Adds 64x64 hybrid dct/dwt transform
This is to add to the 64x64 transform experiment as an alternative to
a 64x64 DCT.
Two levels of wavelet decomposition is used on a 64x64 block, followed
by 16x16 DCT on the four lowest subbands. The highest three subbands
are left untransformed after the first level DWT.
Change-Id: I3d48d5800468d655191933894df6b46e15adca56
diff --git a/configure b/configure
index 57a6145..396cee7 100755
--- a/configure
+++ b/configure
@@ -249,7 +249,8 @@
newbintramodes
comp_interintra_pred
tx32x32
- dwt32x32hybrid
+ tx64x64
+ dwtdcthybrid
cnvcontext
newcoefcontext
"
diff --git a/vp9/common/vp9_entropy.c b/vp9/common/vp9_entropy.c
index cdc8bc1..bc87384 100644
--- a/vp9/common/vp9_entropy.c
+++ b/vp9/common/vp9_entropy.c
@@ -143,7 +143,7 @@
};
#if CONFIG_TX32X32
-#if CONFIG_DWT32X32HYBRID
+#if CONFIG_DWTDCTHYBRID
DECLARE_ALIGNED(16, const int, vp9_coef_bands_32x32[1024]) = {
0, 1, 2, 3, 5, 4, 4, 5, 5, 3, 6, 3, 5, 4, 6,
6, 6, 6,
@@ -458,7 +458,7 @@
951, 920, 889, 858, 827, 796, 765, 734, 703, 735, 766, 797, 828, 859, 890, 921, 952, 983, 1014, 1015, 984, 953, 922, 891, 860, 829, 798, 767, 799, 830, 861, 892,
923, 954, 985, 1016, 1017, 986, 955, 924, 893, 862, 831, 863, 894, 925, 956, 987, 1018, 1019, 988, 957, 926, 895, 927, 958, 989, 1020, 1021, 990, 959, 991, 1022, 1023,
};
-#endif // CONFIG_DWT32X32HYBRID
+#endif // CONFIG_DWTDCTHYBRID
#endif
/* Array indices are identical to previously-existing CONTEXT_NODE indices */
diff --git a/vp9/common/vp9_idctllm.c b/vp9/common/vp9_idctllm.c
index 4dd540e..baa2245 100644
--- a/vp9/common/vp9_idctllm.c
+++ b/vp9/common/vp9_idctllm.c
@@ -1534,7 +1534,7 @@
#endif
#if CONFIG_TX32X32
-#if !CONFIG_DWT32X32HYBRID
+#if !CONFIG_DWTDCTHYBRID
#define DownshiftMultiplyBy2(x) x * 2
#define DownshiftMultiply(x) x
static void idct16(double *input, double *output, int stride) {
@@ -1879,7 +1879,7 @@
vp9_clear_system_state(); // Make it simd safe : __asm emms;
}
-#else // CONFIG_DWT32X32HYBRID
+#else // CONFIG_DWTDCTHYBRID
#define DWT_MAX_LENGTH 32
#define DWT_TYPE 26 // 26/53/97
@@ -1940,8 +1940,8 @@
*x++ = ((*b) << 1) + *a;
}
-void dyadic_synthesize_53(int levels, int width, int height, int16_t *c,
- int pitch_c, int16_t *x, int pitch_x) {
+static void dyadic_synthesize_53(int levels, int width, int height, int16_t *c,
+ int pitch_c, int16_t *x, int pitch_x) {
int th[16], tw[16], lv, i, j, nh, nw, hh = height, hw = width;
short buffer[2 * DWT_MAX_LENGTH];
@@ -2031,8 +2031,8 @@
}
}
-void dyadic_synthesize_26(int levels, int width, int height, int16_t *c,
- int pitch_c, int16_t *x, int pitch_x) {
+static void dyadic_synthesize_26(int levels, int width, int height, int16_t *c,
