job->submit.bcl_start = job->bcl.bo->offset;
v3d_job_add_bo(job, job->bcl.bo);
- job->tile_alloc = v3d_bo_alloc(v3d->screen, 1024 * 1024, "tile alloc");
+ /* The PTB will request the tile alloc initial size per tile at start
+ * of tile binning.
+ */
+ uint32_t tile_alloc_size = (job->draw_tiles_x *
+ job->draw_tiles_y) * 64;
+ /* The PTB allocates in aligned 4k chunks after the initial setup. */
+ tile_alloc_size = align(tile_alloc_size, 4096);
+
+ /* Include the first two chunk allocations that the PTB does so that
+ * we definitely clear the OOM condition before triggering one (the HW
+ * won't trigger OOM during the first allocations).
+ */
+ tile_alloc_size += 8192;
+
+ /* For performance, allocate some extra initial memory after the PTB's
+ * minimal allocations, so that we hopefully don't have to block the
+ * GPU on the kernel handling an OOM signal.
+ */
+ tile_alloc_size += 512 * 1024;
+
+ job->tile_alloc = v3d_bo_alloc(v3d->screen, tile_alloc_size,
+ "tile_alloc");
uint32_t tsda_per_tile_size = v3d->screen->devinfo.ver >= 40 ? 256 : 64;
job->tile_state = v3d_bo_alloc(v3d->screen,
job->draw_tiles_y *
tsda_per_tile_size,
"TSDA");
-#if V3D_VERSION < 40
+#if V3D_VERSION >= 40
+ cl_emit(&job->bcl, TILE_BINNING_MODE_CFG, config) {
+ config.width_in_pixels = v3d->framebuffer.width;
+ config.height_in_pixels = v3d->framebuffer.height;
+ config.number_of_render_targets =
+ MAX2(v3d->framebuffer.nr_cbufs, 1);
+
+ config.multisample_mode_4x = job->msaa;
+
+ config.maximum_bpp_of_all_render_targets = job->internal_bpp;
+ }
+#else /* V3D_VERSION < 40 */
/* "Binning mode lists start with a Tile Binning Mode Configuration
* item (120)"
*
* Part1 signals the end of binning config setup.
*/
- cl_emit(&job->bcl, TILE_BINNING_MODE_CONFIGURATION_PART2, config) {
+ cl_emit(&job->bcl, TILE_BINNING_MODE_CFG_PART2, config) {
config.tile_allocation_memory_address =
cl_address(job->tile_alloc, 0);
config.tile_allocation_memory_size = job->tile_alloc->size;
}
-#endif
- cl_emit(&job->bcl, TILE_BINNING_MODE_CONFIGURATION_PART1, config) {
-#if V3D_VERSION >= 40
- config.width_in_pixels_minus_1 = v3d->framebuffer.width - 1;
- config.height_in_pixels_minus_1 = v3d->framebuffer.height - 1;
- config.number_of_render_targets_minus_1 =
- MAX2(v3d->framebuffer.nr_cbufs, 1) - 1;
-#else /* V3D_VERSION < 40 */
+ cl_emit(&job->bcl, TILE_BINNING_MODE_CFG_PART1, config) {
config.tile_state_data_array_base_address =
cl_address(job->tile_state, 0);
/* Must be >= 1 */
config.number_of_render_targets =
MAX2(v3d->framebuffer.nr_cbufs, 1);
-#endif /* V3D_VERSION < 40 */
config.multisample_mode_4x = job->msaa;
config.maximum_bpp_of_all_render_targets = job->internal_bpp;
}
+#endif /* V3D_VERSION < 40 */
/* There's definitely nothing in the VCD cache we want. */
cl_emit(&job->bcl, FLUSH_VCD_CACHE, bin);
}
static void
-v3d_predraw_check_textures(struct pipe_context *pctx,
- struct v3d_texture_stateobj *stage_tex)
+v3d_predraw_check_stage_inputs(struct pipe_context *pctx,
+ enum pipe_shader_type s)
+{
+ struct v3d_context *v3d = v3d_context(pctx);
+
+ /* Flush writes to textures we're sampling. */
+ for (int i = 0; i < v3d->tex[s].num_textures; i++) {
+ struct pipe_sampler_view *pview = v3d->tex[s].textures[i];
+ if (!pview)
+ continue;
+ struct v3d_sampler_view *view = v3d_sampler_view(pview);
+
+ if (view->texture != view->base.texture &&
+ view->base.format != PIPE_FORMAT_X32_S8X24_UINT)
+ v3d_update_shadow_texture(pctx, &view->base);
+
