SPECIAL data gen mode for FP16 and FP32

Signed-off-by: evacha01 <evan.chandler@arm.com>
Change-Id: I5a9a1c63345bd83ca04bc6c2a99b0ef3612971ee

16 files changed, 529 insertions, 27 deletions

diff --git a/reference_model/CMakeLists.txt b/reference_model/CMakeLists.txt
index b780781..cf1e9fe 100644
--- a/reference_model/CMakeLists.txt
+++ b/reference_model/CMakeLists.txt
@@ -75,6 +75,7 @@ set(CXX_SOURCE
     src/generate/generate_pseudo_random.cc
     src/generate/generate_fixed_data.cc
     src/generate/generate_full_range.cc
+    src/generate/generate_fp_special.cc
     src/generate/generate_entry.cc
     src/generate/generate_utils.cc
     src/verify/verify_abs_error.cc
@@ -84,6 +85,7 @@ set(CXX_SOURCE
     src/verify/verify_reduce_product.cc
     src/verify/verify_relative.cc
     src/verify/verify_ulp.cc
+    src/verify/verify_fp_special.cc
     src/verify/verify_utils.cc
     src/ops/op_factory.cc
     src/ops/tensor_ops.cc
@@ -161,6 +163,7 @@ add_library(tosa_reference_verify_lib SHARED
   src/verify/verify_reduce_product.cc
   src/verify/verify_relative.cc
   src/verify/verify_ulp.cc
+  src/verify/verify_fp_special.cc
   src/verify/verify_utils.cc
   src/verify/verify_config.cc
   src/func_debug.cc
@@ -179,6 +182,7 @@ add_library(tosa_reference_generate_lib SHARED
   src/generate/generate_pseudo_random.cc
   src/generate/generate_fixed_data.cc
   src/generate/generate_full_range.cc
+  src/generate/generate_fp_special.cc
   src/generate/generate_entry.cc
   src/generate/generate_utils.cc
   src/generate/generate_config.cc
diff --git a/reference_model/src/generate/generate_entry.cc b/reference_model/src/generate/generate_entry.cc
index 6f797b3..36d99ad 100644
--- a/reference_model/src/generate/generate_entry.cc
+++ b/reference_model/src/generate/generate_entry.cc
@@ -16,6 +16,7 @@
 
