[Libva] [PATCH 2/7] VP8 HWEnc: Build vp8 gen9 encoding vme pipeline
Zhong Li
zhong.li at intel.com
Mon Dec 22 23:46:19 PST 2014
Signed-off-by: Zhong Li <zhong.li at intel.com>
---
src/gen9_vme.c | 262 +++++++++++++++++++++++++++++++++++++++----------
src/i965_device_info.c | 1 +
src/i965_drv_video.h | 2 +
src/i965_encoder.c | 4 +
4 files changed, 215 insertions(+), 54 deletions(-)
diff --git a/src/gen9_vme.c b/src/gen9_vme.c
index b4310f2..34e0586 100644
--- a/src/gen9_vme.c
+++ b/src/gen9_vme.c
@@ -120,6 +120,31 @@ static struct i965_kernel gen9_vme_mpeg2_kernels[] = {
},
};
+static const uint32_t gen9_vme_vp8_intra_frame[][4] = {
+#include "shaders/vme/vp8_intra_frame_gen9.g9b"
+};
+
+static const uint32_t gen9_vme_vp8_inter_frame[][4] = {
+#include "shaders/vme/inter_frame_gen9.g9b"
+};
+
+static struct i965_kernel gen9_vme_vp8_kernels[] = {
+ {
+ "VME Intra Frame",
+ VME_INTRA_SHADER, /*index*/
+ gen9_vme_vp8_intra_frame,
+ sizeof(gen9_vme_vp8_intra_frame),
+ NULL
+ },
+ {
+ "VME inter Frame",
+ VME_INTER_SHADER,
+ gen9_vme_vp8_inter_frame,
+ sizeof(gen9_vme_vp8_inter_frame),
+ NULL
+ },
+};
+
/* only used for VME source surface state */
static void
gen9_vme_source_surface_state(VADriverContextP ctx,
@@ -170,16 +195,14 @@ static void
gen9_vme_output_buffer_setup(VADriverContextP ctx,
struct encode_state *encode_state,
int index,
- struct intel_encoder_context *encoder_context)
+ struct intel_encoder_context *encoder_context,
+ int is_intra,
+ int width_in_mbs,
+ int height_in_mbs)
{
struct i965_driver_data *i965 = i965_driver_data(ctx);
struct gen6_vme_context *vme_context = encoder_context->vme_context;
- VAEncSequenceParameterBufferH264 *pSequenceParameter = (VAEncSequenceParameterBufferH264 *)encode_state->seq_param_ext->buffer;
- VAEncSliceParameterBufferH264 *pSliceParameter = (VAEncSliceParameterBufferH264 *)encode_state->slice_params_ext[0]->buffer;
- int is_intra = pSliceParameter->slice_type == SLICE_TYPE_I;
- int width_in_mbs = pSequenceParameter->picture_width_in_mbs;
- int height_in_mbs = pSequenceParameter->picture_height_in_mbs;
vme_context->vme_output.num_blocks = width_in_mbs * height_in_mbs;
vme_context->vme_output.pitch = 16; /* in bytes, always 16 */
@@ -207,17 +230,31 @@ gen9_vme_output_buffer_setup(VADriverContextP ctx,
}
static void
+gen9_vme_avc_output_buffer_setup(VADriverContextP ctx,
+ struct encode_state *encode_state,
+ int index,
+ struct intel_encoder_context *encoder_context)
+{
+ VAEncSequenceParameterBufferH264 *pSequenceParameter = (VAEncSequenceParameterBufferH264 *)encode_state->seq_param_ext->buffer;
+ VAEncSliceParameterBufferH264 *pSliceParameter = (VAEncSliceParameterBufferH264 *)encode_state->slice_params_ext[0]->buffer;
+ int is_intra = pSliceParameter->slice_type == SLICE_TYPE_I;
+ int width_in_mbs = pSequenceParameter->picture_width_in_mbs;
+ int height_in_mbs = pSequenceParameter->picture_height_in_mbs;
+
+ gen9_vme_output_buffer_setup(ctx, encode_state, index, encoder_context, is_intra, width_in_mbs, height_in_mbs);
+
+}
+
+static void
gen9_vme_output_vme_batchbuffer_setup(VADriverContextP ctx,
