[Libva] [PATCH 5/7] Enabling HSW avc encoding configuration
Zhao, Yakui
yakui.zhao at intel.com
Mon Jul 28 17:54:23 PDT 2014
On Mon, 2014-07-28 at 01:02 -0600, Li, Zhong wrote:
> On 07/28/2014 02:32 PM, Zhao, Yakui wrote:
> > On Thu, 2014-07-24 at 03:47 -0600, Zhong Li wrote:
> >> Signed-off-by: Zhong Li <zhong.li at intel.com>
> >> ---
> >> src/gen75_vme.c | 22 ++++++----
> >> src/shaders/vme/inter_frame_haswell.asm | 69 +++++++++++++++++++-----------
> >> src/shaders/vme/inter_frame_haswell.g75b | 52 ++++++++++++++--------
> >> src/shaders/vme/vme75.inc | 8 +++-
> >> 4 files changed, 97 insertions(+), 54 deletions(-)
> >>
> >> diff --git a/src/gen75_vme.c b/src/gen75_vme.c
> >> index e8527c3..576e91a 100644
> >> --- a/src/gen75_vme.c
> >> +++ b/src/gen75_vme.c
> >> @@ -534,7 +534,7 @@ gen75_vme_fill_vme_batchbuffer(VADriverContextP ctx,
> >>
> >> /*inline data */
> >> *command_ptr++ = (mb_width << 16 | mb_y << 8 | mb_x);
> >> - *command_ptr++ = ( (1 << 16) | transform_8x8_mode_flag | (mb_intra_ub << 8));
> >> + *command_ptr++ = ((encoder_context->quality_level << 24) | (1 << 16) | transform_8x8_mode_flag | (mb_intra_ub << 8));
> >>
> >> i += 1;
> >> }
> >> @@ -578,14 +578,20 @@ static void gen75_vme_pipeline_programing(VADriverContextP ctx,
> >> int kernel_shader;
> >> bool allow_hwscore = true;
> >> int s;
> >> -
> >> - for (s = 0; s < encode_state->num_slice_params_ext; s++) {
> >> - pSliceParameter = (VAEncSliceParameterBufferH264 *)encode_state->slice_params_ext[s]->buffer;
> >> - if ((pSliceParameter->macroblock_address % width_in_mbs)) {
> >> - allow_hwscore = false;
> >> - break;
> >> - }
> >> + unsigned int is_low_quality = (encoder_context->quality_level == ENCODER_LOW_QUALITY);
> >> +
> >> + if (is_low_quality)
> >> + allow_hwscore = false;
> >> + else {
> >> + for (s = 0; s < encode_state->num_slice_params_ext; s++) {
> >> + pSliceParameter = (VAEncSliceParameterBufferH264 *)encode_state->slice_params_ext[s]->buffer;
> >> + if ((pSliceParameter->macroblock_address % width_in_mbs)) {
> >> + allow_hwscore = false;
> >> + break;
> >> + }
> >> + }
> >> }
> >> +
> >> if ((pSliceParameter->slice_type == SLICE_TYPE_I) ||
> >> (pSliceParameter->slice_type == SLICE_TYPE_I)) {
> >> kernel_shader = VME_INTRA_SHADER;
> >> diff --git a/src/shaders/vme/inter_frame_haswell.asm b/src/shaders/vme/inter_frame_haswell.asm
> >> index bcfd260..399125a 100644
> >> --- a/src/shaders/vme/inter_frame_haswell.asm
> >> +++ b/src/shaders/vme/inter_frame_haswell.asm
> >> @@ -475,24 +475,58 @@ send (16)
> >> {align1};
> >>
> >> /* IME search */
> >> -mov (1) vme_m0.12<1>:UD SEARCH_CTRL_SINGLE + INTER_PART_MASK + INTER_SAD_HAAR:UD {align1}; /* 16x16 Source, harr */
> >> -mov (1) vme_m0.22<1>:UW REF_REGION_SIZE {align1}; /* Reference Width&Height, 48x40 */
> >> -
> >> -mov (1) vme_m0.0<1>:UD vme_m0.8<0,1,0>:UD {align1};
> >> +cmp.z.f0.0 (1) null<1>:uw quality_level_ub<0,1,0>:ub LOW_QUALITY_LEVEL:uw {align1};
> > Although this is not wrong, it will be better that the and operation is
> > firstly executed with the corresponding bit-masks. It can avoid that too
> > many bits are wasted.
> Thanks for your comments, but can you explain which bits are wasted? Do
> you mean register bits ?