+ int pitch_c, int16_t *x, int pitch_x) {
int th[16], tw[16], lv, i, j, nh, nw, hh = height, hw = width;
int16_t buffer[2 * DWT_MAX_LENGTH];
@@ -2111,8 +2111,8 @@
x[length - 1] -= 2 * a_predict1 * x[length - 2];
}
-void dyadic_synthesize_97(int levels, int width, int height, int16_t *c,
- int pitch_c, int16_t *x, int pitch_x) {
+static void dyadic_synthesize_97(int levels, int width, int height, int16_t *c,
+ int pitch_c, int16_t *x, int pitch_x) {
int th[16], tw[16], lv, i, j, nh, nw, hh = height, hw = width;
double buffer[2 * DWT_MAX_LENGTH];
double y[DWT_MAX_LENGTH * DWT_MAX_LENGTH];
@@ -2358,7 +2358,8 @@
vp9_clear_system_state(); // Make it simd safe : __asm emms;
}
-void vp9_short_idct16x16_c_f(int16_t *input, int16_t *output, int pitch) {
+static void vp9_short_idct16x16_c_f(int16_t *input, int16_t *output, int pitch,
+ int scale) {
vp9_clear_system_state(); // Make it simd safe : __asm emms;
{
double out[16*16], out2[16*16];
@@ -2383,13 +2384,13 @@
out2[j*16 + i] = temp_out[j];
}
for (i = 0; i < 16*16; ++i)
- output[i] = round(out2[i] / (64 >> DWT_PRECISION_BITS));
+ output[i] = round(out2[i] / (128 >> scale));
}
vp9_clear_system_state(); // Make it simd safe : __asm emms;
}
void vp9_short_idct32x32_c(int16_t *input, int16_t *output, int pitch) {
- // assume out is a 32x32 buffer
+ // assume output is a 32x32 buffer
// Temporary buffer to hold a 16x16 block for 16x16 inverse dct
int16_t buffer[16 * 16];
// Temporary buffer to hold a 32x32 block for inverse 32x32 dwt
@@ -2400,20 +2401,24 @@
// TODO(debargha): Implement more efficiently by adding output pitch
// argument to the idct16x16 function
- vp9_short_idct16x16_c_f(input, buffer, pitch);
+ vp9_short_idct16x16_c_f(input, buffer, pitch,
+ 1 + DWT_PRECISION_BITS);
for (i = 0; i < 16; ++i) {
vpx_memcpy(buffer2 + i * 32, buffer + i * 16, sizeof(*buffer2) * 16);
}
- vp9_short_idct16x16_c_f(input + 16, buffer, pitch);
+ vp9_short_idct16x16_c_f(input + 16, buffer, pitch,
+ 1 + DWT_PRECISION_BITS);
for (i = 0; i < 16; ++i) {
vpx_memcpy(buffer2 + i * 32 + 16, buffer + i * 16, sizeof(*buffer2) * 16);
}
- vp9_short_idct16x16_c_f(input + 16 * short_pitch, buffer, pitch);
+ vp9_short_idct16x16_c_f(input + 16 * short_pitch, buffer, pitch,
+ 1 + DWT_PRECISION_BITS);
for (i = 0; i < 16; ++i) {
vpx_memcpy(buffer2 + i * 32 + 16 * 32, buffer + i * 16,
sizeof(*buffer2) * 16);
}
- vp9_short_idct16x16_c_f(input + 16 * short_pitch + 16, buffer, pitch);
+ vp9_short_idct16x16_c_f(input + 16 * short_pitch + 16, buffer, pitch,
+ 1 + DWT_PRECISION_BITS);
for (i = 0; i < 16; ++i) {
vpx_memcpy(buffer2 + i * 32 + 16 * 33, buffer + i * 16,