+ v3d_flush_jobs_writing_resource(v3d, view->texture,
+ V3D_FLUSH_DEFAULT);
+ }
+
+ /* Flush writes to UBOs. */
+ foreach_bit(i, v3d->constbuf[s].enabled_mask) {
+ struct pipe_constant_buffer *cb = &v3d->constbuf[s].cb[i];
+ if (cb->buffer) {
+ v3d_flush_jobs_writing_resource(v3d, cb->buffer,
+ V3D_FLUSH_DEFAULT);
+ }
+ }
+
+ /* Flush reads/writes to our SSBOs */
+ foreach_bit(i, v3d->ssbo[s].enabled_mask) {
+ struct pipe_shader_buffer *sb = &v3d->ssbo[s].sb[i];
+ if (sb->buffer) {
+ v3d_flush_jobs_reading_resource(v3d, sb->buffer,
+ V3D_FLUSH_NOT_CURRENT_JOB);
+ }
+ }
+
+ /* Flush reads/writes to our image views */
+ foreach_bit(i, v3d->shaderimg[s].enabled_mask) {
+ struct v3d_image_view *view = &v3d->shaderimg[s].si[i];
+
+ v3d_flush_jobs_reading_resource(v3d, view->base.resource,
+ V3D_FLUSH_NOT_CURRENT_JOB);
+ }
+
+ /* Flush writes to our vertex buffers (i.e. from transform feedback) */
+ if (s == PIPE_SHADER_VERTEX) {
+ foreach_bit(i, v3d->vertexbuf.enabled_mask) {
+ struct pipe_vertex_buffer *vb = &v3d->vertexbuf.vb[i];
+
+ v3d_flush_jobs_writing_resource(v3d, vb->buffer.resource,
+ V3D_FLUSH_DEFAULT);
+ }
+ }
+}
+
+static void
+v3d_predraw_check_outputs(struct pipe_context *pctx)
{
struct v3d_context *v3d = v3d_context(pctx);
- for (int i = 0; i < stage_tex->num_textures; i++) {
- struct pipe_sampler_view *view = stage_tex->textures[i];
- if (!view)
+ /* Flush jobs reading from TF buffers that we are about to write. */
+ if (v3d_transform_feedback_enabled(v3d)) {
+ struct v3d_streamout_stateobj *so = &v3d->streamout;
+
+ for (int i = 0; i < so->num_targets; i++) {
+ if (!so->targets[i])
+ continue;
+
+ const struct pipe_stream_output_target *target =
+ so->targets[i];
+ v3d_flush_jobs_reading_resource(v3d, target->buffer,
+ V3D_FLUSH_DEFAULT);
+ }
+ }
+}
+
+/**
+ * Checks if the state for the current draw reads a particular resource in
+ * in the given shader stage.
+ */
+static bool
+v3d_state_reads_resource(struct v3d_context *v3d,
+ struct pipe_resource *prsc,
+ enum pipe_shader_type s)
+{
+ struct v3d_resource *rsc = v3d_resource(prsc);
+
+ /* Vertex buffers */
+ if (s == PIPE_SHADER_VERTEX) {
+ foreach_bit(i, v3d->vertexbuf.enabled_mask) {
+ struct pipe_vertex_buffer *vb = &v3d->vertexbuf.vb[i];
+ if (!vb->buffer.resource)
+ continue;
+
+ struct v3d_resource *vb_rsc =
+ v3d_resource(vb->buffer.resource);
+ if (rsc->bo == vb_rsc->bo)
+ return true;
+ }
+ }
+
+ /* Constant buffers */
+ foreach_bit(i, v3d->constbuf[s].enabled_mask) {
+ struct pipe_constant_buffer *cb = &v3d->constbuf[s].cb[i];
+ if (!cb->buffer)
+ continue;
+
+ struct v3d_resource *cb_rsc = v3d_resource(cb->buffer);
+ if (rsc->bo == cb_rsc->bo)
+ return true;
+ }
+
+ /* Shader storage buffers */
+ foreach_bit(i, v3d->ssbo[s].enabled_mask) {
+ struct pipe_shader_buffer *sb = &v3d->ssbo[s].sb[i];
+ if (!sb->buffer)
continue;
- v3d_flush_jobs_writing_resource(v3d, view->texture);
+ struct v3d_resource *sb_rsc = v3d_resource(sb->buffer);
+ if (rsc->bo == sb_rsc->bo)
+ return true;
+ }
+
+ /* Textures */
+ for (int i = 0; i < v3d->tex[s].num_textures; i++) {
+ struct pipe_sampler_view *pview = v3d->tex[s].textures[i];
+ if (!pview)
+ continue;
+
+ struct v3d_sampler_view *view = v3d_sampler_view(pview);
+ struct v3d_resource *v_rsc = v3d_resource(view->texture);
+ if (rsc->bo == v_rsc->bo)
+ return true;
+ }
+
+ return false;
+}
+
+static void
+v3d_emit_wait_for_tf(struct v3d_job *job)
+{
+ /* XXX: we might be able to skip this in some cases, for now we
+ * always emit it.
