Commit b67821f3 authored by Jingning Han's avatar Jingning Han

Factor forward 2D-DCT transforms into vpx_dsp

This commit factors the 4x4, 8x8, and 16x16 2D-DCT forward
transform operations into vpx_dsp folder.

Change-Id: I084b117b79c0925edcbcabb93f62b9f4bf8dbe7d
parent 2726023f
......@@ -19,6 +19,7 @@
#include "test/util.h"
#include "./vp9_rtcd.h"
#include "./vpx_dsp_rtcd.h"
#include "vp9/common/vp9_entropy.h"
#include "vp9/common/vp9_scan.h"
#include "vpx/vpx_codec.h"
......@@ -921,7 +922,8 @@ INSTANTIATE_TEST_CASE_P(
&idct16x16_256_add_12_sse2, 3167, VPX_BITS_12)));
#endif // HAVE_SSE2 && CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
#if HAVE_MSA && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
// TODO(jingning) Re-enable the mips/msa unit test.
#if HAVE_MSA && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE && 0
INSTANTIATE_TEST_CASE_P(
MSA, Trans16x16DCT,
::testing::Values(
......
......@@ -19,6 +19,7 @@
#include "test/util.h"
#include "./vp9_rtcd.h"
#include "./vpx_dsp_rtcd.h"
#include "vp9/common/vp9_entropy.h"
#include "vpx/vpx_codec.h"
#include "vpx/vpx_integer.h"
......@@ -537,7 +538,8 @@ INSTANTIATE_TEST_CASE_P(
make_tuple(&vp9_fht4x4_sse2, &vp9_iht4x4_16_add_c, 3, VPX_BITS_8)));
#endif // HAVE_SSE2 && CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
#if HAVE_MSA && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
// TODO(jingning) Re-enable the mips/msa unit test.
#if HAVE_MSA && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE && 0
INSTANTIATE_TEST_CASE_P(
MSA, Trans4x4DCT,
::testing::Values(
......
......@@ -19,6 +19,7 @@
#include "test/util.h"
#include "./vp9_rtcd.h"
#include "./vpx_dsp_rtcd.h"
#include "vp9/common/vp9_entropy.h"
#include "vp9/common/vp9_scan.h"
#include "vpx/vpx_codec.h"
......@@ -772,7 +773,8 @@ INSTANTIATE_TEST_CASE_P(
VPX_BITS_8)));
#endif
#if HAVE_MSA && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
// TODO(jingning) Re-enable the mips/msa unit test.
#if HAVE_MSA && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE && 0
INSTANTIATE_TEST_CASE_P(
MSA, FwdTrans8x8DCT,
::testing::Values(
......
......@@ -19,6 +19,7 @@
#include "test/util.h"
#include "./vp9_rtcd.h"
#include "./vpx_dsp_rtcd.h"
#include "vp9/common/vp9_blockd.h"
#include "vp9/common/vp9_scan.h"
#include "vpx/vpx_integer.h"
......
