Merge pull request #1885 from KhronosGroup/helper-invocation-rework

Helper invocation rework

15 files changed, 218 insertions, 27 deletions

diff --git a/reference/opt/shaders-msl/frag/helper-invocation.msl21.frag b/reference/opt/shaders-msl/frag/helper-invocation.msl21.frag
index a833fa08..9d876df1 100644
--- a/reference/opt/shaders-msl/frag/helper-invocation.msl21.frag
+++ b/reference/opt/shaders-msl/frag/helper-invocation.msl21.frag
@@ -16,9 +16,8 @@ struct main0_in
 fragment main0_out main0(main0_in in [[stage_in]], texture2d<float> uSampler [[texture(0)]], sampler uSamplerSmplr [[sampler(0)]])
 {
     main0_out out = {};
-    bool gl_HelperInvocation = simd_is_helper_thread();
     float4 _52;
-    if (!gl_HelperInvocation)
+    if (!simd_is_helper_thread())
     {
         _52 = uSampler.sample(uSamplerSmplr, in.vUV, level(0.0));
     }
diff --git a/reference/shaders-hlsl-no-opt/frag/helper-invocation.fxconly.nofxc.frag b/reference/shaders-hlsl-no-opt/frag/helper-invocation.fxconly.nofxc.frag
new file mode 100644
index 00000000..ad61b21e
--- /dev/null
+++ b/reference/shaders-hlsl-no-opt/frag/helper-invocation.fxconly.nofxc.frag
@@ -0,0 +1,26 @@
+static float FragColor;
+
+struct SPIRV_Cross_Input
+{
+};
+
+struct SPIRV_Cross_Output
+{
+    float FragColor : SV_Target0;
+};
+
+void frag_main()
+{
+    FragColor = float(IsHelperLane());
+    discard;
+    bool _16 = IsHelperLane();
+    FragColor = float(_16);
+}
+
+SPIRV_Cross_Output main(SPIRV_Cross_Input stage_input)
+{
+    frag_main();
+    SPIRV_Cross_Output stage_output;
+    stage_output.FragColor = FragColor;
+    return stage_output;
+}
diff --git a/reference/shaders-hlsl-no-opt/frag/volatile-helper-invocation.fxconly.nofxc.spv16.frag b/reference/shaders-hlsl-no-opt/frag/volatile-helper-invocation.fxconly.nofxc.spv16.frag
new file mode 100644
index 00000000..1311c863
--- /dev/null
+++ b/reference/shaders-hlsl-no-opt/frag/volatile-helper-invocation.fxconly.nofxc.spv16.frag
@@ -0,0 +1,29 @@
+static float FragColor;
+
+struct SPIRV_Cross_Input
+{
+};
+
+struct SPIRV_Cross_Output
+{
+    float FragColor : SV_Target0;
+};
+
+void frag_main()
+{
+    bool _12 = IsHelperLane();
+    float _15 = float(_12);
+    FragColor = _15;
+    discard;
+    bool _16 = IsHelperLane();
+    float _17 = float(_16);
+    FragColor = _17;
+}
+
+SPIRV_Cross_Output main(SPIRV_Cross_Input stage_input)
+{
+    frag_main();
+    SPIRV_Cross_Output stage_output;
+    stage_output.FragColor = FragColor;
+    return stage_output;
+}
diff --git a/reference/shaders-msl-no-opt/frag/volatile-helper-invocation.msl23.spv16.frag b/reference/shaders-msl-no-opt/frag/volatile-helper-invocation.msl23.spv16.frag
new file mode 100644
index 00000000..eef92efe
--- /dev/null
+++ b/reference/shaders-msl-no-opt/frag/volatile-helper-invocation.msl23.spv16.frag
@@ -0,0 +1,23 @@
+#include <metal_stdlib>
+#include <simd/simd.h>
+
+using namespace metal;
+
+struct main0_out
+{
+    float FragColor [[color(0)]];
+};
+
+fragment main0_out main0()
+{
+    main0_out out = {};
+    bool _12 = simd_is_helper_thread();
+    float _15 = float(_12);
+    out.FragColor = _15;
+    discard_fragment();
+    bool _16 = simd_is_helper_thread();
+    float _17 = float(_16);
+    out.FragColor = _17;
+    return out;
+}
+
diff --git a/reference/shaders-msl/frag/helper-invocation.msl21.frag b/reference/shaders-msl/frag/helper-invocation.msl21.frag
index 7cf97a25..36d14239 100644
--- a/reference/shaders-msl/frag/helper-invocation.msl21.frag
+++ b/reference/shaders-msl/frag/helper-invocation.msl21.frag
@@ -16,10 +16,10 @@ struct main0_in
 };
 
