Merge pull request #1319 from cdavis5e/msl-frag-outputs

MSL: Add options to control emission of fragment outputs.

CMakeLists.txt

@@ -323,7 +323,7 @@ if (SPIRV_CROSS_STATIC)
 endif()
 
 set(spirv-cross-abi-major 0)
-set(spirv-cross-abi-minor 30)
+set(spirv-cross-abi-minor 31)
 set(spirv-cross-abi-patch 0)
 
 if (SPIRV_CROSS_SHARED)

main.cpp

@@ -136,6 +136,25 @@ struct CLIParser
 		return uint32_t(val);
 	}
 
+	uint32_t next_hex_uint()
+	{
+		if (!argc)
+		{
+			THROW("Tried to parse uint, but nothing left in arguments");
+		}
+
+		uint64_t val = stoul(*argv, nullptr, 16);
+		if (val > numeric_limits<uint32_t>::max())
+		{
+			THROW("next_uint() out of range");
+		}
+
+		argc--;
+		argv++;
+
+		return uint32_t(val);
+	}
+
 	double next_double()
 	{
 		if (!argc)
@@ -535,6 +554,9 @@ struct CLIArguments
 	bool msl_decoration_binding = false;
 	bool msl_force_active_argument_buffer_resources = false;
 	bool msl_force_native_arrays = false;
+	bool msl_enable_frag_depth_builtin = true;
+	bool msl_enable_frag_stencil_ref_builtin = true;
+	uint32_t msl_enable_frag_output_mask = 0xffffffff;
 	bool glsl_emit_push_constant_as_ubo = false;
 	bool glsl_emit_ubo_as_plain_uniforms = false;
 	SmallVector<pair<uint32_t, uint32_t>> glsl_ext_framebuffer_fetch;
@@ -638,6 +660,9 @@ static void print_help()
 	                "\t[--msl-decoration-binding]\n"
 	                "\t[--msl-force-active-argument-buffer-resources]\n"
 	                "\t[--msl-force-native-arrays]\n"
+	                "\t[--msl-disable-frag-depth-builtin]\n"
+	                "\t[--msl-disable-frag-stencil-ref-builtin]\n"
+	                "\t[--msl-enable-frag-output-mask <mask>]\n"
 	                "\t[--hlsl]\n"
 	                "\t[--reflect]\n"
 	                "\t[--shader-model]\n"
@@ -831,6 +856,9 @@ static string compile_iteration(const CLIArguments &args, std::vector<uint32_t>
 		msl_opts.enable_decoration_binding = args.msl_decoration_binding;
 		msl_opts.force_active_argument_buffer_resources = args.msl_force_active_argument_buffer_resources;
 		msl_opts.force_native_arrays = args.msl_force_native_arrays;
+		msl_opts.enable_frag_depth_builtin = args.msl_enable_frag_depth_builtin;
+		msl_opts.enable_frag_stencil_ref_builtin = args.msl_enable_frag_stencil_ref_builtin;
+		msl_opts.enable_frag_output_mask = args.msl_enable_frag_output_mask;
 		msl_comp->set_msl_options(msl_opts);
 		for (auto &v : args.msl_discrete_descriptor_sets)
 			msl_comp->add_discrete_descriptor_set(v);
@@ -1213,6 +1241,11 @@ static int main_inner(int argc, char *argv[])
 		args.msl_inline_uniform_blocks.push_back(make_pair(desc_set, binding));
 	});
 	cbs.add("--msl-force-native-arrays", [&args](CLIParser &) { args.msl_force_native_arrays = true; });
+	cbs.add("--msl-disable-frag-depth-builtin", [&args](CLIParser &) { args.msl_enable_frag_depth_builtin = false; });
+	cbs.add("--msl-disable-frag-stencil-ref-builtin",
+	        [&args](CLIParser &) { args.msl_enable_frag_stencil_ref_builtin = false; });
+	cbs.add("--msl-enable-frag-output-mask",
+	        [&args](CLIParser &parser) { args.msl_enable_frag_output_mask = parser.next_hex_uint(); });
 	cbs.add("--extension", [&args](CLIParser &parser) { args.extensions.push_back(parser.next_string()); });
 	cbs.add("--rename-entry-point", [&args](CLIParser &parser) {
 		auto old_name = parser.next_string();

