TheRock/rocm-libraries/projects/hipblaslt/tensilelite/tests/DataInit_test.cpp

// Copyright Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT

#include <gtest/gtest.h>
#include "DataInitialization.hpp"             // isMXTensor / Problem
#include "DataInitializationHelpers.hpp"    // detail::* (MX-only, internally guarded)
#include <Tensile/ContractionProblem.hpp>
#include <Tensile/DataTypes.hpp>
#include <Tensile/TensorDescriptor.hpp>
#include <Tensile/Utils.hpp>
#include <algorithm>
#include <cmath>
#include <cstdint>
#include <cstring>
#include <filesystem>
#include <fstream>
#include <limits>
#include <sstream>
#include <stdexcept>
#include <string>
#include <vector>
#if HIPBLASLT_ENABLE_MXDATAGENERATOR
#include <hip/hip_runtime.h>
#include <mxDataGenerator/dataTypeInfo.hpp>
#include <mxDataGenerator/ocp_e2m1_mxfp4.hpp>
#include <mxDataGenerator/ocp_e2m3_mxfp6.hpp>
#include <mxDataGenerator/ocp_e3m2_mxfp6.hpp>
#include <mxDataGenerator/ocp_e4m3_mxfp8.hpp>
#include <mxDataGenerator/ocp_e5m2_mxfp8.hpp>
#endif

using TensileLite::ContractionProblemGemm;
using TensileLite::DataTypeInfo;
using TensileLite::TensorDescriptor;
using TensileLite::Client::isMXProblem;
using TensileLite::Client::isMXTensor;

// Shorthand for the production helper namespace under test (MX builds only).
#if HIPBLASLT_ENABLE_MXDATAGENERATOR
namespace dt = TensileLite::Client::detail;
#endif
namespace
{
    // -----------------------------------------------------------------------
    // Helper: build a ContractionProblemGemm with the requested A/B dtypes.
    // Mirrors tests/MXScalePadding_test.cpp::makeMXProblem so the geometry
    // matches what the real client produces, then enables MX scaling on each
    // side independently. mxBlock==0 means "do NOT call setMXScale*", so the
    // problem's mxBlockA() / mxBlockB() stays 0 and isMXTensor returns
    // false on that side. This is exactly the lever needed to drive every
    // branch of isMXProblem.
    // -----------------------------------------------------------------------
    ContractionProblemGemm makeProblem(rocisa::DataType aType,
                                       rocisa::DataType bType,
                                       int              mxBlockA,
                                       int              mxBlockB,
                                       size_t           M = 128,
                                       size_t           N = 128,
                                       size_t           K = 256,
                                       size_t           batch  = 1,
                                       bool             transA = true,
                                       bool             transB = false)
    {
        auto problem = ContractionProblemGemm::GEMM_Strides(
            transA, transB,
            aType, bType,
            rocisa::DataType::BFloat16, rocisa::DataType::BFloat16,
            M, N, K, batch,
            transA ? K : M,                 // lda
            transA ? K * M : M * K,         // strideA
            transB ? N : K,                 // ldb
            transB ? N * K : K * N,         // strideB
            M, M * N,                       // ldc, strideC
            M, M * N,                       // ldd, strideD
            0.0);                           // beta
        if(mxBlockA > 0) problem.setMXScaleA(rocisa::DataType::E8, mxBlockA);
        if(mxBlockB > 0) problem.setMXScaleB(rocisa::DataType::E8, mxBlockB);
        return problem;
    }
} // namespace

// =============================================================================
//   Section 1 - TensileLite::Client::isMXTensor
//
//       bool isMXTensor(t, mxBlock) {
//           if(mxBlock == 0) return false;            // (a) short-circuit
//           return dt in {Float4, Float6, BFloat6, Float8, BFloat8}; // (b) dtype gate
//       }
// =============================================================================
struct TensorParam
{
    rocisa::DataType dtype;
    size_t           mxBlock;
    bool             expected;
    char const*      name;
};
class IsMXTensorTest : public ::testing::TestWithParam<TensorParam>
{
};
TEST_P(IsMXTensorTest, MatchesContract)
{
    auto const& p = GetParam();
    // 1x1 descriptor is enough; the helper only inspects .dataType().
    TensorDescriptor t("t", p.dtype, {1, 1}, {1, 1});
    EXPECT_EQ(isMXTensor(t, p.mxBlock), p.expected)
        << "case=" << p.name
        << " dtype=" << static_cast<int>(p.dtype)
        << " mxBlock=" << p.mxBlock;
}

