Commit bcbd3c8f authored by Jingning Han's avatar Jingning Han
Browse files

Arrange 1D forward transform order in vp9_dct.c

Remove the redundant function declarations therein.

Change-Id: I27731fb70bb1abce63da761a5812f518c62f590f
parent 1279d3ba
......@@ -20,34 +20,6 @@
#include "vp9/common/vp9_systemdependent.h"
#include "vp9/encoder/vp9_dct.h"
static void fdct4(const tran_low_t *input, tran_low_t *output);
static void fadst4(const tran_low_t *input, tran_low_t *output);
static void fdct8(const tran_low_t *input, tran_low_t *output);
static void fadst8(const tran_low_t *input, tran_low_t *output);
static void fdct16(const tran_low_t in[16], tran_low_t out[16]);
static void fadst16(const tran_low_t *input, tran_low_t *output);
static const transform_2d FHT_4[] = {
{ fdct4, fdct4 }, // DCT_DCT = 0
{ fadst4, fdct4 }, // ADST_DCT = 1
{ fdct4, fadst4 }, // DCT_ADST = 2
{ fadst4, fadst4 } // ADST_ADST = 3
};
static const transform_2d FHT_8[] = {
{ fdct8, fdct8 }, // DCT_DCT = 0
{ fadst8, fdct8 }, // ADST_DCT = 1
{ fdct8, fadst8 }, // DCT_ADST = 2
{ fadst8, fadst8 } // ADST_ADST = 3
};
static const transform_2d FHT_16[] = {
{ fdct16, fdct16 }, // DCT_DCT = 0
{ fadst16, fdct16 }, // ADST_DCT = 1
{ fdct16, fadst16 }, // DCT_ADST = 2
{ fadst16, fadst16 } // ADST_ADST = 3
};
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
......@@ -75,156 +47,6 @@ static void fdct4(const tran_low_t *input, tran_low_t *output) {
output[3] = (tran_low_t)fdct_round_shift(temp2);
}
void vp9_fdct4x4_1_c(const int16_t *input, tran_low_t *output, int stride) {
int r, c;
tran_low_t sum = 0;
for (r = 0; r < 4; ++r)
for (c = 0; c < 4; ++c)
sum += input[r * stride + c];
output[0] = sum << 1;
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;
}
}
}
static void fadst4(const tran_low_t *input, tran_low_t *output) {
tran_high_t x0, x1, x2, x3;
tran_high_t s0, s1, s2, s3, s4, s5, s6, s7;
x0 = input[0];
x1 = input[1];
x2 = input[2];
x3 = input[3];
if (!(x0 | x1 | x2 | x3)) {
output[0] = output[1] = output[2] = output[3] = 0;
return;
}
s0 = sinpi_1_9 * x0;
s1 = sinpi_4_9 * x0;
s2 = sinpi_2_9 * x1;
s3 = sinpi_1_9 * x1;
s4 = sinpi_3_9 * x2;
s5 = sinpi_4_9 * x3;
s6 = sinpi_2_9 * x3;
s7 = x0 + x1 - x3;
x0 = s0 + s2 + s5;
x1 = sinpi_3_9 * s7;
x2 = s1 - s3 + s6;
x3 = s4;
s0 = x0 + x3;
s1 = x1;
s2 = x2 - x3;
s3 = x2 - x0 + x3;
// 1-D transform scaling factor is sqrt(2).
output[0] = (tran_low_t)fdct_round_shift(s0);
output[1] = (tran_low_t)fdct_round_shift(s1);
output[2] = (tran_low_t)fdct_round_shift(s2);
output[3] = (tran_low_t)fdct_round_shift(s3);
}
void vp9_fht4x4_c(const int16_t *input, tran_low_t *output,
int stride, int tx_type) {
if (tx_type == DCT_DCT) {
vp9_fdct4x4_c(input, output, stride);
} else {
tran_low_t out[4 * 4];
int i, j;
tran_low_t temp_in[4], temp_out[4];
const transform_2d ht = FHT_4[tx_type];
// Columns
for (i = 0; i < 4; ++i) {
for (j = 0; j < 4; ++j)
temp_in[j] = input[j * stride + i] * 16;
if (i == 0 && temp_in[0])
temp_in[0] += 1;
ht.cols(temp_in, temp_out);
for (j = 0; j < 4; ++j)
out[j * 4 + i] = temp_out[j];
}
// Rows
for (i = 0; i < 4; ++i) {
for (j = 0; j < 4; ++j)
temp_in[j] = out[j + i * 4];
ht.rows(temp_in, temp_out);
for (j = 0; j < 4; ++j)
output[j + i * 4] = (temp_out[j] + 1) >> 2;
}
}
}
static void fdct8(const tran_low_t *input, tran_low_t *output) {
tran_high_t s0, s1, s2, s3, s4, s5, s6, s7; // canbe16
tran_high_t t0, t1, t2, t3; // needs32
......@@ -277,402 +99,198 @@ static void fdct8(const tran_low_t *input, tran_low_t *output) {
output[7] = (tran_low_t)fdct_round_shift(t3);
}