> I use cmp because quality_level is not always 2^N, it may be random
> value. I think cmp is convenient, and more robust than and.
What I mean is that and operation is firstly executed and then followed
by cmp. In such case it is enough to define several bits for the
supported configurable level. The remaining bits can still be reserved
for future use.
> >> +(f0.0) jmpi (1) __low_quality_search;
> >>
> >> -add (1) vme_m0.0<1>:W vme_m0.0<0,1,0>:W -16:W {align1}; /* Reference = (x-16,y-12)-(x+32,y+28) */
> >> -add (1) vme_m0.2<1>:W vme_m0.2<0,1,0>:W -12:W {align1};
> >> +__high_quality_search:
> >> +/* M3/M4 search path */
> >> +mov (1) vme_msg_3.0<1>:UD 0x01010101:UD {align1};
> >> +mov (1) vme_msg_3.4<1>:UD 0x10010101:UD {align1};
> >> +mov (1) vme_msg_3.8<1>:UD 0x0F0F0F0F:UD {align1};
> >> +mov (1) vme_msg_3.12<1>:UD 0x100F0F0F:UD {align1};
> >> +mov (1) vme_msg_3.16<1>:UD 0x01010101:UD {align1};
> >> +mov (1) vme_msg_3.20<1>:UD 0x10010101:UD {align1};
> >> +mov (1) vme_msg_3.24<1>:UD 0x0F0F0F0F:UD {align1};
> >> +mov (1) vme_msg_3.28<1>:UD 0x100F0F0F:UD {align1};
> >> +mov (1) vme_msg_4.0<1>:UD 0x01010101:UD {align1};
> >> +mov (1) vme_msg_4.4<1>:UD 0x10010101:UD {align1};
> >> +mov (1) vme_msg_4.8<1>:UD 0x0F0F0F0F:UD {align1};
> >> +mov (1) vme_msg_4.12<1>:UD 0x000F0F0F:UD {align1};
> >> +mov (4) vme_msg_4.16<1>:UD 0x0:UD {align1};
> >>
> >> +mov (1) vme_m0.12<1>:UD SEARCH_CTRL_SINGLE + INTER_PART_MASK + INTER_SAD_HAAR:UD {align1}; /* 16x16 Source, harr */
> >> +mov (1) vme_m0.22<1>:UW REF_REGION_SIZE {align1}; /* Reference Width&Height, 48x40 */
> >> mov (1) vme_m0.0<1>:W -16:W {align1};
> >> mov (1) vme_m0.2<1>:W -12:W {align1};
> >>
> >> -mov (1) vme_m0.4<1>:UD vme_m0.0<0,1,0>:UD {align1};
> >> -
> >> and.z.f0.0 (1) null:uw input_mb_intra_ub<0,1,0>:ub INTRA_PRED_AVAIL_FLAG_AE:uw {align1};
> >> (f0.0) add (1) vme_m0.0<1>:w vme_m0.0<0,1,0>:w 12:w {align1};
> >> and.z.f0.0 (1) null:uw input_mb_intra_ub<0,1,0>:ub INTRA_PRED_AVAIL_FLAG_B:uw {align1};
> >> (f0.0) add (1) vme_m0.2<1>:w vme_m0.2<0,1,0>:w 8:w {align1};
> >> -
> >> +
> >> +jmpi (1) __vme_msg;
> >> +
> >> +__low_quality_search:
> >> +/* M3/M4 search path */
> >> +mov (1) vme_msg_3.0<1>:UD 0x10010101:UD {align1};
> >> +mov (1) vme_msg_3.4<1>:UD 0x100F0F0F:UD {align1};
> >> +mov (1) vme_msg_3.8<1>:UD 0x10010101:UD {align1};
> >> +mov (1) vme_msg_3.12<1>:UD 0x000F0F0F:UD {align1};
> >> +mov (4) vme_msg_3.16<1>:UD 0x0:UD {align1};
> >> +mov (8) vme_msg_4.16<1>:UD 0x0:UD {align1};
> >> +
> >> +mov (1) vme_m0.12<1>:UD SEARCH_CTRL_SINGLE + INTER_PART_MASK + INTER_SAD_HAAR:UD {align1}; /* 16x16 Source, harr */
> >> +mov (1) vme_m0.22<1>:UW MIN_REF_REGION_SIZE {align1}; /* Reference Width&Height, 32x32 */
> >> +mov (1) vme_m0.0<1>:W -8:W {align1};