 #include "generate_dot_product.h"
 #include "generate_fixed_data.h"
+#include "generate_fp_special.h"
 #include "generate_full_range.h"
 #include "generate_pseudo_random.h"
 #include "generate_utils.h"
@@ -46,6 +47,10 @@ bool generate(const GenerateConfig& cfg, void* data, size_t size)
             return generateFullRange(cfg, data, size);
             break;
         }
+        case GeneratorType::FpSpecial: {
+            return generateFpSpecial(cfg, data, size);
+            break;
+        }
         default: {
             WARNING("[Generator] Unsupported generation mode.");
             break;
diff --git a/reference_model/src/generate/generate_fp_special.cc b/reference_model/src/generate/generate_fp_special.cc
new file mode 100644
index 0000000..3602f51
--- /dev/null
+++ b/reference_model/src/generate/generate_fp_special.cc
@@ -0,0 +1,180 @@
+// Copyright (c) 2024, ARM Limited.
+//
+//    Licensed under the Apache License, Version 2.0 (the "License");
+//    you may not use this file except in compliance with the License.
+//    You may obtain a copy of the License at
+//
+//         http://www.apache.org/licenses/LICENSE-2.0
+//
+//    Unless required by applicable law or agreed to in writing, software
+//    distributed under the License is distributed on an "AS IS" BASIS,
+//    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+//    See the License for the specific language governing permissions and
+//    limitations under the License.
+
+#include "generate_fp_special.h"
+#include "half.hpp"
+
+#include <map>
+
+namespace
+{
+
+class SpecialValue
+{
+public:
+    enum SpecialValsEnum
+    {
+        Zero,
+        Inf,
+        NaN,
+        Min,
+        Max,
+        One,
+    };
+
+    SpecialValue() = default;
+    SpecialValue(SpecialValsEnum v)
+        : value(v)
+    {}
+    operator SpecialValsEnum() const
+    {
+        return value;
+    }
+    SpecialValue& operator=(SpecialValsEnum v)
+    {
+        value = v;
+        return *this;
+    }
+    bool operator==(const SpecialValsEnum v) const
+    {
+        return value == v;
+    }
+    bool operator!=(const SpecialValsEnum v) const
+    {
+        return value != v;
+    }
+    SpecialValue operator-()
+    {
+        negative = !negative;
+        return *this;
+    }
+
+    template <typename DataType>
+    DataType evaluate()
+    {
+        switch (value)
+        {
+            case Zero:
+                return static_cast<DataType>(negative ? -0.0 : 0.0);
+            case Inf:
+                return negative ? -std::numeric_limits<DataType>::infinity()
+                                : std::numeric_limits<DataType>::infinity();
+            case NaN:
+                return std::numeric_limits<DataType>::quiet_NaN();
+            case Min:
+                return negative ? -std::numeric_limits<DataType>::min() : std::numeric_limits<DataType>::min();
+            case Max:
+                return negative ? -std::numeric_limits<DataType>::max() : std::numeric_limits<DataType>::max();
+            case One:
+                return static_cast<DataType>(negative ? -1.0 : 1.0);
+            default:
+                WARNING("[Generator][FS] Uninitialised special value.");
+                return static_cast<DataType>(0.0);
+        }
+    }
+
+private:
+    SpecialValsEnum value;
+    bool negative = false;
+};
+
+/*
+Test vals format
+
+I: Number of inputs to an op - referenced by cfg.inputPos
+T: Number of test cases defined for the op
+
+vector of test inputs: {
+    vector of values for test 0:   { valueForinputPos0, valueForinputPos1, ..., valueForinputPosI-1 },
+    vector of values for test 1:   { valueForinputPos0, valueForinputPos1, ..., valueForinputPosI-1 },
+    ...
+    vector of values for test T-1: { valueForinputPos0, valueForinputPos1, ..., valueForinputPosI-1 },
+}
+*/
+using TestValues = std::vector<std::vector<SpecialValue>>;
+
+TestValues equalOpsTestVals{ { SpecialValue(SpecialValue::Zero), -SpecialValue(SpecialValue::Zero) },
+                             { SpecialValue(SpecialValue::Inf), -SpecialValue(SpecialValue::Inf) } };
+
+TestValues addTestVals{ { SpecialValue(SpecialValue::Max), SpecialValue(SpecialValue::One) },
+                        { SpecialValue(SpecialValue::Inf), -SpecialValue(SpecialValue::Inf) } };
+
+TestValues defaultTestVals{ { SpecialValue(SpecialValue::Zero) }, { -SpecialValue(SpecialValue::Zero) },
+                            { SpecialValue(SpecialValue::Inf) },  { -SpecialValue(SpecialValue::Inf) },
