X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Famd%2Fvulkan%2Fradv_meta_resolve.c;h=49326fe9d10693005a01b7e4909ddce14ca63d64;hb=b462ceb482471dd4aacdd5fa7cef9fe25f823d70;hp=105270cb74219ee1dd9217efc0da23602303fdb1;hpb=870032978512576f5ac882b2c88c7657c9f741f4;p=mesa.git diff --git a/src/amd/vulkan/radv_meta_resolve.c b/src/amd/vulkan/radv_meta_resolve.c index 105270cb742..49326fe9d10 100644 --- a/src/amd/vulkan/radv_meta_resolve.c +++ b/src/amd/vulkan/radv_meta_resolve.c @@ -28,40 +28,8 @@ #include "radv_private.h" #include "nir/nir_builder.h" #include "sid.h" -/** - * Vertex attributes used by all pipelines. - */ -struct vertex_attrs { - float position[2]; /**< 3DPRIM_RECTLIST */ -}; - -/* passthrough vertex shader */ -static nir_shader * -build_nir_vs(void) -{ - const struct glsl_type *vec4 = glsl_vec4_type(); - - nir_builder b; - nir_variable *a_position; - nir_variable *v_position; - nir_builder_init_simple_shader(&b, NULL, MESA_SHADER_VERTEX, NULL); - b.shader->info->name = ralloc_strdup(b.shader, "meta_resolve_vs"); - - a_position = nir_variable_create(b.shader, nir_var_shader_in, vec4, - "a_position"); - a_position->data.location = VERT_ATTRIB_GENERIC0; - - v_position = nir_variable_create(b.shader, nir_var_shader_out, vec4, - "gl_Position"); - v_position->data.location = VARYING_SLOT_POS; - - nir_copy_var(&b, v_position, a_position); - - return b.shader; -} - -/* simple passthrough shader */ +/* emit 0, 0, 0, 1 */ static nir_shader * build_nir_fs(void) { @@ -70,7 +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, @@ -153,6 +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) { @@ -174,24 +157,8 @@ create_pipeline(struct radv_device *device, }, .pVertexInputState = &(VkPipelineVertexInputStateCreateInfo) { .sType = VK_STRUCTURE_TYPE_PIPELINE_VERTEX_INPUT_STATE_CREATE_INFO, - .vertexBindingDescriptionCount = 1, - .pVertexBindingDescriptions = (VkVertexInputBindingDescription[]) { - { - .binding = 0, - .stride = sizeof(struct vertex_attrs), - .inputRate = VK_VERTEX_INPUT_RATE_VERTEX - }, - }, - .vertexAttributeDescriptionCount = 1, - .pVertexAttributeDescriptions = (VkVertexInputAttributeDescription[]) { - { - /* Position */ - .location = 0, - .binding = 0, - .format = VK_FORMAT_R32G32_SFLOAT, - .offset = offsetof(struct vertex_attrs, position), - }, - }, + .vertexBindingDescriptionCount = 0, + .vertexAttributeDescriptionCount = 0, }, .pInputAssemblyState = &(VkPipelineInputAssemblyStateCreateInfo) { .sType = VK_STRUCTURE_TYPE_PIPELINE_INPUT_ASSEMBLY_STATE_CREATE_INFO, @@ -200,8 +167,8 @@ create_pipeline(struct radv_device *device, }, .pViewportState = &(VkPipelineViewportStateCreateInfo) { .sType = VK_STRUCTURE_TYPE_PIPELINE_VIEWPORT_STATE_CREATE_INFO, - .viewportCount = 0, - .scissorCount = 0, + .viewportCount = 1, + .scissorCount = 1, }, .pRasterizationState = &(VkPipelineRasterizationStateCreateInfo) { .sType = VK_STRUCTURE_TYPE_PIPELINE_RASTERIZATION_STATE_CREATE_INFO, @@ -236,7 +203,15 @@ create_pipeline(struct radv_device *device, } }, }, - .pDynamicState = NULL, + .pDynamicState = &(VkPipelineDynamicStateCreateInfo) { + .sType = VK_STRUCTURE_TYPE_PIPELINE_DYNAMIC_STATE_CREATE_INFO, + .dynamicStateCount = 2, + .pDynamicStates = (VkDynamicState[]) { + VK_DYNAMIC_STATE_VIEWPORT, + VK_DYNAMIC_STATE_SCISSOR, + }, + }, + .layout = device->meta_state.resolve.p_layout, .renderPass = device->meta_state.resolve.pass, .subpass = 0, }, @@ -260,18 +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 @@ -279,9 +249,7 @@ 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 = build_nir_vs() }; + struct radv_shader_module vs_module = { .nir = radv_meta_build_nir_vs_generate_vertices() }; if (!vs_module.nir) { /* XXX: Need more accurate error */ res = VK_ERROR_OUT_OF_HOST_MEMORY; @@ -315,57 +283,53 @@ emit_resolve(struct radv_cmd_buffer *cmd_buffer, { struct radv_device *device = cmd_buffer->device; VkCommandBuffer cmd_buffer_h = radv_cmd_buffer_to_handle(cmd_buffer); - uint32_t offset; - const struct vertex_attrs vertex_data[3] = { - { - .position = { - dest_offset->x, - dest_offset->y, - }, - }, - { - .position = { - dest_offset->x, - dest_offset->y + resolve_extent->height, - }, - }, - { - .position = { - dest_offset->x + resolve_extent->width, - dest_offset->y, - }, - }, - }; cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_CB; - radv_cmd_buffer_upload_data(cmd_buffer, sizeof(vertex_data), 16, vertex_data, &offset); - struct radv_buffer vertex_buffer = { - .device = device, - .size = sizeof(vertex_data), - .bo = cmd_buffer->upload.upload_bo, - .offset = offset, - }; - VkBuffer vertex_buffer_h = radv_buffer_to_handle(&vertex_buffer); + radv_CmdBindPipeline(cmd_buffer_h, VK_PIPELINE_BIND_POINT_GRAPHICS, + device->meta_state.resolve.pipeline); - radv_CmdBindVertexBuffers(cmd_buffer_h, - /*firstBinding*/ 0, - /*bindingCount*/ 1, - (VkBuffer[]) { vertex_buffer_h }, - (VkDeviceSize[]) { 0 }); + radv_CmdSetViewport(radv_cmd_buffer_to_handle(cmd_buffer), 0, 1, &(VkViewport) { + .x = dest_offset->x, + .y = dest_offset->y, + .width = resolve_extent->width, + .height = resolve_extent->height, + .minDepth = 0.0f, + .maxDepth = 1.0f + }); - 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_CmdSetScissor(radv_cmd_buffer_to_handle(cmd_buffer), 0, 1, &(VkRect2D) { + .offset = *dest_offset, + .extent = *resolve_extent, + }); radv_CmdDraw(cmd_buffer_h, 3, 1, 0, 0); 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, @@ -381,28 +345,40 @@ 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; - + 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->extent.width || - regions[0].extent.height != src_image->extent.height || - regions[0].extent.depth != src_image->extent.depth) - use_compute_resolve = true; + 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) + resolve_method = RESOLVE_COMPUTE; } else - use_compute_resolve = true; - - if (use_compute_resolve) { + resolve_method = RESOLVE_COMPUTE; + + radv_pick_resolve_method_images(src_image, dest_image, + dest_image_layout, cmd_buffer, + &resolve_method); + + if (resolve_method == RESOLVE_FRAGMENT) { + radv_meta_resolve_fragment_image(cmd_buffer, + src_image, + src_image_layout, + dest_image, + dest_image_layout, + region_count, regions); + return; + } + if (resolve_method == RESOLVE_COMPUTE) { radv_meta_resolve_compute_image(cmd_buffer, src_image, src_image_layout, @@ -412,12 +388,18 @@ void radv_CmdResolveImage( return; } - radv_meta_save_graphics_reset_vport_scissor(&saved_state, cmd_buffer); + radv_meta_save(&saved_state, cmd_buffer, + RADV_META_SAVE_GRAPHICS_PIPELINE); - assert(src_image->samples > 1); - assert(dest_image->samples == 1); + 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->samples >= 16) { + if (src_image->info.samples >= 16) { /* See commit aa3f9aaf31e9056a255f9e0472ebdfdaa60abe54 for the * glBlitFramebuffer workaround for samples >= 16. */ @@ -425,7 +407,7 @@ void radv_CmdResolveImage( "samples >= 16"); } - if (src_image->array_size > 1) + if (src_image->info.array_size > 1) radv_finishme("vkCmdResolveImage: multisample array images"); if (dest_image->surface.dcc_size) { @@ -491,8 +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, @@ -508,8 +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, @@ -520,9 +500,9 @@ void radv_CmdResolveImage( radv_image_view_to_handle(&src_iview), radv_image_view_to_handle(&dest_iview), }, - .width = radv_minify(dest_image->extent.width, + .width = radv_minify(dest_image->info.width, region->dstSubresource.mipLevel), - .height = radv_minify(dest_image->extent.height, + .height = radv_minify(dest_image->info.height, region->dstSubresource.mipLevel), .layers = 1 }, @@ -578,6 +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. * @@ -591,15 +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; @@ -613,13 +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 });