Merge pull request #626 from billhollings/master

CompilerMSL support larger texel buffers by using 2D Metal textures.

reference/opt/shaders-msl/vert/texture_buffer.vert

@@ -1,3 +1,5 @@
+#pragma clang diagnostic ignored "-Wmissing-prototypes"
+
 #include <metal_stdlib>
 #include <simd/simd.h>
 
@@ -8,10 +10,16 @@ struct main0_out
     float4 gl_Position [[position]];
 };
 
+// Returns 2D texture coords corresponding to 1D texel buffer coords
+uint2 spvTexelBufferCoord(uint tc)
+{
+    return uint2(tc % 4096, tc / 4096);
+}
+
 vertex main0_out main0(texture2d<float> uSamp [[texture(4)]], texture2d<float> uSampo [[texture(5)]])
 {
     main0_out out = {};
-    out.gl_Position = uSamp.read(uint2(10, 0)) + uSampo.read(uint2(100, 0));
+    out.gl_Position = uSamp.read(spvTexelBufferCoord(10)) + uSampo.read(spvTexelBufferCoord(100));
     return out;
 }
 

reference/shaders-msl-no-opt/vert/functions_nested.vert

@@ -36,6 +36,12 @@ struct main0_out
     float4 gl_Position [[position]];
 };
 
+// Returns 2D texture coords corresponding to 1D texel buffer coords
+uint2 spvTexelBufferCoord(uint tc)
+{
+    return uint2(tc % 4096, tc / 4096);
+}
+
 attr_desc fetch_desc(thread const int& location, constant VertexBuffer& v_227)
 {
     int attribute_flags = v_227.input_attributes[location].w;
@@ -76,10 +82,10 @@ float4 fetch_attr(thread const attr_desc& desc, thread const int& vertex_id, thr
             {
                 int _131 = first_byte;
                 first_byte = _131 + 1;
-                tmp.x = input_stream.read(uint2(_131, 0)).x;
+                tmp.x = input_stream.read(spvTexelBufferCoord(_131)).x;
                 int _138 = first_byte;
                 first_byte = _138 + 1;
-                tmp.y = input_stream.read(uint2(_138, 0)).x;
+                tmp.y = input_stream.read(spvTexelBufferCoord(_138)).x;
                 uint4 param = tmp;
                 int param_1 = desc.swap_bytes;
                 result[n] = float(get_bits(param, param_1));
@@ -89,16 +95,16 @@ float4 fetch_attr(thread const attr_desc& desc, thread const int& vertex_id, thr
             {
                 int _156 = first_byte;
                 first_byte = _156 + 1;
-                tmp.x = input_stream.read(uint2(_156, 0)).x;
+                tmp.x = input_stream.read(spvTexelBufferCoord(_156)).x;
                 int _163 = first_byte;
                 first_byte = _163 + 1;
-                tmp.y = input_stream.read(uint2(_163, 0)).x;
+                tmp.y = input_stream.read(spvTexelBufferCoord(_163)).x;
                 int _170 = first_byte;
                 first_byte = _170 + 1;
-                tmp.z = input_stream.read(uint2(_170, 0)).x;
+                tmp.z = input_stream.read(spvTexelBufferCoord(_170)).x;
                 int _177 = first_byte;
                 first_byte = _177 + 1;
-                tmp.w = input_stream.read(uint2(_177, 0)).x;
+                tmp.w = input_stream.read(spvTexelBufferCoord(_177)).x;
                 uint4 param_2 = tmp;
                 int param_3 = desc.swap_bytes;
                 result[n] = as_type<float>(get_bits(param_2, param_3));
@@ -108,7 +114,7 @@ float4 fetch_attr(thread const attr_desc& desc, thread const int& vertex_id, thr
             {
                 int _195 = first_byte;
                 first_byte = _195 + 1;
-                result[n] = float(input_stream.read(uint2(_195, 0)).x);
+                result[n] = float(input_stream.read(spvTexelBufferCoord(_195)).x);
                 reverse_order = desc.swap_bytes != 0;
                 break;
             }

reference/shaders-msl/vert/texture_buffer.vert

@@ -1,3 +1,5 @@
+#pragma clang diagnostic ignored "-Wmissing-prototypes"
+
 #include <metal_stdlib>
 #include <simd/simd.h>
 
@@ -8,10 +10,16 @@ struct main0_out
     float4 gl_Position [[position]];
 };
 
