Allows initialisation of float16 values for comparison tests by casting.

Affects:
dEQP-VK.glsl.builtin.precision_fp16_storage32b.*

Components: Vulkan
VK-GL-CTS issue: 1860

Change-Id: I0e211dacf31a91d2cff2912440d4283bb5a2e492

diff --git a/external/vulkancts/modules/vulkan/shaderexecutor/vktShaderExecutor.cpp b/external/vulkancts/modules/vulkan/shaderexecutor/vktShaderExecutor.cpp
index e445c6f..9d09197 100644 (file)
--- a/external/vulkancts/modules/vulkan/shaderexecutor/vktShaderExecutor.cpp
+++ b/external/vulkancts/modules/vulkan/shaderexecutor/vktShaderExecutor.cpp
@@ -44,6 +44,7 @@
 #include "deUniquePtr.hpp"
 #include "deStringUtil.hpp"
 #include "deSharedPtr.hpp"
+#include "deFloat16.h"
 
 #include <map>
 #include <sstream>
@@ -1535,7 +1536,7 @@ protected:
        deUint32                                getInputStride          (void) const            { return getLayoutStride(m_inputLayout);        }
        deUint32                                getOutputStride         (void) const            { return getLayoutStride(m_outputLayout);       }
 
-       void                                    uploadInputBuffer       (const void* const* inputPtrs, int numValues);
+       void                                    uploadInputBuffer       (const void* const* inputPtrs, int numValues, bool packFloat16Bit);
        void                                    readOutputBuffer        (void* const* outputPtrs, int numValues);
 
        static void                             declareBufferBlocks     (std::ostream& src, const ShaderSpec& spec);
@@ -1558,7 +1559,7 @@ private:
        static void                             computeVarLayout        (const std::vector<Symbol>& symbols, std::vector<VarLayout>* layout);
        static deUint32                 getLayoutStride         (const vector<VarLayout>& layout);
 
-       static void                             copyToBuffer            (const glu::VarType& varType, const VarLayout& layout, int numValues, const void* srcBasePtr, void* dstBasePtr);
+       static void                             copyToBuffer            (const glu::VarType& varType, const VarLayout& layout, int numValues, const void* srcBasePtr, void* dstBasePtr, bool packFloat16Bit);
        static void                             copyFromBuffer          (const glu::VarType& varType, const VarLayout& layout, int numValues, const void* srcBasePtr, void* dstBasePtr);
 
        de::MovePtr<Allocation> m_inputAlloc;
@@ -1734,7 +1735,7 @@ void BufferIoExecutor::generateExecBufferIo (std::ostream& src, const ShaderSpec
        }
 }
 
-void BufferIoExecutor::copyToBuffer (const glu::VarType& varType, const VarLayout& layout, int numValues, const void* srcBasePtr, void* dstBasePtr)
+void BufferIoExecutor::copyToBuffer (const glu::VarType& varType, const VarLayout& layout, int numValues, const void* srcBasePtr, void* dstBasePtr, bool packFloat16Bit)
 {
        if (varType.isBasicType())
        {
@@ -1754,7 +1755,20 @@ void BufferIoExecutor::copyToBuffer (const glu::VarType& varType, const VarLayou
                                const deUint8*  srcPtr                  = (const deUint8*)srcBasePtr + srcOffset;
                                deUint8*                dstPtr                  = (deUint8*)dstBasePtr + dstOffset;
 
-                               deMemcpy(dstPtr, srcPtr, size * numComps);
+                               if (packFloat16Bit)
+                               {
+                                       // Convert the float values to 16 bit and store in the lower 16 bits of 32 bit ints.
+                                       for (int cmpNdx=0; cmpNdx < numComps; ++cmpNdx)
+                                       {
+                                               deFloat16 f16vals[2] = {};
+                                               f16vals[0] = deFloat32To16Round(((float*)srcPtr)[cmpNdx], DE_ROUNDINGMODE_TO_ZERO);
+                                               deMemcpy(dstPtr + cmpNdx * size, &f16vals[0], size);
+                                       }
+                               }
+                               else
+                               {
+                                       deMemcpy(dstPtr, srcPtr, size * numComps);
+                               }
                        }
                }
        }
@@ -1790,7 +1804,7 @@ void BufferIoExecutor::copyFromBuffer (const glu::VarType& varType, const VarLay
                throw tcu::InternalError("Unsupported type");
 }
 
-void BufferIoExecutor::uploadInputBuffer (const void* const* inputPtrs, int numValues)
+void BufferIoExecutor::uploadInputBuffer (const void* const* inputPtrs, int numValues, bool packFloat16Bit)
 {
        const VkDevice                  vkDevice                        = m_context.getDevice();
        const DeviceInterface&  vk                                      = m_context.getDeviceInterface();
@@ -1807,7 +1821,7 @@ void BufferIoExecutor::uploadInputBuffer (const void* const* inputPtrs, int numV
                const glu::VarType&             varType         = m_shaderSpec.inputs[inputNdx].varType;
                const VarLayout&                layout          = m_inputLayout[inputNdx];
 
