From: Samuel Pitoiset Date: Mon, 12 Nov 2018 16:57:12 +0000 (+0100) Subject: radv: implement fast HTILE clears for depth or stencil only on GFX9 X-Git-Url: https://git.libre-soc.org/?a=commitdiff_plain;h=724107553ccca64bddae45288c15dc0dfd6baf59;p=mesa.git radv: implement fast HTILE clears for depth or stencil only on GFX9 This allows to fast clear the depth part (or the stencil part) of a depth+stencil surface when HTILE is enabled. I didn't test on GFX8, so it's disabled currently. This gives a very nice boost, for example when clearing the depth aspect of a 4096x4096 D32_SFLOAT_S8_UINT image (18x faster). BEFORE: 235 us AFTER: 13 us Signed-off-by: Samuel Pitoiset Reviewed-by: Bas Nieuwenhuizen --- diff --git a/src/amd/vulkan/radv_meta_clear.c b/src/amd/vulkan/radv_meta_clear.c index 9f74936e988..7fdd374afa9 100644 --- a/src/amd/vulkan/radv_meta_clear.c +++ b/src/amd/vulkan/radv_meta_clear.c @@ -303,6 +303,22 @@ create_color_pipeline(struct radv_device *device, return result; } +static void +finish_meta_clear_htile_mask_state(struct radv_device *device) +{ + struct radv_meta_state *state = &device->meta_state; + + radv_DestroyPipeline(radv_device_to_handle(device), + state->clear_htile_mask_pipeline, + &state->alloc); + radv_DestroyPipelineLayout(radv_device_to_handle(device), + state->clear_htile_mask_p_layout, + &state->alloc); + radv_DestroyDescriptorSetLayout(radv_device_to_handle(device), + state->clear_htile_mask_ds_layout, + &state->alloc); +} + void radv_device_finish_meta_clear_state(struct radv_device *device) { @@ -339,6 +355,8 @@ radv_device_finish_meta_clear_state(struct radv_device *device) radv_DestroyPipelineLayout(radv_device_to_handle(device), state->clear_depth_p_layout, &state->alloc); + + finish_meta_clear_htile_mask_state(device); } static void @@ -746,6 +764,69 @@ emit_depthstencil_clear(struct radv_cmd_buffer *cmd_buffer, } } +static uint32_t +clear_htile_mask(struct radv_cmd_buffer *cmd_buffer, + struct radeon_winsys_bo *bo, uint64_t offset, uint64_t size, + uint32_t htile_value, uint32_t htile_mask) +{ + struct radv_device *device = cmd_buffer->device; + struct radv_meta_state *state = &device->meta_state; + uint64_t block_count = round_up_u64(size, 1024); + struct radv_meta_saved_state saved_state; + + radv_meta_save(&saved_state, cmd_buffer, + RADV_META_SAVE_COMPUTE_PIPELINE | + RADV_META_SAVE_CONSTANTS | + RADV_META_SAVE_DESCRIPTORS); + + struct radv_buffer dst_buffer = { + .bo = bo, + .offset = offset, + .size = size + }; + + radv_CmdBindPipeline(radv_cmd_buffer_to_handle(cmd_buffer), + VK_PIPELINE_BIND_POINT_COMPUTE, + state->clear_htile_mask_pipeline); + + radv_meta_push_descriptor_set(cmd_buffer, VK_PIPELINE_BIND_POINT_COMPUTE, + state->clear_htile_mask_p_layout, + 0, /* set */ + 1, /* descriptorWriteCount */ + (VkWriteDescriptorSet[]) { + { + .sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET, + .dstBinding = 0, + .dstArrayElement = 0, + .descriptorCount = 1, + .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER, + .pBufferInfo = &(VkDescriptorBufferInfo) { + .buffer = radv_buffer_to_handle(&dst_buffer), + .offset = 0, + .range = size + } + } + }); + + const unsigned constants[2] = { + htile_value & htile_mask, + ~htile_mask, + }; + + radv_CmdPushConstants(radv_cmd_buffer_to_handle(cmd_buffer), + state->clear_htile_mask_p_layout, + VK_SHADER_STAGE_COMPUTE_BIT, 0, 8, + constants); + + radv_CmdDispatch(radv_cmd_buffer_to_handle(cmd_buffer), block_count, 1, 1); + + radv_meta_restore(&saved_state, cmd_buffer); + + return RADV_CMD_FLAG_CS_PARTIAL_FLUSH | + RADV_CMD_FLAG_INV_VMEM_L1 | + RADV_CMD_FLAG_WRITEBACK_GLOBAL_L2; +} + static uint32_t radv_get_htile_fast_clear_value(const struct radv_image *image, VkClearDepthStencilValue value) @@ -761,6 +842,24 @@ radv_get_htile_fast_clear_value(const struct radv_image *image, return clear_value; } +static uint32_t +radv_get_htile_mask(const struct radv_image *image, VkImageAspectFlags aspects) +{ + uint32_t mask = 0; + + if (!image->surface.has_stencil) { + /* All the HTILE buffer is used when there is no stencil. */ + mask = UINT32_MAX; + } else { + if (aspects & VK_IMAGE_ASPECT_DEPTH_BIT) + mask |= 0xfffffc0f; + if (aspects & VK_IMAGE_ASPECT_STENCIL_BIT) + mask |= 0x000003f0; + } + + return mask; +} + static bool radv_is_fast_clear_depth_allowed(VkClearDepthStencilValue value) { @@ -788,6 +887,7 @@ emit_fast_htile_clear(struct radv_cmd_buffer *cmd_buffer, VkClearDepthStencilValue clear_value = clear_att->clearValue.depthStencil; VkImageAspectFlags aspects = clear_att->aspectMask; uint32_t clear_word, flush_bits; + uint32_t htile_mask; if (!radv_image_has_htile(iview->image)) return false; @@ -821,9 +921,10 @@ emit_fast_htile_clear(struct radv_cmd_buffer *cmd_buffer, if (clear_rect->layerCount != iview->image->info.array_size) return false; - if (!(aspects & VK_IMAGE_ASPECT_DEPTH_BIT) || + if (cmd_buffer->device->physical_device->rad_info.chip_class < GFX9 && + (!(aspects & VK_IMAGE_ASPECT_DEPTH_BIT) || ((vk_format_aspects(iview->image->vk_format) & VK_IMAGE_ASPECT_STENCIL_BIT) && - !(aspects & VK_IMAGE_ASPECT_STENCIL_BIT))) + !(aspects & VK_IMAGE_ASPECT_STENCIL_BIT)))) return false; if (((aspects & VK_IMAGE_ASPECT_DEPTH_BIT) && @@ -841,6 +942,7 @@ emit_fast_htile_clear(struct radv_cmd_buffer *cmd_buffer, return false; clear_word = radv_get_htile_fast_clear_value(iview->image, clear_value); + htile_mask = radv_get_htile_mask(iview->image, aspects); if (pre_flush) { cmd_buffer->state.flush_bits |= (RADV_CMD_FLAG_FLUSH_AND_INV_DB | @@ -850,9 +952,19 @@ emit_fast_htile_clear(struct radv_cmd_buffer *cmd_buffer, cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_DB | RADV_CMD_FLAG_FLUSH_AND_INV_DB_META; - flush_bits = radv_fill_buffer(cmd_buffer, iview->image->bo, - iview->image->offset + iview->image->htile_offset, - iview->image->surface.htile_size, clear_word); + if (htile_mask == UINT_MAX) { + /* Clear the whole HTILE buffer. */ + flush_bits = radv_fill_buffer(cmd_buffer, iview->image->bo, + iview->image->offset + iview->image->htile_offset, + iview->image->surface.htile_size, clear_word); + } else { + /* Only clear depth or stencil bytes in the HTILE buffer. */ + assert(cmd_buffer->device->physical_device->rad_info.chip_class >= GFX9); + flush_bits = clear_htile_mask(cmd_buffer, iview->image->bo, + iview->image->offset + iview->image->htile_offset, + iview->image->surface.htile_size, clear_word, + htile_mask); + } radv_update_ds_clear_metadata(cmd_buffer, iview->image, clear_value, aspects); if (post_flush) { @@ -864,6 +976,148 @@ emit_fast_htile_clear(struct radv_cmd_buffer *cmd_buffer, return true; } +static nir_shader * +build_clear_htile_mask_shader() +{ + nir_builder b; + + nir_builder_init_simple_shader(&b, NULL, MESA_SHADER_COMPUTE, NULL); + b.shader->info.name = ralloc_strdup(b.shader, "meta_clear_htile_mask"); + b.shader->info.cs.local_size[0] = 64; + b.shader->info.cs.local_size[1] = 1; + b.shader->info.cs.local_size[2] = 1; + + nir_ssa_def *invoc_id = nir_load_system_value(&b, nir_intrinsic_load_local_invocation_id, 0); + nir_ssa_def *wg_id = nir_load_system_value(&b, nir_intrinsic_load_work_group_id, 0); + nir_ssa_def *block_size = nir_imm_ivec4(&b, + b.shader->info.cs.local_size[0], + b.shader->info.cs.local_size[1], + b.shader->info.cs.local_size[2], 0); + + nir_ssa_def *global_id = nir_iadd(&b, nir_imul(&b, wg_id, block_size), invoc_id); + + nir_ssa_def *offset = nir_imul(&b, global_id, nir_imm_int(&b, 16)); + offset = nir_channel(&b, offset, 0); + + nir_intrinsic_instr *buf = + nir_intrinsic_instr_create(b.shader, + nir_intrinsic_vulkan_resource_index); + + buf->src[0] = nir_src_for_ssa(nir_imm_int(&b, 0)); + nir_intrinsic_set_desc_set(buf, 0); + nir_intrinsic_set_binding(buf, 0); + nir_ssa_dest_init(&buf->instr, &buf->dest, 1, 32, NULL); + nir_builder_instr_insert(&b, &buf->instr); + + nir_intrinsic_instr *constants = + nir_intrinsic_instr_create(b.shader, + nir_intrinsic_load_push_constant); + nir_intrinsic_set_base(constants, 0); + nir_intrinsic_set_range(constants, 8); + constants->src[0] = nir_src_for_ssa(nir_imm_int(&b, 0)); + constants->num_components = 2; + nir_ssa_dest_init(&constants->instr, &constants->dest, 