INSTANTIATE_TEST_SUITE_P(
    MXFP4OrFP8Coverage,
    IsMXTensorTest,
    ::testing::Values(
        // ----- (a) mxBlock==0 must short-circuit even for MX dtypes --------
        TensorParam{rocisa::DataType::Float4,   0, false, "Float4_block0"},
        TensorParam{rocisa::DataType::Float6,   0, false, "Float6_block0"},
        TensorParam{rocisa::DataType::BFloat6,  0, false, "BFloat6_block0"},
        TensorParam{rocisa::DataType::Float8,   0, false, "Float8_block0"},
        TensorParam{rocisa::DataType::BFloat8,  0, false, "BFloat8_block0"},
        // ----- (b) supported MX dtypes with mxBlock>0 -> true --------------
        TensorParam{rocisa::DataType::Float4,  32, true,  "Float4_block32"},
        TensorParam{rocisa::DataType::Float6,  32, true,  "Float6_block32"},
        TensorParam{rocisa::DataType::BFloat6, 32, true,  "BFloat6_block32"},
        TensorParam{rocisa::DataType::Float8,  32, true,  "Float8_block32"},
        TensorParam{rocisa::DataType::BFloat8, 32, true,  "BFloat8_block32"},
        // ----- (b') unsupported dtypes with mxBlock>0 -> false -------------
        TensorParam{rocisa::DataType::Float,   32, false, "Float_block32"},
        TensorParam{rocisa::DataType::Half,    32, false, "Half_block32"},
        TensorParam{rocisa::DataType::BFloat16,32, false, "BFloat16_block32"},
        TensorParam{rocisa::DataType::Int8,    32, false, "Int8_block32"},
        TensorParam{rocisa::DataType::Int32,   32, false, "Int32_block32"},
        // ----- mxBlock not equal to 32 (any positive value works) ----------
        TensorParam{rocisa::DataType::Float8,    1, true, "Float8_block1"},
        TensorParam{rocisa::DataType::BFloat8, 128, true, "BFloat8_block128"}
    ),
    [](::testing::TestParamInfo<TensorParam> const& info) {
        return std::string(info.param.name);
    }
);

// =============================================================================
//   Section 2 - TensileLite::Client::isMXProblem
//
//   Contract:
//       isMXProblem(P)
//         = isMXTensor(P.a, P.mxBlockA)
//            || isMXTensor(P.b, P.mxBlockB)
// =============================================================================
TEST(IsMXProblem, BothFP4)
{
    auto p = makeProblem(rocisa::DataType::Float4, rocisa::DataType::Float4,
                         /*mxBlockA=*/32, /*mxBlockB=*/32);
    EXPECT_TRUE(isMXProblem(p));
}
TEST(IsMXProblem, BothFP6)
{
    auto p = makeProblem(rocisa::DataType::Float6, rocisa::DataType::Float6,
                         /*mxBlockA=*/32, /*mxBlockB=*/32);
    EXPECT_TRUE(isMXProblem(p));
}
TEST(IsMXProblem, BothBFloat6)
{
    auto p = makeProblem(rocisa::DataType::BFloat6, rocisa::DataType::BFloat6,
                         /*mxBlockA=*/32, /*mxBlockB=*/32);
    EXPECT_TRUE(isMXProblem(p));
}
TEST(IsMXProblem, BothFP8)
{
    auto p = makeProblem(rocisa::DataType::Float8, rocisa::DataType::Float8,
                         /*mxBlockA=*/32, /*mxBlockB=*/32);
    EXPECT_TRUE(isMXProblem(p));
}
TEST(IsMXProblem, BothBFloat8)
{
    auto p = makeProblem(rocisa::DataType::BFloat8, rocisa::DataType::BFloat8,
                         /*mxBlockA=*/32, /*mxBlockB=*/32);
    EXPECT_TRUE(isMXProblem(p));
}
TEST(IsMXProblem, MixedFP4AandFP8B)
{
    auto p = makeProblem(rocisa::DataType::Float4, rocisa::DataType::Float8,
                         /*mxBlockA=*/32, /*mxBlockB=*/32);
    EXPECT_TRUE(isMXProblem(p));
}
TEST(IsMXProblem, MixedBFloat8AandFP4B)
{
    auto p = makeProblem(rocisa::DataType::BFloat8, rocisa::DataType::Float4,
                         /*mxBlockA=*/32, /*mxBlockB=*/32);
    EXPECT_TRUE(isMXProblem(p));
}
TEST(IsMXProblem, OnlyA_isMX_BIsBF16)
{
    // First disjunct true, second disjunct short-circuits false (mxBlockB=0).
    auto p = makeProblem(rocisa::DataType::Float8, rocisa::DataType::BFloat16,
                         /*mxBlockA=*/32, /*mxBlockB=*/0);
    EXPECT_TRUE(isMXProblem(p));
}
TEST(IsMXProblem, OnlyB_isMX_AIsBF16)
{
    auto p = makeProblem(rocisa::DataType::BFloat16, rocisa::DataType::Float4,
                         /*mxBlockA=*/0, /*mxBlockB=*/32);
    EXPECT_TRUE(isMXProblem(p));
}
TEST(IsMXProblem, NeitherIsMX)
{
    auto p = makeProblem(rocisa::DataType::BFloat16, rocisa::DataType::BFloat16,
                         /*mxBlockA=*/0, /*mxBlockB=*/0);
    EXPECT_FALSE(isMXProblem(p));
}
TEST(IsMXProblem, FloatABIsFalse)
{
    auto p = makeProblem(rocisa::DataType::Float, rocisa::DataType::Float,
                         /*mxBlockA=*/0, /*mxBlockB=*/0);
    EXPECT_FALSE(isMXProblem(p));
}