+ */
+ cl_emit(&job->bcl, FLUSH_TRANSFORM_FEEDBACK_DATA, flush);
+
+ cl_emit(&job->bcl, WAIT_FOR_TRANSFORM_FEEDBACK, wait) {
+ /* XXX: Wait for all outstanding writes... maybe we can do
+ * better in some cases.
+ */
+ wait.block_count = 255;
+ }
+
+ /* We have just flushed all our outstanding TF work in this job so make
+ * sure we don't emit TF flushes again for any of it again.
+ */
+ _mesa_set_clear(job->tf_write_prscs, NULL);
+}
+
+static void
+v3d_emit_wait_for_tf_if_needed(struct v3d_context *v3d, struct v3d_job *job)
+{
+ if (!job->tf_enabled)
+ return;
+
+ set_foreach(job->tf_write_prscs, entry) {
+ struct pipe_resource *prsc = (struct pipe_resource *)entry->key;
+ for (int s = 0; s < PIPE_SHADER_COMPUTE; s++) {
+ /* Fragment shaders can only start executing after all
+ * binning (and thus TF) is complete.
+ *
+ * XXX: For VS/GS/TES, if the binning shader does not
+ * read the resource then we could also avoid emitting
+ * the wait.
+ */
+ if (s == PIPE_SHADER_FRAGMENT)
+ continue;
+
+ if (v3d_state_reads_resource(v3d, prsc, s)) {
+ v3d_emit_wait_for_tf(job);
+ return;
+ }
+ }
}
}
/* Upload the uniforms to the indirect CL first */
struct v3d_cl_reloc fs_uniforms =
v3d_write_uniforms(v3d, v3d->prog.fs,
- &v3d->constbuf[PIPE_SHADER_FRAGMENT],
- &v3d->fragtex);
+ PIPE_SHADER_FRAGMENT);
struct v3d_cl_reloc vs_uniforms =
v3d_write_uniforms(v3d, v3d->prog.vs,
- &v3d->constbuf[PIPE_SHADER_VERTEX],
- &v3d->verttex);
+ PIPE_SHADER_VERTEX);
struct v3d_cl_reloc cs_uniforms =
v3d_write_uniforms(v3d, v3d->prog.cs,
- &v3d->constbuf[PIPE_SHADER_VERTEX],
- &v3d->verttex);
+ PIPE_SHADER_VERTEX);
+
+ /* Update the cache dirty flag based on the shader progs data */
+ job->tmu_dirty_rcl |= v3d->prog.cs->prog_data.vs->base.tmu_dirty_rcl;
+ job->tmu_dirty_rcl |= v3d->prog.vs->prog_data.vs->base.tmu_dirty_rcl;
+ job->tmu_dirty_rcl |= v3d->prog.fs->prog_data.fs->base.tmu_dirty_rcl;
/* See GFXH-930 workaround below */
uint32_t num_elements_to_emit = MAX2(vtx->num_elements, 1);
cl_packet_length(GL_SHADER_STATE_ATTRIBUTE_RECORD),
32);
+ /* XXX perf: We should move most of the SHADER_STATE_RECORD setup to
+ * compile time, so that we mostly just have to OR the VS and FS
+ * records together at draw time.
+ */
cl_emit(&job->indirect, GL_SHADER_STATE_RECORD, shader) {
shader.enable_clipping = true;
/* VC5_DIRTY_PRIM_MODE | VC5_DIRTY_RASTERIZER */
* shader needs to write the Z value (even just discards).
