diff options
Diffstat (limited to 'src/gen7_mfd.c')
-rwxr-xr-x | src/gen7_mfd.c | 145 |
1 files changed, 79 insertions, 66 deletions
diff --git a/src/gen7_mfd.c b/src/gen7_mfd.c index 8e0d503b..bfb95bf6 100755 --- a/src/gen7_mfd.c +++ b/src/gen7_mfd.c @@ -65,6 +65,7 @@ gen7_mfd_init_avc_surface(VADriverContextP ctx, if (!gen7_avc_surface) { gen7_avc_surface = calloc(sizeof(GenAvcSurface), 1); + gen7_avc_surface->frame_store_id = -1; assert((obj_surface->size & 0x3f) == 0); obj_surface->private_data = gen7_avc_surface; } @@ -135,12 +136,16 @@ gen7_mfd_surface_state(VADriverContextP ctx, struct object_surface *obj_surface = decode_state->render_object; unsigned int y_cb_offset; unsigned int y_cr_offset; + unsigned int surface_format; assert(obj_surface); y_cb_offset = obj_surface->y_cb_offset; y_cr_offset = obj_surface->y_cr_offset; + surface_format = obj_surface->fourcc == VA_FOURCC_Y800 ? + MFX_SURFACE_MONOCHROME : MFX_SURFACE_PLANAR_420_8; + BEGIN_BCS_BATCH(batch, 6); OUT_BCS_BATCH(batch, MFX_SURFACE_STATE | (6 - 2)); OUT_BCS_BATCH(batch, 0); @@ -148,7 +153,7 @@ gen7_mfd_surface_state(VADriverContextP ctx, ((obj_surface->orig_height - 1) << 18) | ((obj_surface->orig_width - 1) << 4)); OUT_BCS_BATCH(batch, - (MFX_SURFACE_PLANAR_420_8 << 28) | /* 420 planar YUV surface */ + (surface_format << 28) | /* 420 planar YUV surface */ ((standard_select != MFX_FORMAT_JPEG) << 27) | /* interleave chroma, set to 0 for JPEG */ (0 << 22) | /* surface object control state, ignored */ ((obj_surface->width - 1) << 3) | /* pitch */ @@ -353,7 +358,7 @@ gen7_mfd_avc_img_state(VADriverContextP ctx, BEGIN_BCS_BATCH(batch, 16); OUT_BCS_BATCH(batch, MFX_AVC_IMG_STATE | (16 - 2)); OUT_BCS_BATCH(batch, - width_in_mbs * height_in_mbs); + (width_in_mbs * height_in_mbs - 1)); OUT_BCS_BATCH(batch, ((height_in_mbs - 1) << 16) | ((width_in_mbs - 1) << 0)); @@ -425,7 +430,7 @@ gen7_mfd_avc_directmode_state(VADriverContextP ctx, struct object_surface *obj_surface; GenAvcSurface *gen7_avc_surface; VAPictureH264 *va_pic; - int i, j; + int i; BEGIN_BCS_BATCH(batch, 69); OUT_BCS_BATCH(batch, MFX_AVC_DIRECTMODE_STATE | (69 - 2)); @@ -477,26 +482,14 @@ gen7_mfd_avc_directmode_state(VADriverContextP ctx, /* POC List */ for (i = 0; i < ARRAY_ELEMS(gen7_mfd_context->reference_surface); i++) { - if (gen7_mfd_context->reference_surface[i].surface_id != VA_INVALID_ID) { - int found = 0; - - assert(gen7_mfd_context->reference_surface[i].obj_surface != NULL); - - for (j = 0; j < ARRAY_ELEMS(pic_param->ReferenceFrames); j++) { - va_pic = &pic_param->ReferenceFrames[j]; - - if (va_pic->flags & VA_PICTURE_H264_INVALID) - continue; + obj_surface = gen7_mfd_context->reference_surface[i].obj_surface; - if (va_pic->picture_id == gen7_mfd_context->reference_surface[i].surface_id) { - found = 1; - break; - } - } + if (obj_surface) { + const VAPictureH264 * const va_pic = avc_find_picture( + obj_surface->base.id, pic_param->ReferenceFrames, + ARRAY_ELEMS(pic_param->ReferenceFrames)); - assert(found == 1); - assert(!