+                            { SpecialValue(SpecialValue::NaN) },  { SpecialValue(SpecialValue::Min) },
+                            { SpecialValue(SpecialValue::Max) } };
+
+std::map<Op, TestValues> testValues = { { Op::Op_EQUAL, equalOpsTestVals },
+                                        { Op::Op_GREATER, equalOpsTestVals },
+                                        { Op::Op_GREATER_EQUAL, equalOpsTestVals },
+                                        { Op::Op_ADD, addTestVals } };
+
+template <typename DataType>
+bool generate(const TosaReference::GenerateConfig& cfg, DataType* data, size_t size)
+{
+    const TosaReference::FpSpecialInfo& fsinfo = cfg.fpSpecialInfo;
+    uint8_t startIndex                         = fsinfo.startIndex;
+
+    std::vector<DataType> values;
+    auto testValuesResult = testValues.find(cfg.opType);
+    TestValues opTestVals = defaultTestVals;
+    size_t inputIndex     = 0;
+    if (testValuesResult != testValues.end())
+    {
+        // When an op has an entry in testValues we use its op specific special test values, otherwise default values are used
+        opTestVals = testValuesResult->second;
+        inputIndex = cfg.inputPos;
+    }
+
+    for (std::vector<SpecialValue> inputs : opTestVals)
+    {
+        values.push_back(inputs[inputIndex].evaluate<DataType>());
+    }
+
+    const auto T = TosaReference::numElementsFromShape(cfg.shape);
+    for (auto t = 0; t < T; ++t)
+    {
+        data[t] = values[(t + startIndex) % values.size()];
+    }
+    return true;
+}
+}    // namespace
+
+namespace TosaReference
+{
+bool generateFpSpecial(const GenerateConfig& cfg, void* data, size_t size)
+{
+    // Check we support the operator
+    if (cfg.opType == Op::Op_UNKNOWN)
+    {
+        WARNING("[Generator][FS] Unknown operator.");
+        return false;
+    }
+
+    switch (cfg.dataType)
+    {
+        case DType::DType_FP16: {
+            half_float::half* outData = reinterpret_cast<half_float::half*>(data);
+            return generate(cfg, outData, size);
+        }
+        case DType::DType_FP32: {
+            float* outData = reinterpret_cast<float*>(data);
+            return generate(cfg, outData, size);
+        }
+        default:
+            WARNING("[Generator][FS] Unsupported type.");
+            return false;
+    }
+}
+}    // namespace TosaReference
\ No newline at end of file
diff --git a/reference_model/src/generate/generate_fp_special.h b/reference_model/src/generate/generate_fp_special.h
new file mode 100644
index 0000000..9a7315b
--- /dev/null
+++ b/reference_model/src/generate/generate_fp_special.h
@@ -0,0 +1,34 @@
+// Copyright (c) 2024, ARM Limited.
+//
+//    Licensed under the Apache License, Version 2.0 (the "License");
+//    you may not use this file except in compliance with the License.
+//    You may obtain a copy of the License at
+//
+//         http://www.apache.org/licenses/LICENSE-2.0
+//
+//    Unless required by applicable law or agreed to in writing, software
+//    distributed under the License is distributed on an "AS IS" BASIS,
+//    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+//    See the License for the specific language governing permissions and
+//    limitations under the License.
+
+#ifndef GENERATE_FP_SPECIAL_H_
+#define GENERATE_FP_SPECIAL_H_
+
+#include "generate_utils.h"
+
+namespace TosaReference
+{
+
+/// \brief Perform FP special data generation
+///
+/// \param cfg Generator related meta-data
+/// \param data Buffer to generate the data to
+/// \param size Size of the buffer
+///
+/// \return True on successful generation
+bool generateFpSpecial(const GenerateConfig& cfg, void* data, size_t size);
+
+};    // namespace TosaReference
+
+#endif    // GENERATE_FP_SPECIAL_H_
\ No newline at end of file
diff --git a/reference_model/src/generate/generate_utils.cc b/reference_model/src/generate/generate_utils.cc
index f31b443..d62c247 100644
--- a/reference_model/src/generate/generate_utils.cc
+++ b/reference_model/src/generate/generate_utils.cc
@@ -107,7 +107,7 @@ NLOHMANN_JSON_SERIALIZE_ENUM(GeneratorType,
                                  { GeneratorType::DotProduct, "DOT_PRODUCT" },
                                  { GeneratorType::FullRange, "FULL_RANGE" },
                                  { GeneratorType::Boundary, "BOUNDARY" },
-                                 { GeneratorType::Special, "SPECIAL" },
+                                 { GeneratorType::FpSpecial, "FP_SPECIAL" },
                                  { GeneratorType::FixedData, "FIXED_DATA" },
                              })
 