// =============================================================================
//   Section 3 — Byte-stride formula
//
//   For FP8 / BFloat8 the OCP standard packs one element per byte. The
//   DataTypeInfo for these dtypes therefore reports elementSize == 1, and the
//   formula must be the identity on strides[2]. These tests pin BOTH facts:
//   if anyone ever changes elementSize for FP8, or breaks multiplyElementSize,
//   the failure surfaces here instead of as a silent multi-batch FP8 bug.
// =============================================================================
TEST(InitializeMXDataForFP4OrFP8_BatchStrideFormula, FP8_OneBytePerElement)
{
    auto const info = DataTypeInfo::Get(rocisa::DataType::Float8);
    ASSERT_EQ(info.elementSize, 1u)
        << "OCP E4M3 must pack 1 byte per element; if this assertion fires the "
           "patch 3/3 batch-stride formula needs to be revisited.";
    constexpr size_t kStrideElems = 12345; // arbitrary, prime-ish
    size_t const     bytes        = TensileLite::multiplyElementSize(
        kStrideElems, static_cast<float>(info.elementSize));
    EXPECT_EQ(bytes, kStrideElems);
}

TEST(InitializeMXDataForFP4OrFP8_BatchStrideFormula, BFloat8_OneBytePerElement)
{
    auto const info = DataTypeInfo::Get(rocisa::DataType::BFloat8);
    ASSERT_EQ(info.elementSize, 1u) << "OCP E5M2 must pack 1 byte per element.";
    constexpr size_t kStrideElems = 1u << 20; // 1 Mi elements
    size_t const     bytes        = TensileLite::multiplyElementSize(
        kStrideElems, static_cast<float>(info.elementSize));
    EXPECT_EQ(bytes, kStrideElems);
}