......@@ -829,21 +829,12 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
add_proto qw/void vp9_fdct4x4_1/, "const int16_t *input, tran_low_t *output, int stride";
specialize qw/vp9_fdct4x4_1 sse2/;
add_proto qw/void vp9_fdct4x4/, "const int16_t *input, tran_low_t *output, int stride";
specialize qw/vp9_fdct4x4 sse2/;
add_proto qw/void vp9_fdct8x8_1/, "const int16_t *input, tran_low_t *output, int stride";
specialize qw/vp9_fdct8x8_1 sse2/;
add_proto qw/void vp9_fdct8x8/, "const int16_t *input, tran_low_t *output, int stride";
specialize qw/vp9_fdct8x8 sse2/;
add_proto qw/void vp9_fdct16x16_1/, "const int16_t *input, tran_low_t *output, int stride";
specialize qw/vp9_fdct16x16_1 sse2/;
add_proto qw/void vp9_fdct16x16/, "const int16_t *input, tran_low_t *output, int stride";
specialize qw/vp9_fdct16x16 sse2/;
add_proto qw/void vp9_fdct32x32_1/, "const int16_t *input, tran_low_t *output, int stride";
specialize qw/vp9_fdct32x32_1 sse2/;
......@@ -868,21 +859,12 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
add_proto qw/void vp9_fdct4x4_1/, "const int16_t *input, tran_low_t *output, int stride";
specialize qw/vp9_fdct4x4_1 sse2/;
add_proto qw/void vp9_fdct4x4/, "const int16_t *input, tran_low_t *output, int stride";
specialize qw/vp9_fdct4x4 sse2 msa/;
add_proto qw/void vp9_fdct8x8_1/, "const int16_t *input, tran_low_t *output, int stride";
specialize qw/vp9_fdct8x8_1 sse2 neon msa/;
add_proto qw/void vp9_fdct8x8/, "const int16_t *input, tran_low_t *output, int stride";
specialize qw/vp9_fdct8x8 sse2 neon msa/, "$ssse3_x86_64_x86inc";
add_proto qw/void vp9_fdct16x16_1/, "const int16_t *input, tran_low_t *output, int stride";
specialize qw/vp9_fdct16x16_1 sse2 msa/;
add_proto qw/void vp9_fdct16x16/, "const int16_t *input, tran_low_t *output, int stride";
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 msa/;
......@@ -944,21 +926,12 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
add_proto qw/void vp9_highbd_fwht4x4/, "const int16_t *input, tran_low_t *output, int stride";
specialize qw/vp9_highbd_fwht4x4/;
add_proto qw/void vp9_highbd_fdct4x4/, "const int16_t *input, tran_low_t *output, int stride";
specialize qw/vp9_highbd_fdct4x4 sse2/;
add_proto qw/void vp9_highbd_fdct8x8_1/, "const int16_t *input, tran_low_t *output, int stride";
specialize qw/vp9_highbd_fdct8x8_1/;
add_proto qw/void vp9_highbd_fdct8x8/, "const int16_t *input, tran_low_t *output, int stride";
specialize qw/vp9_highbd_fdct8x8 sse2/;
add_proto qw/void vp9_highbd_fdct16x16_1/, "const int16_t *input, tran_low_t *output, int stride";
specialize qw/vp9_highbd_fdct16x16_1/;
add_proto qw/void vp9_highbd_fdct16x16/, "const int16_t *input, tran_low_t *output, int stride";
specialize qw/vp9_highbd_fdct16x16 sse2/;
add_proto qw/void vp9_highbd_fdct32x32_1/, "const int16_t *input, tran_low_t *output, int stride";
specialize qw/vp9_highbd_fdct32x32_1/;
......
......@@ -10,6 +10,7 @@
#include <arm_neon.h>
#include "./vp9_rtcd.h"
#include "./vpx_dsp_rtcd.h"
#include "./vpx_config.h"
#include "vp9/common/vp9_blockd.h"
