diff options
| author | Jingning Han <jingning@google.com> | 2013-09-05 14:09:40 -0700 |
|---|---|---|
| committer | Gerrit Code Review <gerrit@gerrit.golo.chromium.org> | 2013-09-05 14:09:40 -0700 |
| commit | 1c263d69188ea5b707093b6e1f099ab2aca9dfae (patch) | |
| tree | 824348b2a9d5dfd5dbeaaf6f7322a257ee0db7d7 /vp9 | |
| parent | 2156ccaa4a7a6fb25be1d34ad611f6e9b9ad3867 (diff) | |
| parent | 458c2833c00de1d52f466eea3fb8f29067274006 (diff) | |
| download | libvpx-1c263d69188ea5b707093b6e1f099ab2aca9dfae.tar libvpx-1c263d69188ea5b707093b6e1f099ab2aca9dfae.tar.gz libvpx-1c263d69188ea5b707093b6e1f099ab2aca9dfae.tar.bz2 libvpx-1c263d69188ea5b707093b6e1f099ab2aca9dfae.zip | |
Merge "Use saturated addition in SSSE3 of 32x32 quant"
Diffstat (limited to 'vp9')
| -rw-r--r-- | vp9/common/vp9_rtcd_defs.sh | 2 | ||||
| -rw-r--r-- | vp9/encoder/vp9_quantize.c | 1 | ||||
| -rw-r--r-- | vp9/encoder/x86/vp9_quantize_ssse3.asm | 11 |
3 files changed, 12 insertions, 2 deletions
diff --git a/vp9/common/vp9_rtcd_defs.sh b/vp9/common/vp9_rtcd_defs.sh index c2777aa51..f5eeb2c51 100644 --- a/vp9/common/vp9_rtcd_defs.sh +++ b/vp9/common/vp9_rtcd_defs.sh @@ -701,7 +701,7 @@ prototype void vp9_quantize_b "int16_t *coeff_ptr, intptr_t n_coeffs, int skip_b specialize vp9_quantize_b $ssse3_x86_64 prototype void vp9_quantize_b_32x32 "int16_t *coeff_ptr, intptr_t n_coeffs, int skip_block, int16_t *zbin_ptr, int16_t *round_ptr, int16_t *quant_ptr, int16_t *quant_shift_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, int16_t *dequant_ptr, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan" -specialize vp9_quantize_b_32x32 # $ssse3_x86_64 FIXME(jingning): need a unit test on thisbefore enabled +specialize vp9_quantize_b_32x32 $ssse3_x86_64 # # Structured Similarity (SSIM) diff --git a/vp9/encoder/vp9_quantize.c b/vp9/encoder/vp9_quantize.c index fb0e4707a..96abeff38 100644 --- a/vp9/encoder/vp9_quantize.c +++ b/vp9/encoder/vp9_quantize.c @@ -135,6 +135,7 @@ void vp9_quantize_b_32x32_c(int16_t *coeff_ptr, intptr_t n_coeffs, if (x >= zbin) { x += ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1); + x = clamp(x, INT16_MIN, INT16_MAX); y = (((int)(((int)(x * quant_ptr[rc != 0]) >> 16) + x)) * quant_shift_ptr[rc != 0]) >> 15; // quantize (x) diff --git a/vp9/encoder/x86/vp9_quantize_ssse3.asm b/vp9/encoder/x86/vp9_quantize_ssse3.asm index 7deb9815a..ae0d6cd3c 100644 --- a/vp9/encoder/x86/vp9_quantize_ssse3.asm +++ b/vp9/encoder/x86/vp9_quantize_ssse3.asm @@ -70,9 +70,15 @@ cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, zbin, round, quant, \ pcmpgtw m7, m6, m0 ; m7 = c[i] >= zbin punpckhqdq m0, m0 pcmpgtw m12, m11, m0 ; m12 = c[i] >= zbin +%ifidn %1, b_32x32 + paddsw m6, m1 + punpckhqdq m1, m1 + paddsw m11, m1 +%else paddw m6, m1 ; m6 += round punpckhqdq m1, m1 paddw m11, m1 ; m11 += round +%endif pmulhw m8, m6, m2 ; m8 = m6*q>>16 punpckhqdq m2, m2 pmulhw m13, m11, m2 ; m13 = m11*q>>16 @@ -126,9 +132,12 @@ cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, zbin, round, quant, \ pmovmskb r2, m12 or r6, r2 jz .skip_iter -%endif + paddsw m6, m1 + paddsw m11, m1 +%else paddw m6, m1 ; m6 += round paddw m11, m1 ; m11 += round +%endif pmulhw m14, m6, m2 ; m14 = m6*q>>16 pmulhw m13, m11, m2 ; m13 = m11*q>>16 paddw m14, m6 ; m14 += m6 |