struct encode_state *encode_state,
int index,
- struct intel_encoder_context *encoder_context)
-
+ struct intel_encoder_context *encoder_context,
+ int width_in_mbs,
+ int height_in_mbs)
{
struct i965_driver_data *i965 = i965_driver_data(ctx);
struct gen6_vme_context *vme_context = encoder_context->vme_context;
- VAEncSequenceParameterBufferH264 *pSequenceParameter = (VAEncSequenceParameterBufferH264 *)encode_state->seq_param_ext->buffer;
- int width_in_mbs = pSequenceParameter->picture_width_in_mbs;
- int height_in_mbs = pSequenceParameter->picture_height_in_mbs;
vme_context->vme_batchbuffer.num_blocks = width_in_mbs * height_in_mbs + 1;
vme_context->vme_batchbuffer.size_block = 64; /* 4 OWORDs */
@@ -226,15 +263,27 @@ gen9_vme_output_vme_batchbuffer_setup(VADriverContextP ctx,
"VME batchbuffer",
vme_context->vme_batchbuffer.num_blocks * vme_context->vme_batchbuffer.size_block,
0x1000);
- /*
vme_context->vme_buffer_suface_setup(ctx,
&vme_context->gpe_context,
&vme_context->vme_batchbuffer,
BINDING_TABLE_OFFSET(index),
SURFACE_STATE_OFFSET(index));
- */
}
+static void
+gen9_vme_avc_output_vme_batchbuffer_setup(VADriverContextP ctx,
+ struct encode_state *encode_state,
+ int index,
+ struct intel_encoder_context *encoder_context)
+{
+ VAEncSequenceParameterBufferH264 *pSequenceParameter = (VAEncSequenceParameterBufferH264 *)encode_state->seq_param_ext->buffer;
+ int width_in_mbs = pSequenceParameter->picture_width_in_mbs;
+ int height_in_mbs = pSequenceParameter->picture_height_in_mbs;
+
+ gen9_vme_output_vme_batchbuffer_setup(ctx, encode_state, index, encoder_context, width_in_mbs, height_in_mbs);
+}
+
+
static VAStatus
gen9_vme_surface_setup(VADriverContextP ctx,
struct encode_state *encode_state,
@@ -264,8 +313,8 @@ gen9_vme_surface_setup(VADriverContextP ctx,
}
/* VME output */
- gen9_vme_output_buffer_setup(ctx, encode_state, 3, encoder_context);
- gen9_vme_output_vme_batchbuffer_setup(ctx, encode_state, 5, encoder_context);
+ gen9_vme_avc_output_buffer_setup(ctx, encode_state, 3, encoder_context);
+ gen9_vme_avc_output_vme_batchbuffer_setup(ctx, encode_state, 5, encoder_context);
return VA_STATUS_SUCCESS;
}
@@ -727,35 +776,11 @@ gen9_vme_mpeg2_output_buffer_setup(VADriverContextP ctx,
struct intel_encoder_context *encoder_context)
{
- struct i965_driver_data *i965 = i965_driver_data(ctx);
- struct gen6_vme_context *vme_context = encoder_context->vme_context;
VAEncSequenceParameterBufferMPEG2 *seq_param = (VAEncSequenceParameterBufferMPEG2 *)encode_state->seq_param_ext->buffer;
int width_in_mbs = ALIGN(seq_param->picture_width, 16) / 16;
int height_in_mbs = ALIGN(seq_param->picture_height, 16) / 16;
- vme_context->vme_output.num_blocks = width_in_mbs * height_in_mbs;
- vme_context->vme_output.pitch = 16; /* in bytes, always 16 */
-
- if (is_intra)
- vme_context->vme_output.size_block = INTRA_VME_OUTPUT_IN_BYTES * 2;
- else
- vme_context->vme_output.size_block = INTRA_VME_OUTPUT_IN_BYTES * 24;
- /*
- * Inter MV . 32-byte Intra search + 16 IME info + 128 IME MV + 32 IME Ref
- * + 16 FBR Info + 128 FBR MV + 32 FBR Ref.