@@ -159,6 +159,14 @@ void from_json(const nlohmann::json& j, FullRangeInfo& fullRangeInfo)
     }
 }
 
+void from_json(const nlohmann::json& j, FpSpecialInfo& fpSpecialInfo)
+{
+    if (j.contains("start_idx"))
+    {
+        j.at("start_idx").get_to(fpSpecialInfo.startIndex);
+    }
+}
+
 void from_json(const nlohmann::json& j, GenerateConfig& cfg)
 {
     j.at("data_type").get_to(cfg.dataType);
@@ -201,6 +209,13 @@ void from_json(const nlohmann::json& j, GenerateConfig& cfg)
     {
         j.at("full_range_info").get_to(cfg.fullRangeInfo);
     }
+
+    //Set up defaults for fpSpecialInfo
+    cfg.fpSpecialInfo.startIndex = 0;
+    if (j.contains("fp_special_info"))
+    {
+        j.at("fp_special_info").get_to(cfg.fpSpecialInfo);
+    }
 }
 
 std::optional<GenerateConfig> parseGenerateConfig(const char* json, const char* tensorName)
diff --git a/reference_model/src/generate/generate_utils.h b/reference_model/src/generate/generate_utils.h
index 8ce9b0e..0428fd8 100644
--- a/reference_model/src/generate/generate_utils.h
+++ b/reference_model/src/generate/generate_utils.h
@@ -33,7 +33,7 @@ enum class GeneratorType
     DotProduct,
     FullRange,
     Boundary,
-    Special,
+    FpSpecial,
     FixedData,
 };
 
@@ -82,6 +82,14 @@ struct FullRangeInfo
     uint16_t startVal;
 };
 
+/// \brief Op specific generator meta-data
+struct FpSpecialInfo
+{
+    FpSpecialInfo() = default;
+
+    uint8_t startIndex;
+};
+
 /// \brief Generator configuration
 struct GenerateConfig
 {
@@ -95,6 +103,7 @@ struct GenerateConfig
     PseudoRandomInfo pseudoRandomInfo;
     FixedDataInfo fixedDataInfo;
     FullRangeInfo fullRangeInfo;
+    FpSpecialInfo fpSpecialInfo;
 };
 
