radv: fix a performance regression with graphics depth/stencil clears

I recently changed the slow depth/stencil clear path to make sure
depth values are explicitly exported by the fragment shader. This
is actually only useful when VK_EXT_depth_range_unrestricted is
enabled.

While this path is correct, it introduced a performance regression
with Heroes of the Storm, Shadow of Mordor (Vulkan beta) and
probably more titles. This is because it prevents the hardware
to do some optimizations like discarding fragments.

This commit re-introduces the previous (a bit faster) slow
depth/stencil clear path and it selects the unrestricted path
only if VK_EXT_depth_range_unrestricted is enabled.

Closes: https://gitlab.freedesktop.org/mesa/mesa/issues/863
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

2 files changed, 123 insertions, 25 deletions

diff --git a/src/amd/vulkan/radv_meta_clear.c b/src/amd/vulkan/radv_meta_clear.c
index 636a9643843..d96fd4a7a1c 100644
--- a/src/amd/vulkan/radv_meta_clear.c
+++ b/src/amd/vulkan/radv_meta_clear.c
@@ -344,6 +344,16 @@ radv_device_finish_meta_clear_state(struct radv_device *device)
 			radv_DestroyPipeline(radv_device_to_handle(device),
 					     state->clear[i].depthstencil_pipeline[j],
 					     &state->alloc);
+
+			radv_DestroyPipeline(radv_device_to_handle(device),
+					     state->clear[i].depth_only_unrestricted_pipeline[j],
+					     &state->alloc);
+			radv_DestroyPipeline(radv_device_to_handle(device),
+					     state->clear[i].stencil_only_unrestricted_pipeline[j],
+					     &state->alloc);
+			radv_DestroyPipeline(radv_device_to_handle(device),
+					     state->clear[i].depthstencil_unrestricted_pipeline[j],
+					     &state->alloc);
 		}
 		radv_DestroyRenderPass(radv_device_to_handle(device),
 				      state->clear[i].depthstencil_rp,
@@ -355,6 +365,9 @@ radv_device_finish_meta_clear_state(struct radv_device *device)
 	radv_DestroyPipelineLayout(radv_device_to_handle(device),
 				   state->clear_depth_p_layout,
 				   &state->alloc);
+	radv_DestroyPipelineLayout(radv_device_to_handle(device),
+				   state->clear_depth_unrestricted_p_layout,
+				   &state->alloc);
 
 	finish_meta_clear_htile_mask_state(device);
 }
@@ -470,7 +483,9 @@ emit_color_clear(struct radv_cmd_buffer *cmd_buffer,
 
 
 static void
-build_depthstencil_shader(struct nir_shader **out_vs, struct nir_shader **out_fs)
+build_depthstencil_shader(struct nir_shader **out_vs,
+			  struct nir_shader **out_fs,
+			  bool unrestricted)
 {
 	nir_builder vs_b, fs_b;
 
@@ -486,21 +501,36 @@ build_depthstencil_shader(struct nir_shader **out_vs, struct nir_shader **out_fs
 				    "gl_Position");
 	vs_out_pos->data.location = VARYING_SLOT_POS;
 
