(in18, in17, cospi_11_64, cospi_21_64, vec5, vec4);
+ VP9_FDCT32_POSTPROC_2V_POS_H(vec5, vec4);
+ ST_SH(vec5, temp_ptr + 320);
+ ST_SH(vec4, temp_ptr + 640);
+ ADD4(in22, in21, in23, in20, in24, in27, in25, in26, in16, in29, in30, in19);
+ VP9_DOTP_CONST_PAIR(-in16, in19, cospi_12_64, cospi_20_64, in28, in31);
+ SUB2(in29, in28, in30, in31, in16, in19);
+ VP9_DOTP_CONST_PAIR(in19, in16, cospi_19_64, cospi_13_64, vec5, vec4);
+ VP9_FDCT32_POSTPROC_2V_POS_H(vec5, vec4);
+ ST_SH(vec5, temp_ptr + 576);
+ ST_SH(vec4, temp_ptr + 384);
+
+ ADD2(in29, in28, in30, in31, in17, in18);
+ VP9_DOTP_CONST_PAIR(in18, in17, cospi_3_64, cospi_29_64, vec5, vec4);
+ VP9_FDCT32_POSTPROC_2V_POS_H(vec5, vec4);
+ ST_SH(vec5, temp_ptr + 64);
+ ST_SH(vec4, temp_ptr + 896);
+}
+
+static void fdct8x32_1d_column(const int16_t *input, int32_t src_stride,
+ int16_t *tmp_buf, int16_t *tmp_buf_big) {+ fdct8x32_1d_column_load_butterfly(input, src_stride, tmp_buf);
+ fdct8x32_1d_column_even_store(tmp_buf, tmp_buf_big);
+ fdct8x32_1d_column_odd_store(tmp_buf + 128, (tmp_buf_big + 32));
+}
+
+static void fdct8x32_1d_row_load_butterfly(int16_t *temp_buff,
+ int16_t *output) {+ v8i16 in0, in1, in2, in3, in4, in5, in6, in7;
+ v8i16 in8, in9, in10, in11, in12, in13, in14, in15;
+ v8i16 step0, step1, step2, step3, step4, step5, step6, step7;
+
+ LD_SH8(temp_buff, 32, in0, in1, in2, in3, in4, in5, in6, in7);
+ LD_SH8(temp_buff + 24, 32, in8, in9, in10, in11, in12, in13, in14, in15);
+ TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7,
+ in0, in1, in2, in3, in4, in5, in6, in7);
+ TRANSPOSE8x8_SH_SH(in8, in9, in10, in11, in12, in13, in14, in15,
+ in8, in9, in10, in11, in12, in13, in14, in15);
+ BUTTERFLY_16(in0, in1, in2, in3, in4, in5, in6, in7,
+ in8, in9, in10, in11, in12, in13, in14, in15,
+ step0, step1, step2, step3, step4, step5, step6, step7,
+ in8, in9, in10, in11, in12, in13, in14, in15);
+ ST_SH8(step0, step1, step2, step3, step4, step5, step6, step7, output, 8);
+ ST_SH8(in8, in9, in10, in11, in12, in13, in14, in15, (output + 24 * 8), 8);
+
+ /* 2nd set */
+ LD_SH8(temp_buff + 8, 32, in0, in1, in2, in3, in4, in5, in6, in7);
+ LD_SH8(temp_buff + 16, 32, in8, in9, in10, in11, in12, in13, in14, in15);
+ TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7,
+ in0, in1, in2, in3, in4, in5, in6, in7);
+ TRANSPOSE8x8_SH_SH(in8, in9, in10, in11, in12, in13, in14, in15,
+ in8, in9, in10, in11, in12, in13, in14, in15);
+ BUTTERFLY_16(in0, in1, in2, in3, in4, in5, in6, in7,
+ in8, in9, in10, in11, in12, in13, in14, in15,
+ step0, step1, step2, step3, step4, step5, step6, step7,
+ in8, in9, in10, in11, in12, in13, in14, in15);
+ ST_SH8(step0, step1, step2, step3, step4, step5, step6, step7,
+ (output + 8 * 8), 8);
+ ST_SH8(in8, in9, in10, in11, in12, in13, in14, in15, (output + 16 * 8), 8);
+}
+
+static void fdct8x32_1d_row_even_4x(int16_t *input, int16_t *interm_ptr,
+ int16_t *out) {+ v8i16 in0, in1, in2, in3, in4, in5, in6, in7;
+ v8i16 in8, in9, in10, in11, in12, in13, in14, in15;
+ v8i16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+ v4i32 vec0_l, vec1_l, vec2_l, vec3_l, vec4_l, vec5_l, vec6_l, vec7_l;
+ v4i32 vec0_r, vec1_r, vec2_r, vec3_r, vec4_r, vec5_r, vec6_r, vec7_r;
+ v4i32 tmp0_w, tmp1_w, tmp2_w, tmp3_w;
+
+ /* fdct32 even */
+ /* stage 2 */
+ LD_SH8(input, 8, in0, in1, in2, in3, in4, in5, in6, in7);
+ LD_SH8(input + 64, 8, in8, in9, in10, in11, in12, in13, in14, in15);
+
+ BUTTERFLY_16(in0, in1, in2, in3, in4, in5, in6, in7,
+ in8, in9, in10, in11, in12, in13, in14, in15,
+ vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7,
+ in8, in9, in10, in11, in12, in13, in14, in15);
+ ST_SH8(vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, interm_ptr, 8);
+ ST_SH8(in8, in9, in10, in11, in12, in13, in14, in15, interm_ptr + 64, 8);
+
+ /* Stage 3 */
+ UNPCK_SH_SW(vec0, vec0_l, vec0_r);