reference/opt/shaders-msl/frag/disable-frag-output.frag-output.frag

@@ -0,0 +1,35 @@
+#include <metal_stdlib>
+#include <simd/simd.h>
+
+using namespace metal;
+
+struct main0_out
+{
+    float4 buf1 [[color(1)]];
+    float4 buf3 [[color(3)]];
+    float4 buf6 [[color(6)]];
+    float4 buf7 [[color(7)]];
+};
+
+fragment main0_out main0()
+{
+    float4 buf0;
+    float4 buf2;
+    float4 buf4;
+    float4 buf5;
+    float gl_FragDepth;
+    int gl_FragStencilRefARB;
+    main0_out out = {};
+    buf0 = float4(0.0, 0.0, 0.0, 1.0);
+    out.buf1 = float4(1.0, 0.0, 0.0, 1.0);
+    buf2 = float4(0.0, 1.0, 0.0, 1.0);
+    out.buf3 = float4(0.0, 0.0, 1.0, 1.0);
+    buf4 = float4(1.0, 0.0, 1.0, 0.5);
+    buf5 = float4(0.25);
+    out.buf6 = float4(0.75);
+    out.buf7 = float4(1.0);
+    gl_FragDepth = 0.89999997615814208984375;
+    gl_FragStencilRefARB = uint(127);
+    return out;
+}
+

reference/shaders-msl/frag/disable-frag-output.frag-output.frag

@@ -0,0 +1,35 @@
+#include <metal_stdlib>
+#include <simd/simd.h>
+
+using namespace metal;
+
+struct main0_out
+{
+    float4 buf1 [[color(1)]];
+    float4 buf3 [[color(3)]];
+    float4 buf6 [[color(6)]];
+    float4 buf7 [[color(7)]];
+};
+
+fragment main0_out main0()
+{
+    float4 buf0;
+    float4 buf2;
+    float4 buf4;
+    float4 buf5;
+    float gl_FragDepth;
+    int gl_FragStencilRefARB;
+    main0_out out = {};
+    buf0 = float4(0.0, 0.0, 0.0, 1.0);
+    out.buf1 = float4(1.0, 0.0, 0.0, 1.0);
+    buf2 = float4(0.0, 1.0, 0.0, 1.0);
+    out.buf3 = float4(0.0, 0.0, 1.0, 1.0);
+    buf4 = float4(1.0, 0.0, 1.0, 0.5);
+    buf5 = float4(0.25);
+    out.buf6 = float4(0.75);
+    out.buf7 = float4(1.0);
+    gl_FragDepth = 0.89999997615814208984375;
+    gl_FragStencilRefARB = uint(127);
+    return out;
+}
+

shaders-msl/frag/disable-frag-output.frag-output.frag

@@ -0,0 +1,25 @@
+#version 450
+#extension GL_ARB_shader_stencil_export : require
+
+layout(location = 0) out vec4 buf0;
+layout(location = 1) out vec4 buf1;
+layout(location = 2) out vec4 buf2;
+layout(location = 3) out vec4 buf3;
+layout(location = 4) out vec4 buf4;
+layout(location = 5) out vec4 buf5;
+layout(location = 6) out vec4 buf6;
+layout(location = 7) out vec4 buf7;
+
+void main() {
+	buf0 = vec4(0, 0, 0, 1);
+	buf1 = vec4(1, 0, 0, 1);
+	buf2 = vec4(0, 1, 0, 1);
+	buf3 = vec4(0, 0, 1, 1);
+	buf4 = vec4(1, 0, 1, 0.5);
+	buf5 = vec4(0.25, 0.25, 0.25, 0.25);
+	buf6 = vec4(0.75, 0.75, 0.75, 0.75);
+	buf7 = vec4(1, 1, 1, 1);
+	gl_FragDepth = 0.9;
+	gl_FragStencilRefARB = 127;
+}
+

spirv_cross_c.cpp

@@ -615,6 +615,18 @@ spvc_result spvc_compiler_options_set_uint(spvc_compiler_options options, spvc_c
 	case SPVC_COMPILER_OPTION_MSL_FORCE_NATIVE_ARRAYS:
 		options->msl.force_native_arrays = value != 0;
 		break;
+
+	case SPVC_COMPILER_OPTION_MSL_ENABLE_FRAG_OUTPUT_MASK:
+		options->msl.enable_frag_output_mask = value;
+		break;
+
+	case SPVC_COMPILER_OPTION_MSL_ENABLE_FRAG_DEPTH_BUILTIN:
+		options->msl.enable_frag_depth_builtin = value != 0;
+		break;
+
+	case SPVC_COMPILER_OPTION_MSL_ENABLE_FRAG_STENCIL_REF_BUILTIN:
+		options->msl.enable_frag_stencil_ref_builtin = value != 0;
+		break;
 #endif
 
 	default:

spirv_cross_c.h

@@ -33,7 +33,7 @@ extern "C" {
 /* Bumped if ABI or API breaks backwards compatibility. */
 #define SPVC_C_API_VERSION_MAJOR 0
 /* Bumped if APIs or enumerations are added in a backwards compatible way. */
-#define SPVC_C_API_VERSION_MINOR 30
+#define SPVC_C_API_VERSION_MINOR 31
 /* Bumped if internal implementation details change. */
 #define SPVC_C_API_VERSION_PATCH 0
 
