Merge pull request #84 from HansKristian-Work/spv-khr-integer-dot-pro…

…duct Add support for SPV_KHR_integer_dot_product.
HansKristian-Work · Oct 5, 2021 · d53f231 · d53f231
2 parents bff5ab1 + 7b4c1fe
commit d53f231
Show file tree

Hide file tree

Showing 12 changed files with 344 additions and 21 deletions.
diff --git a/dxil_converter.cpp b/dxil_converter.cpp
@@ -4954,6 +4954,10 @@ void Converter::Impl::set_option(const OptionBase &cap)
 		break;
 	}
 
+	case Option::ShaderI8Dot:
+		options.shader_i8_dot_enabled = static_cast<const OptionShaderI8Dot &>(cap).supported;
+		break;
+
 	default:
 		break;
 	}
@@ -5000,6 +5004,7 @@ bool Converter::recognizes_option(Option cap)
 	case Option::StorageInputOutput16:
 	case Option::DescriptorQA:
 	case Option::MinPrecisionNative16Bit:
+	case Option::ShaderI8Dot:
 		return true;
 
 	default:

diff --git a/dxil_converter.hpp b/dxil_converter.hpp
@@ -206,7 +206,8 @@ enum class Option : uint32_t
 	BindlessOffsetBufferLayout = 13,
 	StorageInputOutput16 = 14,
 	DescriptorQA = 15,
-	MinPrecisionNative16Bit = 16
+	MinPrecisionNative16Bit = 16,
+	ShaderI8Dot = 17
 };
 
 enum class ResourceClass : uint32_t
@@ -391,6 +392,16 @@ struct OptionMinPrecisionNative16Bit : OptionBase
 	bool enabled = false;
 };
 
+struct OptionShaderI8Dot : OptionBase
+{
+	OptionShaderI8Dot()
+		: OptionBase(Option::ShaderI8Dot)
+	{
+	}
+
+	bool supported = false;
+};
+
 struct DescriptorTableEntry
 {
 	ResourceClass type;

diff --git a/dxil_spirv.cpp b/dxil_spirv.cpp
@@ -127,6 +127,7 @@ static void print_help()
 	     "\t[--vertex-input semantic location]\n"
 	     "\t[--stream-output semantic index offset stride buffer-index]\n"
 	     "\t[--enable-shader-demote]\n"
+	     "\t[--enable-shader-i8-dot]\n"
 	     "\t[--enable-dual-source-blending]\n"
 	     "\t[--bindless]\n"
 	     "\t[--no-bda]\n"
@@ -157,6 +158,7 @@ struct Arguments
 	bool emit_asm = false;
 	bool validate = false;
 	bool shader_demote = false;
+	bool shader_i8_dot = false;
 	bool dual_source_blending = false;
 	bool debug_all_entry_points = false;
 	bool storage_input_output_16bit = false;
@@ -530,6 +532,7 @@ int main(int argc, char **argv)
 		remapper.stream_outputs.push_back({ std::string(sem), index, offset, stride, buffer_index });
 	});
 	cbs.add("--enable-shader-demote", [&](CLIParser &) { args.shader_demote = true; });
+	cbs.add("--enable-shader-i8-dot", [&](CLIParser &parser) { args.shader_i8_dot = true; });
 	cbs.add("--enable-dual-source-blending", [&](CLIParser &) { args.dual_source_blending = true; });
 	cbs.add("--bindless", [&](CLIParser &) {
 		remapper.bindless = true;
@@ -734,6 +737,13 @@ int main(int argc, char **argv)
 		dxil_spv_converter_add_option(converter, &helper.base);
 	}
 
+	if (args.shader_i8_dot)
+	{
+		const dxil_spv_option_shader_i8_dot helper = { { DXIL_SPV_OPTION_SHADER_I8_DOT },
+		                                               DXIL_SPV_TRUE };
+		dxil_spv_converter_add_option(converter, &helper.base);
+	}
+
 	if (args.dual_source_blending)
 	{
 		const dxil_spv_option_dual_source_blending helper = { { DXIL_SPV_OPTION_DUAL_SOURCE_BLENDING }, DXIL_SPV_TRUE };

diff --git a/dxil_spirv_c.cpp b/dxil_spirv_c.cpp
@@ -794,6 +794,15 @@ dxil_spv_result dxil_spv_converter_add_option(dxil_spv_converter converter, cons
 		break;
 	}
 
+	case DXIL_SPV_OPTION_SHADER_I8_DOT:
+	{
+		OptionShaderI8Dot helper;
+		helper.supported = bool(reinterpret_cast<const dxil_spv_option_shader_i8_dot *>(option)->supported);
+
+		converter->options.emplace_back(duplicate(helper));
+		break;
+	}
+
 	default:
 		return DXIL_SPV_ERROR_UNSUPPORTED_FEATURE;
 	}

diff --git a/dxil_spirv_c.h b/dxil_spirv_c.h
@@ -28,7 +28,7 @@ extern "C" {
 #endif
 
 #define DXIL_SPV_API_VERSION_MAJOR 2
-#define DXIL_SPV_API_VERSION_MINOR 10
+#define DXIL_SPV_API_VERSION_MINOR 11
 #define DXIL_SPV_API_VERSION_PATCH 0
 
