Merge pull request #952 from KhronosGroup/fix-950

MSL: Add native texture buffer support.

CMakeLists.txt

@@ -267,7 +267,7 @@ endif()
 
 if (SPIRV_CROSS_SHARED)
 	set(spirv-cross-abi-major 0)
-	set(spirv-cross-abi-minor 5)
+	set(spirv-cross-abi-minor 6)
 	set(spirv-cross-abi-patch 0)
 	set(SPIRV_CROSS_VERSION ${spirv-cross-abi-major}.${spirv-cross-abi-minor}.${spirv-cross-abi-patch})
 	set(SPIRV_CROSS_INSTALL_LIB_DIR ${CMAKE_INSTALL_PREFIX}/lib)

main.cpp

@@ -509,6 +509,7 @@ struct CLIArguments
 	bool msl_pad_fragment_output = false;
 	bool msl_domain_lower_left = false;
 	bool msl_argument_buffers = false;
+	bool msl_texture_buffer_native = false;
 	bool glsl_emit_push_constant_as_ubo = false;
 	SmallVector<uint32_t> msl_discrete_descriptor_sets;
 	SmallVector<PLSArg> pls_in;
@@ -570,6 +571,7 @@ static void print_help()
 	                "\t[--msl-pad-fragment-output]\n"
 	                "\t[--msl-domain-lower-left]\n"
 	                "\t[--msl-argument-buffers]\n"
+	                "\t[--msl-texture-buffer-native]\n"
 	                "\t[--msl-discrete-descriptor-set <index>]\n"
 	                "\t[--hlsl]\n"
 	                "\t[--reflect]\n"
@@ -727,6 +729,7 @@ static string compile_iteration(const CLIArguments &args, std::vector<uint32_t>
 		msl_opts.pad_fragment_output_components = args.msl_pad_fragment_output;
 		msl_opts.tess_domain_origin_lower_left = args.msl_domain_lower_left;
 		msl_opts.argument_buffers = args.msl_argument_buffers;
+		msl_opts.texture_buffer_native = args.msl_texture_buffer_native;
 		msl_comp->set_msl_options(msl_opts);
 		for (auto &v : args.msl_discrete_descriptor_sets)
 			msl_comp->add_discrete_descriptor_set(v);
@@ -1038,6 +1041,7 @@ static int main_inner(int argc, char *argv[])
 	cbs.add("--msl-argument-buffers", [&args](CLIParser &) { args.msl_argument_buffers = true; });
 	cbs.add("--msl-discrete-descriptor-set",
 	        [&args](CLIParser &parser) { args.msl_discrete_descriptor_sets.push_back(parser.next_uint()); });
+	cbs.add("--msl-texture-buffer-native", [&args](CLIParser &) { args.msl_texture_buffer_native = true; });
 	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/buffer-read-write.frag

@@ -16,10 +16,11 @@ uint2 spvTexelBufferCoord(uint tc)
     return uint2(tc % 4096, tc / 4096);
 }
 
-fragment main0_out main0(texture2d<float> buf [[texture(0)]])
+fragment main0_out main0(texture2d<float> buf [[texture(0)]], texture2d<float, access::write> bufOut [[texture(1)]], float4 gl_FragCoord [[position]])
 {
     main0_out out = {};
     out.FragColor = buf.read(spvTexelBufferCoord(0));
+    bufOut.write(out.FragColor, spvTexelBufferCoord(int(gl_FragCoord.x)));
     return out;
 }
 

reference/opt/shaders-msl/frag/buffer-read-write.texture-buffer-native.msl21.frag

@@ -0,0 +1,18 @@
+#include <metal_stdlib>
+#include <simd/simd.h>
+
+using namespace metal;
+
+struct main0_out
+{
+    float4 FragColor [[color(0)]];
+};
+
+fragment main0_out main0(texture_buffer<float> buf [[texture(0)]], texture_buffer<float, access::write> bufOut [[texture(1)]], float4 gl_FragCoord [[position]])
+{
+    main0_out out = {};
+    out.FragColor = buf.read(uint(0));
+    bufOut.write(out.FragColor, uint(int(gl_FragCoord.x)));
+    return out;
+}
+

reference/opt/shaders-msl/vert/texture_buffer.texture-buffer-native.msl21.vert

@@ -0,0 +1,17 @@
+#include <metal_stdlib>
+#include <simd/simd.h>
+
+using namespace metal;
+
+struct main0_out
+{
+    float4 gl_Position [[position]];
+};
+
+vertex main0_out main0(texture_buffer<float> uSamp [[texture(4)]], texture_buffer<float> uSampo [[texture(5)]])
+{
+    main0_out out = {};
+    out.gl_Position = uSamp.read(uint(10)) + uSampo.read(uint(100));
+    return out;
+}
+

reference/shaders-msl/frag/buffer-read-write.frag

@@ -16,10 +16,11 @@ uint2 spvTexelBufferCoord(uint tc)
     return uint2(tc % 4096, tc / 4096);
 }
 