@@ -582,6 +582,10 @@ typedef enum spvc_compiler_option
 
 	SPVC_COMPILER_OPTION_HLSL_NONWRITABLE_UAV_TEXTURE_AS_SRV = 55 | SPVC_COMPILER_OPTION_HLSL_BIT,
 
+	SPVC_COMPILER_OPTION_MSL_ENABLE_FRAG_OUTPUT_MASK = 56 | SPVC_COMPILER_OPTION_MSL_BIT,
+	SPVC_COMPILER_OPTION_MSL_ENABLE_FRAG_DEPTH_BUILTIN = 57 | SPVC_COMPILER_OPTION_MSL_BIT,
+	SPVC_COMPILER_OPTION_MSL_ENABLE_FRAG_STENCIL_REF_BUILTIN = 58 | SPVC_COMPILER_OPTION_MSL_BIT,
+
 	SPVC_COMPILER_OPTION_INT_MAX = 0x7fffffff
 } spvc_compiler_option;
 

spirv_msl.cpp

@@ -965,6 +965,16 @@ void CompilerMSL::emit_entry_point_declarations()
 	}
 	// For some reason, without this, we end up emitting the arrays twice.
 	buffer_arrays.clear();
+
+	// Emit disabled fragment outputs.
+	std::sort(disabled_frag_outputs.begin(), disabled_frag_outputs.end());
+	for (uint32_t var_id : disabled_frag_outputs)
+	{
+		auto &var = get<SPIRVariable>(var_id);
+		add_local_variable_name(var_id);
+		statement(variable_decl(var), ";");
+		var.deferred_declaration = false;
+	}
 }
 
 string CompilerMSL::compile()
@@ -2442,6 +2452,7 @@ uint32_t CompilerMSL::add_interface_block(StorageClass storage, bool patch)
 
 		bool is_builtin = is_builtin_variable(var);
 		auto bi_type = BuiltIn(get_decoration(var_id, DecorationBuiltIn));
+		uint32_t location = get_decoration(var_id, DecorationLocation);
 
 		// These builtins are part of the stage in/out structs.
 		bool is_interface_block_builtin =
@@ -2467,6 +2478,18 @@ uint32_t CompilerMSL::add_interface_block(StorageClass storage, bool patch)
 		if (bi_type == BuiltInClipDistance)
 			hidden = false;
 
+		// It's not enough to simply avoid marking fragment outputs if the pipeline won't
+		// accept them. We can't put them in the struct at all, or otherwise the compiler
+		// complains that the outputs weren't explicitly marked.
+		if (storage == StorageClassOutput && !patch &&
+			((is_builtin && ((bi_type == BuiltInFragDepth && !msl_options.enable_frag_depth_builtin) ||
+		                    (bi_type == BuiltInFragStencilRefEXT && !msl_options.enable_frag_stencil_ref_builtin))) ||
+		     (!is_builtin && !(msl_options.enable_frag_output_mask & (1 << location)))))
+		{
+			hidden = true;
+			disabled_frag_outputs.push_back(var_id);
+		}
+
 		// Barycentric inputs must be emitted in stage-in, because they can have interpolation arguments.
 		if (is_active && (bi_type == BuiltInBaryCoordNV || bi_type == BuiltInBaryCoordNoPerspNV))
 		{
@@ -2496,7 +2519,6 @@ uint32_t CompilerMSL::add_interface_block(StorageClass storage, bool patch)
 						SPIRV_CROSS_THROW("Component decoration is not supported in tessellation shaders.");
 					else if (pack_components)
 					{
-						uint32_t location = get_decoration(var_id, DecorationLocation);
 						auto &location_meta = meta.location_meta[location];
 						location_meta.num_components = std::max(location_meta.num_components, component + type.vecsize);
 					}
@@ -8877,12 +8899,23 @@ string CompilerMSL::member_attribute_qualifier(const SPIRType &type, uint32_t in
 			switch (builtin)
 			{
 			case BuiltInFragStencilRefEXT:
+				// Similar to PointSize, only mark FragStencilRef if there's a stencil buffer.
+				// Some shaders may include a FragStencilRef builtin even when used to render
+				// without a stencil attachment, and Metal will reject this builtin
+				// when compiling the shader into a render pipeline that does not set
+				// stencilAttachmentPixelFormat.
+				if (!msl_options.enable_frag_stencil_ref_builtin)
+					return "";
 				if (!msl_options.supports_msl_version(2, 1))
 					SPIRV_CROSS_THROW("Stencil export only supported in MSL 2.1 and up.");
 				return string(" [[") + builtin_qualifier(builtin) + "]]";
 