 /// \brief Parse the generator config when given in JSON form
diff --git a/reference_model/src/verify/verifiers.h b/reference_model/src/verify/verifiers.h
index e5f9df1..48c971d 100644
--- a/reference_model/src/verify/verifiers.h
+++ b/reference_model/src/verify/verifiers.h
@@ -85,6 +85,14 @@ bool verifyRelative(const CTensor* referenceTensor,
                     const CTensor* implementationTensor,
                     const RelativeVerifyInfo& rInfo);
 
+/// \brief Perform FP special tests verification
+///
+/// \param referenceTensor      Reference tensor
+/// \param implementationTensor Implementation resulting tensor
+///
+/// \return True if compliant else false
+bool verifyFpSpecial(const CTensor* referenceTensor, const CTensor* implementationTensor);
+
 };    // namespace TosaReference
 
 #endif    // VERIFIERS_H_
diff --git a/reference_model/src/verify/verify_entry.cc b/reference_model/src/verify/verify_entry.cc
index 9702c36..a04c62e 100644
--- a/reference_model/src/verify/verify_entry.cc
+++ b/reference_model/src/verify/verify_entry.cc
@@ -46,6 +46,9 @@ bool verify(const CTensor* ref, const CTensor* refBnd, const CTensor* imp, const
         case VerifyMode::Relative: {
             return verifyRelative(ref, imp, cfg.relativeInfo);
         }
+        case VerifyMode::FpSpecial: {
+            return verifyFpSpecial(ref, imp);
+        }
         default: {
             WARNING("[Verifier] Unsupported verification mode.");
             break;
diff --git a/reference_model/src/verify/verify_fp_special.cc b/reference_model/src/verify/verify_fp_special.cc
new file mode 100644
index 0000000..94550fd
--- /dev/null
+++ b/reference_model/src/verify/verify_fp_special.cc
@@ -0,0 +1,106 @@
+// Copyright (c) 2024, ARM Limited.
+//
+//    Licensed under the Apache License, Version 2.0 (the "License");
+//    you may not use this file except in compliance with the License.
+//    You may obtain a copy of the License at
+//
+//         http://www.apache.org/licenses/LICENSE-2.0
+//
+//    Unless required by applicable law or agreed to in writing, software
+//    distributed under the License is distributed on an "AS IS" BASIS,
+//    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+//    See the License for the specific language governing permissions and
+//    limitations under the License.
+
+#include "half.hpp"
+#include "verifiers.h"
+
+namespace
+{
+template <typename Datatype>
+bool compliant(const double& referenceValue, const Datatype& implementationValue)
+{
+    // Compliant when values are both nans OR when values are both finite/infinite and have the same sign
+    return (std::isnan(referenceValue) && std::isnan(implementationValue)) ||
+           (std::isnan(referenceValue) == std::isnan(implementationValue) &&
+            std::isfinite(referenceValue) == std::isfinite(implementationValue) &&
+            std::signbit(referenceValue) == std::signbit(implementationValue));
+}
+
+template <>
+bool compliant(const double& referenceValue, const half_float::half& implementationValue)
+{
+    // Compliant when values are both nans OR when values are both finite/infinite and have the same sign
+    return (std::isnan(referenceValue) && half_float::isnan(implementationValue)) ||
+           (std::isnan(referenceValue) == half_float::isnan(implementationValue) &&
+            std::isfinite(referenceValue) == half_float::isfinite(implementationValue) &&
+            std::signbit(referenceValue) == half_float::signbit(implementationValue));
+}
+
+template <typename Datatype>
+bool verify(const double* refData,
+            const double* refDataEnd,
+            Datatype* impData,
+            const int64_t elementCount,
+            const std::vector<int32_t> refShape)
+{
+    auto pair = std::mismatch(refData, refDataEnd, impData, std::next(impData, elementCount),
+                              [](const double& referenceValue, const Datatype& implementationValue) {
+                                  return compliant(referenceValue, implementationValue);
+                              });
+
+    if (std::get<0>(pair) == refDataEnd)
+    {
+        // No mismatch found
+        return true;
+    }
+    else
+    {
+        auto pos = TosaReference::indexToPosition(std::get<0>(pair) - refData, refShape);
+        WARNING("[Verfier][FS] Location %s", TosaReference::positionToString(pos).c_str());
+        return false;
+    }
+}
+}    // namespace
+
+namespace TosaReference
+{
+
+bool verifyFpSpecial(const CTensor* referenceTensor, const CTensor* implementationTensor)
+{
+    // Validate that tensors are provided
+    TOSA_REF_REQUIRE(referenceTensor != nullptr, "[FS] Reference tensor is missing");
+    TOSA_REF_REQUIRE(implementationTensor != nullptr, "[FS] Implementation tensor is missing");
+
+    // Get number of elements
+    const std::vector<int32_t> refShape(referenceTensor->shape, referenceTensor->shape + referenceTensor->num_dims);
+    const auto elementCount = numElements(refShape);
+    TOSA_REF_REQUIRE(elementCount > 0, "[FS] Invalid shape for reference tensor");
+
+    TOSA_REF_REQUIRE(referenceTensor->data_type == tosa_datatype_fp64_t, "[FS] Reference tensor is not fp64");
+    const auto* refData = reinterpret_cast<const double*>(referenceTensor->data);
+    TOSA_REF_REQUIRE(refData != nullptr, "[FS] Missing data for reference");
+    const auto* refDataEnd = std::next(refData, elementCount);
+
+    switch (implementationTensor->data_type)
+    {
+        case tosa_datatype_fp32_t: {
+            const auto* impData = reinterpret_cast<const float*>(implementationTensor->data);
+            TOSA_REF_REQUIRE(impData != nullptr, "[FS] Missing data for implementation");
+
+            return verify(refData, refDataEnd, impData, elementCount, refShape);
+        }
+        case tosa_datatype_fp16_t: {
+            const auto* impData = reinterpret_cast<const half_float::half*>(implementationTensor->data);
+            TOSA_REF_REQUIRE(impData != nullptr, "[FS] Missing data for implementation");
+
+            return verify(refData, refDataEnd, impData, elementCount, refShape);
+        }
+        default:
+            WARNING("[Verifier][FS] Data-type not supported.");
+            break;
+    }
+
+    return false;
+}
+}    // namespace TosaReference
diff --git a/reference_model/test/generate_tests.cpp b/reference_model/test/generate_tests.cpp
index 73db631..19b808b 100644
--- a/reference_model/test/generate_tests.cpp
+++ b/reference_model/test/generate_tests.cpp
@@ -1609,4 +1609,101 @@ TEST_CASE("positive - FP16 full range")
     }
 }
 
