Method emitComputeDivergence

GPCS4/Graphics/Gcn/GcnCompilerFlowControl.cpp:172–228

	}

	uint32_t GcnCompiler::emitComputeDivergence()
	{
		// For compute shader, we use even subgroup against low 32 bits of exec
		// and odd subgroup for high 32 bits.
		/*
		 * if (mod(gl_SubgroupID, 2.0) < 1.0)
		 * {
		 *	value = exec_lo
		 * }
		 * else
		 * {
		 *   value = exec_hi
		 * }
		 * result = (value & gl_SubgroupEqMask.x) != 0;
		 */
		const uint32_t utypeId = getScalarTypeId(GcnScalarType::Uint32);
		const uint32_t ftypeId = getScalarTypeId(GcnScalarType::Float32);
		const uint32_t btypeId = m_module.defBoolType();

		auto subgroupId = emitCommonSystemValueLoad(GcnSystemValue::SubgroupID, GcnRegMask::select(0));
		// float modulo should be faster than integer modulo
		auto subFId = m_module.opConvertUtoF(ftypeId, subgroupId.id);
		auto isEven = m_module.opFOrdLessThan(btypeId, subFId, m_module.constf32(1.0));

		uint32_t labelEven  = m_module.allocateId();
		uint32_t labelOdd   = m_module.allocateId();
		uint32_t labelMerge = m_module.allocateId();

		m_module.opSelectionMerge(labelMerge, spv::SelectionControlMaskNone);
		m_module.opBranchConditional(isEven, labelEven, labelOdd);
		m_module.opLabel(labelEven);

		auto execLo = m_state.exec.emitLoad(GcnRegMask::select(0));

		m_module.opBranch(labelMerge);
		m_module.opLabel(labelOdd);

		auto execHi = m_state.exec.emitLoad(GcnRegMask::select(1));

		m_module.opBranch(labelMerge);
		m_module.opLabel(labelMerge);

		// Merge the result with a phi function
		const std::array<SpirvPhiLabel, 2> phiLabels = { {
			{ execLo.low.id, labelEven },
			{ execHi.low.id, labelOdd },
		} };

		uint32_t exec = m_module.opPhi(
			utypeId, phiLabels.size(), phiLabels.data());
		auto mask = emitCommonSystemValueLoad(
			GcnSystemValue::SubgroupEqMask, GcnRegMask::select(0));

		auto     value  = m_module.opBitwiseAnd(utypeId, exec, mask.id);
		uint32_t result = m_module.opINotEqual(btypeId, value, m_module.constu32(0));
		return result;
	}