"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 VAStatus
struct intel_encoder_context *encoder_context)
{
struct object_surface *obj_surface;
- struct i965_driver_data *i965 = i965_driver_data(ctx);
/*Setup surfaces state*/
/* current picture for encoding */
if (!is_intra) {
VAEncSliceParameterBufferH264 *slice_param = (VAEncSliceParameterBufferH264 *)encode_state->slice_params_ext[0]->buffer;
int slice_type;
- struct object_surface *slice_obj_surface;
- int ref_surface_id;
slice_type = intel_avc_enc_slice_type_fixup(slice_param->slice_type);
+ assert(slice_type != SLICE_TYPE_I && slice_type != SLICE_TYPE_SI);
- if (slice_type == SLICE_TYPE_P || slice_type == SLICE_TYPE_B) {
- slice_obj_surface = NULL;
- ref_surface_id = slice_param->RefPicList0[0].picture_id;
- if (ref_surface_id != 0 && ref_surface_id != VA_INVALID_SURFACE) {
- slice_obj_surface = SURFACE(ref_surface_id);
- }
- if (slice_obj_surface && slice_obj_surface->bo) {
- obj_surface = slice_obj_surface;
- } else {
- obj_surface = encode_state->reference_objects[0];
- }
- /* reference 0 */
- if (obj_surface && obj_surface->bo)
- gen8_vme_source_surface_state(ctx, 1, obj_surface, encoder_context);
- }
- if (slice_type == SLICE_TYPE_B) {
- /* reference 1 */
- slice_obj_surface = NULL;
- ref_surface_id = slice_param->RefPicList1[0].picture_id;
- if (ref_surface_id != 0 && ref_surface_id != VA_INVALID_SURFACE) {
- slice_obj_surface = SURFACE(ref_surface_id);
- }
- if (slice_obj_surface && slice_obj_surface->bo) {
- obj_surface = slice_obj_surface;
- } else {
- obj_surface = encode_state->reference_objects[0];
- }
+ intel_avc_vme_reference_state(ctx, encode_state, encoder_context, 0, 1, gen8_vme_source_surface_state);
- obj_surface = encode_state->reference_objects[1];
- if (obj_surface && obj_surface->bo)
- gen8_vme_source_surface_state(ctx, 2, obj_surface, encoder_context);
- }
+ if (slice_type == SLICE_TYPE_B)
+ intel_avc_vme_reference_state(ctx, encode_state, encoder_context, 1, 2, gen8_vme_source_surface_state);
}
/* VME output */
struct gen8_interface_descriptor_data *desc;
int i;
dri_bo *bo;
+ unsigned char *desc_ptr;
- bo = vme_context->gpe_context.idrt.bo;
+ bo = vme_context->gpe_context.dynamic_state.bo;
dri_bo_map(bo, 1);
assert(bo->virtual);
- desc = bo->virtual;
+ desc_ptr = (unsigned char *)bo->virtual + vme_context->gpe_context.idrt_offset;
+
+ desc = (struct gen8_interface_descriptor_data *)desc_ptr;
for (i = 0; i < vme_context->vme_kernel_sum; i++) {
struct i965_kernel *kernel;
assert(sizeof(*desc) == 32);
/*Setup the descritor table*/
memset(desc, 0, sizeof(*desc));
- desc->desc0.kernel_start_pointer = (kernel->bo->offset >> 6);
+ desc->desc0.kernel_start_pointer = kernel->kernel_offset >> 6;
desc->desc3.sampler_count = 0; /* FIXME: */
desc->desc3.sampler_state_pointer = 0;
desc->desc4.binding_table_entry_count = 1; /* FIXME: */
desc->desc4.binding_table_pointer = (BINDING_TABLE_OFFSET(0) >> 5);
desc->desc5.constant_urb_entry_read_offset = 0;
desc->desc5.constant_urb_entry_read_length = CURBE_URB_ENTRY_LENGTH;