sizeof(*buffer2) * 16);
@@ -2426,5 +2431,78 @@
dyadic_synthesize_53(1, 32, 32, buffer2, 32, output, 32);
#endif
}
-#endif // CONFIG_DWT32X32HYBRID
+
+void vp9_short_idct64x64_c(int16_t *input, int16_t *output, int pitch) {
+ // assume output is a 64x64 buffer
+ // Temporary buffer to hold a 16x16 block for 16x16 inverse dct
+ int16_t buffer[16 * 16];
+ // Temporary buffer to hold a 32x32 block for inverse 32x32 dwt
+ int16_t buffer2[64 * 64];
+ // Note: pitch is in bytes, short_pitch is in short units
+ const int short_pitch = pitch >> 1;
+ int i, j;
+
+ // TODO(debargha): Implement more efficiently by adding output pitch
+ // argument to the idct16x16 function
+ vp9_short_idct16x16_c_f(input, buffer, pitch,
+ 2 + DWT_PRECISION_BITS);
+ for (i = 0; i < 16; ++i) {
+ vpx_memcpy(buffer2 + i * 64, buffer + i * 16, sizeof(*buffer2) * 16);
+ }
+ vp9_short_idct16x16_c_f(input + 16, buffer, pitch,
+ 2 + DWT_PRECISION_BITS);
+ for (i = 0; i < 16; ++i) {
+ vpx_memcpy(buffer2 + i * 64 + 16, buffer + i * 16, sizeof(*buffer2) * 16);
+ }
+ vp9_short_idct16x16_c_f(input + 16 * short_pitch, buffer, pitch,
+ 2 + DWT_PRECISION_BITS);
+ for (i = 0; i < 16; ++i) {
+ vpx_memcpy(buffer2 + i * 64 + 16 * 64, buffer + i * 16,
+ sizeof(*buffer2) * 16);
+ }
+ vp9_short_idct16x16_c_f(input + 16 * short_pitch + 16, buffer, pitch,
+ 2 + DWT_PRECISION_BITS);
+ for (i = 0; i < 16; ++i) {
+ vpx_memcpy(buffer2 + i * 64 + 16 * 65, buffer + i * 16,
+ sizeof(*buffer2) * 16);
+ }
+
+ // Copying and scaling highest bands into buffer2
+#if DWT_PRECISION_BITS < 1
+ for (i = 0; i < 32; ++i) {
+ for (j = 0; j < 32; ++j) {
+ buffer2[i * 64 + 32 + j] =
+ input[i * short_pitch + 32 + j] >> (1 - DWT_PRECISION_BITS);
+ }
+ }
+ for (i = 0; i < 32; ++i) {
+ for (j = 0; j < 64; ++j) {
+ buffer2[i * 64 + j] =
+ input[(i + 32) * short_pitch + j] >> (1 - DWT_PRECISION_BITS);
+ }
+ }
+#else
+ for (i = 0; i < 32; ++i) {
+ for (j = 0; j < 32; ++j) {
+ buffer2[i * 64 + 32 + j] =
+ input[i * short_pitch + 32 + j] << (DWT_PRECISION_BITS - 1);
+ }
+ }
+ for (i = 0; i < 32; ++i) {
+ for (j = 0; j < 64; ++j) {
+ buffer2[i * 64 + j] =
+ input[(i + 32) * short_pitch + j] << (DWT_PRECISION_BITS - 1);
+ }
+ }
+#endif
+
+#if DWT_TYPE == 26
+ dyadic_synthesize_26(2, 64, 64, buffer2, 64, output, 64);
+#elif DWT_TYPE == 97
+ dyadic_synthesize_97(2, 64, 64, buffer2, 64, output, 64);
+#elif DWT_TYPE == 53
+ dyadic_synthesize_53(2, 64, 64, buffer2, 64, output, 64);
+#endif
+}
+#endif // CONFIG_DWTDCTHYBRID
#endif // CONFIG_TX32X32
diff --git a/vp9/encoder/vp9_dct.c b/vp9/encoder/vp9_dct.c
index e14421d..0de6393 100644
--- a/vp9/encoder/vp9_dct.c