*/
shader.fragment_shader_does_z_writes =
- (v3d->prog.fs->prog_data.fs->writes_z ||
- v3d->prog.fs->prog_data.fs->discard);
+ v3d->prog.fs->prog_data.fs->writes_z;
+ /* Set if the EZ test must be disabled (due to shader side
+ * effects and the early_z flag not being present in the
+ * shader).
+ */
+ shader.turn_off_early_z_test =
+ v3d->prog.fs->prog_data.fs->disable_ez;
shader.fragment_shader_uses_real_pixel_centre_w_in_addition_to_centroid_w2 =
- v3d->prog.fs->prog_data.fs->uses_centroid_and_center_w;
+ v3d->prog.fs->prog_data.fs->uses_center_w;
+
+#if V3D_VERSION >= 40
+ shader.do_scoreboard_wait_on_first_thread_switch =
+ v3d->prog.fs->prog_data.fs->lock_scoreboard_on_first_thrsw;
+ shader.disable_implicit_point_line_varyings =
+ !v3d->prog.fs->prog_data.fs->uses_implicit_point_line_varyings;
+#endif
shader.number_of_varyings_in_fragment_shader =
- v3d->prog.fs->prog_data.base->num_inputs;
+ v3d->prog.fs->prog_data.fs->num_inputs;
shader.coordinate_shader_propagate_nans = true;
shader.vertex_shader_propagate_nans = true;
shader.fragment_shader_propagate_nans = true;
shader.coordinate_shader_code_address =
- cl_address(v3d->prog.cs->bo, 0);
+ cl_address(v3d_resource(v3d->prog.cs->resource)->bo,
+ v3d->prog.cs->offset);
shader.vertex_shader_code_address =
- cl_address(v3d->prog.vs->bo, 0);
+ cl_address(v3d_resource(v3d->prog.vs->resource)->bo,
+ v3d->prog.vs->offset);
shader.fragment_shader_code_address =
- cl_address(v3d->prog.fs->bo, 0);
+ cl_address(v3d_resource(v3d->prog.fs->resource)->bo,
+ v3d->prog.fs->offset);
/* XXX: Use combined input/output size flag in the common
* case.
*/
- shader.coordinate_shader_has_separate_input_and_output_vpm_blocks = true;
- shader.vertex_shader_has_separate_input_and_output_vpm_blocks = true;
+ shader.coordinate_shader_has_separate_input_and_output_vpm_blocks =
+ v3d->prog.cs->prog_data.vs->separate_segments;
+ shader.vertex_shader_has_separate_input_and_output_vpm_blocks =
+ v3d->prog.vs->prog_data.vs->separate_segments;
+
shader.coordinate_shader_input_vpm_segment_size =
- MAX2(v3d->prog.cs->prog_data.vs->vpm_input_size, 1);
+ v3d->prog.cs->prog_data.vs->separate_segments ?
+ v3d->prog.cs->prog_data.vs->vpm_input_size : 1;
shader.vertex_shader_input_vpm_segment_size =
- MAX2(v3d->prog.vs->prog_data.vs->vpm_input_size, 1);
+ v3d->prog.vs->prog_data.vs->separate_segments ?
+ v3d->prog.vs->prog_data.vs->vpm_input_size : 1;
shader.coordinate_shader_output_vpm_segment_size =
v3d->prog.cs->prog_data.vs->vpm_output_size;
shader.fragment_shader_uniforms_address = fs_uniforms;
#if V3D_VERSION >= 41
+ shader.min_coord_shader_input_segments_required_in_play = 1;
+ shader.min_vertex_shader_input_segments_required_in_play = 1;
+
shader.coordinate_shader_4_way_threadable =
v3d->prog.cs->prog_data.vs->base.threads == 4;
shader.vertex_shader_4_way_threadable =
v3d->prog.vs->prog_data.vs->uses_iid;
shader.address_of_default_attribute_values =
- cl_address(vtx->default_attribute_values, 0);
+ cl_address(v3d_resource(vtx->defaults)->bo,
+ vtx->defaults_offset);
}
+ bool cs_loaded_any = false;
for (int i = 0; i < vtx->num_elements; i++) {
struct pipe_vertex_element *elem = &vtx->pipe[i];
struct pipe_vertex_buffer *vb =
v3d->prog.cs->prog_data.vs->vattr_sizes[i];
attr.number_of_values_read_by_vertex_shader =
v3d->prog.vs->prog_data.vs->vattr_sizes[i];
+
+ /* GFXH-930: At least one attribute must be enabled
+ * and read by CS and VS. If we have attributes being
+ * consumed by the VS but not the CS, then set up a
+ * dummy load of the last attribute into the CS's VPM
+ * inputs. (Since CS is just dead-code-elimination
+ * compared to VS, we can't have CS loading but not
+ * VS).