-
+
- /*kernel start*/
- dri_bo_emit_reloc(bo,
- I915_GEM_DOMAIN_INSTRUCTION, 0,
- 0,
- i * sizeof(*desc) + offsetof(struct gen8_interface_descriptor_data, desc0),
- kernel->bo);
desc++;
}
+
dri_bo_unmap(bo);
return VA_STATUS_SUCCESS;
vme_state_message[31] = mv_num;
- dri_bo_map(vme_context->gpe_context.curbe.bo, 1);
- assert(vme_context->gpe_context.curbe.bo->virtual);
- constant_buffer = vme_context->gpe_context.curbe.bo->virtual;
+ dri_bo_map(vme_context->gpe_context.dynamic_state.bo, 1);
+ assert(vme_context->gpe_context.dynamic_state.bo->virtual);
+ constant_buffer = (unsigned char *)vme_context->gpe_context.dynamic_state.bo->virtual +
+ vme_context->gpe_context.curbe_offset;
/* VME MV/Mb cost table is passed by using const buffer */
/* Now it uses the fixed search path. So it is constructed directly
*/
memcpy(constant_buffer, (char *)vme_context->vme_state_message, 128);
- dri_bo_unmap(vme_context->gpe_context.curbe.bo);
+ dri_bo_unmap(vme_context->gpe_context.dynamic_state.bo);
return VA_STATUS_SUCCESS;
}
+#define MB_SCOREBOARD_A (1 << 0)
+#define MB_SCOREBOARD_B (1 << 1)
+#define MB_SCOREBOARD_C (1 << 2)
+
+/* check whether the mb of (x_index, y_index) is out of bound */
+static inline int loop_in_bounds(int x_index, int y_index, int first_mb, int num_mb, int mb_width, int mb_height)
+{
+ int mb_index;
+ if (x_index < 0 || x_index >= mb_width)
+ return -1;
+ if (y_index < 0 || y_index >= mb_height)
+ return -1;
+
+ mb_index = y_index * mb_width + x_index;
+ if (mb_index < first_mb || mb_index > (first_mb + num_mb))
+ return -1;
+ return 0;
+}
+
+static void
+gen8wa_vme_walker_fill_vme_batchbuffer(VADriverContextP ctx,
+ struct encode_state *encode_state,
+ int mb_width, int mb_height,
+ int kernel,
+ int transform_8x8_mode_flag,
+ struct intel_encoder_context *encoder_context)
+{
+ struct gen6_vme_context *vme_context = encoder_context->vme_context;
+ int mb_row;
+ int s;
+ unsigned int *command_ptr;
+
+#define USE_SCOREBOARD (1 << 21)
+
+ dri_bo_map(vme_context->vme_batchbuffer.bo, 1);
+ command_ptr = vme_context->vme_batchbuffer.bo->virtual;
+
+ for (s = 0; s < encode_state->num_slice_params_ext; s++) {
+ VAEncSliceParameterBufferH264 *pSliceParameter = (VAEncSliceParameterBufferH264 *)encode_state->slice_params_ext[s]->buffer;
+ int first_mb = pSliceParameter->macroblock_address;
+ int num_mb = pSliceParameter->num_macroblocks;
+ unsigned int mb_intra_ub, score_dep;
+ int x_outer, y_outer, x_inner, y_inner;
+ int xtemp_outer = 0;
+
+ x_outer = first_mb % mb_width;
+ y_outer = first_mb / mb_width;
+ mb_row = y_outer;
+
+ for (; x_outer < (mb_width -2 ) && !loop_in_bounds(x_outer, y_outer, first_mb, num_mb, mb_width, mb_height); ) {
+ x_inner = x_outer;
+ y_inner = y_outer;
+ for (; !loop_in_bounds(x_inner, y_inner, first_mb, num_mb, mb_width, mb_height);) {
+ mb_intra_ub = 0;
+ score_dep = 0;
+ if (x_inner != 0) {
+ mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_AE;
+ score_dep |= MB_SCOREBOARD_A;
+ }
+ if (y_inner != mb_row) {
+ mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_B;