-fragment main0_out main0(texture2d<float> buf [[texture(0)]])
+fragment main0_out main0(texture2d<float> buf [[texture(0)]], texture2d<float, access::write> bufOut [[texture(1)]], float4 gl_FragCoord [[position]])
 {
     main0_out out = {};
     out.FragColor = buf.read(spvTexelBufferCoord(0));
+    bufOut.write(out.FragColor, spvTexelBufferCoord(int(gl_FragCoord.x)));
     return out;
 }
 

reference/shaders-msl/frag/buffer-read-write.texture-buffer-native.msl21.frag

@@ -0,0 +1,18 @@
+#include <metal_stdlib>
+#include <simd/simd.h>
+
+using namespace metal;
+
+struct main0_out
+{
+    float4 FragColor [[color(0)]];
+};
+
+fragment main0_out main0(texture_buffer<float> buf [[texture(0)]], texture_buffer<float, access::write> bufOut [[texture(1)]], float4 gl_FragCoord [[position]])
+{
+    main0_out out = {};
+    out.FragColor = buf.read(uint(0));
+    bufOut.write(out.FragColor, uint(int(gl_FragCoord.x)));
+    return out;
+}
+

reference/shaders-msl/vert/texture_buffer.texture-buffer-native.msl21.vert

@@ -0,0 +1,17 @@
+#include <metal_stdlib>
+#include <simd/simd.h>
+
+using namespace metal;
+
+struct main0_out
+{
+    float4 gl_Position [[position]];
+};
+
+vertex main0_out main0(texture_buffer<float> uSamp [[texture(4)]], texture_buffer<float> uSampo [[texture(5)]])
+{
+    main0_out out = {};
+    out.gl_Position = uSamp.read(uint(10)) + uSampo.read(uint(100));
+    return out;
+}
+

shaders-msl/frag/buffer-read-write.frag

@@ -1,10 +1,12 @@
 #version 450
 
 layout(rgba8, binding = 0) uniform readonly imageBuffer buf;
+layout(rgba8, binding = 1) uniform writeonly imageBuffer bufOut;
 
 layout(location = 0) out vec4 FragColor;
 
 void main()
 {
 	FragColor = imageLoad(buf, 0);
+	imageStore(bufOut, int(gl_FragCoord.x), FragColor);
 }

shaders-msl/frag/buffer-read-write.texture-buffer-native.msl21.frag

@@ -0,0 +1,12 @@
+#version 450
+
+layout(rgba8, binding = 0) uniform readonly imageBuffer buf;
+layout(rgba8, binding = 1) uniform writeonly imageBuffer bufOut;
+
+layout(location = 0) out vec4 FragColor;
+
+void main()
+{
+	FragColor = imageLoad(buf, 0);
+	imageStore(bufOut, int(gl_FragCoord.x), FragColor);
+}

shaders-msl/vert/texture_buffer.texture-buffer-native.msl21.vert

@@ -0,0 +1,10 @@
+#version 310 es
+#extension GL_OES_texture_buffer : require
+
+layout(binding = 4) uniform highp samplerBuffer uSamp;
+layout(rgba32f, binding = 5) uniform readonly highp imageBuffer uSampo;
+
+void main()
+{
+   gl_Position = texelFetch(uSamp, 10) + imageLoad(uSampo, 100);
+}

spirv_cross_c.cpp

@@ -526,6 +526,10 @@ spvc_result spvc_compiler_options_set_uint(spvc_compiler_options options, spvc_c
 	case SPVC_COMPILER_OPTION_MSL_ARGUMENT_BUFFERS:
 		options->msl.argument_buffers = value != 0;
 		break;
+
+	case SPVC_COMPILER_OPTION_MSL_TEXTURE_BUFFER_NATIVE:
+		options->msl.texture_buffer_native = 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 5
+#define SPVC_C_API_VERSION_MINOR 6
 /* Bumped if internal implementation details change. */
 #define SPVC_C_API_VERSION_PATCH 0
 
@@ -424,6 +424,8 @@ typedef enum spvc_compiler_option
 
 	SPVC_COMPILER_OPTION_GLSL_EMIT_PUSH_CONSTANT_AS_UNIFORM_BUFFER = 33 | SPVC_COMPILER_OPTION_GLSL_BIT,
 
+	SPVC_COMPILER_OPTION_MSL_TEXTURE_BUFFER_NATIVE = 34 | SPVC_COMPILER_OPTION_MSL_BIT,
+
 	SPVC_COMPILER_OPTION_INT_MAX = 0x7fffffff
 } spvc_compiler_option;
 

spirv_msl.cpp

@@ -3729,6 +3729,9 @@ void CompilerMSL::emit_instruction(const Instruction &instruction)
 		break;
 	}
 