-	nir_intrinsic_instr *in_color_load = nir_intrinsic_instr_create(fs_b.shader, nir_intrinsic_load_push_constant);
-	nir_intrinsic_set_base(in_color_load, 0);
-	nir_intrinsic_set_range(in_color_load, 4);
-	in_color_load->src[0] = nir_src_for_ssa(nir_imm_int(&fs_b, 0));
-	in_color_load->num_components = 1;
-	nir_ssa_dest_init(&in_color_load->instr, &in_color_load->dest, 1, 32, "depth value");
-	nir_builder_instr_insert(&fs_b, &in_color_load->instr);
-
-	nir_variable *fs_out_depth =
-		nir_variable_create(fs_b.shader, nir_var_shader_out,
-				    glsl_int_type(), "f_depth");
-	fs_out_depth->data.location = FRAG_RESULT_DEPTH;
-	nir_store_var(&fs_b, fs_out_depth, &in_color_load->dest.ssa, 0x1);
+	nir_ssa_def *z;
+	if (unrestricted) {
+		nir_intrinsic_instr *in_color_load = nir_intrinsic_instr_create(fs_b.shader, nir_intrinsic_load_push_constant);
+		nir_intrinsic_set_base(in_color_load, 0);
+		nir_intrinsic_set_range(in_color_load, 4);
+		in_color_load->src[0] = nir_src_for_ssa(nir_imm_int(&fs_b, 0));
+		in_color_load->num_components = 1;
+		nir_ssa_dest_init(&in_color_load->instr, &in_color_load->dest, 1, 32, "depth value");
+		nir_builder_instr_insert(&fs_b, &in_color_load->instr);
+
+		nir_variable *fs_out_depth =
+			nir_variable_create(fs_b.shader, nir_var_shader_out,
+					    glsl_int_type(), "f_depth");
+		fs_out_depth->data.location = FRAG_RESULT_DEPTH;
+		nir_store_var(&fs_b, fs_out_depth, &in_color_load->dest.ssa, 0x1);
+
+		z = nir_imm_float(&vs_b, 0.0);
+	} else {
+		nir_intrinsic_instr *in_color_load = nir_intrinsic_instr_create(vs_b.shader, nir_intrinsic_load_push_constant);
+		nir_intrinsic_set_base(in_color_load, 0);
+		nir_intrinsic_set_range(in_color_load, 4);
+		in_color_load->src[0] = nir_src_for_ssa(nir_imm_int(&vs_b, 0));
+		in_color_load->num_components = 1;
+		nir_ssa_dest_init(&in_color_load->instr, &in_color_load->dest, 1, 32, "depth value");
+		nir_builder_instr_insert(&vs_b, &in_color_load->instr);
+
+		z = &in_color_load->dest.ssa;
+	}
 
-	nir_ssa_def *outvec = radv_meta_gen_rect_vertices(&vs_b);
+	nir_ssa_def *outvec = radv_meta_gen_rect_vertices_comp2(&vs_b, z);
 	nir_store_var(&vs_b, vs_out_pos, outvec, 0xf);
 
 	const struct glsl_type *layer_type = glsl_int_type();
@@ -567,6 +597,7 @@ create_depthstencil_pipeline(struct radv_device *device,
                              VkImageAspectFlags aspects,
 			     uint32_t samples,
 			     int index,
+			     bool unrestricted,
 			     VkPipeline *pipeline,
 			     VkRenderPass render_pass)
 {
@@ -579,7 +610,7 @@ create_depthstencil_pipeline(struct radv_device *device,
 		return VK_SUCCESS;
 	}
 
-	build_depthstencil_shader(&vs_nir, &fs_nir);
+	build_depthstencil_shader(&vs_nir, &fs_nir, unrestricted);
 
 	const VkPipelineVertexInputStateCreateInfo vi_state = {
 		.sType = VK_STRUCTURE_TYPE_PIPELINE_VERTEX_INPUT_STATE_CREATE_INFO,
@@ -677,6 +708,7 @@ pick_depthstencil_pipeline(struct radv_cmd_buffer *cmd_buffer,
 {
 	bool fast = depth_view_can_fast_clear(cmd_buffer, iview, aspects, layout,
 	                                      in_render_loop, clear_rect, clear_value);
+	bool unrestricted = cmd_buffer->device->enabled_extensions.EXT_depth_range_unrestricted;
 	int index = DEPTH_CLEAR_SLOW;
 	VkPipeline *pipeline;
 
@@ -688,13 +720,19 @@ pick_depthstencil_pipeline(struct radv_cmd_buffer *cmd_buffer,
 