+ */
+ if (v3d->prog.cs->prog_data.vs->vattr_sizes[i])
+ cs_loaded_any = true;
+ if (i == vtx->num_elements - 1 && !cs_loaded_any) {
+ attr.number_of_values_read_by_coordinate_shader = 1;
+ }
#if V3D_VERSION >= 41
attr.maximum_index = 0xffffff;
#endif
}
+ STATIC_ASSERT(sizeof(vtx->attrs) >= V3D_MAX_VS_INPUTS / 4 * size);
}
if (vtx->num_elements == 0) {
}
}
+ cl_emit(&job->bcl, VCM_CACHE_SIZE, vcm) {
+ vcm.number_of_16_vertex_batches_for_binning =
+ v3d->prog.cs->prog_data.vs->vcm_cache_size;
+ vcm.number_of_16_vertex_batches_for_rendering =
+ v3d->prog.vs->prog_data.vs->vcm_cache_size;
+ }
+
cl_emit(&job->bcl, GL_SHADER_STATE, state) {
state.address = cl_address(job->indirect.bo, shader_rec_offset);
state.number_of_attribute_arrays = num_elements_to_emit;
v3d_bo_unreference(&cs_uniforms.bo);
v3d_bo_unreference(&vs_uniforms.bo);
v3d_bo_unreference(&fs_uniforms.bo);
-
- job->shader_rec_count++;
}
/**
- * Computes the various transform feedback statistics, since they can't be
- * recorded by CL packets.
+ * Updates the number of primitvies generated from the number of vertices
+ * to draw. We do this here instead of using PRIMITIVE_COUNTS_FEEDBACK because
+ * using the GPU packet for this might require sync waits and this is trivial
+ * to handle in the CPU instead.
*/
static void
-v3d_tf_statistics_record(struct v3d_context *v3d,
- const struct pipe_draw_info *info,
- bool prim_tf)
+v3d_update_primitives_generated_counter(struct v3d_context *v3d,
+ const struct pipe_draw_info *info)
{
if (!v3d->active_queries)
return;
uint32_t prims = u_prims_for_vertices(info->mode, info->count);
v3d->prims_generated += prims;
-
- if (prim_tf) {
- /* XXX: Only count if we didn't overflow. */
- v3d->tf_prims_generated += prims;
- }
}
static void
job->ez_state = VC5_EZ_DISABLED;
}
- if (job->first_ez_state == VC5_EZ_UNDECIDED)
+ if (job->first_ez_state == VC5_EZ_UNDECIDED &&
+ (job->ez_state != VC5_EZ_DISABLED || job->draw_calls_queued == 0))
job->first_ez_state = job->ez_state;
}
return;
}
- /* Before setting up the draw, flush anything writing to the textures
- * that we read from.
+ /* Before setting up the draw, flush anything writing to the resources
+ * that we read from or reading from resources we write to.
+ */
+ for (int s = 0; s < PIPE_SHADER_COMPUTE; s++)
+ v3d_predraw_check_stage_inputs(pctx, s);
+
+ if (info->indirect) {
+ v3d_flush_jobs_writing_resource(v3d, info->indirect->buffer,
+ V3D_FLUSH_DEFAULT);
+ }
+
+ v3d_predraw_check_outputs(pctx);
+
+ /* If transform feedback is active and we are switching primitive type
+ * we need to submit the job before drawing and update the vertex count
+ * written to TF based on the primitive type since we will need to
+ * know the exact vertex count if the application decides to call
+ * glDrawTransformFeedback() later.
*/
- v3d_predraw_check_textures(pctx, &v3d->verttex);
- v3d_predraw_check_textures(pctx, &v3d->fragtex);
+ if (v3d->streamout.num_targets > 0 &&
+ u_base_prim_type(info->mode) != u_base_prim_type(v3d->prim_mode)) {
+ v3d_tf_update_counters(v3d);
+ }
struct v3d_job *job = v3d_get_job_for_fbo(v3d);
+ /* If vertex texturing depends on the output of rendering, we need to
+ * ensure that that rendering is complete before we run a coordinate
+ * shader that depends on it.