 #define DXIL_SPV_DESCRIPTOR_QA_INTERFACE_VERSION 1
@@ -277,6 +277,7 @@ typedef enum dxil_spv_option
 	DXIL_SPV_OPTION_STORAGE_INPUT_OUTPUT_16BIT = 14,
 	DXIL_SPV_OPTION_DESCRIPTOR_QA = 15,
 	DXIL_SPV_OPTION_MIN_PRECISION_NATIVE_16BIT = 16,
+	DXIL_SPV_OPTION_SHADER_I8_DOT = 17,
 	DXIL_SPV_OPTION_INT_MAX = 0x7fffffff
 } dxil_spv_option;
 
@@ -408,6 +409,12 @@ typedef struct dxil_spv_option_min_precision_native_16bit
 	dxil_spv_bool enabled;
 } dxil_spv_option_min_precision_native_16bit;
 
+typedef struct dxil_spv_option_shader_i8_dot
+{
+	dxil_spv_option_base base;
+	dxil_spv_bool supported;
+} dxil_spv_option_shader_i8_dot;
+
 /* Gets the ABI version used to build this library. Used to detect API/ABI mismatches. */
 DXIL_SPV_PUBLIC_API void dxil_spv_get_version(unsigned *major, unsigned *minor, unsigned *patch);
 

diff --git a/opcodes/converter_impl.hpp b/opcodes/converter_impl.hpp
@@ -405,6 +405,7 @@ struct Converter::Impl
 		bool descriptor_qa_enabled = false;
 		bool descriptor_qa_sink_handles = true;
 		bool min_precision_prefer_native_16bit = false;
+		bool shader_i8_dot_enabled = false;
 	} options;
 
 	struct BindlessInfo

diff --git a/opcodes/dxil/dxil_arithmetic.cpp b/opcodes/dxil/dxil_arithmetic.cpp
@@ -409,31 +409,50 @@ static spv::Id build_bfe(Converter::Impl &impl, spv::Id value_id, unsigned offse
 
 bool emit_i8_dot_instruction(Converter::Impl &impl, const llvm::CallInst *instruction, bool sign_extend)
 {
-	// Can be improved with a specific intrinsic to support this.
-	// This is mostly a thing for machine learning algorithms.
-	// Could potentially be improved a bit with Int8 stuff, but meh ...
-
 	auto &builder = impl.builder();
 	spv::Id acc = impl.get_id_for_value(instruction->getOperand(1));
 	spv::Id a = impl.get_id_for_value(instruction->getOperand(2));
 	spv::Id b = impl.get_id_for_value(instruction->getOperand(3));
-	for (unsigned i = 0; i < 4; i++)
+
+	if (impl.options.shader_i8_dot_enabled)
 	{
-		spv::Id a_component = build_bfe(impl, a, 8 * i, 8, sign_extend);
-		spv::Id b_component = build_bfe(impl, b, 8 * i, 8, sign_extend);
-		auto *mul_op = impl.allocate(spv::OpIMul, builder.makeUintType(32));
-		mul_op->add_id(a_component);
-		mul_op->add_id(b_component);
-		impl.add(mul_op);
+		builder.addExtension("SPV_KHR_integer_dot_product");
+		builder.addCapability(spv::CapabilityDotProductKHR);
+		builder.addCapability(spv::CapabilityDotProductInput4x8BitPackedKHR);
+
+		// Not supposed to saturate.
+		auto *dot_op = impl.allocate(sign_extend ? spv::OpSDotKHR : spv::OpUDotKHR, builder.makeUintType(32));
+		dot_op->add_id(a);
+		dot_op->add_id(b);
+		dot_op->add_literal(spv::PackedVectorFormatPackedVectorFormat4x8BitKHR);
+		impl.add(dot_op);
+
+		auto *acc_op = impl.allocate(spv::OpIAdd, instruction);
+		acc_op->add_id(acc);
+		acc_op->add_id(dot_op->id);
+		impl.add(acc_op);
+	}
+	else
+	{
+		for (unsigned i = 0; i < 4; i++)
+		{
+			spv::Id a_component = build_bfe(impl, a, 8 * i, 8, sign_extend);
+			spv::Id b_component = build_bfe(impl, b, 8 * i, 8, sign_extend);
+			auto *mul_op = impl.allocate(spv::OpIMul, builder.makeUintType(32));
+			mul_op->add_id(a_component);
+			mul_op->add_id(b_component);
+			impl.add(mul_op);
+
+			auto *add_op = impl.allocate(spv::OpIAdd, builder.makeUintType(32));
+			add_op->add_id(acc);
+			add_op->add_id(mul_op->id);
+			acc = add_op->id;
+			impl.add(add_op);
+		}
 
-		auto *add_op = impl.allocate(spv::OpIAdd, builder.makeUintType(32));
-		add_op->add_id(acc);
-		add_op->add_id(mul_op->id);
-		acc = add_op->id;
-		impl.add(add_op);
+		impl.rewrite_value(instruction, acc);
 	}
 
-	impl.rewrite_value(instruction, acc);
 	return true;
 }