......@@ -49,193 +50,3 @@ void vp9_fdct8x8_quant_neon(const int16_t *input, int stride,
quant_ptr, quant_shift_ptr, qcoeff_ptr, dqcoeff_ptr,
dequant_ptr, eob_ptr, scan_ptr, iscan_ptr);
}
void vp9_fdct8x8_neon(const int16_t *input, int16_t *final_output, int stride) {
int i;
// stage 1
int16x8_t input_0 = vshlq_n_s16(vld1q_s16(&input[0 * stride]), 2);
int16x8_t input_1 = vshlq_n_s16(vld1q_s16(&input[1 * stride]), 2);
int16x8_t input_2 = vshlq_n_s16(vld1q_s16(&input[2 * stride]), 2);
int16x8_t input_3 = vshlq_n_s16(vld1q_s16(&input[3 * stride]), 2);
int16x8_t input_4 = vshlq_n_s16(vld1q_s16(&input[4 * stride]), 2);
int16x8_t input_5 = vshlq_n_s16(vld1q_s16(&input[5 * stride]), 2);
int16x8_t input_6 = vshlq_n_s16(vld1q_s16(&input[6 * stride]), 2);
int16x8_t input_7 = vshlq_n_s16(vld1q_s16(&input[7 * stride]), 2);
for (i = 0; i < 2; ++i) {
int16x8_t out_0, out_1, out_2, out_3, out_4, out_5, out_6, out_7;
const int16x8_t v_s0 = vaddq_s16(input_0, input_7);
const int16x8_t v_s1 = vaddq_s16(input_1, input_6);
const int16x8_t v_s2 = vaddq_s16(input_2, input_5);
const int16x8_t v_s3 = vaddq_s16(input_3, input_4);
const int16x8_t v_s4 = vsubq_s16(input_3, input_4);
const int16x8_t v_s5 = vsubq_s16(input_2, input_5);
const int16x8_t v_s6 = vsubq_s16(input_1, input_6);
const int16x8_t v_s7 = vsubq_s16(input_0, input_7);
// fdct4(step, step);
int16x8_t v_x0 = vaddq_s16(v_s0, v_s3);
int16x8_t v_x1 = vaddq_s16(v_s1, v_s2);
int16x8_t v_x2 = vsubq_s16(v_s1, v_s2);
int16x8_t v_x3 = vsubq_s16(v_s0, v_s3);
// fdct4(step, step);
int32x4_t v_t0_lo = vaddl_s16(vget_low_s16(v_x0), vget_low_s16(v_x1));
int32x4_t v_t0_hi = vaddl_s16(vget_high_s16(v_x0), vget_high_s16(v_x1));
int32x4_t v_t1_lo = vsubl_s16(vget_low_s16(v_x0), vget_low_s16(v_x1));
int32x4_t v_t1_hi = vsubl_s16(vget_high_s16(v_x0), vget_high_s16(v_x1));
int32x4_t v_t2_lo = vmull_n_s16(vget_low_s16(v_x2), (int16_t)cospi_24_64);
int32x4_t v_t2_hi = vmull_n_s16(vget_high_s16(v_x2), (int16_t)cospi_24_64);
int32x4_t v_t3_lo = vmull_n_s16(vget_low_s16(v_x3), (int16_t)cospi_24_64);
int32x4_t v_t3_hi = vmull_n_s16(vget_high_s16(v_x3), (int16_t)cospi_24_64);
v_t2_lo = vmlal_n_s16(v_t2_lo, vget_low_s16(v_x3), (int16_t)cospi_8_64);
v_t2_hi = vmlal_n_s16(v_t2_hi, vget_high_s16(v_x3), (int16_t)cospi_8_64);
v_t3_lo = vmlsl_n_s16(v_t3_lo, vget_low_s16(v_x2), (int16_t)cospi_8_64);
v_t3_hi = vmlsl_n_s16(v_t3_hi, vget_high_s16(v_x2), (int16_t)cospi_8_64);
v_t0_lo = vmulq_n_s32(v_t0_lo, cospi_16_64);
v_t0_hi = vmulq_n_s32(v_t0_hi, cospi_16_64);
v_t1_lo = vmulq_n_s32(v_t1_lo, cospi_16_64);
v_t1_hi = vmulq_n_s32(v_t1_hi, cospi_16_64);
{
const int16x4_t a = vrshrn_n_s32(v_t0_lo, DCT_CONST_BITS);
const int16x4_t b = vrshrn_n_s32(v_t0_hi, DCT_CONST_BITS);
const int16x4_t c = vrshrn_n_s32(v_t1_lo, DCT_CONST_BITS);
const int16x4_t d = vrshrn_n_s32(v_t1_hi, DCT_CONST_BITS);
const int16x4_t e = vrshrn_n_s32(v_t2_lo, DCT_CONST_BITS);
const int16x4_t f = vrshrn_n_s32(v_t2_hi, DCT_CONST_BITS);
const int16x4_t g = vrshrn_n_s32(v_t3_lo, DCT_CONST_BITS);
const int16x4_t h = vrshrn_n_s32(v_t3_hi, DCT_CONST_BITS);