+ *
+ * Given that doing that is unusual, for now we just block the binner
+ * on the last submitted render, rather than tracking the last
+ * rendering to each texture's BO.
+ */
+ if (v3d->tex[PIPE_SHADER_VERTEX].num_textures || info->indirect) {
+ perf_debug("Blocking binner on last render "
+ "due to vertex texturing or indirect drawing.\n");
+ job->submit.in_sync_bcl = v3d->out_sync;
+ }
+
+ /* Mark SSBOs and images as being written. We don't actually know
+ * which ones are read vs written, so just assume the worst.
+ */
+ for (int s = 0; s < PIPE_SHADER_COMPUTE; s++) {
+ foreach_bit(i, v3d->ssbo[s].enabled_mask) {
+ v3d_job_add_write_resource(job,
+ v3d->ssbo[s].sb[i].buffer);
+ job->tmu_dirty_rcl = true;
+ }
+
+ foreach_bit(i, v3d->shaderimg[s].enabled_mask) {
+ v3d_job_add_write_resource(job,
+ v3d->shaderimg[s].si[i].base.resource);
+ job->tmu_dirty_rcl = true;
+ }
+ }
+
/* Get space to emit our draw call into the BCL, using a branch to
* jump to a new BO if necessary.
*/
v3d_update_compiled_shaders(v3d, info->mode);
v3d_update_job_ez(v3d, job);
+ /* If this job was writing to transform feedback buffers before this
+ * draw and we are reading from them here, then we need to wait for TF
+ * to complete before we emit this draw.
+ *
+ * Notice this check needs to happen before we emit state for the
+ * current draw call, where we update job->tf_enabled, so we can ensure
+ * that we only check TF writes for prior draws.
+ */
+ v3d_emit_wait_for_tf_if_needed(v3d, job);
+
#if V3D_VERSION >= 41
v3d41_emit_state(pctx);
#else
prim_tf_enable = (V3D_PRIM_POINTS_TF - V3D_PRIM_POINTS);
#endif
- v3d_tf_statistics_record(v3d, info, v3d->streamout.num_targets);
+ v3d_update_primitives_generated_counter(v3d, info);
/* Note that the primitive type fields match with OpenGL/gallium
* definitions, up to but not including QUADS.
}
#endif
- if (info->instance_count > 1) {
- cl_emit(&job->bcl, INDEXED_INSTANCED_PRIMITIVE_LIST, prim) {
+ if (info->indirect) {
+ cl_emit(&job->bcl, INDIRECT_INDEXED_INSTANCED_PRIM_LIST, prim) {
+ prim.index_type = ffs(info->index_size) - 1;
+#if V3D_VERSION < 40
+ prim.address_of_indices_list =
+ cl_address(rsc->bo, offset);
+#endif /* V3D_VERSION < 40 */
+ prim.mode = info->mode | prim_tf_enable;
+ prim.enable_primitive_restarts = info->primitive_restart;
+
+ prim.number_of_draw_indirect_indexed_records = info->indirect->draw_count;
+
+ prim.stride_in_multiples_of_4_bytes = info->indirect->stride >> 2;
+ prim.address = cl_address(v3d_resource(info->indirect->buffer)->bo,
+ info->indirect->offset);
+ }
+ } else if (info->instance_count > 1) {
+ cl_emit(&job->bcl, INDEXED_INSTANCED_PRIM_LIST, prim) {
prim.index_type = ffs(info->index_size) - 1;
#if V3D_VERSION >= 40
prim.index_offset = offset;
prim.instance_length = info->count;
}
} else {
- cl_emit(&job->bcl, INDEXED_PRIMITIVE_LIST, prim) {
+ cl_emit(&job->bcl, INDEXED_PRIM_LIST, prim) {
prim.index_type = ffs(info->index_size) - 1;
prim.length = info->count;
#if V3D_VERSION >= 40
}
}
- job->draw_calls_queued++;
-
if (info->has_user_indices)
pipe_resource_reference(&prsc, NULL);
} else {
- if (info->instance_count > 1) {
- cl_emit(&job->bcl, VERTEX_ARRAY_INSTANCED_PRIMITIVES, prim) {
+ if (info->indirect) {
+ cl_emit(&job->bcl, INDIRECT_VERTEX_ARRAY_INSTANCED_PRIMS, prim) {
+ prim.mode = info->mode | prim_tf_enable;
+ prim.number_of_draw_indirect_array_records = info->indirect->draw_count;
+
+ prim.stride_in_multiples_of_4_bytes = info->indirect->stride >> 2;
+ prim.address = cl_address(v3d_resource(info->indirect->buffer)->bo,
+ info->indirect->offset);
+ }
+ } else if (info->instance_count > 1) {
+ struct pipe_stream_output_target *so =
+ info->count_from_stream_output;
+ uint32_t vert_count = so ?
