diff options
Diffstat (limited to 'libvpx/vp9/common/vp9_idct.c')
| -rw-r--r-- | libvpx/vp9/common/vp9_idct.c | 31 |
1 files changed, 17 insertions, 14 deletions
diff --git a/libvpx/vp9/common/vp9_idct.c b/libvpx/vp9/common/vp9_idct.c index e3a088e28..69069042c 100644 --- a/libvpx/vp9/common/vp9_idct.c +++ b/libvpx/vp9/common/vp9_idct.c @@ -156,6 +156,8 @@ void vp9_idct16x16_add(const tran_low_t *input, uint8_t *dest, int stride, vpx_idct16x16_1_add(input, dest, stride); else if (eob <= 10) vpx_idct16x16_10_add(input, dest, stride); + else if (eob <= 38) + vpx_idct16x16_38_add(input, dest, stride); else vpx_idct16x16_256_add(input, dest, stride); } @@ -203,7 +205,7 @@ void vp9_iht16x16_add(TX_TYPE tx_type, const tran_low_t *input, uint8_t *dest, #if CONFIG_VP9_HIGHBITDEPTH -void vp9_highbd_iht4x4_16_add_c(const tran_low_t *input, uint8_t *dest8, +void vp9_highbd_iht4x4_16_add_c(const tran_low_t *input, uint16_t *dest, int stride, int tx_type, int bd) { const highbd_transform_2d IHT_4[] = { { vpx_highbd_idct4_c, vpx_highbd_idct4_c }, // DCT_DCT = 0 @@ -211,7 +213,6 @@ void vp9_highbd_iht4x4_16_add_c(const tran_low_t *input, uint8_t *dest8, { vpx_highbd_idct4_c, vpx_highbd_iadst4_c }, // DCT_ADST = 2 { vpx_highbd_iadst4_c, vpx_highbd_iadst4_c } // ADST_ADST = 3 }; - uint16_t *dest = CONVERT_TO_SHORTPTR(dest8); int i, j; tran_low_t out[4 * 4]; @@ -243,14 +244,13 @@ static const highbd_transform_2d HIGH_IHT_8[] = { { vpx_highbd_iadst8_c, vpx_highbd_iadst8_c } // ADST_ADST = 3 }; -void vp9_highbd_iht8x8_64_add_c(const tran_low_t *input, uint8_t *dest8, +void vp9_highbd_iht8x8_64_add_c(const tran_low_t *input, uint16_t *dest, int stride, int tx_type, int bd) { int i, j; tran_low_t out[8 * 8]; tran_low_t *outptr = out; tran_low_t temp_in[8], temp_out[8]; const highbd_transform_2d ht = HIGH_IHT_8[tx_type]; - uint16_t *dest = CONVERT_TO_SHORTPTR(dest8); // Inverse transform row vectors. for (i = 0; i < 8; ++i) { @@ -277,14 +277,13 @@ static const highbd_transform_2d HIGH_IHT_16[] = { { vpx_highbd_iadst16_c, vpx_highbd_iadst16_c } // ADST_ADST = 3 }; -void vp9_highbd_iht16x16_256_add_c(const tran_low_t *input, uint8_t *dest8, +void vp9_highbd_iht16x16_256_add_c(const tran_low_t *input, uint16_t *dest, int stride, int tx_type, int bd) { int i, j; tran_low_t out[16 * 16]; tran_low_t *outptr = out; tran_low_t temp_in[16], temp_out[16]; const highbd_transform_2d ht = HIGH_IHT_16[tx_type]; - uint16_t *dest = CONVERT_TO_SHORTPTR(dest8); // Rows for (i = 0; i < 16; ++i) { @@ -305,7 +304,7 @@ void vp9_highbd_iht16x16_256_add_c(const tran_low_t *input, uint8_t *dest8, } // idct -void vp9_highbd_idct4x4_add(const tran_low_t *input, uint8_t *dest, int stride, +void vp9_highbd_idct4x4_add(const tran_low_t *input, uint16_t *dest, int stride, int eob, int bd) { if (eob > 1) vpx_highbd_idct4x4_16_add(input, dest, stride, bd); @@ -313,7 +312,7 @@ void vp9_highbd_idct4x4_add(const tran_low_t *input, uint8_t *dest, int stride, vpx_highbd_idct4x4_1_add(input, dest, stride, bd); } -void vp9_highbd_iwht4x4_add(const