+ UNPCK_SH_SW(vec1, vec1_l, vec1_r);
+ UNPCK_SH_SW(vec2, vec2_l, vec2_r);
+ UNPCK_SH_SW(vec3, vec3_l, vec3_r);
+ UNPCK_SH_SW(vec4, vec4_l, vec4_r);
+ UNPCK_SH_SW(vec5, vec5_l, vec5_r);
+ UNPCK_SH_SW(vec6, vec6_l, vec6_r);
+ UNPCK_SH_SW(vec7, vec7_l, vec7_r);
+ ADD4(vec0_r, vec7_r, vec1_r, vec6_r, vec2_r, vec5_r, vec3_r, vec4_r,
+ tmp0_w, tmp1_w, tmp2_w, tmp3_w);
+ BUTTERFLY_4(tmp0_w, tmp1_w, tmp2_w, tmp3_w, vec4_r, vec6_r, vec7_r, vec5_r);
+ ADD4(vec0_l, vec7_l, vec1_l, vec6_l, vec2_l, vec5_l, vec3_l, vec4_l,
+ vec0_r, vec1_r, vec2_r, vec3_r);
+
+ tmp3_w = vec0_r + vec3_r;
+ vec0_r = vec0_r - vec3_r;
+ vec3_r = vec1_r + vec2_r;
+ vec1_r = vec1_r - vec2_r;
+
+ VP9_DOTP_CONST_PAIR_W(vec4_r, vec6_r, tmp3_w, vec3_r, cospi_16_64,
+ cospi_16_64, vec4_r, tmp3_w, vec6_r, vec3_r);
+ VP9_FDCT32_POSTPROC_NEG_W(vec4_r);
+ VP9_FDCT32_POSTPROC_NEG_W(tmp3_w);
+ VP9_FDCT32_POSTPROC_NEG_W(vec6_r);
+ VP9_FDCT32_POSTPROC_NEG_W(vec3_r);
+ PCKEV_H2_SH(vec4_r, tmp3_w, vec6_r, vec3_r, vec4, vec5);
+ ST_SH2(vec5, vec4, out, 8);
+
+ VP9_DOTP_CONST_PAIR_W(vec5_r, vec7_r, vec0_r, vec1_r, cospi_24_64,
+ cospi_8_64, vec4_r, tmp3_w, vec6_r, vec3_r);
+ VP9_FDCT32_POSTPROC_NEG_W(vec4_r);
+ VP9_FDCT32_POSTPROC_NEG_W(tmp3_w);
+ VP9_FDCT32_POSTPROC_NEG_W(vec6_r);
+ VP9_FDCT32_POSTPROC_NEG_W(vec3_r);
+ PCKEV_H2_SH(vec4_r, tmp3_w, vec6_r, vec3_r, vec4, vec5);
+ ST_SH2(vec5, vec4, out + 16, 8);
+
+ LD_SH8(interm_ptr, 8, vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7);
+ SUB4(vec3, vec4, vec2, vec5, vec1, vec6, vec0, vec7, vec4, vec5, vec6, vec7);
+ VP9_DOTP_CONST_PAIR(vec6, vec5, cospi_16_64, cospi_16_64, vec5, vec6);
+ ADD2(vec4, vec5, vec7, vec6, vec0, vec1);
+ VP9_DOTP_CONST_PAIR(vec1, vec0, cospi_28_64, cospi_4_64, in5, in4);
+ VP9_FDCT_POSTPROC_2V_NEG_H(in4, in5);
+ ST_SH(in4, out + 32);
+ ST_SH(in5, out + 56);
+
+ SUB2(vec4, vec5, vec7, vec6, vec4, vec7);
+ VP9_DOTP_CONST_PAIR(vec7, vec4, cospi_12_64, cospi_20_64, in5, in4);
+ VP9_FDCT_POSTPROC_2V_NEG_H(in4, in5);
+ ST_SH(in4, out + 40);
+ ST_SH(in5, out + 48);
+
+ LD_SH8(interm_ptr + 64, 8, in8, in9, in10, in11, in12, in13, in14, in15);
+ VP9_DOTP_CONST_PAIR(in13, in10, cospi_16_64, cospi_16_64, vec2, vec5);
+ VP9_DOTP_CONST_PAIR(in12, in11, cospi_16_64, cospi_16_64, vec3, vec4);
+ ADD4(in8, vec3, in9, vec2, in14, vec5, in15, vec4, in0, vec1, vec6, in2);
+ VP9_DOTP_CONST_PAIR(vec6, vec1, cospi_24_64, cospi_8_64, in1, in3);
+ ADD2(in0, in1, in2, in3, vec0, vec7);
+ VP9_DOTP_CONST_PAIR(vec7, vec0, cospi_30_64, cospi_2_64, in5, in4);
+ VP9_FDCT_POSTPROC_2V_NEG_H(in4, in5);
+ ST_SH(in4, out + 64);
+ ST_SH(in5, out + 120);
+
+ SUB2(in0, in1, in2, in3, in0, in2);
+ VP9_DOTP_CONST_PAIR(in2, in0, cospi_14_64, cospi_18_64, in5, in4);
+ VP9_FDCT_POSTPROC_2V_NEG_H(in4, in5);
+ ST_SH(in4, out + 72);
+ ST_SH(in5, out + 112);
+
+ SUB2(in9, vec2, in14, vec5, vec2, vec5);
+ VP9_DOTP_CONST_PAIR((-vec2), vec5, cospi_24_64, cospi_8_64, in2, in1);
+ SUB4(in8, vec3, in15, vec4, in3, in2, in0, in1, in3, in0, vec2, vec5);
+ VP9_DOTP_CONST_PAIR(vec5, vec2, cospi_22_64, cospi_10_64, in5, in4);
+ VP9_FDCT_POSTPROC_2V_NEG_H(in4, in5);
+ ST_SH(in4, out + 80);
+ ST_SH(in5, out + 104);
+
+ ADD2(in3, in2, in0, in1, vec3, vec4);
+ VP9_DOTP_CONST_PAIR(vec4, vec3, cospi_6_64, cospi_26_64, in4, in5);
+ VP9_FDCT_POSTPROC_2V_NEG_H(in4, in5);
+ ST_SH(in4, out + 96);
+ ST_SH(in5, out + 88);
+}
+
+static void fdct8x32_1d_row_even(int16_t *temp, int16_t *out) {+ v8i16 in0, in1, in2, in3, in4, in5, in6, in7;
+ v8i16 in8, in9, in10, in11, in12, in13, in14, in15;
+ v8i16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, temp0, temp1;
+
+ /* fdct32 even */
+ /* stage 2 */
+ LD_SH8(temp, 8, in0, in1, in2, in3, in4, in5, in6, in7);
+ LD_SH8(temp + 64, 8, in8, in9, in10, in11, in12, in13, in14, in15);
+
+ BUTTERFLY_16(in0, in1, in2, in3, in4, in5, in6, in7,
+ in8, in9, in10, in11, in12, in13, in14, in15,
+ vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7,
+ in8, in9, in10, in11, in12, in13, in14, in15);
+
+ /* Stage 3 */
+ ADD4(vec0, vec7, vec1, vec6, vec2, vec5, vec3, vec4, in0, in1, in2, in3);
+ BUTTERFLY_4(in0, in1, in2, in3, temp0, in4, in1, in0);
+ VP9_DOTP_CONST_PAIR(temp0, in4, cospi_16_64, cospi_16_64, temp1, temp0);
+ VP9_FDCT_POSTPROC_2V_NEG_H(temp0, temp1);
+ ST_SH(temp0, out);
+ ST_SH(temp1, out + 8);
+
+ VP9_DOTP_CONST_PAIR(in0, in1, cospi_24_64, cospi_8_64, temp1, temp0);
+ VP9_FDCT_POSTPROC_2V_NEG_H(temp0, temp1);
+ ST_SH(temp0, out + 16);
+ ST_SH(temp1, out + 24);
+
+ SUB4(vec3, vec4, vec2, vec5, vec1, vec6, vec0, vec7, vec4, vec5, vec6, vec7);
+ VP9_DOTP_CONST_PAIR(vec6, vec5, cospi_16_64, cospi_16_64, vec5, vec6);
+ ADD2(vec4, vec5, vec7, vec6, vec0, vec1);
+ VP9_DOTP_CONST_PAIR(vec1, vec0, cospi_28_64, cospi_4_64, temp1, temp0);
+ VP9_FDCT_POSTPROC_2V_NEG_H(temp0, temp1);
+ ST_SH(temp0, out + 32);
+ ST_SH(temp1, out + 56);
+
+ SUB2(vec4, vec5, vec7, vec6, vec4, vec7);
+ VP9_DOTP_CONST_PAIR(vec7, vec4, cospi_12_64, cospi_20_64, temp1, temp0);
+ VP9_FDCT_POSTPROC_2V_NEG_H(temp0, temp1);
+ ST_SH(temp0, out + 40);
+ ST_SH(temp1, out + 48);
+
+ VP9_DOTP_CONST_PAIR(in13, in10, cospi_16_64, cospi_16_64, vec2, vec5);
+ VP9_DOTP_CONST_PAIR(in12, in11, cospi_16_64, cospi_16_64, vec3, vec4);
+ ADD4(in8, vec3, in9, vec2, in14, vec5, in15, vec4, in0, vec1, vec6, in2);
+ VP9_DOTP_CONST_PAIR(vec6, vec1, cospi_24_64, cospi_8_64, in1, in3);
+ ADD2(in0, in1, in2, in3, vec0, vec7);
+ VP9_DOTP_CONST_PAIR(vec7, vec0, cospi_30_64, cospi_2_64, temp1, temp0);
+ VP9_FDCT_POSTPROC_2V_NEG_H(temp0, temp1);
+ ST_SH(temp0, out + 64);
+ ST_SH(temp1, out + 120);
+
+ SUB2(in0, in1, in2, in3, in0, in2);
+ VP9_DOTP_CONST_PAIR(in2, in0, cospi_14_64, cospi_18_64, temp1, temp0);
+ VP9_FDCT_POSTPROC_2V_NEG_H(temp0, temp1);
+ ST_SH(temp0, out + 72);
+ ST_SH(temp1, out + 112);
+
+ SUB2(in9, vec2, in14, vec5, vec2, vec5);
+ VP9_DOTP_CONST_PAIR((-vec2), vec5, cospi_24_64, cospi_8_64, in2, in1);
+ SUB4(in8, vec3, in15, vec4, in3, in2, in0, in1, in3, in0, vec2, vec5)
+ VP9_DOTP_CONST_PAIR(vec5, vec2, cospi_22_64, cospi_10_64, temp1, temp0);
+ VP9_FDCT_POSTPROC_2V_NEG_H(temp0, temp1);
+ ST_SH(temp0, out + 80);
+ ST_SH(temp1, out + 104);
+
+ ADD2(in3, in2, in0, in1, vec3, vec4);
+ VP9_DOTP_CONST_PAIR(vec4, vec3, cospi_6_64, cospi_26_64, temp0, temp1);
+ VP9_FDCT_POSTPROC_2V_NEG_H(temp0, temp1);
+ ST_SH(temp0, out + 96);
+ ST_SH(temp1, out + 88);
+}
+
+static void fdct8x32_1d_row_odd(int16_t *temp, int16_t *interm_ptr,
+ int16_t *out) {+ v8i16 in16, in17, in18, in19, in20, in21, in22, in23;
+ v8i16 in24, in25, in26, in27, in28, in29, in30, in31, vec4, vec5;
+
+ in20 = LD_SH(temp + 32);
+ in21 = LD_SH(temp + 40);
+ in26 = LD_SH(temp + 80);
+ in27 = LD_SH(temp + 88);
+
+ VP9_DOTP_CONST_PAIR(in27, in20, cospi_16_64, cospi_16_64, in20, in27);
+ VP9_DOTP_CONST_PAIR(in26, in21, cospi_16_64, cospi_16_64, in21, in26);
+
+ in18 = LD_SH(temp + 16);
+ in19 = LD_SH(temp + 24);
+ in28 = LD_SH(temp + 96);
+ in29 = LD_SH(temp + 104);
+
+ vec4 = in19 - in20;
+ ST_SH(vec4, interm_ptr + 32);
+ vec4 = in18 - in21;
+ ST_SH(vec4, interm_ptr + 88);
+ vec4 = in28 - in27;
+ ST_SH(vec4, interm_ptr + 56);
+ vec4 = in29 - in26;
+ ST_SH(vec4, interm_ptr + 64);
+
+ ADD4(in18, in21, in19, in20, in28, in27, in29, in26, in21, in20, in27, in26);
+
+ in22 = LD_SH(temp + 48);
+ in23 = LD_SH(temp + 56);
+ in24 = LD_SH(temp + 64);
+ in25 = LD_SH(temp + 72);
+
+ VP9_DOTP_CONST_PAIR(in25, in22, cospi_16_64, cospi_16_64, in22, in25);
+ VP9_DOTP_CONST_PAIR(in24, in23, cospi_16_64, cospi_16_64, in23, in24);
+
+ in16 = LD_SH(temp);
+ in17 = LD_SH(temp + 8);
+ in30 = LD_SH(temp + 112);
+ in31 = LD_SH(temp + 120);
+
+ vec4 = in17 - in22;
+ ST_SH(vec4, interm_ptr + 40);
+ vec4 = in30 - in25;
+ ST_SH(vec4, interm_ptr + 48);