 static inline __attribute__((always_inline))
-float4 foo(thread bool& gl_HelperInvocation, texture2d<float> uSampler, sampler uSamplerSmplr, thread float2& vUV)
+float4 foo(texture2d<float> uSampler, sampler uSamplerSmplr, thread float2& vUV)
 {
     float4 color;
-    if (!gl_HelperInvocation)
+    if (!simd_is_helper_thread())
     {
         color = uSampler.sample(uSamplerSmplr, vUV, level(0.0));
     }
@@ -33,8 +33,7 @@ float4 foo(thread bool& gl_HelperInvocation, texture2d<float> uSampler, sampler
 fragment main0_out main0(main0_in in [[stage_in]], texture2d<float> uSampler [[texture(0)]], sampler uSamplerSmplr [[sampler(0)]])
 {
     main0_out out = {};
-    bool gl_HelperInvocation = simd_is_helper_thread();
-    out.FragColor = foo(gl_HelperInvocation, uSampler, uSamplerSmplr, in.vUV);
+    out.FragColor = foo(uSampler, uSamplerSmplr, in.vUV);
     return out;
 }
 
diff --git a/reference/shaders-no-opt/vulkan/frag/volatile-helper-invocation.vk.nocompat.spv16.frag.vk b/reference/shaders-no-opt/vulkan/frag/volatile-helper-invocation.vk.nocompat.spv16.frag.vk
new file mode 100644
index 00000000..73b0f9b0
--- /dev/null
+++ b/reference/shaders-no-opt/vulkan/frag/volatile-helper-invocation.vk.nocompat.spv16.frag.vk
@@ -0,0 +1,16 @@
+#version 450
+#extension GL_EXT_demote_to_helper_invocation : require
+
+layout(location = 0) out float FragColor;
+
+void main()
+{
+    bool _12 = gl_HelperInvocation;
+    float _15 = float(_12);
+    FragColor = _15;
+    demote;
+    bool _16 = gl_HelperInvocation;
+    float _17 = float(_16);
+    FragColor = _17;
+}
+
diff --git a/shaders-hlsl-no-opt/frag/helper-invocation.fxconly.nofxc.frag b/shaders-hlsl-no-opt/frag/helper-invocation.fxconly.nofxc.frag
new file mode 100644
index 00000000..6f70c772
--- /dev/null
+++ b/shaders-hlsl-no-opt/frag/helper-invocation.fxconly.nofxc.frag
@@ -0,0 +1,11 @@
+#version 450
+#extension GL_EXT_demote_to_helper_invocation : require
+
+layout(location = 0) out float FragColor;
+
+void main()
+{
+	FragColor = float(gl_HelperInvocation);
+	demote;
+	FragColor = float(helperInvocationEXT());
+}
diff --git a/shaders-hlsl-no-opt/frag/volatile-helper-invocation.fxconly.nofxc.spv16.frag b/shaders-hlsl-no-opt/frag/volatile-helper-invocation.fxconly.nofxc.spv16.frag
new file mode 100644
index 00000000..9a8d9d20
--- /dev/null
+++ b/shaders-hlsl-no-opt/frag/volatile-helper-invocation.fxconly.nofxc.spv16.frag
@@ -0,0 +1,11 @@
+#version 450
+#extension GL_EXT_demote_to_helper_invocation : require
+
+layout(location = 0) out float FragColor;
+
+void main()
+{
+	FragColor = float(gl_HelperInvocation);
+	demote;
+	FragColor = float(gl_HelperInvocation);
+}
diff --git a/shaders-msl-no-opt/frag/volatile-helper-invocation.msl23.spv16.frag b/shaders-msl-no-opt/frag/volatile-helper-invocation.msl23.spv16.frag
new file mode 100644
index 00000000..9a8d9d20
--- /dev/null
+++ b/shaders-msl-no-opt/frag/volatile-helper-invocation.msl23.spv16.frag
@@ -0,0 +1,11 @@
+#version 450
+#extension GL_EXT_demote_to_helper_invocation : require
+
+layout(location = 0) out float FragColor;
+
+void main()
+{
+	FragColor = float(gl_HelperInvocation);
+	demote;
+	FragColor = float(gl_HelperInvocation);
+}
diff --git a/shaders-no-opt/vulkan/frag/volatile-helper-invocation.vk.nocompat.spv16.frag b/shaders-no-opt/vulkan/frag/volatile-helper-invocation.vk.nocompat.spv16.frag
new file mode 100644
index 00000000..9a8d9d20
--- /dev/null
+++ b/shaders-no-opt/vulkan/frag/volatile-helper-invocation.vk.nocompat.spv16.frag
@@ -0,0 +1,11 @@
+#version 450
+#extension GL_EXT_demote_to_helper_invocation : require
+
+layout(location = 0) out float FragColor;
+
+void main()
+{
+	FragColor = float(gl_HelperInvocation);
+	demote;
+	FragColor = float(gl_HelperInvocation);
+}
diff --git a/spirv_common.hpp b/spirv_common.hpp
index f568d7dc..4aaa7148 100644
--- a/spirv_common.hpp
+++ b/spirv_common.hpp
@@ -1072,7 +1072,6 @@ struct SPIRVariable : IVariant
 
