GLSL: Implement GL_EXT_mesh_shader

checkout_glslang_spirv_tools.sh

@@ -2,7 +2,7 @@
 # Copyright 2016-2021 The Khronos Group Inc.
 # SPDX-License-Identifier: Apache-2.0
 
-GLSLANG_REV=69ae9e7460499b488cb2d32edae623a95264db14
+GLSLANG_REV=a53aa3e94f8b4bf22a6eb6e8a207fe91ab02b989
 SPIRV_TOOLS_REV=4c456f7da67c5437a6fb7d4d20d78e2a5ae2acf2
 SPIRV_HEADERS_REV=87d5b782bec60822aa878941e6b13c0a9a954c9b
 PROTOCOL=https

main.cpp

@@ -1082,6 +1082,10 @@ static ExecutionModel stage_to_execution_model(const std::string &stage)
 		return ExecutionModelMissKHR;
 	else if (stage == "rcall")
 		return ExecutionModelCallableKHR;
+	else if (stage == "mesh")
+		return spv::ExecutionModelMeshEXT;
+	else if (stage == "task")
+		return spv::ExecutionModelTaskEXT;
 	else
 		SPIRV_CROSS_THROW("Invalid stage.");
 }

reference/opt/shaders/frag/barycentric-khr.frag

@@ -3,7 +3,7 @@
 
 layout(location = 0) out vec2 value;
 layout(location = 0) pervertexEXT in vec2 vUV[3];
-layout(location = 1) pervertexEXT in vec2 vUV2[3];
+layout(location = 3) pervertexEXT in vec2 vUV2[3];
 
 void main()
 {

reference/opt/shaders/mesh/mesh-shader-basic-lines.spv14.vk.nocompat.mesh.vk

@@ -0,0 +1,66 @@
+#version 450
+#extension GL_EXT_mesh_shader : require
+#extension GL_EXT_fragment_shading_rate : require
+layout(local_size_x = 2, local_size_y = 3, local_size_z = 4) in;
+layout(max_vertices = 24, max_primitives = 22, lines) out;
+
+out gl_MeshPerVertexEXT
+{
+    vec4 gl_Position;
+    float gl_PointSize;
+    float gl_ClipDistance[1];
+    float gl_CullDistance[2];
+} gl_MeshVerticesEXT[];
+
+struct TaskPayload
+{
+    float a;
+    float b;
+    int c;
+};
+
+layout(location = 0) out vec4 vOut[24];
+layout(location = 2) out BlockOut
+{
+    vec4 a;
+    vec4 b;
+} outputs[24];
+
+layout(location = 1) perprimitiveEXT out vec4 vPrim[22];
+layout(location = 4) perprimitiveEXT out BlockOutPrim
+{
+    vec4 a;
+    vec4 b;
+} prim_outputs[22];
+
+taskPayloadSharedEXT TaskPayload payload;
+shared float shared_float[16];
+
+void main()
+{
+    SetMeshOutputsEXT(24u, 22u);
+    vec3 _29 = vec3(gl_GlobalInvocationID);
+    float _31 = _29.x;
+    gl_MeshVerticesEXT[gl_LocalInvocationIndex].gl_Position = vec4(_31, _29.yz, 1.0);
+    gl_MeshVerticesEXT[gl_LocalInvocationIndex].gl_PointSize = 2.0;
+    gl_MeshVerticesEXT[gl_LocalInvocationIndex].gl_ClipDistance[0] = 4.0;
+    gl_MeshVerticesEXT[gl_LocalInvocationIndex].gl_CullDistance[1] = 5.0;
+    vOut[gl_LocalInvocationIndex] = vec4(_31, _29.yz, 2.0);
+    outputs[gl_LocalInvocationIndex].a = vec4(5.0);
+    outputs[gl_LocalInvocationIndex].b = vec4(6.0);
+    barrier();
+    if (gl_LocalInvocationIndex < 22u)
+    {
+        vPrim[gl_LocalInvocationIndex] = vec4(vec3(gl_WorkGroupID), 3.0);
+        prim_outputs[gl_LocalInvocationIndex].a = vec4(payload.a);
+        prim_outputs[gl_LocalInvocationIndex].b = vec4(payload.b);
+        gl_PrimitiveLineIndicesEXT[gl_LocalInvocationIndex] = uvec2(0u, 1u) + uvec2(gl_LocalInvocationIndex);
+        int _129 = int(gl_GlobalInvocationID.x);
+        gl_MeshPrimitivesEXT[gl_LocalInvocationIndex].gl_PrimitiveID = _129;
+        gl_MeshPrimitivesEXT[gl_LocalInvocationIndex].gl_Layer = _129 + 1;
+        gl_MeshPrimitivesEXT[gl_LocalInvocationIndex].gl_ViewportIndex = _129 + 2;
+        gl_MeshPrimitivesEXT[gl_LocalInvocationIndex].gl_CullPrimitiveEXT = (gl_GlobalInvocationID.x & 1u) != 0u;
+        gl_MeshPrimitivesEXT[gl_LocalInvocationIndex].gl_PrimitiveShadingRateEXT = _129 + 3;
+    }
+}
+