+ vec4 = in31 - in24;
+ ST_SH(vec4, interm_ptr + 72);
+ vec4 = in16 - in23;
+ ST_SH(vec4, interm_ptr + 80);
+
+ ADD4(in16, in23, in17, in22, in30, in25, in31, in24, in16, in17, in30, in31);
+ VP9_DOTP_CONST_PAIR(in26, in21, cospi_24_64, cospi_8_64, in18, in29);
+ VP9_DOTP_CONST_PAIR(in27, in20, cospi_24_64, cospi_8_64, in19, in28);
+
+ ADD4(in16, in19, in17, in18, in30, in29, in31, in28, in27, in22, in21, in25);
+ VP9_DOTP_CONST_PAIR(in21, in22, cospi_28_64, cospi_4_64, in26, in24);
+ ADD2(in27, in26, in25, in24, in23, in20);
+
+ VP9_DOTP_CONST_PAIR(in20, in23, cospi_31_64, cospi_1_64, vec4, vec5);
+ VP9_FDCT_POSTPROC_2V_NEG_H(vec5, vec4);
+ ST_SH(vec5, out);
+ ST_SH(vec4, out + 120);
+
+ SUB2(in27, in26, in25, in24, in22, in21);
+
+ VP9_DOTP_CONST_PAIR(in21, in22, cospi_15_64, cospi_17_64, vec5, vec4);
+ VP9_FDCT_POSTPROC_2V_NEG_H(vec5, vec4);
+ ST_SH(vec5, out + 112);
+ ST_SH(vec4, out + 8);
+
+ SUB4(in17, in18, in16, in19, in31, in28, in30, in29, in23, in26, in24, in20);
+ VP9_DOTP_CONST_PAIR((-in23), in20, cospi_28_64, cospi_4_64, in27, in25);
+ SUB2(in26, in27, in24, in25, in23, in20);
+
+ VP9_DOTP_CONST_PAIR(in20, in23, cospi_23_64, cospi_9_64, vec4, vec5);
+ VP9_FDCT_POSTPROC_2V_NEG_H(vec5, vec4);
+ ST_SH(vec4, out + 16);
+ ST_SH(vec5, out + 104);
+
+ ADD2(in26, in27, in24, in25, in22, in21);
+ VP9_DOTP_CONST_PAIR(in21, in22, cospi_7_64, cospi_25_64, vec4, vec5);
+ VP9_FDCT_POSTPROC_2V_NEG_H(vec5, vec4);
+ ST_SH(vec4, out + 24);
+ ST_SH(vec5, out + 96);
+
+ in20 = LD_SH(interm_ptr + 32);
+ in21 = LD_SH(interm_ptr + 88);
+ in27 = LD_SH(interm_ptr + 56);
+ in26 = LD_SH(interm_ptr + 64);
+
+ in16 = in20;
+ in17 = in21;
+ VP9_DOTP_CONST_PAIR(-in16, in27, cospi_24_64, cospi_8_64, in20, in27);
+ VP9_DOTP_CONST_PAIR(-in17, in26, cospi_24_64, cospi_8_64, in21, in26);
+
+ in22 = LD_SH(interm_ptr + 40);
+ in25 = LD_SH(interm_ptr + 48);
+ in24 = LD_SH(interm_ptr + 72);
+ in23 = LD_SH(interm_ptr + 80);
+
+ SUB4(in23, in20, in22, in21, in25, in26, in24, in27, in28, in17, in18, in31);
+ VP9_DOTP_CONST_PAIR(in18, in17, cospi_12_64, cospi_20_64, in29, in30);
+ ADD2(in28, in29, in31, in30, in16, in19);
+ VP9_DOTP_CONST_PAIR(in19, in16, cospi_27_64, cospi_5_64, vec5, vec4);
+ VP9_FDCT_POSTPROC_2V_NEG_H(vec5, vec4);
+ ST_SH(vec5, out + 32);
+ ST_SH(vec4, out + 88);
+
+ SUB2(in28, in29, in31, in30, in17, in18);
+ VP9_DOTP_CONST_PAIR(in18, in17, cospi_11_64, cospi_21_64, vec5, vec4);
+ VP9_FDCT_POSTPROC_2V_NEG_H(vec5, vec4);
+ ST_SH(vec5, out + 40);
+ ST_SH(vec4, out + 80);
+
+ ADD4(in22, in21, in23, in20, in24, in27, in25, in26, in16, in29, in30, in19);
+ VP9_DOTP_CONST_PAIR(-in16, in19, cospi_12_64, cospi_20_64, in28, in31);
+ SUB2(in29, in28, in30, in31, in16, in19);
+
+ VP9_DOTP_CONST_PAIR(in19, in16, cospi_19_64, cospi_13_64, vec5, vec4);
+ VP9_FDCT_POSTPROC_2V_NEG_H(vec5, vec4);
+ ST_SH(vec5, out + 72);
+ ST_SH(vec4, out + 48);
+
+ ADD2(in29, in28, in30, in31, in17, in18);
+
+ VP9_DOTP_CONST_PAIR(in18, in17, cospi_3_64, cospi_29_64, vec5, vec4);
+ VP9_FDCT_POSTPROC_2V_NEG_H(vec5, vec4);
+ ST_SH(vec4, out + 56);
+ ST_SH(vec5, out + 64);
+}
+
+static void fdct8x32_1d_row_transpose_store(int16_t *temp, int16_t *output) {+ v8i16 in0, in1, in2, in3, in4, in5, in6, in7;
+ v8i16 in0_1, in1_1, in2_1, in3_1, in4_1, in5_1, in6_1, in7_1;
+
+ /* 1st set */
+ in0 = LD_SH(temp);
+ in4 = LD_SH(temp + 32);
+ in2 = LD_SH(temp + 64);
+ in6 = LD_SH(temp + 96);
+ in1 = LD_SH(temp + 128);
+ in7 = LD_SH(temp + 152);
+ in3 = LD_SH(temp + 192);
+ in5 = LD_SH(temp + 216);
+
+ TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7,
+ in0, in1, in2, in3, in4, in5, in6, in7);
+
+ /* 2nd set */
+ in0_1 = LD_SH(temp + 16);
+ in1_1 = LD_SH(temp + 232);
+ in2_1 = LD_SH(temp + 80);
+ in3_1 = LD_SH(temp + 168);
+ in4_1 = LD_SH(temp + 48);
+ in5_1 = LD_SH(temp + 176);
+ in6_1 = LD_SH(temp + 112);