-               copyToBuffer(varType, layout, numValues, inputPtrs[inputNdx], m_inputAlloc->getHostPtr());
+               copyToBuffer(varType, layout, numValues, inputPtrs[inputNdx], m_inputAlloc->getHostPtr(), packFloat16Bit);
        }
 
        flushAlloc(vk, vkDevice, *m_inputAlloc);
@@ -1910,7 +1924,7 @@ ComputeShaderExecutor::~ComputeShaderExecutor     (void)
 {
 }
 
-std::string getTypeSpirv(const glu::DataType type)
+std::string getTypeSpirv(const glu::DataType type, const bool packFloat16Bit = false)
 {
        switch(type)
        {
@@ -1923,13 +1937,13 @@ std::string getTypeSpirv(const glu::DataType type)
        case glu::TYPE_FLOAT16_VEC4:
                return "%v4f16";
        case glu::TYPE_FLOAT:
-               return "%f32";
+               return packFloat16Bit ? "%u32" : "%f32";                // f16 values will be bitcast from ui32.
        case glu::TYPE_FLOAT_VEC2:
-               return "%v2f32";
+               return packFloat16Bit ? "%v2u32" : "%v2f32";    // f16 values will be bitcast from ui32.
        case glu::TYPE_FLOAT_VEC3:
-               return "%v3f32";
+               return packFloat16Bit ? "%v3u32" : "%v3f32";    // f16 values will be bitcast from ui32.
        case glu::TYPE_FLOAT_VEC4:
-               return "%v4f32";
+               return packFloat16Bit ? "%v4u32" : "%v4f32";    // f16 values will be bitcast from ui32.
        case glu::TYPE_INT:
                return "%i32";
        case glu::TYPE_INT_VEC2:
@@ -2017,8 +2031,8 @@ std::string generateSpirv(const ShaderSpec& spec, const bool are16Bit, const boo
 {
        const int                       operationAmount = 10;
        int                                     moveBitNdx              = 0;
-       const std::string       inputType1              = getTypeSpirv(spec.inputs[0].varType.getBasicType());
-       const std::string       inputType2              = getTypeSpirv(spec.inputs[1].varType.getBasicType());
+       const std::string       inputType1              = getTypeSpirv(spec.inputs[0].varType.getBasicType(), spec.packFloat16Bit);
+       const std::string       inputType2              = getTypeSpirv(spec.inputs[1].varType.getBasicType(), spec.packFloat16Bit);
        const std::string       outputType              = getTypeSpirv(spec.outputs[0].varType.getBasicType());
        const std::string       packType                = spec.packFloat16Bit ? getTypeSpirv(getDataTypeFloat16Scalars(spec.inputs[0].varType.getBasicType())) : "";
 
@@ -2138,7 +2152,9 @@ std::string generateSpirv(const ShaderSpec& spec, const bool are16Bit, const boo
                "%v2i32 = OpTypeVector %i32 2\n"
                "%v3i32 = OpTypeVector %i32 3\n"
                "%v4i32 = OpTypeVector %i32 4\n"
+               "%v2u32 = OpTypeVector %u32 2\n"
                "%v3u32 = OpTypeVector %u32 3\n"
+               "%v4u32 = OpTypeVector %u32 4\n"
                "\n"
                "%ip_u32   = OpTypePointer Input %u32\n"
                "%ip_v3u32 = OpTypePointer Input %v3u32\n"
@@ -2217,8 +2233,38 @@ std::string generateSpirv(const ShaderSpec& spec, const bool are16Bit, const boo
                "%src_val_0_0 = OpLoad " << inputType1 << " %src_ptr_0_0\n";
 