> >> +mov (1) vme_m0.2<1>:W -8:W {align1};
> >> +
> >> +and.z.f0.0 (1) null:uw input_mb_intra_ub<0,1,0>:ub INTRA_PRED_AVAIL_FLAG_AE:uw {align1};
> >> +(f0.0) add (1) vme_m0.0<1>:w vme_m0.0<0,1,0>:w 4:w {align1};
> >> +and.z.f0.0 (1) null:uw input_mb_intra_ub<0,1,0>:ub INTRA_PRED_AVAIL_FLAG_B:uw {align1};
> >> +(f0.0) add (1) vme_m0.2<1>:w vme_m0.2<0,1,0>:w 4:w {align1};
> >> +
> >> +__vme_msg:
> >> +mov (1) vme_m0.4<1>:UD vme_m0.0<0,1,0>:UD {align1};
> >> add (2) vme_m0.0<1>:w vme_m0.0<2,2,1>:w mb_ref_win.16<2,2,1>:w {align1};
> >> add (2) vme_m0.4<1>:w vme_m0.4<2,2,1>:w mb_ref_win.16<2,2,1>:w {align1};
> >> mov (8) vme_msg_0.0<1>:UD vme_m0.0<8,8,1>:UD {align1};
> >> @@ -507,23 +541,6 @@ mov (1) vme_m1.20<1>:ud mb_mvp_ref.0<0,1,0>:ud {align1};
> >> mov (8) vme_msg_1.0<1>:UD vme_m1.0<8,8,1>:UD {align1};
> >>
> >> mov (8) vme_msg_2<1>:UD vme_m2.0<8,8,1>:UD {align1};
> >> -/* M3/M4 search path */
> >> -
> >> -mov (1) vme_msg_3.0<1>:UD 0x01010101:UD {align1};
> >> -mov (1) vme_msg_3.4<1>:UD 0x10010101:UD {align1};
> >> -mov (1) vme_msg_3.8<1>:UD 0x0F0F0F0F:UD {align1};
> >> -mov (1) vme_msg_3.12<1>:UD 0x100F0F0F:UD {align1};
> >> -mov (1) vme_msg_3.16<1>:UD 0x01010101:UD {align1};
> >> -mov (1) vme_msg_3.20<1>:UD 0x10010101:UD {align1};
> >> -mov (1) vme_msg_3.24<1>:UD 0x0F0F0F0F:UD {align1};
> >> -mov (1) vme_msg_3.28<1>:UD 0x100F0F0F:UD {align1};
> >> -
> >> -mov (1) vme_msg_4.0<1>:UD 0x01010101:UD {align1};
> >> -mov (1) vme_msg_4.4<1>:UD 0x10010101:UD {align1};
> >> -mov (1) vme_msg_4.8<1>:UD 0x0F0F0F0F:UD {align1};
> >> -mov (1) vme_msg_4.12<1>:UD 0x000F0F0F:UD {align1};
> >> -
> >> -mov (4) vme_msg_4.16<1>:UD 0x0:UD {align1};
> >>
> >> send (8)
> >> vme_msg_ind
> >> diff --git a/src/shaders/vme/inter_frame_haswell.g75b b/src/shaders/vme/inter_frame_haswell.g75b
> >> index 61551f9..1a60c51 100644
> >> --- a/src/shaders/vme/inter_frame_haswell.g75b
> >> +++ b/src/shaders/vme/inter_frame_haswell.g75b
> >> @@ -145,13 +145,13 @@
> >> { 0x00000001, 0x2fa401ad, 0x00000b04, 0x00000000 },
> >> { 0x00000001, 0x2fa801ad, 0x00000b24, 0x00000000 },
> >> { 0x00000040, 0x2fe00c01, 0x00001400, 0x00000020 },
> >> - { 0x00000020, 0x34001c00, 0x00001400, 0x00000850 },
> >> + { 0x00000020, 0x34001c00, 0x00001400, 0x00000930 },
> >> { 0x00000001, 0x2ac001ad, 0x00000fe4, 0x00000000 },
> >> { 0x00000001, 0x2fa001ad, 0x00000ae6, 0x00000000 },
> >> { 0x00000001, 0x2fa401ad, 0x00000b06, 0x00000000 },
> >> { 0x00000001, 0x2fa801ad, 0x00000b26, 0x00000000 },
> >> { 0x00000040, 0x2fe00c01, 0x00001400, 0x00000020 },
> >> - { 0x00000020, 0x34001c00, 0x00001400, 0x000007f0 },
> >> + { 0x00000020, 0x34001c00, 0x00001400, 0x000008d0 },
> >> { 0x00000001, 0x2ac201ad, 0x00000fe4, 0x00000000 },
> >> { 0x0020000c, 0x2a803dad, 0x00450ac0, 0x00020002 },