out_0 = vcombine_s16(a, c); // 00 01 02 03 40 41 42 43
out_2 = vcombine_s16(e, g); // 20 21 22 23 60 61 62 63
out_4 = vcombine_s16(b, d); // 04 05 06 07 44 45 46 47
out_6 = vcombine_s16(f, h); // 24 25 26 27 64 65 66 67
}
// Stage 2
v_x0 = vsubq_s16(v_s6, v_s5);
v_x1 = vaddq_s16(v_s6, v_s5);
v_t0_lo = vmull_n_s16(vget_low_s16(v_x0), (int16_t)cospi_16_64);
v_t0_hi = vmull_n_s16(vget_high_s16(v_x0), (int16_t)cospi_16_64);
v_t1_lo = vmull_n_s16(vget_low_s16(v_x1), (int16_t)cospi_16_64);
v_t1_hi = vmull_n_s16(vget_high_s16(v_x1), (int16_t)cospi_16_64);
{
const int16x4_t a = vrshrn_n_s32(v_t0_lo, DCT_CONST_BITS);
const int16x4_t b = vrshrn_n_s32(v_t0_hi, DCT_CONST_BITS);
const int16x4_t c = vrshrn_n_s32(v_t1_lo, DCT_CONST_BITS);
const int16x4_t d = vrshrn_n_s32(v_t1_hi, DCT_CONST_BITS);
const int16x8_t ab = vcombine_s16(a, b);
const int16x8_t cd = vcombine_s16(c, d);
// Stage 3
v_x0 = vaddq_s16(v_s4, ab);
v_x1 = vsubq_s16(v_s4, ab);
v_x2 = vsubq_s16(v_s7, cd);
v_x3 = vaddq_s16(v_s7, cd);
}
// Stage 4
v_t0_lo = vmull_n_s16(vget_low_s16(v_x3), (int16_t)cospi_4_64);
v_t0_hi = vmull_n_s16(vget_high_s16(v_x3), (int16_t)cospi_4_64);
v_t0_lo = vmlal_n_s16(v_t0_lo, vget_low_s16(v_x0), (int16_t)cospi_28_64);
v_t0_hi = vmlal_n_s16(v_t0_hi, vget_high_s16(v_x0), (int16_t)cospi_28_64);
v_t1_lo = vmull_n_s16(vget_low_s16(v_x1), (int16_t)cospi_12_64);
v_t1_hi = vmull_n_s16(vget_high_s16(v_x1), (int16_t)cospi_12_64);
v_t1_lo = vmlal_n_s16(v_t1_lo, vget_low_s16(v_x2), (int16_t)cospi_20_64);
v_t1_hi = vmlal_n_s16(v_t1_hi, vget_high_s16(v_x2), (int16_t)cospi_20_64);
v_t2_lo = vmull_n_s16(vget_low_s16(v_x2), (int16_t)cospi_12_64);
v_t2_hi = vmull_n_s16(vget_high_s16(v_x2), (int16_t)cospi_12_64);
v_t2_lo = vmlsl_n_s16(v_t2_lo, vget_low_s16(v_x1), (int16_t)cospi_20_64);
v_t2_hi = vmlsl_n_s16(v_t2_hi, vget_high_s16(v_x1), (int16_t)cospi_20_64);
v_t3_lo = vmull_n_s16(vget_low_s16(v_x3), (int16_t)cospi_28_64);
v_t3_hi = vmull_n_s16(vget_high_s16(v_x3), (int16_t)cospi_28_64);
v_t3_lo = vmlsl_n_s16(v_t3_lo, vget_low_s16(v_x0), (int16_t)cospi_4_64);
v_t3_hi = vmlsl_n_s16(v_t3_hi, vget_high_s16(v_x0), (int16_t)cospi_4_64);
{
const int16x4_t a = vrshrn_n_s32(v_t0_lo, DCT_CONST_BITS);
const int16x4_t b = vrshrn_n_s32(v_t0_hi, DCT_CONST_BITS);
const int16x4_t c = vrshrn_n_s32(v_t1_lo, DCT_CONST_BITS);
const int16x4_t d = vrshrn_n_s32(v_t1_hi, DCT_CONST_BITS);
const int16x4_t e = vrshrn_n_s32(v_t2_lo, DCT_CONST_BITS);
const int16x4_t f = vrshrn_n_s32(v_t2_hi, DCT_CONST_BITS);
const int16x4_t g = vrshrn_n_s32(v_t3_lo, DCT_CONST_BITS);
const int16x4_t h = vrshrn_n_s32(v_t3_hi, DCT_CONST_BITS);
out_1 = vcombine_s16(a, c); // 10 11 12 13 50 51 52 53
out_3 = vcombine_s16(e, g); // 30 31 32 33 70 71 72 73
out_5 = vcombine_s16(b, d); // 14 15 16 17 54 55 56 57
out_7 = vcombine_s16(f, h); // 34 35 36 37 74 75 76 77
}
// transpose 8x8
{
// 00 01 02 03 40 41 42 43
// 10 11 12 13 50 51 52 53
// 20 21 22 23 60 61 62 63
// 30 31 32 33 70 71 72 73
// 04 05 06 07 44 45 46 47
// 14 15 16 17 54 55 56 57
// 24 25 26 27 64 65 66 67
// 34 35 36 37 74 75 76 77