(va_pic->flags & VA_PICTURE_H264_INVALID)); - + assert(va_pic != NULL); OUT_BCS_BATCH(batch, va_pic->TopFieldOrderCnt); OUT_BCS_BATCH(batch, va_pic->BottomFieldOrderCnt); } else { @@ -513,6 +506,15 @@ gen7_mfd_avc_directmode_state(VADriverContextP ctx, } static void +gen7_mfd_avc_phantom_slice_first(VADriverContextP ctx, + VAPictureParameterBufferH264 *pic_param, + VASliceParameterBufferH264 *next_slice_param, + struct gen7_mfd_context *gen7_mfd_context) +{ + gen6_mfd_avc_phantom_slice(ctx, pic_param, next_slice_param, gen7_mfd_context->base.batch); +} + +static void gen7_mfd_avc_slice_state(VADriverContextP ctx, VAPictureParameterBufferH264 *pic_param, VASliceParameterBufferH264 *slice_param, @@ -748,7 +750,8 @@ gen7_mfd_avc_decode_init(VADriverContextP ctx, assert(decode_state->pic_param && decode_state->pic_param->buffer); pic_param = (VAPictureParameterBufferH264 *)decode_state->pic_param->buffer; - intel_update_avc_frame_store_index(ctx, decode_state, pic_param, gen7_mfd_context->reference_surface); + intel_update_avc_frame_store_index(ctx, decode_state, pic_param, + gen7_mfd_context->reference_surface, &gen7_mfd_context->fs_ctx); width_in_mbs = pic_param->picture_width_in_mbs_minus1 + 1; height_in_mbs = pic_param->picture_height_in_mbs_minus1 + 1; assert(width_in_mbs > 0 && width_in_mbs <= 256); /* 4K */ @@ -756,20 +759,12 @@ gen7_mfd_avc_decode_init(VADriverContextP ctx, /* Current decoded picture */ obj_surface = decode_state->render_object; - obj_surface->flags &= ~SURFACE_REF_DIS_MASK; - obj_surface->flags |= (pic_param->pic_fields.bits.reference_pic_flag ? SURFACE_REFERENCED : 0); - i965_check_alloc_surface_bo(ctx, obj_surface, 1, VA_FOURCC('N','V','1','2'), SUBSAMPLE_YUV420); - - /* initial uv component for YUV400 case */ - if (pic_param->seq_fields.bits.chroma_format_idc == 0) { - unsigned int uv_offset = obj_surface->width * obj_surface->height; - unsigned int uv_size = obj_surface->width * obj_surface->height / 2; - - drm_intel_gem_bo_map_gtt(obj_surface->bo); - memset(obj_surface->bo->virtual + uv_offset, 0x80, uv_size); - drm_intel_gem_bo_unmap_gtt(obj_surface->bo); - } + if (pic_param->pic_fields.bits.reference_pic_flag) + obj_surface->flags |= SURFACE_REFERENCED; + else + obj_surface->flags &= ~SURFACE_REFERENCED; + avc_ensure_surface_bo(ctx, decode_state, obj_surface, pic_param); gen7_mfd_init_avc_surface(ctx, pic_param, obj_surface); dri_bo_unreference(gen7_mfd_context->post_deblocking_output.bo); @@ -856,6 +851,9 @@ gen7_mfd_avc_decode_picture(VADriverContextP ctx, else next_slice_group_param = (VASliceParameterBufferH264 *)decode_state->slice_params[j + 1]->buffer; + if (j == 0 && slice_param->first_mb_in_slice) + gen7_mfd_avc_phantom_slice_first(ctx, pic_param, slice_param, gen7_mfd_context); + for (i = 0; i < decode_state->slice_params[j]->num_elements; i++) { assert(slice_param->slice_data_flag == VA_SLICE_DATA_FLAG_ALL); assert((slice_param->slice_type == SLICE_TYPE_I) || @@ -906,7 +904,7 @@ gen7_mfd_mpeg2_decode_init(VADriverContextP ctx, /* Current decoded picture */ obj_surface = decode_state->render_object; - i965_check_alloc_surface_bo(ctx, obj_surface, 1, VA_FOURCC('N','V','1','2'), SUBSAMPLE_YUV420); + i965_check_alloc_surface_bo(ctx, obj_surface, 1, VA_FOURCC_NV12, SUBSAMPLE_YUV420); dri_bo_unreference(gen7_mfd_context->pre_deblocking_output.bo); gen7_mfd_context->pre_deblocking_output.bo = obj_surface->bo; @@ -942,7 +940,7 @@ gen7_mfd_mpeg2_pic_state(VADriverContextP ctx, assert(decode_state->pic_param && decode_state->pic_param->buffer); pic_param = (VAPictureParameterBufferMPEG2 *)decode_state->pic_param->buffer; - if (IS_HASWELL(i965->intel.device_id)) { + if (IS_HASWELL(i965->intel.device_info)) { /* XXX: disable concealment for now */ slice_concealment_disable_bit = 1; } @@ -1040,10 +1038,35 @@ gen7_mfd_mpeg2_qm_state(VADriverContextP ctx, } } +uint32_t mpeg2_get_slice_data_length(dri_bo *slice_data_bo, VASliceParameterBufferMPEG2 *slice_param) +{ + uint8_t *buf; + uint32_t buf_offset = slice_param->slice_data_offset + (slice_param->macroblock_offset >> 3); + uint32_t buf_size = slice_param->slice_data_size - (slice_param->macroblock_offset >> 3); + uint32_t i; + + dri_bo_map(slice_data_bo, 0); + buf = (uint8_t *)slice_data_bo->virtual + buf_offset; + + for (i = 3; i < buf_size; i++) { + if (buf[i - 3] && + !buf[i - 2] && + !buf[i - 1] && + !buf[i]) { + dri_bo_unmap(slice_data_bo); + return i - 3 + 1; + } + } + + dri_bo_unmap(slice_data_bo); + return buf_size; +} + static void gen7_mfd_mpeg2_bsd_object(VADriverContextP ctx, VAPictureParameterBufferMPEG2 *pic_param, VASliceParameterBufferMPEG2 *slice_param, + dri_bo *slice_data_bo, VASliceParameterBufferMPEG2 *next_slice_param, struct gen7_mfd_context *gen7_mfd_context) { @@ -1074,7 +1097,7 @@ gen7_mfd_mpeg2_bsd_object(VADriverContextP ctx, BEGIN_BCS_BATCH(batch, 5); OUT_BCS_BATCH(batch, MFD_MPEG2_BSD_OBJECT | (5 - 2)); OUT_BCS_BATCH(batch, - slice_param->slice_data_size - (slice_param->macroblock_offset >> 3)); + mpeg2_get_slice_data_length(slice_data_bo, slice_param)); OUT_BCS_BATCH(batch, slice_param->slice_data_offset + (slice_param->macroblock_offset >> 3)); OUT_BCS_BATCH(batch, @@ -1086,7 +1109,7 @@ gen7_mfd_mpeg2_bsd_object(VADriverContextP ctx, (slice_param->macroblock_offset & 0x7)); OUT_BCS_BATCH(batch, (slice_param->quantiser_scale_code << 24) | - (IS_HASWELL(i965->intel.device_id) ? (vpos1 << 8 | hpos1) : 0)); + (IS_HASWELL(i965->intel.device_info) ? (vpos1 << 8 | hpos1) : 0)); ADVANCE_BCS_BATCH(batch); } @@ -1137,7 +1160,7 @@ gen7_mfd_mpeg2_decode_picture(VADriverContextP ctx, else next_slice_param = next_slice_group_param; - gen7_mfd_mpeg2_bsd_object(ctx, pic_param, slice_param, next_slice_param, gen7_mfd_context); + gen7_mfd_mpeg2_bsd_object(ctx, pic_param, slice_param, slice_data_bo, next_slice_param, gen7_mfd_context); slice_param++; } } @@ -1247,7 +1270,7 @@ gen7_mfd_vc1_decode_init(VADriverContextP ctx, /* Current decoded picture */ obj_surface = decode_state->render_object; - i965_check_alloc_surface_bo(ctx, obj_surface, 1, VA_FOURCC('N','V','1','2'), SUBSAMPLE_YUV420); + i965_check_alloc_surface_bo(ctx, obj_surface, 1, VA_FOURCC_NV12, SUBSAMPLE_YUV420); gen7_mfd_init_vc1_surface(ctx, pic_param, obj_surface); dri_bo_unreference(gen7_mfd_context->post_deblocking_output.bo); @@ -1798,13 +1821,13 @@ gen7_mfd_jpeg_decode_init(VADriverContextP ctx, struct object_surface *obj_surface; VAPictureParameterBufferJPEGBaseline *pic_param; int subsampling = SUBSAMPLE_YUV420; - int fourcc = VA_FOURCC('I', 'M', 'C', '3'); + int fourcc = VA_FOURCC_IMC3; pic_param = (VAPictureParameterBufferJPEGBaseline *)decode_state->pic_param->buffer; if (pic_param->num_components == 1) { subsampling = SUBSAMPLE_YUV400; - fourcc = VA_FOURCC('Y', '8', '0', '0'); + fourcc = VA_FOURCC_Y800; } else if (pic_param->num_components == 3) { int h1 = pic_param->components[0].h_sampling_factor; int