+ v3d_stream_output_target_get_vertex_count(so) :
+ info->count;
+ cl_emit(&job->bcl, VERTEX_ARRAY_INSTANCED_PRIMS, prim) {
prim.mode = info->mode | prim_tf_enable;
prim.index_of_first_vertex = info->start;
prim.number_of_instances = info->instance_count;
- prim.instance_length = info->count;
+ prim.instance_length = vert_count;
}
} else {
- cl_emit(&job->bcl, VERTEX_ARRAY_PRIMITIVES, prim) {
+ struct pipe_stream_output_target *so =
+ info->count_from_stream_output;
+ uint32_t vert_count = so ?
+ v3d_stream_output_target_get_vertex_count(so) :
+ info->count;
+ cl_emit(&job->bcl, VERTEX_ARRAY_PRIMS, prim) {
prim.mode = info->mode | prim_tf_enable;
- prim.length = info->count;
+ prim.length = vert_count;
prim.index_of_first_vertex = info->start;
}
}
}
+
+ /* A flush is required in between a TF draw and any following TF specs
+ * packet, or the GPU may hang. Just flush each time for now.
+ */
+ if (v3d->streamout.num_targets)
+ cl_emit(&job->bcl, TRANSFORM_FEEDBACK_FLUSH_AND_COUNT, flush);
+
job->draw_calls_queued++;
+ if (v3d->streamout.num_targets)
+ job->tf_draw_calls_queued++;
+
+ /* Increment the TF offsets by how many verts we wrote. XXX: This
+ * needs some clamping to the buffer size.
+ */
+ for (int i = 0; i < v3d->streamout.num_targets; i++)
+ v3d->streamout.offsets[i] += info->count;
- if (v3d->zsa && job->zsbuf &&
- (v3d->zsa->base.depth.enabled ||
- v3d->zsa->base.stencil[0].enabled)) {
+ if (v3d->zsa && job->zsbuf && v3d->zsa->base.depth.enabled) {
struct v3d_resource *rsc = v3d_resource(job->zsbuf->texture);
v3d_job_add_bo(job, rsc->bo);
- if (v3d->zsa->base.depth.enabled) {
- job->resolve |= PIPE_CLEAR_DEPTH;
- rsc->initialized_buffers = PIPE_CLEAR_DEPTH;
- }
+ job->load |= PIPE_CLEAR_DEPTH & ~job->clear;
+ if (v3d->zsa->base.depth.writemask)
+ job->store |= PIPE_CLEAR_DEPTH;
+ rsc->initialized_buffers = PIPE_CLEAR_DEPTH;
+ }
- if (v3d->zsa->base.stencil[0].enabled) {
- job->resolve |= PIPE_CLEAR_STENCIL;
- rsc->initialized_buffers |= PIPE_CLEAR_STENCIL;
+ if (v3d->zsa && job->zsbuf && v3d->zsa->base.stencil[0].enabled) {
+ struct v3d_resource *rsc = v3d_resource(job->zsbuf->texture);
+ if (rsc->separate_stencil)
+ rsc = rsc->separate_stencil;
+
+ v3d_job_add_bo(job, rsc->bo);
+
+ job->load |= PIPE_CLEAR_STENCIL & ~job->clear;
+ if (v3d->zsa->base.stencil[0].writemask ||
+ v3d->zsa->base.stencil[1].writemask) {
+ job->store |= PIPE_CLEAR_STENCIL;
}
+ rsc->initialized_buffers |= PIPE_CLEAR_STENCIL;
}
- for (int i = 0; i < VC5_MAX_DRAW_BUFFERS; i++) {
+ for (int i = 0; i < V3D_MAX_DRAW_BUFFERS; i++) {
uint32_t bit = PIPE_CLEAR_COLOR0 << i;
+ int blend_rt = v3d->blend->base.independent_blend_enable ? i : 0;
- if (job->resolve & bit || !job->cbufs[i])
+ if (job->store & bit || !job->cbufs[i])
continue;
struct v3d_resource *rsc = v3d_resource(job->cbufs[i]->texture);
- job->resolve |= bit;
+ job->load |= bit & ~job->clear;
+ if (v3d->blend->base.rt[blend_rt].colormask)
+ job->store |= bit;
v3d_job_add_bo(job, rsc->bo);
}
v3d_flush(pctx);
}
+/**
+ * Implements gallium's clear() hook (glClear()) by drawing a pair of triangles.
