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 *
{
struct v3d_context *v3d = v3d_context(pctx);
+ /* XXX perf: If we're reading from the output of TF in this job, we
+ * should instead be using the wait for transform feedback
+ * functionality.
+ */
+
/* Flush writes to textures we're sampling. */
for (int i = 0; i < v3d->tex[s].num_textures; i++) {
- struct pipe_sampler_view *view = v3d->tex[s].textures[i];
- if (!view)
+ 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);
}
if (cb->buffer)
v3d_flush_jobs_writing_resource(v3d, cb->buffer);
}
+
+ /* Flush 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_writing_resource(v3d, view->base.resource);
+ }
+
+ /* 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);
+ }
+ }
+}
+
+static void
+v3d_predraw_check_outputs(struct pipe_context *pctx)
+{
+ struct v3d_context *v3d = v3d_context(pctx);
+
+ /* 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);
+ }
+ }
}
static void
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_center_w;
+#if V3D_VERSION >= 40
+ 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;
v3d->prog.vs->prog_data.vs->separate_segments;
shader.coordinate_shader_input_vpm_segment_size =
- v3d->prog.cs->prog_data.vs->vpm_input_size;
+ 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 =
- v3d->prog.vs->prog_data.vs->vpm_input_size;
+ 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;
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) >= VC5_MAX_ATTRIBUTES * size);
+ STATIC_ASSERT(sizeof(vtx->attrs) >= V3D_MAX_VS_INPUTS / 4 * size);
}
if (vtx->num_elements == 0) {
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_TYPES; s++)
+ 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_predraw_check_outputs(pctx);
+
struct v3d_job *job = v3d_get_job_for_fbo(v3d);
/* If vertex texturing depends on the output of rendering, we need to
* on the last submitted render, rather than tracking the last
* rendering to each texture's BO.
*/
- if (v3d->tex[PIPE_SHADER_VERTEX].num_textures) {
+ if (v3d->tex[PIPE_SHADER_VERTEX].num_textures || info->indirect) {
perf_debug("Blocking binner on last render "
- "due to vertex texturing.\n");
+ "due to vertex texturing or indirect drawing.\n");
job->submit.in_sync_bcl = v3d->out_sync;
}
+ /* Mark SSBOs 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.
*/
}
#endif
- if (info->instance_count > 1) {
+ 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
if (info->has_user_indices)
pipe_resource_reference(&prsc, NULL);
} else {
- if (info->instance_count > 1) {
+ 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) {
cl_emit(&job->bcl, VERTEX_ARRAY_INSTANCED_PRIMS, prim) {
prim.mode = info->mode | prim_tf_enable;
prim.index_of_first_vertex = info->start;
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;
buffers &= ~PIPE_CLEAR_DEPTHSTENCIL;
}
- 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;
if (!(buffers & bit))
continue;