void vp9_fdct8x8_1_c(const int16_t *input, tran_low_t *output, int stride) {
int r, c;
tran_low_t sum = 0;
for (r = 0; r < 8; ++r)
for (c = 0; c < 8; ++c)
sum += input[r * stride + c];
static void fdct16(const tran_low_t in[16], tran_low_t out[16]) {
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
output[0] = sum;
output[1] = 0;
}
// step 1
input[0] = in[0] + in[15];
input[1] = in[1] + in[14];
input[2] = in[2] + in[13];
input[3] = in[3] + in[12];
input[4] = in[4] + in[11];
input[5] = in[5] + in[10];
input[6] = in[6] + in[ 9];
input[7] = in[7] + in[ 8];
void vp9_fdct8x8_c(const int16_t *input, tran_low_t *final_output, int stride) {
int i, j;
tran_low_t intermediate[64];
step1[0] = in[7] - in[ 8];
step1[1] = in[6] - in[ 9];
step1[2] = in[5] - in[10];
step1[3] = in[4] - in[11];
step1[4] = in[3] - in[12];
step1[5] = in[2] - in[13];
step1[6] = in[1] - in[14];
step1[7] = in[0] - in[15];
// Transform columns
// fdct8(step, step);
{
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;
// stage 1
s0 = input[0] + input[7];
s1 = input[1] + input[6];
s2 = input[2] + input[5];
s3 = input[3] + input[4];
s4 = input[3] - input[4];
s5 = input[2] - input[5];
s6 = input[1] - input[6];
s7 = input[0] - input[7];
// 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);
// 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 = x3 * cospi_8_64 + x2 * cospi_24_64;
t3 = x3 * cospi_24_64 - x2 * cospi_8_64;
out[0] = (tran_low_t)fdct_round_shift(t0);
out[4] = (tran_low_t)fdct_round_shift(t2);
out[8] = (tran_low_t)fdct_round_shift(t1);
out[12] = (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 2
t0 = (s6 - s5) * cospi_16_64;
t1 = (s6 + s5) * cospi_16_64;
t2 = fdct_round_shift(t0);
t3 = fdct_round_shift(t1);
// 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++;
}
}
// Stage 3
x0 = s4 + t2;
x1 = s4 - t2;
x2 = s7 - t3;
x3 = s7 + t3;
// 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;
// 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;
out[2] = (tran_low_t)fdct_round_shift(t0);
out[6] = (tran_low_t)fdct_round_shift(t2);
out[10] = (tran_low_t)fdct_round_shift(t1);
out[14] = (tran_low_t)fdct_round_shift(t3);
}
}
void vp9_fdct8x8_quant_c(const int16_t *input, int stride,
tran_low_t *coeff_ptr, intptr_t n_coeffs,
int skip_block,
const int16_t *zbin_ptr, const int16_t *round_ptr,
const int16_t *quant_ptr,
const int16_t *quant_shift_ptr,
tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
const int16_t *dequant_ptr,
uint16_t *eob_ptr,
const int16_t *scan, const int16_t *iscan) {
int eob = -1;
// step 2
temp1 = (step1[5] - step1[2]) * cospi_16_64;
temp2 = (step1[4] - step1[3]) * cospi_16_64;
step2[2] = fdct_round_shift(temp1);
step2[3] = fdct_round_shift(temp2);
temp1 = (step1[4] + step1[3]) * cospi_16_64;
temp2 = (step1[5] + step1[2]) * cospi_16_64;
step2[4] = fdct_round_shift(temp1);
step2[5] = fdct_round_shift(temp2);
int i, j;
tran_low_t intermediate[64];
// step 3
step3[0] = step1[0] + step2[3];
step3[1] = step1[1] + step2[2];
step3[2] = step1[1] - step2[2];
step3[3] = step1[0] - step2[3];
step3[4] = step1[7] - step2[4];
step3[5] = step1[6] - step2[5];
step3[6] = step1[6] + step2[5];
step3[7] = step1[7] + step2[4];
// 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
// step 4
temp1 = step3[1] * -cospi_8_64 + step3[6] * cospi_24_64;
temp2 = step3[2] * cospi_24_64 + step3[5] * cospi_8_64;
step2[1] = fdct_round_shift(temp1);
step2[2] = fdct_round_shift(temp2);
temp1 = step3[2] * cospi_8_64 - step3[5] * cospi_24_64;
temp2 = step3[1] * cospi_24_64 + step3[6] * cospi_8_64;
step2[5] = fdct_round_shift(temp1);
step2[6] = fdct_round_shift(temp2);