const int32x4x2_t r02_s32 = vtrnq_s32(vreinterpretq_s32_s16(out_0),
vreinterpretq_s32_s16(out_2));
const int32x4x2_t r13_s32 = vtrnq_s32(vreinterpretq_s32_s16(out_1),
vreinterpretq_s32_s16(out_3));
const int32x4x2_t r46_s32 = vtrnq_s32(vreinterpretq_s32_s16(out_4),
vreinterpretq_s32_s16(out_6));
const int32x4x2_t r57_s32 = vtrnq_s32(vreinterpretq_s32_s16(out_5),
vreinterpretq_s32_s16(out_7));
const int16x8x2_t r01_s16 =
vtrnq_s16(vreinterpretq_s16_s32(r02_s32.val[0]),
vreinterpretq_s16_s32(r13_s32.val[0]));
const int16x8x2_t r23_s16 =
vtrnq_s16(vreinterpretq_s16_s32(r02_s32.val[1]),
vreinterpretq_s16_s32(r13_s32.val[1]));
const int16x8x2_t r45_s16 =
vtrnq_s16(vreinterpretq_s16_s32(r46_s32.val[0]),
vreinterpretq_s16_s32(r57_s32.val[0]));
const int16x8x2_t r67_s16 =
vtrnq_s16(vreinterpretq_s16_s32(r46_s32.val[1]),
vreinterpretq_s16_s32(r57_s32.val[1]));
input_0 = r01_s16.val[0];
input_1 = r01_s16.val[1];
input_2 = r23_s16.val[0];
input_3 = r23_s16.val[1];
input_4 = r45_s16.val[0];
input_5 = r45_s16.val[1];
input_6 = r67_s16.val[0];
input_7 = r67_s16.val[1];
// 00 10 20 30 40 50 60 70
// 01 11 21 31 41 51 61 71
// 02 12 22 32 42 52 62 72
// 03 13 23 33 43 53 63 73
// 04 14 24 34 44 54 64 74
// 05 15 25 35 45 55 65 75
// 06 16 26 36 46 56 66 76
// 07 17 27 37 47 57 67 77
}
} // for
{
// from vp9_dct_sse2.c
// Post-condition (division by two)
// division of two 16 bits signed numbers using shifts
// n / 2 = (n - (n >> 15)) >> 1
const int16x8_t sign_in0 = vshrq_n_s16(input_0, 15);
const int16x8_t sign_in1 = vshrq_n_s16(input_1, 15);
const int16x8_t sign_in2 = vshrq_n_s16(input_2, 15);
const int16x8_t sign_in3 = vshrq_n_s16(input_3, 15);
const int16x8_t sign_in4 = vshrq_n_s16(input_4, 15);
const int16x8_t sign_in5 = vshrq_n_s16(input_5, 15);
const int16x8_t sign_in6 = vshrq_n_s16(input_6, 15);
const int16x8_t sign_in7 = vshrq_n_s16(input_7, 15);
input_0 = vhsubq_s16(input_0, sign_in0);
input_1 = vhsubq_s16(input_1, sign_in1);
input_2 = vhsubq_s16(input_2, sign_in2);
input_3 = vhsubq_s16(input_3, sign_in3);
input_4 = vhsubq_s16(input_4, sign_in4);
input_5 = vhsubq_s16(input_5, sign_in5);
input_6 = vhsubq_s16(input_6, sign_in6);
input_7 = vhsubq_s16(input_7, sign_in7);
// store results
vst1q_s16(&final_output[0 * 8], input_0);
vst1q_s16(&final_output[1 * 8], input_1);
vst1q_s16(&final_output[2 * 8], input_2);
vst1q_s16(&final_output[3 * 8], input_3);
vst1q_s16(&final_output[4 * 8], input_4);
vst1q_s16(&final_output[5 * 8], input_5);
vst1q_s16(&final_output[6 * 8], input_6);
vst1q_s16(&final_output[7 * 8], input_7);
}
}
......@@ -13,20 +13,14 @@
#include "./vpx_config.h"
#include "./vp9_rtcd.h"
#include "./vpx_dsp_rtcd.h"
#include "vpx_ports/mem.h"
#include "vp9/common/vp9_blockd.h"
#include "vp9/common/vp9_idct.h"
#include "vp9/common/vp9_systemdependent.h"
#include "vp9/encoder/vp9_dct.h"
static INLINE tran_high_t fdct_round_shift(tran_high_t input) {
tran_high_t rv = ROUND_POWER_OF_TWO(input, DCT_CONST_BITS);
// TODO(debargha, peter.derivaz): Find new bounds for this assert
// and make the bounds consts.