2, 32, "constants"); + nir_builder_instr_insert(&b, &constants->instr); + + nir_intrinsic_instr *load = + nir_intrinsic_instr_create(b.shader, nir_intrinsic_load_ssbo); + load->src[0] = nir_src_for_ssa(&buf->dest.ssa); + load->src[1] = nir_src_for_ssa(offset); + nir_ssa_dest_init(&load->instr, &load->dest, 4, 32, NULL); + load->num_components = 4; + nir_builder_instr_insert(&b, &load->instr); + + /* data = (data & ~htile_mask) | (htile_value & htile_mask) */ + nir_ssa_def *data = + nir_iand(&b, &load->dest.ssa, + nir_channel(&b, &constants->dest.ssa, 1)); + data = nir_ior(&b, data, nir_channel(&b, &constants->dest.ssa, 0)); + + nir_intrinsic_instr *store = + nir_intrinsic_instr_create(b.shader, nir_intrinsic_store_ssbo); + store->src[0] = nir_src_for_ssa(data); + store->src[1] = nir_src_for_ssa(&buf->dest.ssa); + store->src[2] = nir_src_for_ssa(offset); + nir_intrinsic_set_write_mask(store, 0xf); + store->num_components = 4; + nir_builder_instr_insert(&b, &store->instr); + + return b.shader; +} + +static VkResult +init_meta_clear_htile_mask_state(struct radv_device *device) +{ + struct radv_meta_state *state = &device->meta_state; + struct radv_shader_module cs = { .nir = NULL }; + VkResult result; + + cs.nir = build_clear_htile_mask_shader(); + + VkDescriptorSetLayoutCreateInfo ds_layout_info = { + .sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO, + .flags = VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR, + .bindingCount = 1, + .pBindings = (VkDescriptorSetLayoutBinding[]) { + { + .binding = 0, + .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER, + .descriptorCount = 1, + .stageFlags = VK_SHADER_STAGE_COMPUTE_BIT, + .pImmutableSamplers = NULL + }, + } + }; + + result = radv_CreateDescriptorSetLayout(radv_device_to_handle(device), + &ds_layout_info, &state->alloc, + &state->clear_htile_mask_ds_layout); + if (result != VK_SUCCESS) + goto fail; + + VkPipelineLayoutCreateInfo p_layout_info = { + .sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO, + .setLayoutCount = 1, + .pSetLayouts = &state->clear_htile_mask_ds_layout, + .pushConstantRangeCount = 1, + .pPushConstantRanges = &(VkPushConstantRange){ + VK_SHADER_STAGE_COMPUTE_BIT, 0, 8, + }, + }; + + result = radv_CreatePipelineLayout(radv_device_to_handle(device), + &p_layout_info, &state->alloc, + &state->clear_htile_mask_p_layout); + if (result != VK_SUCCESS) + goto fail; + + VkPipelineShaderStageCreateInfo shader_stage = { + .sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO, + .stage = VK_SHADER_STAGE_COMPUTE_BIT, + .module = radv_shader_module_to_handle(&cs), + .pName = "main", + .pSpecializationInfo = NULL, + }; + + VkComputePipelineCreateInfo pipeline_info = { + .sType = VK_STRUCTURE_TYPE_COMPUTE_PIPELINE_CREATE_INFO, + .stage = shader_stage, + .flags = 0, + .layout = state->clear_htile_mask_p_layout, + }; + + result = radv_CreateComputePipelines(radv_device_to_handle(device), + radv_pipeline_cache_to_handle(&state->cache), + 1, &pipeline_info, NULL, + &state->clear_htile_mask_pipeline); + + ralloc_free(cs.nir); + return result; +fail: + ralloc_free(cs.nir); + return result; +} + VkResult radv_device_init_meta_clear_state(struct radv_device *device, bool on_demand) { @@ -898,6 +1152,10 @@ radv_device_init_meta_clear_state(struct radv_device *device, bool on_demand) if (res != VK_SUCCESS) goto fail; + res = init_meta_clear_htile_mask_state(device); + if (res != VK_SUCCESS) + goto fail; + if (on_demand) return VK_SUCCESS; diff --git a/src/amd/vulkan/radv_private.h b/src/amd/vulkan/radv_private.h index 24e922da5e8..ac756f2c247 100644 --- a/src/amd/vulkan/radv_private.h +++ b/src/amd/vulkan/radv_private.h @@ -457,6 +457,12 @@ struct radv_meta_state { VkPipelineLayout clear_color_p_layout; VkPipelineLayout clear_depth_p_layout; + + /* Optimized compute fast HTILE clear for stencil or depth only. */ + VkPipeline clear_htile_mask_pipeline; + VkPipelineLayout clear_htile_mask_p_layout; + VkDescriptorSetLayout clear_htile_mask_ds_layout; + struct { VkRenderPass render_pass[NUM_META_FS_KEYS][RADV_META_DST_LAYOUT_COUNT];