 	// Temporaries which can remain forwarded as long as this variable is not modified.
 	SmallVector<ID> dependees;
-	bool forwardable = true;
 
 	bool deferred_declaration = false;
 	bool phi_variable = false;
diff --git a/spirv_glsl.cpp b/spirv_glsl.cpp
index 4c465133..52cef4ec 100644
--- a/spirv_glsl.cpp
+++ b/spirv_glsl.cpp
@@ -9585,8 +9585,11 @@ bool CompilerGLSL::should_forward(uint32_t id) const
 	// This is important because otherwise we'll get local sampler copies (highp sampler2D foo = bar) that are invalid in OpenGL GLSL
 
 	auto *var = maybe_get<SPIRVariable>(id);
-	if (var && var->forwardable)
-		return true;
+	if (var)
+	{
+		// Never forward volatile variables, e.g. SPIR-V 1.6 IsHelperInvocation.
+		return !has_decoration(id, DecorationVolatile);
+	}
 
 	// For debugging emit temporary variables for all expressions
 	if (options.force_temporary)
@@ -9599,6 +9602,12 @@ bool CompilerGLSL::should_forward(uint32_t id) const
 	if (expr && expr->expression_dependencies.size() >= max_expression_dependencies)
 		return false;
 
+	if (expr && expr->loaded_from && has_decoration(expr->loaded_from, DecorationVolatile))
+	{
+		// Never forward volatile variables.
+		return false;
+	}
+
 	// Immutable expression can always be forwarded.
 	if (is_immutable(id))
 		return true;
@@ -12089,7 +12098,7 @@ void CompilerGLSL::emit_instruction(const Instruction &instruction)
 			pure = false;
 		}
 
-		if (var && var->forwardable)
+		if (var)
 		{
 			bool forward = forced_temporaries.find(id) == end(forced_temporaries);
 			auto &e = emit_op(result_type, id, imgexpr, forward);
@@ -12850,6 +12859,8 @@ void CompilerGLSL::emit_instruction(const Instruction &instruction)
 		if (!options.vulkan_semantics)
 			SPIRV_CROSS_THROW("GL_EXT_demote_to_helper_invocation is only supported in Vulkan GLSL.");
 		require_extension_internal("GL_EXT_demote_to_helper_invocation");
+		// Helper lane state with demote is volatile by nature.
+		// Do not forward this.
 		emit_op(ops[0], ops[1], "helperInvocationEXT()", false);
 		break;
 
diff --git a/spirv_hlsl.cpp b/spirv_hlsl.cpp
index 646fce33..253d6ad0 100644
--- a/spirv_hlsl.cpp
+++ b/spirv_hlsl.cpp
@@ -728,6 +728,11 @@ void CompilerHLSL::emit_builtin_inputs_in_struct()
 			// Handled specially.
 			break;
 
+		case BuiltInHelperInvocation:
+			if (hlsl_options.shader_model < 50 || (get_entry_point().model != ExecutionModelFragment && get_entry_point().model != ExecutionModelGLCompute))
+				SPIRV_CROSS_THROW("Helper Invocation input is only supported in PS 5.0 or higher.");
+			break;
+
 		case BuiltInClipDistance:
 			// HLSL is a bit weird here, use SV_ClipDistance0, SV_ClipDistance1 and so on with vectors.
 			for (uint32_t clip = 0; clip < clip_distance_count; clip += 4)
@@ -984,6 +989,8 @@ std::string CompilerHLSL::builtin_to_glsl(spv::BuiltIn builtin, spv::StorageClas
 		return "WaveGetLaneIndex()";
 	case BuiltInSubgroupSize:
 		return "WaveGetLaneCount()";
+	case BuiltInHelperInvocation:
+		return "IsHelperLane()";
 
 	default:
 		return CompilerGLSL::builtin_to_glsl(builtin, storage);
@@ -1103,6 +1110,11 @@ void CompilerHLSL::emit_builtin_variables()
 			type = "uint4";
 			break;
 