+++ b/vp9/encoder/vp9_dct.c
@@ -1332,8 +1332,9 @@
#undef ROUNDING
#endif
+#if CONFIG_TX32X32 || CONFIG_TX64X64
+#if !CONFIG_DWTDCTHYBRID
#if CONFIG_TX32X32
-#if !CONFIG_DWT32X32HYBRID
static void dct32_1d(double *input, double *output, int stride) {
static const double C1 = 0.998795456205; // cos(pi * 1 / 64)
static const double C2 = 0.995184726672; // cos(pi * 2 / 64)
@@ -1684,8 +1685,9 @@
vp9_clear_system_state(); // Make it simd safe : __asm emms;
}
+#endif // CONFIG_TX32X32
-#else // CONFIG_DWT32X32HYBRID
+#else // CONFIG_DWTDCTHYBRID
#define DWT_MAX_LENGTH 64
#define DWT_TYPE 26 // 26/53/97
@@ -2108,7 +2110,8 @@
vp9_clear_system_state(); // Make it simd safe : __asm emms;
}
-void vp9_short_fdct16x16_c_f(short *input, short *out, int pitch) {
+static void vp9_short_fdct16x16_c_f(short *input, short *out, int pitch,
+ int scale) {
vp9_clear_system_state(); // Make it simd safe : __asm emms;
{
int shortpitch = pitch >> 1;
@@ -2134,11 +2137,12 @@
}
// Scale by some magic number
for (i = 0; i < 256; i++)
- out[i] = (short)round(output[i] / (4 << DWT_PRECISION_BITS));
+ out[i] = (short)round(output[i] / (2 << scale));
}
vp9_clear_system_state(); // Make it simd safe : __asm emms;
}
+#if CONFIG_TX32X32
void vp9_short_fdct32x32_c(short *input, short *out, int pitch) {
// assume out is a 32x32 buffer
short buffer[16 * 16];
@@ -2153,21 +2157,82 @@
#endif
// TODO(debargha): Implement more efficiently by adding output pitch
// argument to the dct16x16 function
- vp9_short_fdct16x16_c_f(out, buffer, 64);
+ vp9_short_fdct16x16_c_f(out, buffer, 64, 1 + DWT_PRECISION_BITS);
for (i = 0; i < 16; ++i)
vpx_memcpy(out + i * 32, buffer + i * 16, sizeof(short) * 16);
- vp9_short_fdct16x16_c_f(out + 16, buffer, 64);
+ vp9_short_fdct16x16_c_f(out + 16, buffer, 64, 1 + DWT_PRECISION_BITS);
for (i = 0; i < 16; ++i)
vpx_memcpy(out + i * 32 + 16, buffer + i * 16, sizeof(short) * 16);
- vp9_short_fdct16x16_c_f(out + 32 * 16, buffer, 64);
+ vp9_short_fdct16x16_c_f(out + 32 * 16, buffer, 64, 1 + DWT_PRECISION_BITS);
for (i = 0; i < 16; ++i)
vpx_memcpy(out + i * 32 + 32 * 16, buffer + i * 16, sizeof(short) * 16);
- vp9_short_fdct16x16_c_f(out + 33 * 16, buffer, 64);
+ vp9_short_fdct16x16_c_f(out + 33 * 16, buffer, 64, 1 + DWT_PRECISION_BITS);
for (i = 0; i < 16; ++i)
vpx_memcpy(out + i * 32 + 33 * 16, buffer + i * 16, sizeof(short) * 16);
}
-#endif // CONFIG_DWT32X32HYBRID
#endif // CONFIG_TX32X32
+
+#if CONFIG_TX64X64
+void vp9_short_fdct64x64_c(short *input, short *out, int pitch) {
+ // assume out is a 64x64 buffer
+ short buffer[16 * 16];
+ int i, j;
+ const int short_pitch = pitch >> 1;
+#if DWT_TYPE == 26