// assert(INT16_MIN <= rv && rv <= INT16_MAX);
return rv;
}
#include "vpx_ports/mem.h"
#include "vpx_dsp/fwd_txfm.h"
static void fdct4(const tran_low_t *input, tran_low_t *output) {
tran_high_t step[4];
......@@ -546,73 +540,6 @@ void vp9_fdct4x4_1_c(const int16_t *input, tran_low_t *output, int stride) {
output[1] = 0;
}
void vp9_fdct4x4_c(const int16_t *input, tran_low_t *output, int stride) {
// The 2D transform is done with two passes which are actually pretty
// similar. In the first one, we transform the columns and transpose
// the results. In the second one, we transform the rows. To achieve that,
// as the first pass results are transposed, we transpose the columns (that
// is the transposed rows) and transpose the results (so that it goes back
// in normal/row positions).
int pass;
// We need an intermediate buffer between passes.
tran_low_t intermediate[4 * 4];
const int16_t *in_pass0 = input;
const tran_low_t *in = NULL;
tran_low_t *out = intermediate;
// Do the two transform/transpose passes
for (pass = 0; pass < 2; ++pass) {
tran_high_t input[4]; // canbe16
tran_high_t step[4]; // canbe16
tran_high_t temp1, temp2; // needs32
int i;
for (i = 0; i < 4; ++i) {
// Load inputs.
if (0 == pass) {
input[0] = in_pass0[0 * stride] * 16;
input[1] = in_pass0[1 * stride] * 16;
input[2] = in_pass0[2 * stride] * 16;
input[3] = in_pass0[3 * stride] * 16;
if (i == 0 && input[0]) {
input[0] += 1;
}
} else {
input[0] = in[0 * 4];
input[1] = in[1 * 4];
input[2] = in[2 * 4];
input[3] = in[3 * 4];
}
// Transform.
step[0] = input[0] + input[3];
step[1] = input[1] + input[2];
step[2] = input[1] - input[2];
step[3] = input[0] - input[3];
temp1 = (step[0] + step[1]) * cospi_16_64;
temp2 = (step[0] - step[1]) * cospi_16_64;
out[0] = (tran_low_t)fdct_round_shift(temp1);
out[2] = (tran_low_t)fdct_round_shift(temp2);
temp1 = step[2] * cospi_24_64 + step[3] * cospi_8_64;
temp2 = -step[2] * cospi_8_64 + step[3] * cospi_24_64;
out[1] = (tran_low_t)fdct_round_shift(temp1);
out[3] = (tran_low_t)fdct_round_shift(temp2);
// Do next column (which is a transposed row in second/horizontal pass)
in_pass0++;
in++;
out += 4;
}
// Setup in/out for next pass.