> >> { 0x00200040, 0x2a883dad, 0x00450a80, 0x00030003 },
> >> @@ -192,18 +192,45 @@
> >> { 0x00000001, 0x28380021, 0x0000019c, 0x00000000 },
> >> { 0x00000001, 0x283c0021, 0x00000488, 0x00000000 },
> >> { 0x0a800031, 0x20001cac, 0x00000800, 0x040a0203 },
> >> + { 0x01000010, 0x20002e28, 0x000000a7, 0x00020002 },
> >> + { 0x00010020, 0x34001c00, 0x00001400, 0x00000160 },
> >> + { 0x00000001, 0x28600061, 0x00000000, 0x01010101 },
> >> + { 0x00000001, 0x28640061, 0x00000000, 0x10010101 },
> >> + { 0x00000001, 0x28680061, 0x00000000, 0x0f0f0f0f },
> >> + { 0x00000001, 0x286c0061, 0x00000000, 0x100f0f0f },
> >> + { 0x00000001, 0x28700061, 0x00000000, 0x01010101 },
> >> + { 0x00000001, 0x28740061, 0x00000000, 0x10010101 },
> >> + { 0x00000001, 0x28780061, 0x00000000, 0x0f0f0f0f },
> >> + { 0x00000001, 0x287c0061, 0x00000000, 0x100f0f0f },
> >> + { 0x00000001, 0x28800061, 0x00000000, 0x01010101 },
> >> + { 0x00000001, 0x28840061, 0x00000000, 0x10010101 },
> >> + { 0x00000001, 0x28880061, 0x00000000, 0x0f0f0f0f },
> >> + { 0x00000001, 0x288c0061, 0x00000000, 0x000f0f0f },
> >> + { 0x00400001, 0x28900061, 0x00000000, 0x00000000 },
> >> { 0x00000001, 0x244c0061, 0x00000000, 0x00200000 },
> >> { 0x00000001, 0x24560169, 0x00000000, 0x28302830 },
> >> - { 0x00000001, 0x24400021, 0x00000448, 0x00000000 },
> >> - { 0x00000040, 0x24403dad, 0x00000440, 0xfff0fff0 },
> >> - { 0x00000040, 0x24423dad, 0x00000442, 0xfff4fff4 },
> >> { 0x00000001, 0x244001ed, 0x00000000, 0xfff0fff0 },
> >> { 0x00000001, 0x244201ed, 0x00000000, 0xfff4fff4 },
> >> - { 0x00000001, 0x24440021, 0x00000440, 0x00000000 },
> >> { 0x01000005, 0x20002e28, 0x000000a5, 0x00600060 },
> >> { 0x00010040, 0x24403dad, 0x00000440, 0x000c000c },
> >> { 0x01000005, 0x20002e28, 0x000000a5, 0x00100010 },
> >> { 0x00010040, 0x24423dad, 0x00000442, 0x00080008 },
> >> + { 0x00000020, 0x34001c00, 0x00001400, 0x000000e0 },
> >> + { 0x00000001, 0x28600061, 0x00000000, 0x10010101 },
> >> + { 0x00000001, 0x28640061, 0x00000000, 0x100f0f0f },
> >> + { 0x00000001, 0x28680061, 0x00000000, 0x10010101 },
> >> + { 0x00000001, 0x286c0061, 0x00000000, 0x000f0f0f },
> >> + { 0x00400001, 0x28700061, 0x00000000, 0x00000000 },
> >> + { 0x00600001, 0x28900061, 0x00000000, 0x00000000 },
> >> + { 0x00000001, 0x244c0061, 0x00000000, 0x00200000 },
> >> + { 0x00000001, 0x24560169, 0x00000000, 0x20202020 },
> >> + { 0x00000001, 0x244001ed, 0x00000000, 0xfff8fff8 },
> >> + { 0x00000001, 0x244201ed, 0x00000000, 0xfff8fff8 },
> >> + { 0x01000005, 0x20002e28, 0x000000a5, 0x00600060 },
> >> + { 0x00010040, 0x24403dad, 0x00000440, 0x00040004 },
> >> + { 0x01000005, 0x20002e28, 0x000000a5, 0x00100010 },
> >> + { 0x00010040, 0x24423dad, 0x00000442, 0x00040004 },
> >> + { 0x00000001, 0x24440021, 0x00000440, 0x00000000 },