+void fp_special_test_FP32(const std::string tosaName,
+                          const size_t tosaElements,
+                          const std::string templateJsonCfg,
+                          const std::string opStr,
+                          const std::string startIndexStr,
+                          const std::vector<uint32_t> expected)
+{
+    std::string jsonCfg = templateJsonCfg;
+    update_json_template(jsonCfg, "_OP_", opStr);
+    update_json_template(jsonCfg, "_START_", startIndexStr);
+
+    std::vector<float> buffer(tosaElements);
+    REQUIRE(tgd_generate_data(jsonCfg.c_str(), tosaName.c_str(), (void*)buffer.data(), tosaElements * 4));
+    check_output<float>(buffer, expected);
+}
+
+TEST_CASE("positive - FP32 FP Special")
+{
+    std::string templateJsonCfg = R"({
+        "tensors" : {
+            "input0" : {
+                "generator": "FP_SPECIAL",
+                "data_type": "FP32",
+                "input_type": "VARIABLE",
+                "shape" : [ 5, 6, 7 ],
+                "input_pos": 0,
+                "op" : "_OP_",
+                "fp_special_info": {
+                    "start_idx": _START_
+                }
+            },
+            "input1" : {
+                "generator": "FP_SPECIAL",
+                "data_type": "FP32",
+                "input_type": "VARIABLE",
+                "shape" : [ 5, 6, 7 ],
+                "input_pos": 1,
+                "op" : "_OP_",
+                "fp_special_info": {
+                    "start_idx": _START_
+                }
+            }
+        }
+    })";
+
+    const std::string tosaName0 = "input0";
+    const std::string tosaName1 = "input1";
+    const size_t tosaElements   = 5 * 6 * 7;
+
+    SUBCASE("equal, input 0")
+    {
+        std::vector<uint32_t> expected = { 0x0, 0x7F800000, 0x0 };
+        fp_special_test_FP32(tosaName0, tosaElements, templateJsonCfg, "EQUAL", "0", expected);
+    }
+    SUBCASE("equal, input 1")
+    {
+        std::vector<uint32_t> expected = { 0x80000000, 0xFF800000, 0x80000000 };
+        fp_special_test_FP32(tosaName1, tosaElements, templateJsonCfg, "EQUAL", "0", expected);
+    }
+    SUBCASE("greater, input 0")
+    {
+        std::vector<uint32_t> expected = { 0x0, 0x7F800000, 0x0 };
+        fp_special_test_FP32(tosaName0, tosaElements, templateJsonCfg, "GREATER", "0", expected);
+    }
+    SUBCASE("greater, input 1")
+    {
+        std::vector<uint32_t> expected = { 0x80000000, 0xFF800000, 0x80000000 };
+        fp_special_test_FP32(tosaName1, tosaElements, templateJsonCfg, "GREATER", "0", expected);
+    }
+    SUBCASE("add, input 0")
+    {
+        std::vector<uint32_t> expected = { 0x7F7FFFFF, 0x7F800000, 0x7F7FFFFF };
+        fp_special_test_FP32(tosaName0, tosaElements, templateJsonCfg, "ADD", "0", expected);
+    }
+    SUBCASE("add, input 1")
+    {
+        std::vector<uint32_t> expected = { 0x3F800000, 0xFF800000, 0x3F800000 };
+        fp_special_test_FP32(tosaName1, tosaElements, templateJsonCfg, "ADD", "0", expected);
+    }
+    SUBCASE("maximum, input 0")
+    {
+        std::vector<uint32_t> expected = { 0x0, 0x80000000, 0x7F800000 };
+        fp_special_test_FP32(tosaName0, tosaElements, templateJsonCfg, "MAXIMUM", "0", expected);
+    }
+    SUBCASE("maximum, input 1")
+    {
+        std::vector<uint32_t> expected = { 0x0, 0x80000000, 0x7F800000 };
+        fp_special_test_FP32(tosaName1, tosaElements, templateJsonCfg, "MAXIMUM", "0", expected);
+    }
+    SUBCASE("maximum, startIndex 100")
+    {
+        // A startIndex of 100 creates an offset of 2 in the MAXIMUM op's test data (size: 7) 100 % 7 = 2
+        std::vector<uint32_t> expected = { 0x7F800000, 0xFF800000, 0x7FC00000 };
+        fp_special_test_FP32(tosaName0, tosaElements, templateJsonCfg, "MAXIMUM", "100", expected);
+    }
+}
+
 TEST_SUITE_END();    // generate
diff --git a/scripts/schemavalidation/datagen-config.schema.json b/scripts/schemavalidation/datagen-config.schema.json
index 19e8b62..6046873 100644
--- a/scripts/schemavalidation/datagen-config.schema.json
+++ b/scripts/schemavalidation/datagen-config.schema.json
@@ -22,7 +22,7 @@
                     "type": "object",
                     "properties": {
                         "generator": {
-                            "description": "data generator name - PSEUDO_RANDOM, DOT_PRODUCT, FULL_RANGE, BOUNDARY, or SPECIAL",
+                            "description": "data generator name - PSEUDO_RANDOM, DOT_PRODUCT, FULL_RANGE, BOUNDARY, or FP_SPECIAL",
                             "type": "string"
                         },
                         "data_type": {
@@ -134,6 +134,20 @@
                             "additionalProperties": false,
                             "required": [ ]
                         },
+                        "fp_special_info": {
+                            "description": "info required for the FP_SPECIAL generator",
+                            "type": "object",
+                            "properties":
+                            {
+                                "start_idx": {
+                                    "description": "starting index for the test data",
+                                    "type": "integer",
+                                    "minimum": 0
+                                }
+                            },
+                            "additionalProperties": false,
+                            "required": [ ]
+                        },
                         "fixed_data_info": {
                             "description": "info required for FIXED_DATA generator",
                             "type": "object",
diff --git a/verif/conformance/test_select.py b/verif/conformance/test_select.py
index e3a8ffb..e3f1738 100644
--- a/verif/conformance/test_select.py
+++ b/verif/conformance/test_select.py
@@ -259,9 +259,9 @@ class Operator:
                     negative and "ERRORIF" in str(path)
                 ):
                     # Check for test set paths
-                    match = re.match(r"(.*)_(s[0-9]+|full)", path.name)
+                    match = re.match(r"(.*)_(s[0-9]+|full|fs)", path.name)
                     if match:
-                        if match.group(2) in ["s0", "full"]:
+                        if match.group(2) in ["s0", "full", "fs"]:
                             # Only return the truncated test name
                             # of the first test of a set, and for full tests
                             yield path.with_name(match.group(1))
@@ -317,7 +317,7 @@ class Operator:
     def _get_extra_test_paths(path):
         """Expand a path to find extra tests."""
         paths = []
-        for suffix in ["full"]:
+        for suffix in ["full", "fs"]:
             suffix_path = path.with_name(f"{path.name}_{suffix}")
             if suffix_path.exists():
                 paths.append(suffix_path)
diff --git a/verif/generator/tosa_arg_gen.py b/verif/generator/tosa_arg_gen.py
index 8d6c8d7..5957a33 100644
--- a/verif/generator/tosa_arg_gen.py
+++ b/verif/generator/tosa_arg_gen.py
@@ -264,6 +264,9 @@ class TosaTensorGen:
             return [[]] * num_shapes
 