+	case OpImageTexelPointer:
+		SPIRV_CROSS_THROW("MSL does not support atomic operations on images or texel buffers.");
+
 	// Casting
 	case OpQuantizeToF16:
 	{
@@ -4483,9 +4486,16 @@ string CompilerMSL::to_function_args(uint32_t img, const SPIRType &imgtype, bool
 		if (coord_type.vecsize > 1)
 			tex_coords = enclose_expression(tex_coords) + ".x";
 
-		// Metal texel buffer textures are 2D, so convert 1D coord to 2D.
-		if (is_fetch)
-			tex_coords = "spvTexelBufferCoord(" + round_fp_tex_coords(tex_coords, coord_is_fp) + ")";
+		if (msl_options.texture_buffer_native)
+		{
+			tex_coords = "uint(" + round_fp_tex_coords(tex_coords, coord_is_fp) + ")";
+		}
+		else
+		{
+			// Metal texel buffer textures are 2D, so convert 1D coord to 2D.
+			if (is_fetch)
+				tex_coords = "spvTexelBufferCoord(" + round_fp_tex_coords(tex_coords, coord_is_fp) + ")";
+		}
 
 		alt_coord_component = 1;
 		break;
@@ -5613,14 +5623,14 @@ void CompilerMSL::entry_point_args_builtin(string &ep_args)
 			if (!ep_args.empty())
 				ep_args += ", ";
 			ep_args +=
-				join("constant uint* spvIndirectParams [[buffer(", msl_options.indirect_params_buffer_index, ")]]");
+			    join("constant uint* spvIndirectParams [[buffer(", msl_options.indirect_params_buffer_index, ")]]");
 		}
 		else if (stage_out_var_id)
 		{
 			if (!ep_args.empty())
 				ep_args += ", ";
 			ep_args +=
-				join("device uint* spvIndirectParams [[buffer(", msl_options.indirect_params_buffer_index, ")]]");
+			    join("device uint* spvIndirectParams [[buffer(", msl_options.indirect_params_buffer_index, ")]]");
 		}
 
 		// Tessellation control shaders get three additional parameters:
@@ -6643,6 +6653,18 @@ string CompilerMSL::image_type_glsl(const SPIRType &type, uint32_t id)
 			img_type_name += (img_type.arrayed ? "texture1d_array" : "texture1d");
 			break;
 		case DimBuffer:
+			if (img_type.ms || img_type.arrayed)
+				SPIRV_CROSS_THROW("Cannot use texel buffers with multisampling or array layers.");
+
+			if (msl_options.texture_buffer_native)
+			{
+				if (!msl_options.supports_msl_version(2, 1))
+					SPIRV_CROSS_THROW("Native texture_buffer type is only supported in MSL 2.1.");
+				img_type_name = "texture_buffer";
+			}
+			else
+				img_type_name += "texture2d";
+			break;
 		case Dim2D:
 		case DimSubpassData:
 			if (img_type.ms && img_type.arrayed)
@@ -7328,7 +7350,7 @@ CompilerMSL::SPVFuncImpl CompilerMSL::OpCodePreprocessor::get_spv_func_impl(Op o
 	{
 		// Retrieve the image type, and if it's a Buffer, emit a texel coordinate function
 		uint32_t tid = result_types[args[opcode == OpImageWrite ? 0 : 2]];
-		if (tid && compiler.get<SPIRType>(tid).image.dim == DimBuffer)
+		if (tid && compiler.get<SPIRType>(tid).image.dim == DimBuffer && !compiler.msl_options.texture_buffer_native)
 			return SPVFuncImplTexelBufferCoords;
 
 		if (opcode == OpImageFetch && compiler.msl_options.swizzle_texture_samples)

spirv_msl.hpp

@@ -194,6 +194,9 @@ public:
 		// Add support to explicit pad out components.
 		bool pad_fragment_output_components = false;
 
+		// Requires MSL 2.1, use the native support for texel buffers.
+		bool texture_buffer_native = false;
+
 		bool is_ios()
 		{
 			return platform == iOS;

test_shaders.py

@@ -172,6 +172,8 @@ def cross_compile_msl(shader, spirv, opt, iterations, paths):
         msl_args.append('--msl-domain-lower-left')
     if '.argument.' in shader:
         msl_args.append('--msl-argument-buffers')
+    if '.texture-buffer-native.' in shader:
+        msl_args.append('--msl-texture-buffer-native')
     if '.discrete.' in shader:
         # Arbitrary for testing purposes.
         msl_args.append('--msl-discrete-descriptor-set')