- * 16 * (2 + 2 * (1 + 8 + 2))= 16 * 24.
- */
-
- vme_context->vme_output.bo = dri_bo_alloc(i965->intel.bufmgr,
- "VME output buffer",
- vme_context->vme_output.num_blocks * vme_context->vme_output.size_block,
- 0x1000);
- assert(vme_context->vme_output.bo);
- vme_context->vme_buffer_suface_setup(ctx,
- &vme_context->gpe_context,
- &vme_context->vme_output,
- BINDING_TABLE_OFFSET(index),
- SURFACE_STATE_OFFSET(index));
+ gen9_vme_output_buffer_setup(ctx, encode_state, index, encoder_context, is_intra, width_in_mbs, height_in_mbs);
}
static void
@@ -765,24 +790,11 @@ gen9_vme_mpeg2_output_vme_batchbuffer_setup(VADriverContextP ctx,
struct intel_encoder_context *encoder_context)
{
- struct i965_driver_data *i965 = i965_driver_data(ctx);
- struct gen6_vme_context *vme_context = encoder_context->vme_context;
VAEncSequenceParameterBufferMPEG2 *seq_param = (VAEncSequenceParameterBufferMPEG2 *)encode_state->seq_param_ext->buffer;
int width_in_mbs = ALIGN(seq_param->picture_width, 16) / 16;
int height_in_mbs = ALIGN(seq_param->picture_height, 16) / 16;
- vme_context->vme_batchbuffer.num_blocks = width_in_mbs * height_in_mbs + 1;
- vme_context->vme_batchbuffer.size_block = 64; /* 4 OWORDs */
- vme_context->vme_batchbuffer.pitch = 16;
- vme_context->vme_batchbuffer.bo = dri_bo_alloc(i965->intel.bufmgr,
- "VME batchbuffer",
- vme_context->vme_batchbuffer.num_blocks * vme_context->vme_batchbuffer.size_block,
- 0x1000);
- vme_context->vme_buffer_suface_setup(ctx,
- &vme_context->gpe_context,
- &vme_context->vme_batchbuffer,
- BINDING_TABLE_OFFSET(index),
- SURFACE_STATE_OFFSET(index));
+ gen9_vme_output_vme_batchbuffer_setup(ctx, encode_state, index, encoder_context, width_in_mbs, height_in_mbs);
}
static VAStatus
@@ -1131,6 +1143,143 @@ gen9_vme_mpeg2_pipeline(VADriverContextP ctx,
}
static void
+gen9_vme_vp8_output_buffer_setup(VADriverContextP ctx,
+ struct encode_state *encode_state,
+ int index,
+ int is_intra,
+ struct intel_encoder_context *encoder_context)
+
+{
+ VAEncSequenceParameterBufferVP8 *seq_param = (VAEncSequenceParameterBufferVP8 *)encode_state->seq_param_ext->buffer;
+ int width_in_mbs = ALIGN(seq_param->frame_width, 16) / 16;
+ int height_in_mbs = ALIGN(seq_param->frame_height, 16) / 16;
+
+ gen9_vme_output_buffer_setup(ctx, encode_state, index, encoder_context, is_intra, width_in_mbs, height_in_mbs);
+}
+
+static void
+gen9_vme_vp8_output_vme_batchbuffer_setup(VADriverContextP ctx,
+ struct encode_state *encode_state,
+ int index,
+ struct intel_encoder_context *encoder_context)
+
+{
+ VAEncSequenceParameterBufferVP8 *seq_param = (VAEncSequenceParameterBufferVP8 *)encode_state->seq_param_ext->buffer;
+ int width_in_mbs = ALIGN(seq_param->frame_width, 16) / 16;
+ int height_in_mbs = ALIGN(seq_param->frame_height, 16) / 16;
+
+ gen9_vme_output_vme_batchbuffer_setup(ctx, encode_state, index, encoder_context, width_in_mbs, height_in_mbs);