h2 = pic_param->components[1].h_sampling_factor; @@ -1816,31 +1839,31 @@ gen7_mfd_jpeg_decode_init(VADriverContextP ctx, if (h1 == 2 && h2 == 1 && h3 == 1 && v1 == 2 && v2 == 1 && v3 == 1) { subsampling = SUBSAMPLE_YUV420; - fourcc = VA_FOURCC('I', 'M', 'C', '3'); + fourcc = VA_FOURCC_IMC3; } else if (h1 == 2 && h2 == 1 && h3 == 1 && v1 == 1 && v2 == 1 && v3 == 1) { subsampling = SUBSAMPLE_YUV422H; - fourcc = VA_FOURCC('4', '2', '2', 'H'); + fourcc = VA_FOURCC_422H; } else if (h1 == 1 && h2 == 1 && h3 == 1 && v1 == 1 && v2 == 1 && v3 == 1) { subsampling = SUBSAMPLE_YUV444; - fourcc = VA_FOURCC('4', '4', '4', 'P'); + fourcc = VA_FOURCC_444P; } else if (h1 == 4 && h2 == 1 && h3 == 1 && v1 == 1 && v2 == 1 && v3 == 1) { subsampling = SUBSAMPLE_YUV411; - fourcc = VA_FOURCC('4', '1', '1', 'P'); + fourcc = VA_FOURCC_411P; } else if (h1 == 1 && h2 == 1 && h3 == 1 && v1 == 2 && v2 == 1 && v3 == 1) { subsampling = SUBSAMPLE_YUV422V; - fourcc = VA_FOURCC('4', '2', '2', 'V'); + fourcc = VA_FOURCC_422V; } else if (h1 == 2 && h2 == 1 && h3 == 1 && v1 == 2 && v2 == 2 && v3 == 2) { subsampling = SUBSAMPLE_YUV422H; - fourcc = VA_FOURCC('4', '2', '2', 'H'); + fourcc = VA_FOURCC_422H; } else if (h2 == 2 && h2 == 2 && h3 == 2 && v1 == 2 && v2 == 1 && v3 == 1) { subsampling = SUBSAMPLE_YUV422V; - fourcc = VA_FOURCC('4', '2', '2', 'V'); + fourcc = VA_FOURCC_422V; } else assert(0); } else { @@ -2091,18 +2114,6 @@ gen7_mfd_jpeg_bsd_object(VADriverContextP ctx, /* Workaround for JPEG decoding on Ivybridge */ -VAStatus -i965_DestroySurfaces(VADriverContextP ctx, - VASurfaceID *surface_list, - int num_surfaces); -VAStatus -i965_CreateSurfaces(VADriverContextP ctx, - int width, - int height, - int format, - int num_surfaces, - VASurfaceID *surfaces); - static struct { int width; int height; @@ -2145,7 +2156,7 @@ gen7_jpeg_wa_init(VADriverContextP ctx, obj_surface = SURFACE(gen7_mfd_context->jpeg_wa_surface_id); assert(obj_surface); - i965_check_alloc_surface_bo(ctx, obj_surface, 1, VA_FOURCC('N', 'V', '1', '2'), SUBSAMPLE_YUV420); + i965_check_alloc_surface_bo(ctx, obj_surface, 1, VA_FOURCC_NV12, SUBSAMPLE_YUV420); gen7_mfd_context->jpeg_wa_surface_object = obj_surface; if (!gen7_mfd_context->jpeg_wa_slice_data_bo) { @@ -2320,7 +2331,7 @@ gen7_jpeg_wa_avc_img_state(VADriverContextP ctx, BEGIN_BCS_BATCH(batch, 16); OUT_BCS_BATCH(batch, MFX_AVC_IMG_STATE | (16 - 2)); OUT_BCS_BATCH(batch, - width_in_mbs * height_in_mbs); + (width_in_mbs * height_in_mbs - 1)); OUT_BCS_BATCH(batch, ((height_in_mbs - 1) << 16) | ((width_in_mbs - 1) << 0)); @@ -2614,9 +2625,10 @@ gen7_mfd_decode_picture(VADriverContextP ctx, gen7_mfd_mpeg2_decode_picture(ctx, decode_state, gen7_mfd_context); break; - case VAProfileH264Baseline: + case VAProfileH264ConstrainedBaseline: case VAProfileH264Main: case VAProfileH264High: + case VAProfileH264StereoHigh: gen7_mfd_avc_decode_picture(ctx, decode_state, gen7_mfd_context); break; @@ -2708,9 +2720,10 @@ gen7_dec_hw_context_init(VADriverContextP ctx, struct object_config *obj_config) gen7_mfd_mpeg2_context_init(ctx, gen7_mfd_context); break; - case VAProfileH264Baseline: + case VAProfileH264ConstrainedBaseline: case VAProfileH264Main: case VAProfileH264High: + case VAProfileH264StereoHigh: gen7_mfd_avc_context_init(ctx, gen7_mfd_context); break; default: |