+// Returns 2D texture coords corresponding to 1D texel buffer coords
+uint2 spvTexelBufferCoord(uint tc)
+{
+    return uint2(tc % 4096, tc / 4096);
+}
+
 vertex main0_out main0(texture2d<float> uSamp [[texture(4)]], texture2d<float> uSampo [[texture(5)]])
 {
     main0_out out = {};
-    out.gl_Position = uSamp.read(uint2(10, 0)) + uSampo.read(uint2(100, 0));
+    out.gl_Position = uSamp.read(spvTexelBufferCoord(10)) + uSampo.read(spvTexelBufferCoord(100));
     return out;
 }
 

spirv_msl.cpp

@@ -1173,6 +1173,18 @@ void CompilerMSL::emit_custom_functions()
 			statement("");
 			break;
 
+		case SPVFuncImplTexelBufferCoords:
+		{
+			string tex_width_str = convert_to_string(msl_options.texel_buffer_texture_width);
+			statement("// Returns 2D texture coords corresponding to 1D texel buffer coords");
+			statement("uint2 spvTexelBufferCoord(uint tc)");
+			begin_scope();
+			statement(join("return uint2(tc % ", tex_width_str, ", tc / ", tex_width_str, ");"));
+			end_scope();
+			statement("");
+			break;
+		}
+
 		case SPVFuncImplInverse4x4:
 			statement("// Returns the determinant of a 2x2 matrix.");
 			statement("inline float spvDet2x2(float a1, float a2, float b1, float b2)");
@@ -2454,8 +2466,9 @@ string CompilerMSL::to_function_args(uint32_t img, const SPIRType &imgtype, bool
 		if (coord_type.vecsize > 1)
 			tex_coords += ".x";
 
+		// Metal texel buffer textures are 2D, so convert 1D coord to 2D.
 		if (is_fetch)
-			tex_coords = "uint2(" + round_fp_tex_coords(tex_coords, coord_is_fp) + ", 0)"; // Metal textures are 2D
+			tex_coords = "spvTexelBufferCoord(" + round_fp_tex_coords(tex_coords, coord_is_fp) + ")";
 
 		alt_coord = ".y";
 
@@ -4053,8 +4066,8 @@ CompilerMSL::SPVFuncImpl CompilerMSL::OpCodePreprocessor::get_spv_func_impl(Op o
 		auto &return_type = compiler.get<SPIRType>(args[0]);
 		if (!return_type.array.empty())
 			return SPVFuncImplArrayCopy;
-		else
-			return SPVFuncImplNone;
+
+		break;
 	}
 
 	case OpStore:
@@ -4072,14 +4085,24 @@ CompilerMSL::SPVFuncImpl CompilerMSL::OpCodePreprocessor::get_spv_func_impl(Op o
 		else
 		{
 			// Or ... an expression.
-			if (result_types[id_rhs] != 0)
-				type = &compiler.get<SPIRType>(result_types[id_rhs]);
+			uint32_t tid = result_types[id_rhs];
+			if (tid)
+				type = &compiler.get<SPIRType>(tid);
 		}
 
 		if (type && compiler.is_array(*type))
 			return SPVFuncImplArrayCopy;
-		else
-			return SPVFuncImplNone;
+
+		break;
+	}
+
+	case OpImageFetch:
+	{
+		// Retrieve the image type, and if it's a Buffer, emit a texel coordinate function
+		uint32_t tid = result_types[args[2]];
+		if (tid && compiler.get<SPIRType>(tid).image.dim == DimBuffer)
+			return SPVFuncImplTexelBufferCoords;
+
 		break;
 	}
 

spirv_msl.hpp

@@ -152,6 +152,7 @@ public:
 
 		Platform platform = macOS;
 		uint32_t msl_version = make_msl_version(1, 2);
+		uint32_t texel_buffer_texture_width = 4096;		// Width of 2D Metal textures used as 1D texel buffers
 		bool enable_point_size_builtin = true;
 		bool resolve_specialized_array_lengths = true;
 
@@ -215,6 +216,7 @@ public:
 		SPVFuncImplFindSMsb,
 		SPVFuncImplFindUMsb,
 		SPVFuncImplArrayCopy,
+		SPVFuncImplTexelBufferCoords,
 		SPVFuncImplInverse4x4,
 		SPVFuncImplInverse3x3,
 		SPVFuncImplInverse2x2,