+}
+
+static VAStatus
+gen9_vme_vp8_surface_setup(VADriverContextP ctx,
+ struct encode_state *encode_state,
+ int is_intra,
+ struct intel_encoder_context *encoder_context)
+{
+ struct object_surface *obj_surface;
+
+ /*Setup surfaces state*/
+ /* current picture for encoding */
+ obj_surface = encode_state->input_yuv_object;
+ gen9_vme_source_surface_state(ctx, 0, obj_surface, encoder_context);
+ gen9_vme_media_source_surface_state(ctx, 4, obj_surface, encoder_context);
+ gen9_vme_media_chroma_source_surface_state(ctx, 6, obj_surface, encoder_context);
+
+ if (!is_intra) {
+ /* reference 0 */
+ obj_surface = encode_state->reference_objects[0];
+
+ if (obj_surface->bo != NULL)
+ gen9_vme_source_surface_state(ctx, 1, obj_surface, encoder_context);
+
+ /* reference 1 */
+ obj_surface = encode_state->reference_objects[1];
+
+ if (obj_surface && obj_surface->bo != NULL)
+ gen9_vme_source_surface_state(ctx, 2, obj_surface, encoder_context);
+ }
+
+ /* VME output */
+ gen9_vme_vp8_output_buffer_setup(ctx, encode_state, 3, is_intra, encoder_context);
+ gen9_vme_vp8_output_vme_batchbuffer_setup(ctx, encode_state, 5, encoder_context);
+
+ return VA_STATUS_SUCCESS;
+}
+
+static void
+gen9_vme_vp8_pipeline_programing(VADriverContextP ctx,
+ struct encode_state *encode_state,
+ int is_intra,
+ struct intel_encoder_context *encoder_context)
+{
+ struct gen6_vme_context *vme_context = encoder_context->vme_context;
+ struct intel_batchbuffer *batch = encoder_context->base.batch;
+ VAEncSequenceParameterBufferVP8 *seq_param = (VAEncSequenceParameterBufferVP8 *)encode_state->seq_param_ext->buffer;
+ int width_in_mbs = ALIGN(seq_param->frame_width, 16) / 16;
+ int height_in_mbs = ALIGN(seq_param->frame_height, 16) / 16;
+ int kernel_shader = (is_intra ? VME_INTRA_SHADER : VME_INTER_SHADER);
+
+ gen9wa_vme_mpeg2_walker_fill_vme_batchbuffer(ctx,
+ encode_state,
+ width_in_mbs, height_in_mbs,
+ kernel_shader,
+ encoder_context);
+
+ intel_batchbuffer_start_atomic(batch, 0x1000);
+ gen9_gpe_pipeline_setup(ctx, &vme_context->gpe_context, batch);
+ BEGIN_BATCH(batch, 4);
+ OUT_BATCH(batch, MI_BATCH_BUFFER_START | (1 << 8) | (1 << 0));
+ OUT_RELOC(batch,
+ vme_context->vme_batchbuffer.bo,
+ I915_GEM_DOMAIN_COMMAND, 0,
+ 0);
+ OUT_BATCH(batch, 0);
+ OUT_BATCH(batch, 0);
+ ADVANCE_BATCH(batch);
+
+ gen9_gpe_pipeline_end(ctx, &vme_context->gpe_context, batch);
+
+ intel_batchbuffer_end_atomic(batch);
+}
+
+static VAStatus gen9_vme_vp8_prepare(VADriverContextP ctx,
+ struct encode_state *encode_state,
+ struct intel_encoder_context *encoder_context)
+{
+ VAStatus vaStatus = VA_STATUS_SUCCESS;
+ VAEncPictureParameterBufferVP8 *pPicParameter = (VAEncPictureParameterBufferVP8 *)encode_state->pic_param_ext->buffer;
+ int is_intra = !pPicParameter->pic_flags.bits.frame_type;
+
+ /* add update vp8 mbmv cost later */
