From 821a49981ff386559f8a8fdf6bf3526b8deb2415 Mon Sep 17 00:00:00 2001 From: Mike Blumenkrantz Date: Thu, 3 Mar 2022 13:37:51 -0500 Subject: [PATCH] lavapipe: accurately set image/ssbo access based on shader usage Reviewed-by: Dave Airlie Part-of: --- src/gallium/frontends/lavapipe/lvp_execute.c | 24 +++++++++++++++++--- 1 file changed, 21 insertions(+), 3 deletions(-) diff --git a/src/gallium/frontends/lavapipe/lvp_execute.c b/src/gallium/frontends/lavapipe/lvp_execute.c index 815f1eb0270..271a0a9568b 100644 --- a/src/gallium/frontends/lavapipe/lvp_execute.c +++ b/src/gallium/frontends/lavapipe/lvp_execute.c @@ -42,6 +42,7 @@ #include "util/u_prim_restart.h" #include "util/format/u_format_zs.h" #include "util/ptralloc.h" +#include "tgsi/tgsi_from_mesa.h" #include "vk_util.h" @@ -116,6 +117,8 @@ struct rendering_state { struct pipe_vertex_buffer vb[PIPE_MAX_ATTRIBS]; struct cso_velems_state velem; + struct lvp_access_info access[MESA_SHADER_STAGES]; + struct pipe_sampler_view *sv[PIPE_SHADER_TYPES][PIPE_MAX_SAMPLERS]; int num_sampler_views[PIPE_SHADER_TYPES]; struct pipe_sampler_state ss[PIPE_SHADER_TYPES][PIPE_MAX_SAMPLERS]; @@ -335,7 +338,7 @@ static void emit_state(struct rendering_state *state) if (state->sb_dirty[sh]) { state->pctx->set_shader_buffers(state->pctx, sh, 0, state->num_shader_buffers[sh], - state->sb[sh], (1 << state->num_shader_buffers[sh]) - 1); + state->sb[sh], state->access[tgsi_processor_to_shader_stage(sh)].buffers_written); } } @@ -380,6 +383,8 @@ static void handle_compute_pipeline(struct vk_cmd_queue_entry *cmd, { LVP_FROM_HANDLE(lvp_pipeline, pipeline, cmd->u.bind_pipeline.pipeline); + memcpy(&state->access[MESA_SHADER_COMPUTE], &pipeline->access[MESA_SHADER_COMPUTE], sizeof(struct lvp_access_info)); + state->dispatch_info.block[0] = pipeline->pipeline_nir[MESA_SHADER_COMPUTE]->info.workgroup_size[0]; state->dispatch_info.block[1] = pipeline->pipeline_nir[MESA_SHADER_COMPUTE]->info.workgroup_size[1]; state->dispatch_info.block[2] = pipeline->pipeline_nir[MESA_SHADER_COMPUTE]->info.workgroup_size[2]; @@ -489,6 +494,8 @@ static void handle_graphics_pipeline(struct vk_cmd_queue_entry *cmd, unsigned fb_samples = 0; bool clip_halfz = state->rs_state.clip_halfz; + memcpy(state->access, pipeline->access, sizeof(struct lvp_access_info) * 5); //4 vertex stages + fragment + memset(dynamic_states, 0, sizeof(dynamic_states)); if (pipeline->graphics_create_info.pDynamicState) { @@ -1136,8 +1143,19 @@ static void fill_image_view_stage(struct rendering_state *state, state->iv[p_stage][idx].u.tex.last_layer = iv->subresourceRange.baseArrayLayer + lvp_get_layerCount(iv->image, &iv->subresourceRange) - 1; } state->iv[p_stage][idx].u.tex.level = iv->subresourceRange.baseMipLevel; - state->iv[p_stage][idx].access = PIPE_IMAGE_ACCESS_READ_WRITE; - state->iv[p_stage][idx].shader_access = PIPE_IMAGE_ACCESS_READ_WRITE; + + assert(idx < 32); + state->iv[p_stage][idx].access = 0; + state->iv[p_stage][idx].shader_access = 0; + if (state->access[stage].images_read & BITFIELD_BIT(idx)) { + state->iv[p_stage][idx].access |= PIPE_IMAGE_ACCESS_READ; + state->iv[p_stage][idx].shader_access |= PIPE_IMAGE_ACCESS_READ; + } + if (state->access[stage].images_written & BITFIELD_BIT(idx)) { + state->iv[p_stage][idx].access |= PIPE_IMAGE_ACCESS_WRITE; + state->iv[p_stage][idx].shader_access |= PIPE_IMAGE_ACCESS_WRITE; + } + if (state->num_shader_images[p_stage] <= idx) state->num_shader_images[p_stage] = idx + 1;