-			case BuiltInSampleMask:
 			case BuiltInFragDepth:
+				// Ditto FragDepth.
+				if (!msl_options.enable_frag_depth_builtin)
+					return "";
+				/* fallthrough */
+			case BuiltInSampleMask:
 				return string(" [[") + builtin_qualifier(builtin) + "]]";
 
 			default:
@@ -8890,6 +8923,9 @@ string CompilerMSL::member_attribute_qualifier(const SPIRType &type, uint32_t in
 			}
 		}
 		uint32_t locn = get_ordered_member_location(type.self, index);
+		// Metal will likely complain about missing color attachments, too.
+		if (locn != k_unknown_location && !(msl_options.enable_frag_output_mask & (1 << locn)))
+			return "";
 		if (locn != k_unknown_location && has_member_decoration(type.self, index, DecorationIndex))
 			return join(" [[color(", locn, "), index(", get_member_decoration(type.self, index, DecorationIndex),
 			            ")]]");
@@ -11345,13 +11381,17 @@ string CompilerMSL::builtin_to_glsl(BuiltIn builtin, StorageClass storage)
 		if (!msl_options.supports_msl_version(2, 0))
 			SPIRV_CROSS_THROW("ViewportIndex requires Metal 2.0.");
 		/* fallthrough */
+	case BuiltInFragDepth:
+	case BuiltInFragStencilRefEXT:
+		if ((builtin == BuiltInFragDepth && !msl_options.enable_frag_depth_builtin) ||
+		    (builtin == BuiltInFragStencilRefEXT && !msl_options.enable_frag_stencil_ref_builtin))
+			break;
+		/* fallthrough */
 	case BuiltInPosition:
 	case BuiltInPointSize:
 	case BuiltInClipDistance:
 	case BuiltInCullDistance:
 	case BuiltInLayer:
-	case BuiltInFragDepth:
-	case BuiltInFragStencilRefEXT:
 	case BuiltInSampleMask:
 		if (get_execution_model() == ExecutionModelTessellationControl)
 			break;

spirv_msl.hpp

@@ -268,7 +268,10 @@ public:
 		uint32_t dynamic_offsets_buffer_index = 23;
 		uint32_t shader_input_wg_index = 0;
 		uint32_t device_index = 0;
+		uint32_t enable_frag_output_mask = 0xffffffff;
 		bool enable_point_size_builtin = true;
+		bool enable_frag_depth_builtin = true;
+		bool enable_frag_stencil_ref_builtin = true;
 		bool disable_rasterization = false;
 		bool capture_output_to_buffer = false;
 		bool swizzle_texture_samples = false;
@@ -870,6 +873,8 @@ protected:
 	// Must be ordered since array is in a specific order.
 	std::map<SetBindingPair, std::pair<uint32_t, uint32_t>> buffers_requiring_dynamic_offset;
 
+	SmallVector<uint32_t> disabled_frag_outputs;
+
 	std::unordered_set<SetBindingPair, InternalHasher> inline_uniform_blocks;
 
 	uint32_t argument_buffer_ids[kMaxArgumentBuffers];

test_shaders.py

@@ -258,6 +258,12 @@ def cross_compile_msl(shader, spirv, opt, iterations, paths):
         msl_args.append('--msl-force-native-arrays')
     if '.zero-initialize.' in shader:
         msl_args.append('--force-zero-initialized-variables')
+    if '.frag-output.' in shader:
+        # Arbitrary for testing purposes.
+        msl_args.append('--msl-disable-frag-depth-builtin')
+        msl_args.append('--msl-disable-frag-stencil-ref-builtin')
+        msl_args.append('--msl-enable-frag-output-mask')
+        msl_args.append('0x000000ca')
 
     subprocess.check_call(msl_args)