+ score_dep |= MB_SCOREBOARD_B;
+ if (x_inner != 0)
+ mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_D;
+ if (x_inner != (mb_width -1)) {
+ mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_C;
+ score_dep |= MB_SCOREBOARD_C;
+ }
+ }
+
+ *command_ptr++ = (CMD_MEDIA_OBJECT | (8 - 2));
+ *command_ptr++ = kernel;
+ *command_ptr++ = USE_SCOREBOARD;
+ /* Indirect data */
+ *command_ptr++ = 0;
+ /* the (X, Y) term of scoreboard */
+ *command_ptr++ = ((y_inner << 16) | x_inner);
+ *command_ptr++ = score_dep;
+ /*inline data */
+ *command_ptr++ = (mb_width << 16 | y_inner << 8 | x_inner);
+ *command_ptr++ = ((1 << 18) | (1 << 16) | transform_8x8_mode_flag | (mb_intra_ub << 8));
+ *command_ptr++ = CMD_MEDIA_STATE_FLUSH;
+ *command_ptr++ = 0;
+
+ x_inner -= 2;
+ y_inner += 1;
+ }
+ x_outer += 1;
+ }
+
+ xtemp_outer = mb_width - 2;
+ if (xtemp_outer < 0)
+ xtemp_outer = 0;
+ x_outer = xtemp_outer;
+ y_outer = first_mb / mb_width;
+ for (;!loop_in_bounds(x_outer, y_outer, first_mb, num_mb, mb_width, mb_height); ) {
+ y_inner = y_outer;
+ x_inner = x_outer;
+ for (; !loop_in_bounds(x_inner, y_inner, first_mb, num_mb, mb_width, mb_height);) {
+ mb_intra_ub = 0;
+ score_dep = 0;
+ if (x_inner != 0) {
+ mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_AE;
+ score_dep |= MB_SCOREBOARD_A;
+ }
+ if (y_inner != mb_row) {
+ mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_B;
+ score_dep |= MB_SCOREBOARD_B;
+ if (x_inner != 0)
+ mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_D;
+
+ if (x_inner != (mb_width -1)) {
+ mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_C;
+ score_dep |= MB_SCOREBOARD_C;
+ }
+ }
+
+ *command_ptr++ = (CMD_MEDIA_OBJECT | (8 - 2));
+ *command_ptr++ = kernel;
+ *command_ptr++ = USE_SCOREBOARD;
+ /* Indirect data */
+ *command_ptr++ = 0;
+ /* the (X, Y) term of scoreboard */
+ *command_ptr++ = ((y_inner << 16) | x_inner);
+ *command_ptr++ = score_dep;
+ /*inline data */
+ *command_ptr++ = (mb_width << 16 | y_inner << 8 | x_inner);
+ *command_ptr++ = ((1 << 18) | (1 << 16) | transform_8x8_mode_flag | (mb_intra_ub << 8));
+
+ *command_ptr++ = CMD_MEDIA_STATE_FLUSH;
+ *command_ptr++ = 0;
+ x_inner -= 2;
+ y_inner += 1;
+ }
+ x_outer++;
+ if (x_outer >= mb_width) {
+ y_outer += 1;
+ x_outer = xtemp_outer;
+ }
+ }
+ }
+
+ *command_ptr++ = MI_BATCH_BUFFER_END;
+ *command_ptr++ = 0;
+
+ dri_bo_unmap(vme_context->vme_batchbuffer.bo);
+}
static void
gen8_vme_fill_vme_batchbuffer(VADriverContextP ctx,
*command_ptr++ = (mb_width << 16 | mb_y << 8 | mb_x);
*command_ptr++ = ( (1 << 16) | transform_8x8_mode_flag | (mb_intra_ub << 8));
+ *command_ptr++ = CMD_MEDIA_STATE_FLUSH;
+ *command_ptr++ = 0;
i += 1;
}
}
- *command_ptr++ = 0;
*command_ptr++ = MI_BATCH_BUFFER_END;
+ *command_ptr++ = 0;
dri_bo_unmap(vme_context->vme_batchbuffer.bo);
}
{
struct gen6_vme_context *vme_context = encoder_context->vme_context;
- i965_gpe_context_init(ctx, &vme_context->gpe_context);
+ gen8_gpe_context_init(ctx, &vme_context->gpe_context);
/* VME output buffer */