         shape = testGen.makeShape(rng, rank)
+        # Do not broadcast for some tests
+        if error_name is None and rng.randInt(high=100) < 10:
+            return [shape] * num_shapes
         shape_list = []
 
         # Choose any one of the inputs to broadcast
@@ -785,6 +788,10 @@ class TosaTensorValuesGen:
             "tensors": {},
         }
         dg_tens_meta = tens_data["tensors"]
+
+        fp_special_info = {}
+        fp_special_info["start_idx"] = int(rng.randInt())
+
         for idx, shape in enumerate(shapeList):
 
             tens_meta = {}
@@ -858,6 +865,8 @@ class TosaTensorValuesGen:
                     rng.randInt(0, gtu.DTYPE_ATTRIBUTES[dtypeList[idx]]["fullset"])
                 )
                 tens_meta["full_range_info"] = info
+            elif dg_type == gtu.DataGenType.FP_SPECIAL:
+                tens_meta["fp_special_info"] = fp_special_info
             else:
                 # TODO - other data gen type
                 assert False, "TODO: support other data gen types"
@@ -1862,16 +1871,12 @@ class TosaArgGen:
         for dg_type in dataGenTypesList:
             for arg_str, args_dict in arg_list:
                 gen_args_dict = args_dict.copy()
+                # Only create one test by default - no sets of tests
+                num_test_sets = 0
+
                 if dg_type == gtu.DataGenType.PSEUDO_RANDOM:
                     if error_name is None:
-                        num_test_sets = (
-                            args_dict["num_test_sets"]
-                            if "num_test_sets" in args_dict
-                            else 0
-                        )
-                    else:
-                        # Add single test for pseudo random
-                        num_test_sets = 0
+                        num_test_sets = args_dict.get("num_test_sets", 0)
 
                 elif dg_type == gtu.DataGenType.DOT_PRODUCT:
                     # Extra tests for each dot product test set
@@ -1900,13 +1905,23 @@ class TosaArgGen:
                             f"Skipping {opName}{shape_info} as tensor data size too small for full range of values {tensor_size} < {gtu.DTYPE_ATTRIBUTES[dtype]['fullset']}"
                         )
                         continue
-                    # Large enough tensor data size for full range, add a single test
-                    num_test_sets = 0
+                    # Large enough tensor data size for full range, add full test
                     arg_str = f"{arg_str}_full" if arg_str else "full"
                     gen_args_dict["tags"] = args_dict.get("tags", []) + [
                         "non_finite_fp_data"
                     ]
 
+                elif dg_type == gtu.DataGenType.FP_SPECIAL:
+                    shapes_set = {tuple(x) for x in shapeList}
+                    if len(shapes_set) != 1:
+                        logger.info(
+                            f"Changing {opName} input shapes {shapes_set} - broadcasting incompatable with special test"
+                        )
+                        shapeList = [np.int32(np.broadcast_shapes(*shapeList))] * len(
+                            shapeList
+                        )
+                    arg_str = f"{arg_str}_fs" if arg_str else "fs"
+
                 gen_args_dict["dg_type"] = dg_type
                 if num_test_sets > 0:
                     for s in range(0, num_test_sets):
diff --git a/verif/generator/tosa_test_gen.py b/verif/generator/tosa_test_gen.py
index 38ab3f4..40788a2 100644
--- a/verif/generator/tosa_test_gen.py
+++ b/verif/generator/tosa_test_gen.py
@@ -268,7 +268,7 @@ class TosaTestGen:
                 if "ksb" in argsDict
                 else int(argsDict["ks"]),
             }
-        elif argsDict["dg_type"] == gtu.DataGenType.SPECIAL:
+        elif argsDict["dg_type"] == gtu.DataGenType.FP_SPECIAL:
             mode = gtu.ComplianceMode.FP_SPECIAL
         elif "compliance" in op and "ulp" in op["compliance"]:
             mode = gtu.ComplianceMode.ULP
@@ -3352,7 +3352,11 @@ class TosaTestGen:
         DType.FP32: (gtu.DataGenType.DOT_PRODUCT,),
     }
     EW_UNARY_DATAGEN = {
-        DType.FP16: (gtu.DataGenType.PSEUDO_RANDOM, gtu.DataGenType.FULL_RANGE)
+        DType.FP16: (gtu.DataGenType.PSEUDO_RANDOM, gtu.DataGenType.FULL_RANGE),
+    }
+    PR_FS_DATAGEN = {
+        DType.FP16: (gtu.DataGenType.PSEUDO_RANDOM, gtu.DataGenType.FP_SPECIAL),
+        DType.FP32: (gtu.DataGenType.PSEUDO_RANDOM, gtu.DataGenType.FP_SPECIAL),
     }
 
