X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Famd%2Fvulkan%2Fradv_meta_resolve.c;h=49326fe9d10693005a01b7e4909ddce14ca63d64;hb=b462ceb482471dd4aacdd5fa7cef9fe25f823d70;hp=3e6633dec03fa9c127690f1794e5842c2163edc6;hpb=69136f4e633be4eeed0b226b02c9876511414b76;p=mesa.git diff --git a/src/amd/vulkan/radv_meta_resolve.c b/src/amd/vulkan/radv_meta_resolve.c index 3e6633dec03..49326fe9d10 100644 --- a/src/amd/vulkan/radv_meta_resolve.c +++ b/src/amd/vulkan/radv_meta_resolve.c @@ -38,7 +38,7 @@ build_nir_fs(void) nir_variable *f_color; /* vec4, fragment output color */ nir_builder_init_simple_shader(&b, NULL, MESA_SHADER_FRAGMENT, NULL); - b.shader->info->name = ralloc_asprintf(b.shader, + b.shader->info.name = ralloc_asprintf(b.shader, "meta_resolve_fs"); f_color = nir_variable_create(b.shader, nir_var_shader_out, vec4, @@ -121,6 +121,21 @@ create_pipeline(struct radv_device *device, goto cleanup; } + VkPipelineLayoutCreateInfo pl_create_info = { + .sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO, + .setLayoutCount = 0, + .pSetLayouts = NULL, + .pushConstantRangeCount = 0, + .pPushConstantRanges = NULL, + }; + + result = radv_CreatePipelineLayout(radv_device_to_handle(device), + &pl_create_info, + &device->meta_state.alloc, + &device->meta_state.resolve.p_layout); + if (result != VK_SUCCESS) + goto cleanup; + result = radv_graphics_pipeline_create(device_h, radv_pipeline_cache_to_handle(&device->meta_state.cache), &(VkGraphicsPipelineCreateInfo) { @@ -196,6 +211,7 @@ create_pipeline(struct radv_device *device, VK_DYNAMIC_STATE_SCISSOR, }, }, + .layout = device->meta_state.resolve.p_layout, .renderPass = device->meta_state.resolve.pass, .subpass = 0, }, @@ -219,18 +235,13 @@ void radv_device_finish_meta_resolve_state(struct radv_device *device) { struct radv_meta_state *state = &device->meta_state; - VkDevice device_h = radv_device_to_handle(device); - VkRenderPass pass_h = device->meta_state.resolve.pass; - const VkAllocationCallbacks *alloc = &device->meta_state.alloc; - - if (pass_h) - radv_DestroyRenderPass(device_h, pass_h, - &device->meta_state.alloc); - VkPipeline pipeline_h = state->resolve.pipeline; - if (pipeline_h) { - radv_DestroyPipeline(device_h, pipeline_h, alloc); - } + radv_DestroyRenderPass(radv_device_to_handle(device), + state->resolve.pass, &state->alloc); + radv_DestroyPipelineLayout(radv_device_to_handle(device), + state->resolve.p_layout, &state->alloc); + radv_DestroyPipeline(radv_device_to_handle(device), + state->resolve.pipeline, &state->alloc); } VkResult @@ -238,8 +249,6 @@ radv_device_init_meta_resolve_state(struct radv_device *device) { VkResult res = VK_SUCCESS; - zero(device->meta_state.resolve); - struct radv_shader_module vs_module = { .nir = radv_meta_build_nir_vs_generate_vertices() }; if (!vs_module.nir) { /* XXX: Need more accurate error */ @@ -277,13 +286,8 @@ emit_resolve(struct radv_cmd_buffer *cmd_buffer, cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_CB; - VkPipeline pipeline_h = device->meta_state.resolve.pipeline; - RADV_FROM_HANDLE(radv_pipeline, pipeline, pipeline_h); - - if (cmd_buffer->state.pipeline != pipeline) { - radv_CmdBindPipeline(cmd_buffer_h, VK_PIPELINE_BIND_POINT_GRAPHICS, - pipeline_h); - } + radv_CmdBindPipeline(cmd_buffer_h, VK_PIPELINE_BIND_POINT_GRAPHICS, + device->meta_state.resolve.pipeline); radv_CmdSetViewport(radv_cmd_buffer_to_handle(cmd_buffer), 0, 1, &(VkViewport) { .x = dest_offset->x, @@ -303,6 +307,29 @@ emit_resolve(struct radv_cmd_buffer *cmd_buffer, cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_CB; } +enum radv_resolve_method { + RESOLVE_HW, + RESOLVE_COMPUTE, + RESOLVE_FRAGMENT, +}; + +static void radv_pick_resolve_method_images(struct radv_image *src_image, + struct radv_image *dest_image, + VkImageLayout dest_image_layout, + struct radv_cmd_buffer *cmd_buffer, + enum radv_resolve_method *method) + +{ + uint32_t queue_mask = radv_image_queue_family_mask(dest_image, + cmd_buffer->queue_family_index, + cmd_buffer->queue_family_index); + if (radv_layout_dcc_compressed(dest_image, dest_image_layout, queue_mask)) { + *method = RESOLVE_FRAGMENT; + } else if (dest_image->surface.micro_tile_mode != src_image->surface.micro_tile_mode) { + *method = RESOLVE_COMPUTE; + } +} + void radv_CmdResolveImage( VkCommandBuffer cmd_buffer_h, VkImage src_image_h, @@ -318,27 +345,30 @@ void radv_CmdResolveImage( struct radv_device *device = cmd_buffer->device; struct radv_meta_saved_state saved_state; VkDevice device_h = radv_device_to_handle(device); - bool use_compute_resolve = false; - bool use_fragment_resolve = false; + enum radv_resolve_method resolve_method = RESOLVE_HW; /* we can use the hw resolve only for single full resolves */ if (region_count == 1) { if (regions[0].srcOffset.x || regions[0].srcOffset.y || regions[0].srcOffset.z) - use_compute_resolve = true; + resolve_method = RESOLVE_COMPUTE; if (regions[0].dstOffset.x || regions[0].dstOffset.y || regions[0].dstOffset.z) - use_compute_resolve = true; + resolve_method = RESOLVE_COMPUTE; if (regions[0].extent.width != src_image->info.width || regions[0].extent.height != src_image->info.height || regions[0].extent.depth != src_image->info.depth) - use_compute_resolve = true; + resolve_method = RESOLVE_COMPUTE; } else - use_compute_resolve = true; + resolve_method = RESOLVE_COMPUTE; + + radv_pick_resolve_method_images(src_image, dest_image, + dest_image_layout, cmd_buffer, + &resolve_method); - if (use_fragment_resolve) { + if (resolve_method == RESOLVE_FRAGMENT) { radv_meta_resolve_fragment_image(cmd_buffer, src_image, src_image_layout, @@ -348,7 +378,7 @@ void radv_CmdResolveImage( return; } - if (use_compute_resolve) { + if (resolve_method == RESOLVE_COMPUTE) { radv_meta_resolve_compute_image(cmd_buffer, src_image, src_image_layout, @@ -358,9 +388,15 @@ void radv_CmdResolveImage( return; } - radv_meta_save_graphics_reset_vport_scissor_novertex(&saved_state, cmd_buffer); + radv_meta_save(&saved_state, cmd_buffer, + RADV_META_SAVE_GRAPHICS_PIPELINE); assert(src_image->info.samples > 1); + if (src_image->info.samples <= 1) { + /* this causes GPU hangs if we get past here */ + fprintf(stderr, "radv: Illegal resolve operation (src not multisampled), will hang GPU."); + return; + } assert(dest_image->info.samples == 1); if (src_image->info.samples >= 16) { @@ -437,8 +473,7 @@ void radv_CmdResolveImage( .baseArrayLayer = src_base_layer + layer, .layerCount = 1, }, - }, - cmd_buffer, VK_IMAGE_USAGE_SAMPLED_BIT); + }); struct radv_image_view dest_iview; radv_image_view_init(&dest_iview, cmd_buffer->device, @@ -454,8 +489,7 @@ void radv_CmdResolveImage( .baseArrayLayer = dest_base_layer + layer, .layerCount = 1, }, - }, - cmd_buffer, VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT); + }); VkFramebuffer fb_h; radv_CreateFramebuffer(device_h, @@ -524,6 +558,7 @@ radv_cmd_buffer_resolve_subpass(struct radv_cmd_buffer *cmd_buffer) struct radv_framebuffer *fb = cmd_buffer->state.framebuffer; const struct radv_subpass *subpass = cmd_buffer->state.subpass; struct radv_meta_saved_state saved_state; + enum radv_resolve_method resolve_method = RESOLVE_HW; /* FINISHME(perf): Skip clears for resolve attachments. * @@ -537,15 +572,44 @@ radv_cmd_buffer_resolve_subpass(struct radv_cmd_buffer *cmd_buffer) if (!subpass->has_resolve) return; - radv_meta_save_graphics_reset_vport_scissor(&saved_state, cmd_buffer); - for (uint32_t i = 0; i < subpass->color_count; ++i) { VkAttachmentReference src_att = subpass->color_attachments[i]; VkAttachmentReference dest_att = subpass->resolve_attachments[i]; + + if (src_att.attachment == VK_ATTACHMENT_UNUSED || + dest_att.attachment == VK_ATTACHMENT_UNUSED) + continue; + struct radv_image *dst_img = cmd_buffer->state.framebuffer->attachments[dest_att.attachment].attachment->image; - if (dest_att.attachment == VK_ATTACHMENT_UNUSED) + struct radv_image *src_img = cmd_buffer->state.framebuffer->attachments[src_att.attachment].attachment->image; + + radv_pick_resolve_method_images(dst_img, src_img, dest_att.layout, cmd_buffer, &resolve_method); + if (resolve_method == RESOLVE_FRAGMENT) { + break; + } + } + + if (resolve_method == RESOLVE_COMPUTE) { + radv_cmd_buffer_resolve_subpass_cs(cmd_buffer); + return; + } else if (resolve_method == RESOLVE_FRAGMENT) { + radv_cmd_buffer_resolve_subpass_fs(cmd_buffer); + return; + } + + radv_meta_save(&saved_state, cmd_buffer, + RADV_META_SAVE_GRAPHICS_PIPELINE); + + for (uint32_t i = 0; i < subpass->color_count; ++i) { + VkAttachmentReference src_att = subpass->color_attachments[i]; + VkAttachmentReference dest_att = subpass->resolve_attachments[i]; + + if (src_att.attachment == VK_ATTACHMENT_UNUSED || + dest_att.attachment == VK_ATTACHMENT_UNUSED) continue; + struct radv_image *dst_img = cmd_buffer->state.framebuffer->attachments[dest_att.attachment].attachment->image; + if (dst_img->surface.dcc_size) { radv_initialize_dcc(cmd_buffer, dst_img, 0xffffffff); cmd_buffer->state.attachments[dest_att.attachment].current_layout = VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL; @@ -559,13 +623,6 @@ radv_cmd_buffer_resolve_subpass(struct radv_cmd_buffer *cmd_buffer) radv_cmd_buffer_set_subpass(cmd_buffer, &resolve_subpass, false); - /* Subpass resolves must respect the render area. We can ignore the - * render area here because vkCmdBeginRenderPass set the render area - * with 3DSTATE_DRAWING_RECTANGLE. - * - * XXX(chadv): Does the hardware really respect - * 3DSTATE_DRAWING_RECTANGLE when draing a 3DPRIM_RECTLIST? - */ emit_resolve(cmd_buffer, &(VkOffset2D) { 0, 0 }, &(VkExtent2D) { fb->width, fb->height });