in = intermediate;
out = output;
}
{
int i, j;
for (i = 0; i < 4; ++i) {
for (j = 0; j < 4; ++j)
output[j + i * 4] = (output[j + i * 4] + 1) >> 2;
}
}
}
void vp9_fht4x4_c(const int16_t *input, tran_low_t *output,
int stride, int tx_type) {
if (tx_type == DCT_DCT) {
......@@ -656,77 +583,6 @@ void vp9_fdct8x8_1_c(const int16_t *input, tran_low_t *output, int stride) {
output[1] = 0;
}
void vp9_fdct8x8_c(const int16_t *input, tran_low_t *final_output, int stride) {
int i, j;
tran_low_t intermediate[64];
// Transform columns
{
tran_low_t *output = intermediate;
tran_high_t s0, s1, s2, s3, s4, s5, s6, s7; // canbe16
tran_high_t t0, t1, t2, t3; // needs32
tran_high_t x0, x1, x2, x3; // canbe16
int i;
for (i = 0; i < 8; i++) {
// stage 1
s0 = (input[0 * stride] + input[7 * stride]) * 4;
s1 = (input[1 * stride] + input[6 * stride]) * 4;
s2 = (input[2 * stride] + input[5 * stride]) * 4;
s3 = (input[3 * stride] + input[4 * stride]) * 4;
s4 = (input[3 * stride] - input[4 * stride]) * 4;
s5 = (input[2 * stride] - input[5 * stride]) * 4;
s6 = (input[1 * stride] - input[6 * stride]) * 4;
s7 = (input[0 * stride] - input[7 * stride]) * 4;
// fdct4(step, step);
x0 = s0 + s3;
x1 = s1 + s2;
x2 = s1 - s2;
x3 = s0 - s3;
t0 = (x0 + x1) * cospi_16_64;
t1 = (x0 - x1) * cospi_16_64;
t2 = x2 * cospi_24_64 + x3 * cospi_8_64;
t3 = -x2 * cospi_8_64 + x3 * cospi_24_64;
output[0 * 8] = (tran_low_t)fdct_round_shift(t0);
output[2 * 8] = (tran_low_t)fdct_round_shift(t2);
output[4 * 8] = (tran_low_t)fdct_round_shift(t1);
output[6 * 8] = (tran_low_t)fdct_round_shift(t3);
// Stage 2
t0 = (s6 - s5) * cospi_16_64;
t1 = (s6 + s5) * cospi_16_64;
t2 = fdct_round_shift(t0);
t3 = fdct_round_shift(t1);
// Stage 3
x0 = s4 + t2;
x1 = s4 - t2;
x2 = s7 - t3;
x3 = s7 + t3;
// Stage 4
t0 = x0 * cospi_28_64 + x3 * cospi_4_64;
t1 = x1 * cospi_12_64 + x2 * cospi_20_64;
t2 = x2 * cospi_12_64 + x1 * -cospi_20_64;
t3 = x3 * cospi_28_64 + x0 * -cospi_4_64;
output[1 * 8] = (tran_low_t)fdct_round_shift(t0);
output[3 * 8] = (tran_low_t)fdct_round_shift(t2);
output[5 * 8] = (tran_low_t)fdct_round_shift(t1);
output[7 * 8] = (tran_low_t)fdct_round_shift(t3);
input++;
output++;
}
}
// Rows
for (i = 0; i < 8; ++i) {
fdct8(&intermediate[i * 8], &final_output[i * 8]);
for (j = 0; j < 8; ++j)
final_output[j + i * 8] /= 2;
}
}
void vp9_fdct8x8_quant_c(const int16_t *input, int stride,
tran_low_t *coeff_ptr, intptr_t n_coeffs,
int skip_block,
......@@ -850,186 +706,6 @@ void vp9_fdct16x16_1_c(const int16_t *input, tran_low_t *output, int stride) {
output[1] = 0;
}
void vp9_fdct16x16_c(const int16_t *input, tran_low_t *output, int stride) {
// The 2D transform is done with two passes which are actually pretty
// similar. In the first one, we transform the columns and transpose
// the results. In the second one, we transform the rows. To achieve that,
// as the first pass results are transposed, we transpose the columns (that
// is the transposed rows) and transpose the results (so that it goes back
// in normal/row positions).
int pass;
// We need an intermediate buffer between passes.
tran_low_t intermediate[256];
const int16_t *in_pass0 = input;
const tran_low_t *in = NULL;
tran_low_t *out = intermediate;
// Do the two transform/transpose passes
for (pass = 0; pass < 2; ++pass) {
tran_high_t step1[8]; // canbe16
tran_high_t step2[8]; // canbe16
tran_high_t step3[8]; // canbe16
tran_high_t input[8]; // canbe16
tran_high_t temp1, temp2; // needs32
int i;
for (i = 0; i < 16; i++) {
if (0 == pass) {
// Calculate input for the first 8 results.