> >> { 0x00200040, 0x244035ad, 0x00450440, 0x00450a90 },
> >> { 0x00200040, 0x244435ad, 0x00450444, 0x00450a90 },
> >> { 0x00600001, 0x28000021, 0x008d0440, 0x00000000 },
> >> @@ -214,19 +241,6 @@
> >> { 0x00000001, 0x24740021, 0x00000ac0, 0x00000000 },
> >> { 0x00600001, 0x28200021, 0x008d0460, 0x00000000 },
> >> { 0x00600001, 0x28400021, 0x008d0560, 0x00000000 },
> >> - { 0x00000001, 0x28600061, 0x00000000, 0x01010101 },
> >> - { 0x00000001, 0x28640061, 0x00000000, 0x10010101 },
> >> - { 0x00000001, 0x28680061, 0x00000000, 0x0f0f0f0f },
> >> - { 0x00000001, 0x286c0061, 0x00000000, 0x100f0f0f },
> >> - { 0x00000001, 0x28700061, 0x00000000, 0x01010101 },
> >> - { 0x00000001, 0x28740061, 0x00000000, 0x10010101 },
> >> - { 0x00000001, 0x28780061, 0x00000000, 0x0f0f0f0f },
> >> - { 0x00000001, 0x287c0061, 0x00000000, 0x100f0f0f },
> >> - { 0x00000001, 0x28800061, 0x00000000, 0x01010101 },
> >> - { 0x00000001, 0x28840061, 0x00000000, 0x10010101 },
> >> - { 0x00000001, 0x28880061, 0x00000000, 0x0f0f0f0f },
> >> - { 0x00000001, 0x288c0061, 0x00000000, 0x000f0f0f },
> >> - { 0x00400001, 0x28900061, 0x00000000, 0x00000000 },
> >> { 0x08600031, 0x21801ca1, 0x00000800, 0x0a784000 },
> >> { 0x00000001, 0x25740061, 0x00000000, 0x00000000 },
> >> { 0x00000001, 0x25750231, 0x00000199, 0x00000000 },
> >> diff --git a/src/shaders/vme/vme75.inc b/src/shaders/vme/vme75.inc
> >> index be49056..82055db 100644
> >> --- a/src/shaders/vme/vme75.inc
> >> +++ b/src/shaders/vme/vme75.inc
> >> @@ -59,6 +59,8 @@ define(`SEARCH_CTRL_DUAL_RECORD', `0x00000300')
> >> define(`SEARCH_CTRL_DUAL_REFERENCE', `0x00000700')
> >>
> >> define(`REF_REGION_SIZE', `0x2830:UW')
> >> +define(`MIN_REF_REGION_SIZE', `0x2020:UW')
> >> +define(`DREF_REGION_SIZE', `0x2020:UW')
> >>
> >> define(`BI_SUB_MB_PART_MASK', `0x0c000000')
> >> define(`MAX_NUM_MV', `0x00000020')
> >> @@ -133,6 +135,7 @@ define(`orig_y_ub', `inline_reg0.1')
> >> define(`transform_8x8_ub', `inline_reg0.4')
> >> define(`input_mb_intra_ub', `inline_reg0.5')
> >> define(`num_macroblocks', `inline_reg0.6')
> >> +define(`quality_level_ub', `inline_reg0.7')
> >>
> >> /*
> >> * GRF 6~11 -- reserved
> >> @@ -311,7 +314,6 @@ define(`mb_mv3', `r96')
> >> define(`mb_ref', `r97')
> >> define(`mb_ref_win', `r84')
> >>
> >> -define(`DREF_REGION_SIZE', `0x2020:UW')
> >> define(`PRED_L0', `0x0':uw)
> >> define(`PRED_L1', `0x1':uw)
> >> define(`PRED_BI', `0x2':uw)
> >> @@ -337,3 +339,7 @@ define(`INTER_8X16MODE', `0x02')
> >> define(`OBR_MESSAGE_FENCE', `7')
> >> define(`OBR_MF_NOCOMMIT', `0')
> >> define(`OBR_MF_COMMIT', `0x20')
> >> +
> >> +define(`DEFAULT_QUALITY_LEVEL', `0x00')
> >> +define(`HIGH_QUALITY_LEVEL', `0x01')
> >> +define(`LOW_QUALITY_LEVEL', `0x02')
> >
> >
> >
>
More information about the Libva
mailing list