     TOSA_OP_LIST = {
@@ -3716,7 +3720,7 @@ class TosaTestGen:
                 TosaErrorValidator.evDimensionMismatch,
                 TosaErrorValidator.evBroadcastShapesMismatch,
             ),
-            "data_gen": PSEUDO_RANDOM_DATAGEN,
+            "data_gen": PR_FS_DATAGEN,
             "compliance": {"ulp": 0.5},
         },
         "arithmetic_right_shift": {
@@ -3938,7 +3942,7 @@ class TosaTestGen:
                 TosaErrorValidator.evDimensionMismatch,
                 TosaErrorValidator.evBroadcastShapesMismatch,
             ),
-            "data_gen": PSEUDO_RANDOM_DATAGEN,
+            "data_gen": PR_FS_DATAGEN,
         },
         "minimum": {
             "op": Op.MINIMUM,
@@ -4330,7 +4334,7 @@ class TosaTestGen:
                 TosaErrorValidator.evDimensionMismatch,
                 TosaErrorValidator.evBroadcastShapesMismatch,
             ),
-            "data_gen": PSEUDO_RANDOM_DATAGEN,
+            "data_gen": PR_FS_DATAGEN,
         },
         "greater_equal": {
             "op": Op.GREATER_EQUAL,
@@ -4351,7 +4355,7 @@ class TosaTestGen:
                 TosaErrorValidator.evDimensionMismatch,
                 TosaErrorValidator.evBroadcastShapesMismatch,
             ),
-            "data_gen": PSEUDO_RANDOM_DATAGEN,
+            "data_gen": PR_FS_DATAGEN,
         },
         "greater": {
             "op": Op.GREATER,
@@ -4372,7 +4376,7 @@ class TosaTestGen:
                 TosaErrorValidator.evDimensionMismatch,
                 TosaErrorValidator.evBroadcastShapesMismatch,
             ),
-            "data_gen": PSEUDO_RANDOM_DATAGEN,
+            "data_gen": PR_FS_DATAGEN,
         },
         # Reduction operators
         "reduce_all": {
diff --git a/verif/generator/tosa_utils.py b/verif/generator/tosa_utils.py
index a8e321e..478190d 100644
--- a/verif/generator/tosa_utils.py
+++ b/verif/generator/tosa_utils.py
@@ -55,7 +55,7 @@ class DataGenType(IntEnum):
     DOT_PRODUCT = 1
     BOUNDARY = 2
     FULL_RANGE = 3
-    SPECIAL = 4
+    FP_SPECIAL = 4
     FIXED_DATA = 5
 
 
diff --git a/verif/tests/test_tosa_refmodel.py b/verif/tests/test_tosa_refmodel.py
index 24ee9e2..bb52a86 100644
--- a/verif/tests/test_tosa_refmodel.py
+++ b/verif/tests/test_tosa_refmodel.py
@@ -1,5 +1,5 @@
 """Tests for tosa_reference_model."""
-# Copyright (c) 2022-2023, ARM Limited.
+# Copyright (c) 2022-2024, ARM Limited.
 # SPDX-License-Identifier: Apache-2.0
 import json
 import re
@@ -134,9 +134,10 @@ class BuildTosaTest:
 
 
 # Tests - op_name, ref_model_type, num_expected_tests
+# FP Special datagen adds a second expected test to FP16 and FP32 tests for OPs it is added to
 TEST_PARAMS = [
     ("add", "int32", 1),
-    ("add", "fp32", 1),
+    ("add", "fp32", 2),
     ("abs", "int32", 1),
     ("abs", "fp32", 1),
     ("abs", "fp16", 1),
@@ -223,13 +224,20 @@ def test_refmodel_simple_op(tosaTest):
             assert const_file.is_file()
             consts.append(np.load(str(const_file)))
 
+        # Check if the data is from FP special datagen which can give invalid results
+        fp_special_data = test_dir.match("*_fs")
+
         # Perform Numpy operation
         if op_name == "abs":
             assert len(tensors) == 1
             result = np.abs(tensors[0])
         elif op_name == "add":
             assert len(tensors) == 2
-            result = np.add(tensors[0], tensors[1])
+            if fp_special_data:
+                with np.errstate(invalid="ignore"):
+                    result = np.add(tensors[0], tensors[1])
+            else:
+                result = np.add(tensors[0], tensors[1])
         elif op_name == "concat":
             assert len(consts) == 1
             # Get axis from test directory name