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;
// step 5
step1[0] = step3[0] + step2[1];
step1[1] = step3[0] - step2[1];
step1[2] = step3[3] + step2[2];
step1[3] = step3[3] - step2[2];
step1[4] = step3[4] - step2[5];
step1[5] = step3[4] + step2[5];
step1[6] = step3[7] - step2[6];
step1[7] = step3[7] + step2[6];
// 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);
// step 6
temp1 = step1[0] * cospi_30_64 + step1[7] * cospi_2_64;
temp2 = step1[1] * cospi_14_64 + step1[6] * cospi_18_64;
out[1] = (tran_low_t)fdct_round_shift(temp1);
out[9] = (tran_low_t)fdct_round_shift(temp2);
// Stage 2
t0 = (s6 - s5) * cospi_16_64;
t1 = (s6 + s5) * cospi_16_64;
t2 = fdct_round_shift(t0);
t3 = fdct_round_shift(t1);
temp1 = step1[2] * cospi_22_64 + step1[5] * cospi_10_64;
temp2 = step1[3] * cospi_6_64 + step1[4] * cospi_26_64;
out[5] = (tran_low_t)fdct_round_shift(temp1);
out[13] = (tran_low_t)fdct_round_shift(temp2);
// Stage 3
x0 = s4 + t2;
x1 = s4 - t2;
x2 = s7 - t3;
x3 = s7 + t3;
temp1 = step1[3] * -cospi_26_64 + step1[4] * cospi_6_64;
temp2 = step1[2] * -cospi_10_64 + step1[5] * cospi_22_64;
out[3] = (tran_low_t)fdct_round_shift(temp1);
out[11] = (tran_low_t)fdct_round_shift(temp2);
// 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++;
}
}
temp1 = step1[1] * -cospi_18_64 + step1[6] * cospi_14_64;
temp2 = step1[0] * -cospi_2_64 + step1[7] * cospi_30_64;
out[7] = (tran_low_t)fdct_round_shift(temp1);
out[15] = (tran_low_t)fdct_round_shift(temp2);
}
// Rows
for (i = 0; i < 8; ++i) {
fdct8(&intermediate[i * 8], &coeff_ptr[i * 8]);
for (j = 0; j < 8; ++j)
coeff_ptr[j + i * 8] /= 2;
}
static void fadst4(const tran_low_t *input, tran_low_t *output) {
tran_high_t x0, x1, x2, x3;
tran_high_t s0, s1, s2, s3, s4, s5, s6, s7;
// TODO(jingning) Decide the need of these arguments after the
// quantization process is completed.
(void)zbin_ptr;
(void)quant_shift_ptr;
(void)iscan;
x0 = input[0];
x1 = input[1];
x2 = input[2];
x3 = input[3];
memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
if (!(x0 | x1 | x2 | x3)) {
output[0] = output[1] = output[2] = output[3] = 0;
return;
}
if (!skip_block) {
// Quantization pass: All coefficients with index >= zero_flag are
// skippable. Note: zero_flag can be zero.
for (i = 0; i < n_coeffs; i++) {
const int rc = scan[i];
const int coeff = coeff_ptr[rc];
const int coeff_sign = (coeff >> 31);
const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
s0 = sinpi_1_9 * x0;
s1 = sinpi_4_9 * x0;
s2 = sinpi_2_9 * x1;
s3 = sinpi_1_9 * x1;
s4 = sinpi_3_9 * x2;
s5 = sinpi_4_9 * x3;
s6 = sinpi_2_9 * x3;
s7 = x0 + x1 - x3;
int tmp = clamp(abs_coeff + round_ptr[rc != 0], INT16_MIN, INT16_MAX);
tmp = (tmp * quant_ptr[rc != 0]) >> 16;
x0 = s0 + s2 + s5;
x1 = sinpi_3_9 * s7;
x2 = s1 - s3 + s6;
x3 = s4;
qcoeff_ptr[rc] = (tmp ^ coeff_sign) - coeff_sign;
dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0];
s0 = x0 + x3;
s1 = x1;
s2 = x2 - x3;
s3 = x2 - x0 + x3;
if (tmp)
eob = i;
}
}
*eob_ptr = eob + 1;
// 1-D transform scaling factor is sqrt(2).
output[0] = (tran_low_t)fdct_round_shift(s0);
output[1] = (tran_low_t)fdct_round_shift(s1);
output[2] = (tran_low_t)fdct_round_shift(s2);
output[3] = (tran_low_t)fdct_round_shift(s3);
}
void vp9_fdct16x16_1_c(const int16_t *input, tran_low_t *output, int stride) {
int r, c;
tran_low_t sum = 0;
for (r = 0; r < 16; ++r)
for (c = 0; c < 16; ++c)
sum += input[r * stride + c];
static void fadst8(const tran_low_t *input, tran_low_t *output) {
tran_high_t s0, s1, s2, s3, s4, s5, s6, s7;
output[0] = sum >> 1;
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.
<