 	switch (aspects) {
 	case VK_IMAGE_ASPECT_DEPTH_BIT | VK_IMAGE_ASPECT_STENCIL_BIT:
-		pipeline = &meta_state->clear[samples_log2].depthstencil_pipeline[index];
+		pipeline = unrestricted ?
+			   &meta_state->clear[samples_log2].depthstencil_unrestricted_pipeline[index] :
+			   &meta_state->clear[samples_log2].depthstencil_pipeline[index];
 		break;
 	case VK_IMAGE_ASPECT_DEPTH_BIT:
-		pipeline = &meta_state->clear[samples_log2].depth_only_pipeline[index];
+		pipeline = unrestricted ?
+			   &meta_state->clear[samples_log2].depth_only_unrestricted_pipeline[index] :
+			   &meta_state->clear[samples_log2].depth_only_pipeline[index];
 		break;
 	case VK_IMAGE_ASPECT_STENCIL_BIT:
-		pipeline = &meta_state->clear[samples_log2].stencil_only_pipeline[index];
+		pipeline = unrestricted ?
+			   &meta_state->clear[samples_log2].stencil_only_unrestricted_pipeline[index] :
+			   &meta_state->clear[samples_log2].stencil_only_pipeline[index];
 		break;
 	default:
 		unreachable("expected depth or stencil aspect");
@@ -710,7 +748,7 @@ pick_depthstencil_pipeline(struct radv_cmd_buffer *cmd_buffer,
 	}
 
 	if (*pipeline == VK_NULL_HANDLE) {
-		VkResult ret = create_depthstencil_pipeline(cmd_buffer->device, aspects, 1u << samples_log2, index,
+		VkResult ret = create_depthstencil_pipeline(cmd_buffer->device, aspects, 1u << samples_log2, index, unrestricted,
 		                                            pipeline, cmd_buffer->device->meta_state.clear[samples_log2].depthstencil_rp);
 		if (ret != VK_SUCCESS) {
 			cmd_buffer->record_result = ret;
@@ -755,10 +793,17 @@ emit_depthstencil_clear(struct radv_cmd_buffer *cmd_buffer,
 	if (!(aspects & VK_IMAGE_ASPECT_DEPTH_BIT))
 		clear_value.depth = 1.0f;
 
-	radv_CmdPushConstants(radv_cmd_buffer_to_handle(cmd_buffer),
-			      device->meta_state.clear_depth_p_layout,
-			      VK_SHADER_STAGE_FRAGMENT_BIT, 0, 4,
-			      &clear_value.depth);
+	if (cmd_buffer->device->enabled_extensions.EXT_depth_range_unrestricted) {
+		radv_CmdPushConstants(radv_cmd_buffer_to_handle(cmd_buffer),
+				      device->meta_state.clear_depth_unrestricted_p_layout,
+				      VK_SHADER_STAGE_FRAGMENT_BIT, 0, 4,
+				      &clear_value.depth);
+	} else {
+		radv_CmdPushConstants(radv_cmd_buffer_to_handle(cmd_buffer),
+				      device->meta_state.clear_depth_p_layout,
+				      VK_SHADER_STAGE_VERTEX_BIT, 0, 4,
+				      &clear_value.depth);
+	}
 
 	uint32_t prev_reference = cmd_buffer->state.dynamic.stencil_reference.front;
 	if (aspects & VK_IMAGE_ASPECT_STENCIL_BIT) {
@@ -1244,7 +1289,7 @@ radv_device_init_meta_clear_state(struct radv_device *device, bool on_demand)
 		.sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO,
 		.setLayoutCount = 0,
 		.pushConstantRangeCount = 1,
-		.pPushConstantRanges = &(VkPushConstantRange){VK_SHADER_STAGE_FRAGMENT_BIT, 0, 4},
+		.pPushConstantRanges = &(VkPushConstantRange){VK_SHADER_STAGE_VERTEX_BIT, 0, 4},
 	};
 
 	res = radv_CreatePipelineLayout(radv_device_to_handle(device),
@@ -1254,6 +1299,20 @@ radv_device_init_meta_clear_state(struct radv_device *device, bool on_demand)
 	if (res != VK_SUCCESS)
 		goto fail;
 