+ */
static void
-v3d_clear(struct pipe_context *pctx, unsigned buffers,
- const union pipe_color_union *color, double depth, unsigned stencil)
+v3d_draw_clear(struct v3d_context *v3d,
+ unsigned buffers,
+ const union pipe_color_union *color,
+ double depth, unsigned stencil)
{
- struct v3d_context *v3d = v3d_context(pctx);
- struct v3d_job *job = v3d_get_job_for_fbo(v3d);
+ static const union pipe_color_union dummy_color = {};
- /* We can't flag new buffers for clearing once we've queued draws. We
- * could avoid this by using the 3d engine to clear.
+ /* The blitter util dereferences the color regardless, even though the
+ * gallium clear API may not pass one in when only Z/S are cleared.
*/
+ if (!color)
+ color = &dummy_color;
+
+ v3d_blitter_save(v3d);
+ util_blitter_clear(v3d->blitter,
+ v3d->framebuffer.width,
+ v3d->framebuffer.height,
+ util_framebuffer_get_num_layers(&v3d->framebuffer),
+ buffers, color, depth, stencil,
+ util_framebuffer_get_num_samples(&v3d->framebuffer) > 1);
+}
+
+/**
+ * Attempts to perform the GL clear by using the TLB's fast clear at the start
+ * of the frame.
+ */
+static unsigned
+v3d_tlb_clear(struct v3d_job *job, unsigned buffers,
+ const union pipe_color_union *color,
+ double depth, unsigned stencil)
+{
+ struct v3d_context *v3d = job->v3d;
+
if (job->draw_calls_queued) {
- perf_debug("Flushing rendering to process new clear.\n");
- v3d_job_submit(v3d, job);
- job = v3d_get_job_for_fbo(v3d);
+ /* If anything in the CL has drawn using the buffer, then the
+ * TLB clear we're trying to add now would happen before that
+ * drawing.
+ */
+ buffers &= ~(job->load | job->store);
}
- for (int i = 0; i < VC5_MAX_DRAW_BUFFERS; i++) {
+ /* GFXH-1461: If we were to emit a load of just depth or just stencil,
+ * then the clear for the other may get lost. We need to decide now
+ * if it would be possible to need to emit a load of just one after
+ * we've set up our TLB clears.
+ */
+ if (buffers & PIPE_CLEAR_DEPTHSTENCIL &&
+ (buffers & PIPE_CLEAR_DEPTHSTENCIL) != PIPE_CLEAR_DEPTHSTENCIL &&
+ job->zsbuf &&
+ util_format_is_depth_and_stencil(job->zsbuf->texture->format)) {
+ buffers &= ~PIPE_CLEAR_DEPTHSTENCIL;
+ }
+
+ for (int i = 0; i < V3D_MAX_DRAW_BUFFERS; i++) {
uint32_t bit = PIPE_CLEAR_COLOR0 << i;
if (!(buffers & bit))
continue;
job->draw_min_y = 0;
job->draw_max_x = v3d->framebuffer.width;
job->draw_max_y = v3d->framebuffer.height;
- job->cleared |= buffers;
- job->resolve |= buffers;
+ job->clear |= buffers;
+ job->store |= buffers;
v3d_start_draw(v3d);
+
+ return buffers;
+}
+
+static void
+v3d_clear(struct pipe_context *pctx, unsigned buffers,
+ const union pipe_color_union *color, double depth, unsigned stencil)
+{
+ struct v3d_context *v3d = v3d_context(pctx);
+ struct v3d_job *job = v3d_get_job_for_fbo(v3d);
+
+ buffers &= ~v3d_tlb_clear(job, buffers, color, depth, stencil);
+
+ if (buffers)
+ v3d_draw_clear(v3d, buffers, color, depth, stencil);
}
static void
projects / mesa.git / blobdiff