tran_low_t *input, uint8_t *dest, int stride, +void vp9_highbd_iwht4x4_add(const tran_low_t *input, uint16_t *dest, int stride, int eob, int bd) { if (eob > 1) vpx_highbd_iwht4x4_16_add(input, dest, stride, bd); @@ -321,7 +320,7 @@ void vp9_highbd_iwht4x4_add(const tran_low_t *input, uint8_t *dest, int stride, vpx_highbd_iwht4x4_1_add(input, dest, stride, bd); } -void vp9_highbd_idct8x8_add(const tran_low_t *input, uint8_t *dest, int stride, +void vp9_highbd_idct8x8_add(const tran_low_t *input, uint16_t *dest, int stride, int eob, int bd) { // If dc is 1, then input[0] is the reconstructed value, do not need // dequantization. Also, when dc is 1, dc is counted in eobs, namely eobs >=1. @@ -338,7 +337,7 @@ void vp9_highbd_idct8x8_add(const tran_low_t *input, uint8_t *dest, int stride, } } -void vp9_highbd_idct16x16_add(const tran_low_t *input, uint8_t *dest, +void vp9_highbd_idct16x16_add(const tran_low_t *input, uint16_t *dest, int stride, int eob, int bd) { // The calculation can be simplified if there are not many non-zero dct // coefficients. Use eobs to separate different cases. @@ -347,18 +346,22 @@ void vp9_highbd_idct16x16_add(const tran_low_t *input, uint8_t *dest, vpx_highbd_idct16x16_1_add(input, dest, stride, bd); } else if (eob <= 10) { vpx_highbd_idct16x16_10_add(input, dest, stride, bd); + } else if (eob <= 38) { + vpx_highbd_idct16x16_38_add(input, dest, stride, bd); } else { vpx_highbd_idct16x16_256_add(input, dest, stride, bd); } } -void vp9_highbd_idct32x32_add(const tran_low_t *input, uint8_t *dest, +void vp9_highbd_idct32x32_add(const tran_low_t *input, uint16_t *dest, int stride, int eob, int bd) { // Non-zero coeff only in upper-left 8x8 if (eob == 1) { vpx_highbd_idct32x32_1_add(input, dest, stride, bd); } else if (eob <= 34) { vpx_highbd_idct32x32_34_add(input, dest, stride, bd); + } else if (eob <= 135) { + vpx_highbd_idct32x32_135_add(input, dest, stride, bd); } else { vpx_highbd_idct32x32_1024_add(input, dest, stride, bd); } @@ -366,7 +369,7 @@ void vp9_highbd_idct32x32_add(const tran_low_t *input, uint8_t *dest, // iht void vp9_highbd_iht4x4_add(TX_TYPE tx_type, const tran_low_t *input, - uint8_t *dest, int stride, int eob, int bd) { + uint16_t *dest, int stride, int eob, int bd) { if (tx_type == DCT_DCT) vp9_highbd_idct4x4_add(input, dest, stride, eob, bd); else @@ -374,7 +377,7 @@ void vp9_highbd_iht4x4_add(TX_TYPE tx_type, const tran_low_t *input, } void vp9_highbd_iht8x8_add(TX_TYPE tx_type, const tran_low_t *input, - uint8_t *dest, int stride, int eob, int bd) { + uint16_t *dest, int stride, int eob, int bd) { if (tx_type == DCT_DCT) { vp9_highbd_idct8x8_add(input, dest, stride, eob, bd); } else { @@ -383,7 +386,7 @@ void vp9_highbd_iht8x8_add(TX_TYPE tx_type, const tran_low_t *input, } void vp9_highbd_iht16x16_add(TX_TYPE tx_type, const tran_low_t *input, - uint8_t *dest, int stride, int eob, int bd) { + uint16_t *dest, int stride, int eob, int bd) { if (tx_type == DCT_DCT) { vp9_highbd_idct16x16_add(input, dest, stride, eob, bd); } else { |