        if(spec.packFloat16Bit)
-               src << "%val_f16_0_0 = OpFConvert " << packType <<" %src_val_0_0\n"
-                       "OpStore %in0 %val_f16_0_0\n";
+       {
+               if (spec.inputs[0].varType.getScalarSize() > 1)
+               {
+                       // Extract the val0 u32 input channels into individual f16 values.
+                       for (int i=0;i<spec.inputs[0].varType.getScalarSize();++i)
+                       {
+                               src << "%src_val_0_0_" << i << " = OpCompositeExtract %u32 %src_val_0_0 " << i << "\n"
+                                       "%val_v2f16_0_0_" << i << " = OpBitcast %v2f16 %src_val_0_0_" << i << "\n"
+                                       "%val_f16_0_0_" << i << " = OpCompositeExtract %f16 %val_v2f16_0_0_" << i << " 0\n";
+                       }
+
+                       if (spec.inputs[0].varType.getScalarSize() > 1)
+                       {
+                               // Construct the input vector.
+                               src << "%val_f16_0_0   = OpCompositeConstruct " << packType;
+                               for (int i=0;i<spec.inputs[0].varType.getScalarSize();++i)
+                               {
+                                       src << " %val_f16_0_0_" << i;
+                               }
+
+                               src << "\n";
+                               src << "OpStore %in0 %val_f16_0_0\n";
+                       }
+               }
+               else
+               {
+                       src << "%val_v2f16_0_0 = OpBitcast %v2f16 %src_val_0_0\n"
+                               "%val_f16_0_0 = OpCompositeExtract %f16 %val_v2f16_0_0 0\n";
+
+                       src <<  "OpStore %in0 %val_f16_0_0\n";
+               }
+       }
        else
                src << "OpStore %in0 %src_val_0_0\n";
 
@@ -2227,8 +2273,38 @@ std::string generateSpirv(const ShaderSpec& spec, const bool are16Bit, const boo
                "%src_val_0_1 = OpLoad " << inputType2 << " %src_ptr_0_1\n";
 
        if (spec.packFloat16Bit)
-               src << "%val_f16_0_1 = OpFConvert " << packType << " %src_val_0_1\n"
-                       "OpStore %in1 %val_f16_0_1\n";
+       {
+               if (spec.inputs[0].varType.getScalarSize() > 1)
+               {
+                       // Extract the val1 u32 input channels into individual f16 values.
+                       for (int i=0;i<spec.inputs[0].varType.getScalarSize();++i)
+                       {
+                               src << "%src_val_0_1_" << i << " = OpCompositeExtract %u32 %src_val_0_1 " << i << "\n"
+                                       "%val_v2f16_0_1_" << i << " = OpBitcast %v2f16 %src_val_0_1_" << i << "\n"
+                                       "%val_f16_0_1_" << i << " = OpCompositeExtract %f16 %val_v2f16_0_1_" << i << " 0\n";
+                       }
+
+                       if (spec.inputs[0].varType.getScalarSize() > 1)
+                       {
+                               // Construct the input vector.
+                               src << "%val_f16_0_1   = OpCompositeConstruct " << packType;
+                               for (int i=0;i<spec.inputs[0].varType.getScalarSize();++i)
+                               {
+                                       src << " %val_f16_0_1_" << i;
+                               }
+
+                               src << "\n";
+                               src <<  "OpStore %in1 %val_f16_0_1\n";
+                       }
+               }
+               else
+               {
+                       src << "%val_v2f16_0_1 = OpBitcast %v2f16 %src_val_0_1\n"
+                               "%val_f16_0_1 = OpCompositeExtract %f16 %val_v2f16_0_1 0\n";
+
+                       src <<  "OpStore %in1 %val_f16_0_1\n";
+               }
+       }
        else
                src << "OpStore %in1 %src_val_0_1\n";
 
@@ -2262,6 +2338,7 @@ std::string generateSpirv(const ShaderSpec& spec, const bool are16Bit, const boo
                "\n"
                "OpReturn\n"
                "OpFunctionEnd\n";
+
        return src.str();
 }
 
@@ -2344,7 +2421,10 @@ void ComputeShaderExecutor::execute (int numValues, const void* const* inputs, v
        initBuffers(numValues);
 
        // Setup input buffer & copy data
-       uploadInputBuffer(inputs, numValues);
+       // For spirv shaders using packed 16 bit float values as input, the floats are converted to 16 bit before
+       // storing in the lower 16 bits of 32 bit integers in the uniform buffer and cast back to 16 bit floats in
+       // the shader.
+       uploadInputBuffer(inputs, numValues, m_shaderSpec.packFloat16Bit && m_shaderSpec.spirVShader);
 
        // Create command pool
        cmdPool = createCommandPool(vk, vkDevice, VK_COMMAND_POOL_CREATE_TRANSIENT_BIT, queueFamilyIndex);
@@ -2925,7 +3005,7 @@ void TessControlExecutor::execute (int numValues, const void* const* inputs, voi
        initBuffers(numValues);
 
        // Setup input buffer & copy data
-       uploadInputBuffer(inputs, numValues);
+       uploadInputBuffer(inputs, numValues, false);
 
        renderTess(numValues, patchSize * numValues, patchSize, extraResources);
 
@@ -3026,7 +3106,7 @@ void TessEvaluationExecutor::execute (int numValues, const void* const* inputs,
        initBuffers(alignedValues);
 
        // Setup input buffer & copy data
-       uploadInputBuffer(inputs, numValues);
+       uploadInputBuffer(inputs, numValues, false);
 
        renderTess((deUint32)alignedValues, (deUint32)alignedValues, (deUint32)patchSize, extraResources);