// =============================================================================
//   Section 4 — direct calls into TensileLite::Client::detail (MX builds only)
// =============================================================================
#if HIPBLASLT_ENABLE_MXDATAGENERATOR
// -----------------------------------------------------------------------------
// 4.1  detail::hipMxScaleTypeForDataGenerator
// -----------------------------------------------------------------------------
TEST(HipMxScaleTypeForDataGenerator, MapsFloat8ToHIP_R_8F_E4M3)
{
    EXPECT_EQ(dt::hipMxScaleTypeForDataGenerator(rocisa::DataType::Float8),
              HIP_R_8F_E4M3);
}
TEST(HipMxScaleTypeForDataGenerator, MapsE5M3ToHIP_R_8F_E5M3_EXT)
{
    EXPECT_EQ(dt::hipMxScaleTypeForDataGenerator(rocisa::DataType::E5M3),
              static_cast<hipDataType>(HIP_R_8F_E5M3_EXT));
}
TEST(HipMxScaleTypeForDataGenerator, MapsE8AndNoneToHIP_R_8F_UE8M0)
{
    EXPECT_EQ(dt::hipMxScaleTypeForDataGenerator(rocisa::DataType::E8),
              HIP_R_8F_UE8M0);
    EXPECT_EQ(dt::hipMxScaleTypeForDataGenerator(rocisa::DataType::None),
              HIP_R_8F_UE8M0);
}
TEST(HipMxScaleTypeForDataGenerator, ThrowsOnUnsupportedScaleType)
{
    EXPECT_THROW(dt::hipMxScaleTypeForDataGenerator(rocisa::DataType::Float4),
                 std::runtime_error);
    EXPECT_THROW(dt::hipMxScaleTypeForDataGenerator(rocisa::DataType::BFloat8),
                 std::runtime_error);
    EXPECT_THROW(dt::hipMxScaleTypeForDataGenerator(rocisa::DataType::Float),
                 std::runtime_error);
}

// -----------------------------------------------------------------------------
// 4.2  detail::hipMxDataTypeForDataGenerator
// -----------------------------------------------------------------------------
TEST(HipMxDataTypeForDataGenerator, MapsFloat4ToHIP_R_4F_E2M1)
{
    EXPECT_EQ(dt::hipMxDataTypeForDataGenerator(rocisa::DataType::Float4),
              static_cast<hipDataType>(HIP_R_4F_E2M1));
}
TEST(HipMxDataTypeForDataGenerator, MapsFloat8ToHIP_R_8F_E4M3)
{
    EXPECT_EQ(dt::hipMxDataTypeForDataGenerator(rocisa::DataType::Float8),
              HIP_R_8F_E4M3);
}
TEST(HipMxDataTypeForDataGenerator, MapsBFloat8ToHIP_R_8F_E5M2)
{
    EXPECT_EQ(dt::hipMxDataTypeForDataGenerator(rocisa::DataType::BFloat8),
              HIP_R_8F_E5M2);
}
TEST(HipMxDataTypeForDataGenerator, MapsFloat6ToHIP_R_6F_E2M3)
{
    EXPECT_EQ(dt::hipMxDataTypeForDataGenerator(rocisa::DataType::Float6),
              static_cast<hipDataType>(HIP_R_6F_E2M3));
}
TEST(HipMxDataTypeForDataGenerator, MapsBFloat6ToHIP_R_6F_E3M2)
{
    EXPECT_EQ(dt::hipMxDataTypeForDataGenerator(rocisa::DataType::BFloat6),
              static_cast<hipDataType>(HIP_R_6F_E3M2));
}
TEST(HipMxDataTypeForDataGenerator, ThrowsOnUnsupportedDataType)
{
    EXPECT_THROW(dt::hipMxDataTypeForDataGenerator(rocisa::DataType::Float),
                 std::runtime_error);
    EXPECT_THROW(dt::hipMxDataTypeForDataGenerator(rocisa::DataType::Half),
                 std::runtime_error);
    EXPECT_THROW(dt::hipMxDataTypeForDataGenerator(rocisa::DataType::BFloat16),
                 std::runtime_error);
}

// -----------------------------------------------------------------------------
// 4.8  detail::decodeE8M0
// -----------------------------------------------------------------------------
TEST(DecodeE8M0, ZeroByteIsZero)        { EXPECT_EQ(dt::decodeE8M0(0x00), 0.0f); }
TEST(DecodeE8M0, AllOnesByteIsNaN)      { EXPECT_TRUE(std::isnan(dt::decodeE8M0(0xFF))); }
TEST(DecodeE8M0, BiasByteIsOne)         { EXPECT_EQ(dt::decodeE8M0(0x7F), 1.0f); }   // 2^0
TEST(DecodeE8M0, NextAboveBiasIsTwo)    { EXPECT_EQ(dt::decodeE8M0(0x80), 2.0f); }   // 2^1
TEST(DecodeE8M0, NextBelowBiasIsHalf)   { EXPECT_EQ(dt::decodeE8M0(0x7E), 0.5f); }   // 2^-1
TEST(DecodeE8M0, MatchesLdexpFormulaOverWideRange)
{
    for(int byte = 1; byte < 0xFF; ++byte)
    {
        float expected = std::ldexp(1.0f, byte - 127);
        EXPECT_EQ(dt::decodeE8M0(static_cast<uint8_t>(byte)), expected)
            << "byte=" << byte;
    }
}