+ in7_1 = LD_SH(temp + 240);
+
+ ST_SH8(in0, in1, in2, in3, in4, in5, in6, in7, output, 32);
+ TRANSPOSE8x8_SH_SH(in0_1, in1_1, in2_1, in3_1, in4_1, in5_1, in6_1, in7_1,
+ in0_1, in1_1, in2_1, in3_1, in4_1, in5_1, in6_1, in7_1);
+
+ /* 3rd set */
+ in0 = LD_SH(temp + 8);
+ in1 = LD_SH(temp + 136);
+ in2 = LD_SH(temp + 72);
+ in3 = LD_SH(temp + 200);
+ in4 = LD_SH(temp + 40);
+ in5 = LD_SH(temp + 208);
+ in6 = LD_SH(temp + 104);
+ in7 = LD_SH(temp + 144);
+
+ ST_SH8(in0_1, in1_1, in2_1, in3_1, in4_1, in5_1, in6_1, in7_1,
+ output + 8, 32);
+ TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7,
+ in0, in1, in2, in3, in4, in5, in6, in7);
+ ST_SH8(in0, in1, in2, in3, in4, in5, in6, in7, output + 16, 32);
+
+ /* 4th set */
+ in0_1 = LD_SH(temp + 24);
+ in1_1 = LD_SH(temp + 224);
+ in2_1 = LD_SH(temp + 88);
+ in3_1 = LD_SH(temp + 160);
+ in4_1 = LD_SH(temp + 56);
+ in5_1 = LD_SH(temp + 184);
+ in6_1 = LD_SH(temp + 120);
+ in7_1 = LD_SH(temp + 248);
+
+ TRANSPOSE8x8_SH_SH(in0_1, in1_1, in2_1, in3_1, in4_1, in5_1, in6_1, in7_1,
+ in0_1, in1_1, in2_1, in3_1, in4_1, in5_1, in6_1, in7_1);
+ ST_SH8(in0_1, in1_1, in2_1, in3_1, in4_1, in5_1, in6_1, in7_1,
+ output + 24, 32);
+}
+
+static void fdct32x8_1d_row(int16_t *temp, int16_t *temp_buf,
+ int16_t *output) {+ fdct8x32_1d_row_load_butterfly(temp, temp_buf);
+ fdct8x32_1d_row_even(temp_buf, temp_buf);
+ fdct8x32_1d_row_odd(temp_buf + 128, temp, temp_buf + 128);
+ fdct8x32_1d_row_transpose_store(temp_buf, output);
+}
+
+static void fdct32x8_1d_row_4x(int16_t *tmp_buf_big, int16_t *tmp_buf,
+ int16_t *output) {+ fdct8x32_1d_row_load_butterfly(tmp_buf_big, tmp_buf);
+ fdct8x32_1d_row_even_4x(tmp_buf, tmp_buf_big, tmp_buf);
+ fdct8x32_1d_row_odd(tmp_buf + 128, tmp_buf_big, tmp_buf + 128);
+ fdct8x32_1d_row_transpose_store(tmp_buf, output);
+}
+
+void vp9_fdct32x32_msa(const int16_t *input, int16_t *output,
+ int32_t src_stride) {+ int32_t i;
+ DECLARE_ALIGNED(32, int16_t, tmp_buf_big[1024]);
+ DECLARE_ALIGNED(32, int16_t, tmp_buf[256]);
+
+ /* column transform */
+ for (i = 0; i < 4; ++i) {+ fdct8x32_1d_column(input + (8 * i), src_stride, tmp_buf,
+ tmp_buf_big + (8 * i));
+ }
+
+ /* row transform */
+ fdct32x8_1d_row_4x(tmp_buf_big, tmp_buf, output);
+
+ /* row transform */
+ for (i = 1; i < 4; ++i) {+ fdct32x8_1d_row(tmp_buf_big + (i * 256), tmp_buf, output + (i * 256));
+ }
+}
+
+void vp9_fdct32x32_1_msa(const int16_t *input, int16_t *out, int32_t stride) {+ out[1] = 0;
+
+ out[0] = VP9_LD_HADD(input, stride);
+ out[0] += VP9_LD_HADD(input + 8, stride);
+ out[0] += VP9_LD_HADD(input + 16, stride);
+ out[0] += VP9_LD_HADD(input + 24, stride);
+ out[0] += VP9_LD_HADD(input + 32 * 8, stride);
+ out[0] += VP9_LD_HADD(input + 32 * 8 + 8, stride);
+ out[0] += VP9_LD_HADD(input + 32 * 8 + 16, stride);
+ out[0] += VP9_LD_HADD(input + 32 * 8 + 24, stride);
+ out[0] += VP9_LD_HADD(input + 32 * 16, stride);
+ out[0] += VP9_LD_HADD(input + 32 * 16 + 8, stride);
+ out[0] += VP9_LD_HADD(input + 32 * 16 + 16, stride);
+ out[0] += VP9_LD_HADD(input + 32 * 16 + 24, stride);
+ out[0] += VP9_LD_HADD(input + 32 * 24, stride);
+ out[0] += VP9_LD_HADD(input + 32 * 24 + 8, stride);
+ out[0] += VP9_LD_HADD(input + 32 * 24 + 16, stride);
+ out[0] += VP9_LD_HADD(input + 32 * 24 + 24, stride);
+ out[0] >>= 3;
+}
+
+static void fdct8x32_1d_row_even_rd(int16_t *temp, int16_t *out) {+ v8i16 in0, in1, in2, in3, in4, in5, in6, in7;
+ v8i16 in8, in9, in10, in11, in12, in13, in14, in15;
+ v8i16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, temp0, temp1;
+
+ /* fdct32 even */
+ /* stage 2 */
+ LD_SH8(temp, 8, in0, in1, in2, in3, in4, in5, in6, in7);
+ LD_SH8(temp + 64, 8, in8, in9, in10, in11, in12, in13, in14, in15);
+
+ BUTTERFLY_16(in0, in1, in2, in3, in4, in5, in6, in7,
+ in8, in9, in10, in11, in12, in13, in14, in15,
+ vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7,