+	VkPipelineLayoutCreateInfo pl_depth_unrestricted_create_info = {
+		.sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO,
+		.setLayoutCount = 0,
+		.pushConstantRangeCount = 1,
+		.pPushConstantRanges = &(VkPushConstantRange){VK_SHADER_STAGE_FRAGMENT_BIT, 0, 4},
+	};
+
+	res = radv_CreatePipelineLayout(radv_device_to_handle(device),
+					&pl_depth_unrestricted_create_info,
+					&device->meta_state.alloc,
+					&device->meta_state.clear_depth_unrestricted_p_layout);
+	if (res != VK_SUCCESS)
+		goto fail;
+
 	res = init_meta_clear_htile_mask_state(device);
 	if (res != VK_SUCCESS)
 		goto fail;
@@ -1291,6 +1350,7 @@ radv_device_init_meta_clear_state(struct radv_device *device, bool on_demand)
 							   VK_IMAGE_ASPECT_DEPTH_BIT,
 							   samples,
 							   j,
+							   false,
 							   &state->clear[i].depth_only_pipeline[j],
 							   state->clear[i].depthstencil_rp);
 			if (res != VK_SUCCESS)
@@ -1300,6 +1360,7 @@ radv_device_init_meta_clear_state(struct radv_device *device, bool on_demand)
 							   VK_IMAGE_ASPECT_STENCIL_BIT,
 							   samples,
 							   j,
+							   false,
 							   &state->clear[i].stencil_only_pipeline[j],
 							   state->clear[i].depthstencil_rp);
 			if (res != VK_SUCCESS)
@@ -1310,10 +1371,42 @@ radv_device_init_meta_clear_state(struct radv_device *device, bool on_demand)
 							   VK_IMAGE_ASPECT_STENCIL_BIT,
 							   samples,
 							   j,
+							   false,
 							   &state->clear[i].depthstencil_pipeline[j],
 							   state->clear[i].depthstencil_rp);
 			if (res != VK_SUCCESS)
 				goto fail;
+
+			res = create_depthstencil_pipeline(device,
+							   VK_IMAGE_ASPECT_DEPTH_BIT,
+							   samples,
+							   j,
+							   true,
+							   &state->clear[i].depth_only_unrestricted_pipeline[j],
+							   state->clear[i].depthstencil_rp);
+			if (res != VK_SUCCESS)
+				goto fail;
+
+			res = create_depthstencil_pipeline(device,
+							   VK_IMAGE_ASPECT_STENCIL_BIT,
+							   samples,
+							   j,
+							   true,
+							   &state->clear[i].stencil_only_unrestricted_pipeline[j],
+							   state->clear[i].depthstencil_rp);
+			if (res != VK_SUCCESS)
+				goto fail;
+
+			res = create_depthstencil_pipeline(device,
+							   VK_IMAGE_ASPECT_DEPTH_BIT |
+							   VK_IMAGE_ASPECT_STENCIL_BIT,
+							   samples,
+							   j,
+							   true,
+							   &state->clear[i].depthstencil_unrestricted_pipeline[j],
+							   state->clear[i].depthstencil_rp);
+			if (res != VK_SUCCESS)
+				goto fail;
 		}
 	}
 	return VK_SUCCESS;
diff --git a/src/amd/vulkan/radv_private.h b/src/amd/vulkan/radv_private.h
index 0f5aac29484..5b97b09c867 100644
--- a/src/amd/vulkan/radv_private.h
+++ b/src/amd/vulkan/radv_private.h
@@ -475,10 +475,15 @@ struct radv_meta_state {
 		VkPipeline depth_only_pipeline[NUM_DEPTH_CLEAR_PIPELINES];
 		VkPipeline stencil_only_pipeline[NUM_DEPTH_CLEAR_PIPELINES];
 		VkPipeline depthstencil_pipeline[NUM_DEPTH_CLEAR_PIPELINES];
+
+		VkPipeline depth_only_unrestricted_pipeline[NUM_DEPTH_CLEAR_PIPELINES];
+		VkPipeline stencil_only_unrestricted_pipeline[NUM_DEPTH_CLEAR_PIPELINES];
+		VkPipeline depthstencil_unrestricted_pipeline[NUM_DEPTH_CLEAR_PIPELINES];
 	} clear[MAX_SAMPLES_LOG2];
 
 	VkPipelineLayout                          clear_color_p_layout;
 	VkPipelineLayout                          clear_depth_p_layout;
+	VkPipelineLayout                          clear_depth_unrestricted_p_layout;
 
 	/* Optimized compute fast HTILE clear for stencil or depth only. */
 	VkPipeline clear_htile_mask_pipeline;