+ //intel_vme_update_mbmv_cost(ctx, encode_state, encoder_context);
+
+ /*Setup all the memory object*/
+ gen9_vme_vp8_surface_setup(ctx, encode_state, is_intra, encoder_context);
+ gen9_vme_interface_setup(ctx, encode_state, encoder_context);
+ gen9_vme_constant_setup(ctx, encode_state, encoder_context);
+
+ /*Programing media pipeline*/
+ gen9_vme_vp8_pipeline_programing(ctx, encode_state, is_intra, encoder_context);
+
+ return vaStatus;
+}
+
+static VAStatus
+gen9_vme_vp8_pipeline(VADriverContextP ctx,
+ VAProfile profile,
+ struct encode_state *encode_state,
+ struct intel_encoder_context *encoder_context)
+{
+ gen9_vme_media_init(ctx, encoder_context);
+ gen9_vme_vp8_prepare(ctx, encode_state, encoder_context);
+ gen9_vme_run(ctx, encode_state, encoder_context);
+ gen9_vme_stop(ctx, encode_state, encoder_context);
+
+ return VA_STATUS_SUCCESS;
+}
+
+static void
gen9_vme_context_destroy(void *context)
{
struct gen6_vme_context *vme_context = context;
@@ -1172,7 +1321,12 @@ Bool gen9_vme_context_init(VADriverContextP ctx, struct intel_encoder_context *e
vme_kernel_list = gen9_vme_mpeg2_kernels;
encoder_context->vme_pipeline = gen9_vme_mpeg2_pipeline;
i965_kernel_num = sizeof(gen9_vme_mpeg2_kernels) / sizeof(struct i965_kernel);
+ break;
+ case CODEC_VP8:
+ vme_kernel_list = gen9_vme_vp8_kernels;
+ encoder_context->vme_pipeline = gen9_vme_vp8_pipeline;
+ i965_kernel_num = sizeof(gen9_vme_vp8_kernels) / sizeof(struct i965_kernel);
break;
default:
diff --git a/src/i965_device_info.c b/src/i965_device_info.c
index d03aceb..e63f509 100755
--- a/src/i965_device_info.c
+++ b/src/i965_device_info.c
@@ -344,6 +344,7 @@ static struct hw_codec_info skl_hw_codec_info = {
.has_di_motion_adptive = 1,
.has_di_motion_compensated = 1,
.has_vp8_decoding = 1,
+ .has_vp8_encoding = 1,
.has_blending = 1,
.has_h264_mvc_encoding = 1,
.has_hevc_decoding = 1,
diff --git a/src/i965_drv_video.h b/src/i965_drv_video.h
index cb4e8e8..1019f80 100644
--- a/src/i965_drv_video.h
+++ b/src/i965_drv_video.h
@@ -34,6 +34,7 @@
#include <va/va_enc_h264.h>
#include <va/va_enc_mpeg2.h>
#include <va/va_enc_jpeg.h>
+#include <va/va_enc_vp8.h>
#include <va/va_vpp.h>
#include <va/va_backend.h>
#include <va/va_backend_vpp.h>
@@ -454,6 +455,7 @@ va_enc_packed_type_to_idx(int packed_type);
#define CODEC_MPEG2 1
#define CODEC_H264_MVC 2
#define CODEC_JPEG 3
+#define CODEC_VP8 4
#define H264_DELIMITER0 0x00
#define H264_DELIMITER1 0x00
diff --git a/src/i965_encoder.c b/src/i965_encoder.c
index d924f5a..86f8757 100644
--- a/src/i965_encoder.c
+++ b/src/i965_encoder.c
@@ -554,6 +554,10 @@ intel_enc_hw_context_init(VADriverContextP ctx,
encoder_context->codec = CODEC_JPEG;
break;
+ case VAProfileVP8Version0_3:
+ encoder_context->codec = CODEC_VP8;
+ break;
+
default:
/* Never get here */
assert(0);
--
1.9.1
More information about the Libva
mailing list