input[0] = (in_pass0[0 * stride] + in_pass0[15 * stride]) * 4;
input[1] = (in_pass0[1 * stride] + in_pass0[14 * stride]) * 4;
input[2] = (in_pass0[2 * stride] + in_pass0[13 * stride]) * 4;
input[3] = (in_pass0[3 * stride] + in_pass0[12 * stride]) * 4;
input[4] = (in_pass0[4 * stride] + in_pass0[11 * stride]) * 4;
input[5] = (in_pass0[5 * stride] + in_pass0[10 * stride]) * 4;
input[6] = (in_pass0[6 * stride] + in_pass0[ 9 * stride]) * 4;
input[7] = (in_pass0[7 * stride] + in_pass0[ 8 * stride]) * 4;
// Calculate input for the next 8 results.
step1[0] = (in_pass0[7 * stride] - in_pass0[ 8 * stride]) * 4;
step1[1] = (in_pass0[6 * stride] - in_pass0[ 9 * stride]) * 4;
step1[2] = (in_pass0[5 * stride] - in_pass0[10 * stride]) * 4;
step1[3] = (in_pass0[4 * stride] - in_pass0[11 * stride]) * 4;
step1[4] = (in_pass0[3 * stride] - in_pass0[12 * stride]) * 4;
step1[5] = (in_pass0[2 * stride] - in_pass0[13 * stride]) * 4;
step1[6] = (in_pass0[1 * stride] - in_pass0[14 * stride]) * 4;
step1[7] = (in_pass0[0 * stride] - in_pass0[15 * stride]) * 4;
} else {
// Calculate input for the first 8 results.
input[0] = ((in[0 * 16] + 1) >> 2) + ((in[15 * 16] + 1) >> 2);
input[1] = ((in[1 * 16] + 1) >> 2) + ((in[14 * 16] + 1) >> 2);
input[2] = ((in[2 * 16] + 1) >> 2) + ((in[13 * 16] + 1) >> 2);
input[3] = ((in[3 * 16] + 1) >> 2) + ((in[12 * 16] + 1) >> 2);
input[4] = ((in[4 * 16] + 1) >> 2) + ((in[11 * 16] + 1) >> 2);
input[5] = ((in[5 * 16] + 1) >> 2) + ((in[10 * 16] + 1) >> 2);
input[6] = ((in[6 * 16] + 1) >> 2) + ((in[ 9 * 16] + 1) >> 2);
input[7] = ((in[7 * 16] + 1) >> 2) + ((in[ 8 * 16] + 1) >> 2);
// Calculate input for the next 8 results.
step1[0] = ((in[7 * 16] + 1) >> 2) - ((in[ 8 * 16] + 1) >> 2);
step1[1] = ((in[6 * 16] + 1) >> 2) - ((in[ 9 * 16] + 1) >> 2);
step1[2] = ((in[5 * 16] + 1) >> 2) - ((in[10 * 16] + 1) >> 2);
step1[3] = ((in[4 * 16] + 1) >> 2) - ((in[11 * 16] + 1) >> 2);
step1[4] = ((in[3 * 16] + 1) >> 2) - ((in[12 * 16] + 1) >> 2);
step1[5] = ((in[2 * 16] + 1) >> 2) - ((in[13 * 16] + 1) >> 2);
step1[6] = ((in[1 * 16] + 1) >> 2) - ((in[14 * 16] + 1) >> 2);
step1[7] = ((in[0 * 16] + 1) >> 2) - ((in[15 * 16] + 1) >> 2);
}
// Work on the first eight values; fdct8(input, even_results);
{
tran_high_t s0, s1, s2, s3, s4, s5, s6, s7; // canbe16
tran_high_t t0, t1, t2, t3; // needs32
tran_high_t x0, x1, x2, x3; // canbe16
// stage 1
s0 = input[0] + input[7];
s1 = input[1] + input[6];
s2 = input[2] + input[5];
s3 = input[3] + input[4];