reference/opt/shaders/mesh/mesh-shader-basic-points.spv14.vk.nocompat.mesh.vk

@@ -0,0 +1,66 @@
+#version 450
+#extension GL_EXT_mesh_shader : require
+#extension GL_EXT_fragment_shading_rate : require
+layout(local_size_x = 2, local_size_y = 3, local_size_z = 4) in;
+layout(max_vertices = 24, max_primitives = 22, points) out;
+
+out gl_MeshPerVertexEXT
+{
+    vec4 gl_Position;
+    float gl_PointSize;
+    float gl_ClipDistance[1];
+    float gl_CullDistance[2];
+} gl_MeshVerticesEXT[];
+
+struct TaskPayload
+{
+    float a;
+    float b;
+    int c;
+};
+
+layout(location = 0) out vec4 vOut[24];
+layout(location = 2) out BlockOut
+{
+    vec4 a;
+    vec4 b;
+} outputs[24];
+
+layout(location = 1) perprimitiveEXT out vec4 vPrim[22];
+layout(location = 4) perprimitiveEXT out BlockOutPrim
+{
+    vec4 a;
+    vec4 b;
+} prim_outputs[22];
+
+taskPayloadSharedEXT TaskPayload payload;
+shared float shared_float[16];
+
+void main()
+{
+    SetMeshOutputsEXT(24u, 22u);
+    vec3 _29 = vec3(gl_GlobalInvocationID);
+    float _31 = _29.x;
+    gl_MeshVerticesEXT[gl_LocalInvocationIndex].gl_Position = vec4(_31, _29.yz, 1.0);
+    gl_MeshVerticesEXT[gl_LocalInvocationIndex].gl_PointSize = 2.0;
+    gl_MeshVerticesEXT[gl_LocalInvocationIndex].gl_ClipDistance[0] = 4.0;
+    gl_MeshVerticesEXT[gl_LocalInvocationIndex].gl_CullDistance[1] = 5.0;
+    vOut[gl_LocalInvocationIndex] = vec4(_31, _29.yz, 2.0);
+    outputs[gl_LocalInvocationIndex].a = vec4(5.0);
+    outputs[gl_LocalInvocationIndex].b = vec4(6.0);
+    barrier();
+    if (gl_LocalInvocationIndex < 22u)
+    {
+        vPrim[gl_LocalInvocationIndex] = vec4(vec3(gl_WorkGroupID), 3.0);
+        prim_outputs[gl_LocalInvocationIndex].a = vec4(payload.a);
+        prim_outputs[gl_LocalInvocationIndex].b = vec4(payload.b);
+        gl_PrimitivePointIndicesEXT[gl_LocalInvocationIndex] = gl_LocalInvocationIndex;
+        int _124 = int(gl_GlobalInvocationID.x);
+        gl_MeshPrimitivesEXT[gl_LocalInvocationIndex].gl_PrimitiveID = _124;
+        gl_MeshPrimitivesEXT[gl_LocalInvocationIndex].gl_Layer = _124 + 1;
+        gl_MeshPrimitivesEXT[gl_LocalInvocationIndex].gl_ViewportIndex = _124 + 2;
+        gl_MeshPrimitivesEXT[gl_LocalInvocationIndex].gl_CullPrimitiveEXT = (gl_GlobalInvocationID.x & 1u) != 0u;
+        gl_MeshPrimitivesEXT[gl_LocalInvocationIndex].gl_PrimitiveShadingRateEXT = _124 + 3;
+    }
+}
+