+ in8, in9, in10, in11, in12, in13, in14, in15);
+ VP9_FDCT_POSTPROC_2V_NEG_H(vec0, vec1);
+ VP9_FDCT_POSTPROC_2V_NEG_H(vec2, vec3);
+ VP9_FDCT_POSTPROC_2V_NEG_H(vec4, vec5);
+ VP9_FDCT_POSTPROC_2V_NEG_H(vec6, vec7);
+ VP9_FDCT_POSTPROC_2V_NEG_H(in8, in9);
+ VP9_FDCT_POSTPROC_2V_NEG_H(in10, in11);
+ VP9_FDCT_POSTPROC_2V_NEG_H(in12, in13);
+ VP9_FDCT_POSTPROC_2V_NEG_H(in14, in15);
+
+ /* Stage 3 */
+ ADD4(vec0, vec7, vec1, vec6, vec2, vec5, vec3, vec4, in0, in1, in2, in3);
+
+ temp0 = in0 + in3;
+ in0 = in0 - in3;
+ in3 = in1 + in2;
+ in1 = in1 - in2;
+
+ VP9_DOTP_CONST_PAIR(temp0, in3, cospi_16_64, cospi_16_64, temp1, temp0);
+ ST_SH(temp0, out);
+ ST_SH(temp1, out + 8);
+
+ VP9_DOTP_CONST_PAIR(in0, in1, cospi_24_64, cospi_8_64, temp1, temp0);
+ ST_SH(temp0, out + 16);
+ ST_SH(temp1, out + 24);
+
+ SUB4(vec3, vec4, vec2, vec5, vec1, vec6, vec0, vec7, vec4, vec5, vec6, vec7);
+ VP9_DOTP_CONST_PAIR(vec6, vec5, cospi_16_64, cospi_16_64, vec5, vec6);
+ ADD2(vec4, vec5, vec7, vec6, vec0, vec1);
+ VP9_DOTP_CONST_PAIR(vec1, vec0, cospi_28_64, cospi_4_64, temp1, temp0);
+ ST_SH(temp0, out + 32);
+ ST_SH(temp1, out + 56);
+
+ SUB2(vec4, vec5, vec7, vec6, vec4, vec7);
+ VP9_DOTP_CONST_PAIR(vec7, vec4, cospi_12_64, cospi_20_64, temp1, temp0);
+ ST_SH(temp0, out + 40);
+ ST_SH(temp1, out + 48);
+
+ VP9_DOTP_CONST_PAIR(in13, in10, cospi_16_64, cospi_16_64, vec2, vec5);
+ VP9_DOTP_CONST_PAIR(in12, in11, cospi_16_64, cospi_16_64, vec3, vec4);
+ ADD4(in8, vec3, in9, vec2, in14, vec5, in15, vec4, in0, vec1, vec6, in2);
+ VP9_DOTP_CONST_PAIR(vec6, vec1, cospi_24_64, cospi_8_64, in1, in3);
+ ADD2(in0, in1, in2, in3, vec0, vec7);
+ VP9_DOTP_CONST_PAIR(vec7, vec0, cospi_30_64, cospi_2_64, temp1, temp0);
+ ST_SH(temp0, out + 64);
+ ST_SH(temp1, out + 120);
+
+ SUB2(in0, in1, in2, in3, in0, in2);
+ VP9_DOTP_CONST_PAIR(in2, in0, cospi_14_64, cospi_18_64, temp1, temp0);
+ ST_SH(temp0, out + 72);
+ ST_SH(temp1, out + 112);
+
+ SUB2(in9, vec2, in14, vec5, vec2, vec5);
+ VP9_DOTP_CONST_PAIR((-vec2), vec5, cospi_24_64, cospi_8_64, in2, in1);
+ SUB4(in8, vec3, in15, vec4, in3, in2, in0, in1, in3, in0, vec2, vec5);
+ VP9_DOTP_CONST_PAIR(vec5, vec2, cospi_22_64, cospi_10_64, temp1, temp0);
+ ST_SH(temp0, out + 80);
+ ST_SH(temp1, out + 104);
+
+ ADD2(in3, in2, in0, in1, vec3, vec4);
+ VP9_DOTP_CONST_PAIR(vec4, vec3, cospi_6_64, cospi_26_64, temp0, temp1);
+ ST_SH(temp0, out + 96);
+ ST_SH(temp1, out + 88);
+}
+
+static void fdct8x32_1d_row_odd_rd(int16_t *temp, int16_t *interm_ptr,
+ int16_t *out) {+ v8i16 in16, in17, in18, in19, in20, in21, in22, in23;
+ v8i16 in24, in25, in26, in27, in28, in29, in30, in31;
+ v8i16 vec4, vec5;
+
+ in20 = LD_SH(temp + 32);
+ in21 = LD_SH(temp + 40);
+ in26 = LD_SH(temp + 80);
+ in27 = LD_SH(temp + 88);
+
+ VP9_DOTP_CONST_PAIR(in27, in20, cospi_16_64, cospi_16_64, in20, in27);
+ VP9_DOTP_CONST_PAIR(in26, in21, cospi_16_64, cospi_16_64, in21, in26);
+
+ VP9_FDCT_POSTPROC_2V_NEG_H(in20, in21);
+ VP9_FDCT_POSTPROC_2V_NEG_H(in26, in27);
+
+ in18 = LD_SH(temp + 16);
+ in19 = LD_SH(temp + 24);
+ in28 = LD_SH(temp + 96);
+ in29 = LD_SH(temp + 104);
+
+ VP9_FDCT_POSTPROC_2V_NEG_H(in18, in19);
+ VP9_FDCT_POSTPROC_2V_NEG_H(in28, in29);
+
+ vec4 = in19 - in20;
+ ST_SH(vec4, interm_ptr + 32);
+ vec4 = in18 - in21;
+ ST_SH(vec4, interm_ptr + 88);
+ vec4 = in29 - in26;
+ ST_SH(vec4, interm_ptr + 64);
+ vec4 = in28 - in27;
+ ST_SH(vec4, interm_ptr + 56);
+
+ ADD4(in18, in21, in19, in20, in28, in27, in29, in26, in21, in20, in27, in26);
+
+ in22 = LD_SH(temp + 48);
+ in23 = LD_SH(temp + 56);
+ in24 = LD_SH(temp + 64);
+ in25 = LD_SH(temp + 72);
+
+ VP9_DOTP_CONST_PAIR(in25, in22, cospi_16_64, cospi_16_64, in22, in25);
+ VP9_DOTP_CONST_PAIR(in24, in23, cospi_16_64, cospi_16_64, in23, in24);
+ VP9_FDCT_POSTPROC_2V_NEG_H(in22, in23);
+ VP9_FDCT_POSTPROC_2V_NEG_H(in24, in25);
