X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fgallium%2Fdrivers%2Fv3d%2Fv3dx_draw.c;h=629f35a8d7fd523b148908b7722d7bd6af1acde2;hb=9b96ae69bc7f07d267e880fe77a389c2b944f019;hp=5ad4710876f9e1a3b6a4857be94fa48892d2a1da;hpb=533b3f054152f1eab3b1880b14b744afc03294ca;p=mesa.git diff --git a/src/gallium/drivers/v3d/v3dx_draw.c b/src/gallium/drivers/v3d/v3dx_draw.c index 5ad4710876f..629f35a8d7f 100644 --- a/src/gallium/drivers/v3d/v3dx_draw.c +++ b/src/gallium/drivers/v3d/v3dx_draw.c @@ -55,7 +55,28 @@ v3d_start_draw(struct v3d_context *v3d) 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 * @@ -136,7 +157,8 @@ v3d_predraw_check_stage_inputs(struct pipe_context *pctx, continue; struct v3d_sampler_view *view = v3d_sampler_view(pview); - if (view->texture != view->base.texture) + 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); @@ -203,14 +225,24 @@ v3d_emit_gl_shader_state(struct v3d_context *v3d, * 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; @@ -235,9 +267,11 @@ v3d_emit_gl_shader_state(struct v3d_context *v3d, 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; @@ -294,6 +328,7 @@ v3d_emit_gl_shader_state(struct v3d_context *v3d, 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 = @@ -313,11 +348,25 @@ v3d_emit_gl_shader_state(struct v3d_context *v3d, 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_ATTRIBUTES * size); + STATIC_ASSERT(sizeof(vtx->attrs) >= V3D_MAX_VS_INPUTS / 4 * size); } if (vtx->num_elements == 0) { @@ -463,7 +512,7 @@ v3d_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info) /* Before setting up the draw, flush anything writing to the textures * that we read from. */ - 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) @@ -479,16 +528,16 @@ v3d_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info) * 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_TYPES; s++) { + 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);