+		case BuiltInHelperInvocation:
+			if (hlsl_options.shader_model < 50)
+				SPIRV_CROSS_THROW("Need SM 5.0 for Helper Invocation.");
+			break;
+
 		case BuiltInClipDistance:
 			array_size = clip_distance_count;
 			type = "float";
@@ -2521,6 +2533,7 @@ void CompilerHLSL::emit_hlsl_entry_point()
 		case BuiltInPointCoord:
 		case BuiltInSubgroupSize:
 		case BuiltInSubgroupLocalInvocationId:
+		case BuiltInHelperInvocation:
 			break;
 
 		case BuiltInSubgroupEqMask:
@@ -5346,7 +5359,7 @@ void CompilerHLSL::emit_instruction(const Instruction &instruction)
 				                        image_format_to_components(get<SPIRType>(var->basetype).image.format), imgexpr);
 		}
 
-		if (var && var->forwardable)
+		if (var)
 		{
 			bool forward = forced_temporaries.find(id) == end(forced_temporaries);
 			auto &e = emit_op(result_type, id, imgexpr, forward);
@@ -5590,7 +5603,12 @@ void CompilerHLSL::emit_instruction(const Instruction &instruction)
 	}
 
 	case OpIsHelperInvocationEXT:
-		SPIRV_CROSS_THROW("helperInvocationEXT() is not supported in HLSL.");
+		if (hlsl_options.shader_model < 50 || (get_entry_point().model != ExecutionModelFragment && get_entry_point().model != ExecutionModelGLCompute))
+			SPIRV_CROSS_THROW("Helper Invocation input is only supported in PS 5.0 or higher.");
+		// Helper lane state with demote is volatile by nature.
+		// Do not forward this.
+		emit_op(ops[0], ops[1], "IsHelperLane()", false);
+		break;
 
 	case OpBeginInvocationInterlockEXT:
 	case OpEndInvocationInterlockEXT:
diff --git a/spirv_msl.cpp b/spirv_msl.cpp
index 4eeb3b7f..8ea27d5f 100644
--- a/spirv_msl.cpp
+++ b/spirv_msl.cpp
@@ -1548,6 +1548,14 @@ void CompilerMSL::extract_global_variables_from_functions()
 	// Uniforms
 	unordered_set<uint32_t> global_var_ids;
 	ir.for_each_typed_id<SPIRVariable>([&](uint32_t, SPIRVariable &var) {
+		// Some builtins resolve directly to a function call which does not need any declared variables.
+		// Skip these.
+		if (var.storage == StorageClassInput && has_decoration(var.self, DecorationBuiltIn) &&
+		    BuiltIn(get_decoration(var.self, DecorationBuiltIn)) == BuiltInHelperInvocation)
+		{
+			return;
+		}
+
 		if (var.storage == StorageClassInput || var.storage == StorageClassOutput ||
 		    var.storage == StorageClassUniform || var.storage == StorageClassUniformConstant ||
 		    var.storage == StorageClassPushConstant || var.storage == StorageClassStorageBuffer)
@@ -12163,16 +12171,6 @@ void CompilerMSL::fix_up_shader_inputs_outputs()
 					});
 				}
 				break;
-			case BuiltInHelperInvocation:
-				if (msl_options.is_ios() && !msl_options.supports_msl_version(2, 3))
-					SPIRV_CROSS_THROW("simd_is_helper_thread() requires version 2.3 on iOS.");
-				else if (msl_options.is_macos() && !msl_options.supports_msl_version(2, 1))
-					SPIRV_CROSS_THROW("simd_is_helper_thread() requires version 2.1 on macOS.");
-
-				entry_func.fixup_hooks_in.push_back([=]() {
-					statement(builtin_type_decl(bi_type), " ", to_expression(var_id), " = simd_is_helper_thread();");
-				});
-				break;
 			case BuiltInInvocationId:
 				// This is direct-mapped without multi-patch workgroups.
 				if (get_execution_model() != ExecutionModelTessellationControl || !msl_options.multi_patch_workgroup)
@@ -14467,6 +14465,14 @@ string CompilerMSL::builtin_to_glsl(BuiltIn builtin, StorageClass storage)
 		}
 		break;
 