+
+ in16 = LD_SH(temp);
+ in17 = LD_SH(temp + 8);
+ in30 = LD_SH(temp + 112);
+ in31 = LD_SH(temp + 120);
+
+ VP9_FDCT_POSTPROC_2V_NEG_H(in16, in17);
+ VP9_FDCT_POSTPROC_2V_NEG_H(in30, in31);
+
+ vec4 = in17 - in22;
+ ST_SH(vec4, interm_ptr + 40);
+ vec4 = in30 - in25;
+ ST_SH(vec4, interm_ptr + 48);
+ vec4 = in31 - in24;
+ ST_SH(vec4, interm_ptr + 72);
+ vec4 = in16 - in23;
+ ST_SH(vec4, interm_ptr + 80);
+
+ ADD4(in16, in23, in17, in22, in30, in25, in31, in24, in16, in17, in30, in31);
+ VP9_DOTP_CONST_PAIR(in26, in21, cospi_24_64, cospi_8_64, in18, in29);
+ VP9_DOTP_CONST_PAIR(in27, in20, cospi_24_64, cospi_8_64, in19, in28);
+ ADD4(in16, in19, in17, in18, in30, in29, in31, in28, in27, in22, in21, in25);
+ VP9_DOTP_CONST_PAIR(in21, in22, cospi_28_64, cospi_4_64, in26, in24);
+ ADD2(in27, in26, in25, in24, in23, in20);
+ VP9_DOTP_CONST_PAIR(in20, in23, cospi_31_64, cospi_1_64, vec4, vec5);
+ ST_SH(vec5, out);
+ ST_SH(vec4, out + 120);
+
+ SUB2(in27, in26, in25, in24, in22, in21);
+ VP9_DOTP_CONST_PAIR(in21, in22, cospi_15_64, cospi_17_64, vec5, vec4);
+ ST_SH(vec5, out + 112);
+ ST_SH(vec4, out + 8);
+
+ SUB4(in17, in18, in16, in19, in31, in28, in30, in29, in23, in26, in24, in20);
+ VP9_DOTP_CONST_PAIR((-in23), in20, cospi_28_64, cospi_4_64, in27, in25);
+ SUB2(in26, in27, in24, in25, in23, in20);
+ VP9_DOTP_CONST_PAIR(in20, in23, cospi_23_64, cospi_9_64, vec4, vec5);
+ ST_SH(vec4, out + 16);
+ ST_SH(vec5, out + 104);
+
+ ADD2(in26, in27, in24, in25, in22, in21);
+ VP9_DOTP_CONST_PAIR(in21, in22, cospi_7_64, cospi_25_64, vec4, vec5);
+ ST_SH(vec4, out + 24);
+ ST_SH(vec5, out + 96);
+
+ in20 = LD_SH(interm_ptr + 32);
+ in21 = LD_SH(interm_ptr + 88);
+ in27 = LD_SH(interm_ptr + 56);
+ in26 = LD_SH(interm_ptr + 64);
+
+ in16 = in20;
+ in17 = in21;
+ VP9_DOTP_CONST_PAIR(-in16, in27, cospi_24_64, cospi_8_64, in20, in27);
+ VP9_DOTP_CONST_PAIR(-in17, in26, cospi_24_64, cospi_8_64, in21, in26);
+
+ in22 = LD_SH(interm_ptr + 40);
+ in25 = LD_SH(interm_ptr + 48);
+ in24 = LD_SH(interm_ptr + 72);
+ in23 = LD_SH(interm_ptr + 80);
+
+ SUB4(in23, in20, in22, in21, in25, in26, in24, in27, in28, in17, in18, in31);
+ VP9_DOTP_CONST_PAIR(in18, in17, cospi_12_64, cospi_20_64, in29, in30);
+ in16 = in28 + in29;
+ in19 = in31 + in30;
+ VP9_DOTP_CONST_PAIR(in19, in16, cospi_27_64, cospi_5_64, vec5, vec4);
+ ST_SH(vec5, out + 32);
+ ST_SH(vec4, out + 88);
+
+ SUB2(in28, in29, in31, in30, in17, in18);
+ VP9_DOTP_CONST_PAIR(in18, in17, cospi_11_64, cospi_21_64, vec5, vec4);
+ ST_SH(vec5, out + 40);
+ ST_SH(vec4, out + 80);
+
+ ADD4(in22, in21, in23, in20, in24, in27, in25, in26, in16, in29, in30, in19);
+ VP9_DOTP_CONST_PAIR(-in16, in19, cospi_12_64, cospi_20_64, in28, in31);
+ SUB2(in29, in28, in30, in31, in16, in19);
+ VP9_DOTP_CONST_PAIR(in19, in16, cospi_19_64, cospi_13_64, vec5, vec4);
+ ST_SH(vec5, out + 72);
+ ST_SH(vec4, out + 48);
+
+ ADD2(in29, in28, in30, in31, in17, in18);
+ VP9_DOTP_CONST_PAIR(in18, in17, cospi_3_64, cospi_29_64, vec5, vec4);
+ ST_SH(vec4, out + 56);
+ ST_SH(vec5, out + 64);
+}
+
+static void fdct32x8_1d_row_rd(int16_t *tmp_buf_big, int16_t *tmp_buf,
+ int16_t *output) {+ fdct8x32_1d_row_load_butterfly(tmp_buf_big, tmp_buf);
+ fdct8x32_1d_row_even_rd(tmp_buf, tmp_buf);
+ fdct8x32_1d_row_odd_rd((tmp_buf + 128), tmp_buf_big, (tmp_buf + 128));
+ fdct8x32_1d_row_transpose_store(tmp_buf, output);
+}
+
+void vp9_fdct32x32_rd_msa(const int16_t *input, int16_t *out,
+ int32_t src_stride) {+ int32_t i;
+ DECLARE_ALIGNED(32, int16_t, tmp_buf_big[1024]);
+ DECLARE_ALIGNED(32, int16_t, tmp_buf[256]);
+
+ /* column transform */
+ for (i = 0; i < 4; ++i) {+ fdct8x32_1d_column(input + (8 * i), src_stride, &tmp_buf[0],
+ &tmp_buf_big[0] + (8 * i));
+ }
+
+ /* row transform */
+ for (i = 0; i < 4; ++i) {+ fdct32x8_1d_row_rd(&tmp_buf_big[0] + (8 * i * 32), &tmp_buf[0],