// -----------------------------------------------------------------------------
// 4.9  detail::decodeMXElement
//      Calls the production dispatcher AND DGen directly with the same inputs;
//      the two must agree for every supported dtype, and the unsupported
//      branch must return NaN.
// -----------------------------------------------------------------------------
TEST(DecodeMXElement, FP4PathMatchesDGenToFloatPacked)
{
    // Two FP4 elements packed in one byte: low nibble 0x2 (= +1.0),
    // high nibble 0x0 (= +0.0). Scale 0x7F = 1.0.
    uint8_t data [1] = {0x02};
    uint8_t scale[1] = {0x7F};
    for(size_t e : {size_t{0}, size_t{1}})
    {
        float prod = dt::decodeMXElement(rocisa::DataType::Float4,
                                         scale, data, 0, e);
        float dgen = DGen::toFloatPacked<DGen::ocp_e2m1_mxfp4>(
            scale, data, 0, e);
        EXPECT_EQ(prod, dgen) << "elem=" << e;
    }
}
TEST(DecodeMXElement, Float8PathMatchesDGenToFloat_E4M3)
{
    uint8_t data [1] = {0x38};   // 2^0 = +1.0 in E4M3
    uint8_t scale[1] = {0x7F};
    float   prod = dt::decodeMXElement(rocisa::DataType::Float8,
                                       scale, data, 0, 0);
    float   dgen = DGen::toFloat<DGen::ocp_e4m3_mxfp8>(scale, data, 0, 0);
    EXPECT_EQ(prod, dgen);
    EXPECT_EQ(prod, 1.0f);
}
TEST(DecodeMXElement, BFloat8PathMatchesDGenToFloat_E5M2)
{
    uint8_t data [1] = {0x3C};   // 2^0 = +1.0 in E5M2
    uint8_t scale[1] = {0x7F};
    float   prod = dt::decodeMXElement(rocisa::DataType::BFloat8,
                                       scale, data, 0, 0);
    float   dgen = DGen::toFloat<DGen::ocp_e5m2_mxfp8>(scale, data, 0, 0);
    EXPECT_EQ(prod, dgen);
    EXPECT_EQ(prod, 1.0f);
}
TEST(DecodeMXElement, Float6PathMatchesDGenToFloatPacked)
{
    uint8_t data [3] = {0x00, 0x00, 0x00};
    uint8_t scale[1] = {0x7F};
    for(size_t e = 0; e < 4; ++e)
    {
        float prod = dt::decodeMXElement(rocisa::DataType::Float6,
                                         scale, data, 0, e);
        float dgen = DGen::toFloatPacked<DGen::ocp_e2m3_mxfp6>(
            scale, data, 0, e);
        EXPECT_EQ(prod, dgen) << "elem=" << e;
    }
}
TEST(DecodeMXElement, BFloat6PathMatchesDGenToFloatPacked)
{
    uint8_t data [3] = {0x00, 0x00, 0x00};
    uint8_t scale[1] = {0x7F};
    for(size_t e = 0; e < 4; ++e)
    {
        float prod = dt::decodeMXElement(rocisa::DataType::BFloat6,
                                         scale, data, 0, e);
        float dgen = DGen::toFloatPacked<DGen::ocp_e3m2_mxfp6>(
            scale, data, 0, e);
        EXPECT_EQ(prod, dgen) << "elem=" << e;
    }
}
TEST(DecodeMXElement, UnsupportedDataTypeReturnsNaN)
{
    uint8_t data [1] = {0x00};
    uint8_t scale[1] = {0x7F};
    EXPECT_TRUE(std::isnan(dt::decodeMXElement(rocisa::DataType::Float,
                                               scale, data, 0, 0)));
    EXPECT_TRUE(std::isnan(dt::decodeMXElement(rocisa::DataType::Half,
                                               scale, data, 0, 0)));
}

#endif // HIPBLASLT_ENABLE_MXDATAGENERATOR