+ dyadic_analyze_26(2, 64, 64, input, short_pitch, out, 64);
+#elif DWT_TYPE == 97
+ dyadic_analyze_97(2, 64, 64, input, short_pitch, out, 64);
+#elif DWT_TYPE == 53
+ dyadic_analyze_53(2, 64, 64, input, short_pitch, out, 64);
+#endif
+ // TODO(debargha): Implement more efficiently by adding output pitch
+ // argument to the dct16x16 function
+ vp9_short_fdct16x16_c_f(out, buffer, 128, 2 + DWT_PRECISION_BITS);
+ for (i = 0; i < 16; ++i)
+ vpx_memcpy(out + i * 64, buffer + i * 16, sizeof(short) * 16);
+
+ vp9_short_fdct16x16_c_f(out + 16, buffer, 128, 2 + DWT_PRECISION_BITS);
+ for (i = 0; i < 16; ++i)
+ vpx_memcpy(out + i * 64 + 16, buffer + i * 16, sizeof(short) * 16);
+
+ vp9_short_fdct16x16_c_f(out + 64 * 16, buffer, 128, 2 + DWT_PRECISION_BITS);
+ for (i = 0; i < 16; ++i)
+ vpx_memcpy(out + i * 64 + 64 * 16, buffer + i * 16, sizeof(short) * 16);
+
+ vp9_short_fdct16x16_c_f(out + 65 * 16, buffer, 128, 2 + DWT_PRECISION_BITS);
+ for (i = 0; i < 16; ++i)
+ vpx_memcpy(out + i * 64 + 65 * 16, buffer + i * 16, sizeof(short) * 16);
+
+ // There is no dct used on the highest bands for now.
+ // Need to scale these coeffs by a factor of 2/2^DWT_PRECISION_BITS
+ // TODO(debargha): experiment with turning these coeffs to 0
+#if DWT_PRECISION_BITS < 1
+ for (i = 0; i < 32; ++i) {
+ for (j = 0; j < 32; ++j) {
+ out[i * 64 + 32 + j] <<= (1 - DWT_PRECISION_BITS);
+ }
+ }
+ for (i = 0; i < 32; ++i) {
+ for (j = 0; j < 64; ++j) {
+ out[i * 64 + j] <<= (1 - DWT_PRECISION_BITS);
+ }
+ }
+#else
+ for (i = 0; i < 32; ++i) {
+ for (j = 0; j < 32; ++j) {
+ out[i * 64 + 32 + j] >>= (DWT_PRECISION_BITS - 1);
+ }
+ }
+ for (i = 0; i < 32; ++i) {
+ for (j = 0; j < 64; ++j) {
+ out[i * 64 + j] >>= (DWT_PRECISION_BITS - 1);
+ }
+ }
+#endif
+}
+#endif // CONFIG_TX64X64
+#endif // CONFIG_DWTDCTHYBRID
+#endif // CONFIG_TX32X32 || CONFIG_TX64X64
diff --git a/vp9/encoder/vp9_rdopt.c b/vp9/encoder/vp9_rdopt.c
index c695c04..956d8f9 100644
--- a/vp9/encoder/vp9_rdopt.c
+++ b/vp9/encoder/vp9_rdopt.c
@@ -965,17 +965,17 @@
SUPERBLOCK * const x_sb = &x->sb_coeff_data;
MACROBLOCKD * const xd = &x->e_mbd;
SUPERBLOCKD * const xd_sb = &xd->sb_coeff_data;
-#if DEBUG_ERROR || CONFIG_DWT32X32HYBRID
+#if DEBUG_ERROR || CONFIG_DWTDCTHYBRID
int16_t out[1024];
#endif
vp9_transform_sby_32x32(x);
vp9_quantize_sby_32x32(x);
-#if DEBUG_ERROR || CONFIG_DWT32X32HYBRID
+#if DEBUG_ERROR || CONFIG_DWTDCTHYBRID
vp9_short_idct32x32(xd_sb->dqcoeff, out, 64);
#endif
-#if !CONFIG_DWT32X32HYBRID
+#if !CONFIG_DWTDCTHYBRID
*distortion = vp9_sb_block_error_c(x_sb->coeff, xd_sb->dqcoeff, 1024);
#else
*distortion = vp9_block_error_c(x_sb->src_diff, out, 1024) << 4;