+ out + (8 * i * 32));
+ }
+}
--- a/vp9/encoder/mips/msa/vp9_fdct_msa.h
+++ b/vp9/encoder/mips/msa/vp9_fdct_msa.h
@@ -15,6 +15,24 @@
#include "vp9/common/vp9_idct.h"
#include "vp9/common/mips/msa/vp9_macros_msa.h"
+#define VP9_DOTP_CONST_PAIR(reg0, reg1, cnst0, cnst1, out0, out1) { \+ v8i16 k0_m = __msa_fill_h(cnst0); \
+ v4i32 s0_m, s1_m, s2_m, s3_m; \
+ \
+ s0_m = (v4i32)__msa_fill_h(cnst1); \
+ k0_m = __msa_ilvev_h((v8i16)s0_m, k0_m); \
+ \
+ ILVRL_H2_SW((-reg1), reg0, s1_m, s0_m); \
+ ILVRL_H2_SW(reg0, reg1, s3_m, s2_m); \
+ DOTP_SH2_SW(s1_m, s0_m, k0_m, k0_m, s1_m, s0_m); \
+ SRARI_W2_SW(s1_m, s0_m, DCT_CONST_BITS); \
+ out0 = __msa_pckev_h((v8i16)s0_m, (v8i16)s1_m); \
+ \
+ DOTP_SH2_SW(s3_m, s2_m, k0_m, k0_m, s1_m, s0_m); \
+ SRARI_W2_SW(s1_m, s0_m, DCT_CONST_BITS); \
+ out1 = __msa_pckev_h((v8i16)s0_m, (v8i16)s1_m); \
+}
+
#define VP9_DOT_SHIFT_RIGHT_PCK_H(in0, in1, in2) ({ \v8i16 dst_m; \
v4i32 tp0_m, tp1_m; \
@@ -148,7 +166,6 @@
out3 = VP9_DOT_SHIFT_RIGHT_PCK_H(s6_m, s7_m, x2_m); \
}
-/* FDCT16x16 specific */
#define VP9_FDCT8x16_ODD(input0, input1, input2, input3, \
input4, input5, input6, input7, \
out1, out3, out5, out7, \
@@ -249,5 +266,68 @@
cnst1_m = __msa_splati_h(coeff2_m, 3); \
cnst0_m = __msa_ilvev_h(cnst0_m, cnst1_m); \
out3 = VP9_DOT_SHIFT_RIGHT_PCK_H(vec0_m, vec1_m, cnst0_m); \
+}
+
+#define VP9_FDCT32_POSTPROC_NEG_W(vec) { \+ v4i32 temp_m; \
+ v4i32 one_m = __msa_ldi_w(1); \
+ \
+ temp_m = __msa_clti_s_w(vec, 0); \
+ vec += 1; \
+ temp_m = one_m & temp_m; \
+ vec += temp_m; \
+ vec >>= 2; \
+}
+
+#define VP9_FDCT32_POSTPROC_2V_POS_H(vec0, vec1) { \+ v8i16 tp0_m, tp1_m; \
+ v8i16 one = __msa_ldi_h(1); \
+ \
+ tp0_m = __msa_clei_s_h(vec0, 0); \
+ tp1_m = __msa_clei_s_h(vec1, 0); \
+ tp0_m = (v8i16)__msa_xori_b((v16u8)tp0_m, 255); \
+ tp1_m = (v8i16)__msa_xori_b((v16u8)tp1_m, 255); \
+ vec0 += 1; \
+ vec1 += 1; \
+ tp0_m = one & tp0_m; \
+ tp1_m = one & tp1_m; \
+ vec0 += tp0_m; \
+ vec1 += tp1_m; \
+ vec0 >>= 2; \
+ vec1 >>= 2; \
+}
+
+#define VP9_DOTP_CONST_PAIR_W(reg0_left, reg1_left, reg0_right, \
+ reg1_right, const0, const1, \
+ out0, out1, out2, out3) { \+ v4i32 s0_m, s1_m, s2_m, s3_m, s4_m, s5_m, s6_m, s7_m; \
+ v2i64 tp0_m, tp1_m, tp2_m, tp3_m; \
+ v4i32 k0_m = __msa_fill_w((int32_t) const0); \
+ \
+ s0_m = __msa_fill_w((int32_t) const1); \
+ k0_m = __msa_ilvev_w(s0_m, k0_m); \
+ \
+ ILVRL_W2_SW(-reg1_left, reg0_left, s1_m, s0_m); \
+ ILVRL_W2_SW(reg0_left, reg1_left, s3_m, s2_m); \
+ ILVRL_W2_SW(-reg1_right, reg0_right, s5_m, s4_m); \
+ ILVRL_W2_SW(reg0_right, reg1_right, s7_m, s6_m); \
+ \
+ DOTP_SW2_SD(s0_m, s1_m, k0_m, k0_m, tp0_m, tp1_m); \
+ DOTP_SW2_SD(s4_m, s5_m, k0_m, k0_m, tp2_m, tp3_m); \
+ tp0_m = __msa_srari_d(tp0_m, DCT_CONST_BITS); \
+ tp1_m = __msa_srari_d(tp1_m, DCT_CONST_BITS); \
+ tp2_m = __msa_srari_d(tp2_m, DCT_CONST_BITS); \
+ tp3_m = __msa_srari_d(tp3_m, DCT_CONST_BITS); \
+ out0 = __msa_pckev_w((v4i32)tp0_m, (v4i32)tp1_m); \
+ out1 = __msa_pckev_w((v4i32)tp2_m, (v4i32)tp3_m); \
+ \
+ DOTP_SW2_SD(s2_m, s3_m, k0_m, k0_m, tp0_m, tp1_m); \
+ DOTP_SW2_SD(s6_m, s7_m, k0_m, k0_m, tp2_m, tp3_m); \
+ tp0_m = __msa_srari_d(tp0_m, DCT_CONST_BITS); \
+ tp1_m = __msa_srari_d(tp1_m, DCT_CONST_BITS); \
+ tp2_m = __msa_srari_d(tp2_m, DCT_CONST_BITS); \
+ tp3_m = __msa_srari_d(tp3_m, DCT_CONST_BITS); \
+ out2 = __msa_pckev_w((v4i32)tp0_m, (v4i32)tp1_m); \
+ out3 = __msa_pckev_w((v4i32)tp2_m, (v4i32)tp3_m); \
}
#endif /* VP9_ENCODER_MIPS_MSA_VP9_FDCT_MSA_H_ */
--- a/vp9/vp9cx.mk
+++ b/vp9/vp9cx.mk
@@ -153,6 +153,7 @@
VP9_CX_SRCS-$(HAVE_NEON) += encoder/arm/neon/vp9_variance_neon.c
VP9_CX_SRCS-$(HAVE_MSA) += encoder/mips/msa/vp9_fdct16x16_msa.c
+VP9_CX_SRCS-$(HAVE_MSA) += encoder/mips/msa/vp9_fdct32x32_msa.c
VP9_CX_SRCS-$(HAVE_MSA) += encoder/mips/msa/vp9_fdct_msa.h
VP9_CX_SRCS-yes := $(filter-out $(VP9_CX_SRCS_REMOVE-yes),$(VP9_CX_SRCS-yes))