dri_bo_unreference(vme_context->vme_output.bo);
break;
}
}
+
if ((pSliceParameter->slice_type == SLICE_TYPE_I) ||
(pSliceParameter->slice_type == SLICE_TYPE_I)) {
kernel_shader = VME_INTRA_SHADER;
kernel_shader = VME_INTER_SHADER;
}
if (allow_hwscore)
- gen7_vme_walker_fill_vme_batchbuffer(ctx,
+ gen8wa_vme_walker_fill_vme_batchbuffer(ctx,
encode_state,
width_in_mbs, height_in_mbs,
kernel_shader,
}
static void
+gen8wa_vme_mpeg2_walker_fill_vme_batchbuffer(VADriverContextP ctx,
+ struct encode_state *encode_state,
+ int mb_width, int mb_height,
+ int kernel,
+ struct intel_encoder_context *encoder_context)
+{
+ struct gen6_vme_context *vme_context = encoder_context->vme_context;
+ unsigned int *command_ptr;
+
+#define MPEG2_SCOREBOARD (1 << 21)
+
+ dri_bo_map(vme_context->vme_batchbuffer.bo, 1);
+ command_ptr = vme_context->vme_batchbuffer.bo->virtual;
+
+ {
+ unsigned int mb_intra_ub, score_dep;
+ int x_outer, y_outer, x_inner, y_inner;
+ int xtemp_outer = 0;
+ int first_mb = 0;
+ int num_mb = mb_width * mb_height;
+
+ x_outer = 0;
+ y_outer = 0;
+
+
+ for (; x_outer < (mb_width -2 ) && !loop_in_bounds(x_outer, y_outer, first_mb, num_mb, mb_width, mb_height); ) {
+ x_inner = x_outer;
+ y_inner = y_outer;
+ for (; !loop_in_bounds(x_inner, y_inner, first_mb, num_mb, mb_width, mb_height);) {
+ mb_intra_ub = 0;
+ score_dep = 0;
+ if (x_inner != 0) {
+ mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_AE;
+ score_dep |= MB_SCOREBOARD_A;
+ }
+ if (y_inner != 0) {
+ mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_B;
+ score_dep |= MB_SCOREBOARD_B;
+
+ if (x_inner != 0)
+ mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_D;
+
+ if (x_inner != (mb_width -1)) {
+ mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_C;
+ score_dep |= MB_SCOREBOARD_C;
+ }
+ }
+
+ *command_ptr++ = (CMD_MEDIA_OBJECT | (8 - 2));
+ *command_ptr++ = kernel;
+ *command_ptr++ = MPEG2_SCOREBOARD;
+ /* Indirect data */
+ *command_ptr++ = 0;
+ /* the (X, Y) term of scoreboard */
+ *command_ptr++ = ((y_inner << 16) | x_inner);
+ *command_ptr++ = score_dep;
+ /*inline data */
+ *command_ptr++ = (mb_width << 16 | y_inner << 8 | x_inner);
+ *command_ptr++ = ((1 << 18) | (1 << 16) | (mb_intra_ub << 8));
+ *command_ptr++ = CMD_MEDIA_STATE_FLUSH;
+ *command_ptr++ = 0;
+
+ x_inner -= 2;
+ y_inner += 1;
+ }
+ x_outer += 1;
+ }
+
+ xtemp_outer = mb_width - 2;
+ if (xtemp_outer < 0)
+ xtemp_outer = 0;
+ x_outer = xtemp_outer;
+ y_outer = 0;
+ for (;!loop_in_bounds(x_outer, y_outer, first_mb, num_mb, mb_width, mb_height); ) {
+ y_inner = y_outer;
+ x_inner = x_outer;
+ for (; !loop_in_bounds(x_inner, y_inner, first_mb, num_mb, mb_width, mb_height);) {
+ mb_intra_ub = 0;
+ score_dep = 0;
+ if (x_inner != 0) {
+ mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_AE;
+ score_dep |= MB_SCOREBOARD_A;
+ }
+ if (y_inner != 0) {
+ mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_B;
+ score_dep |= MB_SCOREBOARD_B;
+
+ if (x_inner != 0)
+ mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_D;
+