reference/opt/shaders/mesh/mesh-shader-basic-triangle.spv14.vk.nocompat.mesh.vk

@@ -0,0 +1,66 @@
+#version 450
+#extension GL_EXT_mesh_shader : require
+#extension GL_EXT_fragment_shading_rate : require
+layout(local_size_x = 2, local_size_y = 3, local_size_z = 4) in;
+layout(max_vertices = 24, max_primitives = 22, triangles) out;
+
+out gl_MeshPerVertexEXT
+{
+    vec4 gl_Position;
+    float gl_PointSize;
+    float gl_ClipDistance[1];
+    float gl_CullDistance[2];
+} gl_MeshVerticesEXT[];
+
+struct TaskPayload
+{
+    float a;
+    float b;
+    int c;
+};
+
+layout(location = 0) out vec4 vOut[24];
+layout(location = 2) out BlockOut
+{
+    vec4 a;
+    vec4 b;
+} outputs[24];
+
+layout(location = 1) perprimitiveEXT out vec4 vPrim[22];
+layout(location = 4) perprimitiveEXT out BlockOutPrim
+{
+    vec4 a;
+    vec4 b;
+} prim_outputs[22];
+
+taskPayloadSharedEXT TaskPayload payload;
+shared float shared_float[16];
+
+void main()
+{
+    SetMeshOutputsEXT(24u, 22u);
+    vec3 _29 = vec3(gl_GlobalInvocationID);
+    float _31 = _29.x;
+    gl_MeshVerticesEXT[gl_LocalInvocationIndex].gl_Position = vec4(_31, _29.yz, 1.0);
+    gl_MeshVerticesEXT[gl_LocalInvocationIndex].gl_PointSize = 2.0;
+    gl_MeshVerticesEXT[gl_LocalInvocationIndex].gl_ClipDistance[0] = 4.0;
+    gl_MeshVerticesEXT[gl_LocalInvocationIndex].gl_CullDistance[1] = 5.0;
+    vOut[gl_LocalInvocationIndex] = vec4(_31, _29.yz, 2.0);
+    outputs[gl_LocalInvocationIndex].a = vec4(5.0);
+    outputs[gl_LocalInvocationIndex].b = vec4(6.0);
+    barrier();
+    if (gl_LocalInvocationIndex < 22u)
+    {
+        vPrim[gl_LocalInvocationIndex] = vec4(vec3(gl_WorkGroupID), 3.0);
+        prim_outputs[gl_LocalInvocationIndex].a = vec4(payload.a);
+        prim_outputs[gl_LocalInvocationIndex].b = vec4(payload.b);
+        gl_PrimitiveTriangleIndicesEXT[gl_LocalInvocationIndex] = uvec3(0u, 1u, 2u) + uvec3(gl_LocalInvocationIndex);
+        int _128 = int(gl_GlobalInvocationID.x);
+        gl_MeshPrimitivesEXT[gl_LocalInvocationIndex].gl_PrimitiveID = _128;
+        gl_MeshPrimitivesEXT[gl_LocalInvocationIndex].gl_Layer = _128 + 1;
+        gl_MeshPrimitivesEXT[gl_LocalInvocationIndex].gl_ViewportIndex = _128 + 2;
+        gl_MeshPrimitivesEXT[gl_LocalInvocationIndex].gl_CullPrimitiveEXT = (gl_GlobalInvocationID.x & 1u) != 0u;
+        gl_MeshPrimitivesEXT[gl_LocalInvocationIndex].gl_PrimitiveShadingRateEXT = _128 + 3;
+    }
+}
+