+	case BuiltInHelperInvocation:
+		if (msl_options.is_ios() && !msl_options.supports_msl_version(2, 3))
+			SPIRV_CROSS_THROW("simd_is_helper_thread() requires version 2.3 on iOS.");
+		else if (msl_options.is_macos() && !msl_options.supports_msl_version(2, 1))
+			SPIRV_CROSS_THROW("simd_is_helper_thread() requires version 2.1 on macOS.");
+		// In SPIR-V 1.6 with Volatile HelperInvocation, we cannot emit a fixup early.
+		return "simd_is_helper_thread()";
+
 	default:
 		break;
 	}
diff --git a/test_shaders.py b/test_shaders.py
index 49038939..2726ae75 100755
--- a/test_shaders.py
+++ b/test_shaders.py
@@ -192,8 +192,18 @@ def cross_compile_msl(shader, spirv, opt, iterations, paths):
     spirv_path = create_temporary()
     msl_path = create_temporary(os.path.basename(shader))
 
+    spirv_16 = '.spv16.' in shader
     spirv_14 = '.spv14.' in shader
-    spirv_env = 'vulkan1.1spv1.4' if spirv_14 else 'vulkan1.1'
+
+    if spirv_16:
+        spirv_env = 'spv1.6'
+        glslang_env = 'spirv1.6'
+    elif spirv_14:
+        spirv_env = 'vulkan1.1spv1.4'
+        glslang_env = 'spirv1.4'
+    else:
+        spirv_env = 'vulkan1.1'
+        glslang_env = 'vulkan1.1'
 
     spirv_cmd = [paths.spirv_as, '--target-env', spirv_env, '-o', spirv_path, shader]
     if '.preserve.' in shader:
@@ -202,7 +212,6 @@ def cross_compile_msl(shader, spirv, opt, iterations, paths):
     if spirv:
         subprocess.check_call(spirv_cmd)
     else:
-        glslang_env = 'spirv1.4' if spirv_14 else 'vulkan1.1'
         subprocess.check_call([paths.glslang, '--amb' ,'--target-env', glslang_env, '-V', '-o', spirv_path, shader])
 
     if opt and (not shader_is_invalid_spirv(shader)):
@@ -442,8 +451,19 @@ def cross_compile_hlsl(shader, spirv, opt, force_no_external_validation, iterati
     spirv_path = create_temporary()
     hlsl_path = create_temporary(os.path.basename(shader))
 
+    spirv_16 = '.spv16.' in shader
     spirv_14 = '.spv14.' in shader
-    spirv_env = 'vulkan1.1spv1.4' if spirv_14 else 'vulkan1.1'
+
+    if spirv_16:
+        spirv_env = 'spv1.6'
+        glslang_env = 'spirv1.6'
+    elif spirv_14:
+        spirv_env = 'vulkan1.1spv1.4'
+        glslang_env = 'spirv1.4'
+    else:
+        spirv_env = 'vulkan1.1'
+        glslang_env = 'vulkan1.1'
+
     spirv_cmd = [paths.spirv_as, '--target-env', spirv_env, '-o', spirv_path, shader]
     if '.preserve.' in shader:
         spirv_cmd.append('--preserve-numeric-ids')
@@ -451,7 +471,6 @@ def cross_compile_hlsl(shader, spirv, opt, force_no_external_validation, iterati
     if spirv:
         subprocess.check_call(spirv_cmd)
     else:
-        glslang_env = 'spirv1.4' if spirv_14 else 'vulkan1.1'
         subprocess.check_call([paths.glslang, '--amb', '--target-env', glslang_env, '-V', '-o', spirv_path, shader])
 
     if opt and (not shader_is_invalid_spirv(hlsl_path)):
@@ -526,10 +545,13 @@ def cross_compile(shader, vulkan, spirv, invalid_spirv, eliminate, is_legacy, fl
     spirv_14 = '.spv14.' in shader
     if spirv_16:
         spirv_env = 'spv1.6'
+        glslang_env = 'spirv1.6'
     elif spirv_14:
         spirv_env = 'vulkan1.1spv1.4'
+        glslang_env = 'spirv1.4'
     else:
         spirv_env = 'vulkan1.1'
+        glslang_env = 'vulkan1.1'
 
     if vulkan or spirv:
         vulkan_glsl_path = create_temporary('vk' + os.path.basename(shader))
@@ -541,7 +563,6 @@ def cross_compile(shader, vulkan, spirv, invalid_spirv, eliminate, is_legacy, fl
     if spirv:
         subprocess.check_call(spirv_cmd)
     else:
-        glslang_env = 'spirv1.4' if spirv_14 else 'vulkan1.1'
         subprocess.check_call([paths.glslang, '--amb', '--target-env', glslang_env, '-V', '-o', spirv_path, shader])
 
     if opt and (not invalid_spirv):