+ if (x_inner != (mb_width -1)) {
+ mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_C;
+ score_dep |= MB_SCOREBOARD_C;
+ }
+ }
+
+ *command_ptr++ = (CMD_MEDIA_OBJECT | (8 - 2));
+ *command_ptr++ = kernel;
+ *command_ptr++ = MPEG2_SCOREBOARD;
+ /* Indirect data */
+ *command_ptr++ = 0;
+ /* the (X, Y) term of scoreboard */
+ *command_ptr++ = ((y_inner << 16) | x_inner);
+ *command_ptr++ = score_dep;
+ /*inline data */
+ *command_ptr++ = (mb_width << 16 | y_inner << 8 | x_inner);
+ *command_ptr++ = ((1 << 18) | (1 << 16) | (mb_intra_ub << 8));
+
+ *command_ptr++ = CMD_MEDIA_STATE_FLUSH;
+ *command_ptr++ = 0;
+ x_inner -= 2;
+ y_inner += 1;
+ }
+ x_outer++;
+ if (x_outer >= mb_width) {
+ y_outer += 1;
+ x_outer = xtemp_outer;
+ }
+ }
+ }
+
+ *command_ptr++ = MI_BATCH_BUFFER_END;
+ *command_ptr++ = 0;
+
+ dri_bo_unmap(vme_context->vme_batchbuffer.bo);
+ return;
+}
+
+static void
gen8_vme_mpeg2_fill_vme_batchbuffer(VADriverContextP ctx,
struct encode_state *encode_state,
int mb_width, int mb_height,
*command_ptr++ = (mb_width << 16 | mb_y << 8 | mb_x);
*command_ptr++ = ( (1 << 16) | transform_8x8_mode_flag | (mb_intra_ub << 8));
+ *command_ptr++ = CMD_MEDIA_STATE_FLUSH;
+ *command_ptr++ = 0;
i += 1;
}
}
}
- *command_ptr++ = 0;
*command_ptr++ = MI_BATCH_BUFFER_END;
+ *command_ptr++ = 0;
dri_bo_unmap(vme_context->vme_batchbuffer.bo);
}
}
if (allow_hwscore)
- gen7_vme_mpeg2_walker_fill_vme_batchbuffer(ctx,
+ gen8wa_vme_mpeg2_walker_fill_vme_batchbuffer(ctx,
encode_state,
width_in_mbs, height_in_mbs,
kernel_shader,
intel_batchbuffer_start_atomic(batch, 0x1000);
gen8_gpe_pipeline_setup(ctx, &vme_context->gpe_context, batch);
- BEGIN_BATCH(batch, 2);
- OUT_BATCH(batch, MI_BATCH_BUFFER_START | (2 << 6));
+ 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);
intel_batchbuffer_end_atomic(batch);
{
struct gen6_vme_context *vme_context = context;
- i965_gpe_context_destroy(&vme_context->gpe_context);
+ gen8_gpe_context_destroy(&vme_context->gpe_context);
dri_bo_unreference(vme_context->vme_output.bo);
vme_context->vme_output.bo = NULL;
vme_context->vme_kernel_sum = i965_kernel_num;
vme_context->gpe_context.surface_state_binding_table.length = (SURFACE_STATE_PADDED_SIZE + sizeof(unsigned int)) * MAX_MEDIA_SURFACES_GEN6;
- vme_context->gpe_context.idrt.max_entries = MAX_INTERFACE_DESC_GEN6;
- vme_context->gpe_context.idrt.entry_size = sizeof(struct gen8_interface_descriptor_data);
+ vme_context->gpe_context.idrt_size = sizeof(struct gen8_interface_descriptor_data) * MAX_INTERFACE_DESC_GEN6;
+ vme_context->gpe_context.curbe_size = CURBE_TOTAL_DATA_LENGTH;
+ vme_context->gpe_context.sampler_size = 0;
- vme_context->gpe_context.curbe.length = CURBE_TOTAL_DATA_LENGTH;
vme_context->gpe_context.vfe_state.max_num_threads = 60 - 1;
vme_context->gpe_context.vfe_state.num_urb_entries = 16;
gen7_vme_scoreboard_init(ctx, vme_context);
- i965_gpe_load_kernels(ctx,
+ gen8_gpe_load_kernels(ctx,
&vme_context->gpe_context,
vme_kernel_list,
i965_kernel_num);
projects / platform / upstream / libva-intel-driver.git / blobdiff