reference/opt/shaders/task/task-shader-basic-2.vk.spv14.nocompat.task.vk

@@ -0,0 +1,42 @@
+#version 450
+#extension GL_EXT_mesh_shader : require
+layout(local_size_x = 4, local_size_y = 3, local_size_z = 2) in;
+
+struct Payload
+{
+    float v[3];
+};
+
+shared float vs[24];
+taskPayloadSharedEXT Payload p;
+
+void main()
+{
+    vs[gl_LocalInvocationIndex] = 10.0;
+    barrier();
+    if (gl_LocalInvocationIndex < 12u)
+    {
+        vs[gl_LocalInvocationIndex] += vs[gl_LocalInvocationIndex + 12u];
+    }
+    barrier();
+    if (gl_LocalInvocationIndex < 6u)
+    {
+        vs[gl_LocalInvocationIndex] += vs[gl_LocalInvocationIndex + 6u];
+    }
+    barrier();
+    if (gl_LocalInvocationIndex < 3u)
+    {
+        vs[gl_LocalInvocationIndex] += vs[gl_LocalInvocationIndex + 3u];
+    }
+    barrier();
+    p.v[gl_LocalInvocationIndex] = vs[gl_LocalInvocationIndex];
+    if (vs[5] > 20.0)
+    {
+        EmitMeshTasksEXT(uint(int(vs[4])), uint(int(vs[6])), uint(int(vs[8])));
+    }
+    else
+    {
+        EmitMeshTasksEXT(uint(int(vs[6])), 10u, 50u);
+    }
+}
+

reference/opt/shaders/task/task-shader-basic.vk.spv14.nocompat.task.vk

@@ -0,0 +1,35 @@
+#version 450
+#extension GL_EXT_mesh_shader : require
+layout(local_size_x = 4, local_size_y = 3, local_size_z = 2) in;
+
+struct Payload
+{
+    float v[3];
+};
+
+shared float vs[24];
+taskPayloadSharedEXT Payload p;
+
+void main()
+{
+    vs[gl_LocalInvocationIndex] = 10.0;
+    barrier();
+    if (gl_LocalInvocationIndex < 12u)
+    {
+        vs[gl_LocalInvocationIndex] += vs[gl_LocalInvocationIndex + 12u];
+    }
+    barrier();
+    if (gl_LocalInvocationIndex < 6u)
+    {
+        vs[gl_LocalInvocationIndex] += vs[gl_LocalInvocationIndex + 6u];
+    }
+    barrier();
+    if (gl_LocalInvocationIndex < 3u)
+    {
+        vs[gl_LocalInvocationIndex] += vs[gl_LocalInvocationIndex + 3u];
+    }
+    barrier();
+    p.v[gl_LocalInvocationIndex] = vs[gl_LocalInvocationIndex];
+    EmitMeshTasksEXT(uint(int(vs[4])), uint(int(vs[6])), uint(int(vs[8])));
+}
+

reference/shaders-no-opt/asm/task/task-shader.vk.nocompat.invalid.spv14.asm.task.vk

@@ -0,0 +1,35 @@
+#version 450
+#extension GL_EXT_mesh_shader : require
+layout(local_size_x = 4, local_size_y = 3, local_size_z = 2) in;
+
+struct Payload
+{
+    float v[3];
+};
+
+shared float vs[24];
+taskPayloadSharedEXT Payload p;
+
+void main()
+{
+    vs[gl_LocalInvocationIndex] = 10.0;
+    barrier();
+    if (gl_LocalInvocationIndex < 12u)
+    {
+        vs[gl_LocalInvocationIndex] += vs[gl_LocalInvocationIndex + 12u];
+    }
+    barrier();
+    if (gl_LocalInvocationIndex < 6u)
+    {
+        vs[gl_LocalInvocationIndex] += vs[gl_LocalInvocationIndex + 6u];
+    }
+    barrier();
+    if (gl_LocalInvocationIndex < 3u)
+    {
+        vs[gl_LocalInvocationIndex] += vs[gl_LocalInvocationIndex + 3u];
+    }
+    barrier();
+    p.v[gl_LocalInvocationIndex] = vs[gl_LocalInvocationIndex];
+    EmitMeshTasksEXT(uint(int(vs[4])), uint(int(vs[6])), uint(int(vs[8])));
+}
+

reference/shaders/frag/barycentric-khr.frag

@@ -3,7 +3,7 @@
 
 layout(location = 0) out vec2 value;
 layout(location = 0) pervertexEXT in vec2 vUV[3];
-layout(location = 1) pervertexEXT in vec2 vUV2[3];
+layout(location = 3) pervertexEXT in vec2 vUV2[3];
 
 void main()
 {