<table border="1" cellspacing="0" cellpadding="8">
    <tr>
        <th>Issue</th>
        <td>
            <a href=https://github.com/llvm/llvm-project/issues/83626>83626</a>
        </td>
    </tr>

    <tr>
        <th>Summary</th>
        <td>
            [DXIL] implement dot intrinsic lowering
        </td>
    </tr>

    <tr>
      <th>Labels</th>
      <td>
            new issue
      </td>
    </tr>

    <tr>
      <th>Assignees</th>
      <td>
            farzonl
      </td>
    </tr>

    <tr>
      <th>Reporter</th>
      <td>
          farzonl
      </td>
    </tr>
</table>

<pre>
    There are three parts
First to do float\half dot products we need to support three opcodes with varying argument lengths

- `@dx.op.dot2.f32(i32 54 ...)` - 4 arguments a[0], a[1], b[0], b[1]
- `@dx.op.dot3.f32(i32 55 ...)` - 6 arguments a[0], a[1], a[2], b[0], b[1], b[2]
- `@dx.op.dot4.f32(i32 56 ...)` - 8 arguments a[0], a[1], a[2], a[3], b[0], b[1], b[2], b[3]

For each of these we will need to do 4 to 8 extract element before we call the intrinsic.

Part 1 would be to create a pass that  flattens the vectors to scalars into the form shown above
Part2 is to modify DXIL.td to represent DIXIL ops

For good references on behavior see:
- https://godbolt.org/z/TbvshPchs
- [lib/HLSL/HLOperationLower.cpp](https://github.com/microsoft/DirectXShaderCompiler/blob/main/lib/HLSL/HLOperationLower.cpp#L6646C5-L6651C37)

Part 3 is to support integer dot products.

We will need to create a DxilTrinaryOperation which we don't currently have and further support the lowering of DXIL::OpCode::UMad  and  DXIL::OpCode::IMad

See
- https://godbolt.org/z/srbzrhMbq
- [TranslateIDot] (https://github.com/microsoft/DirectXShaderCompiler/blob/main/lib/HLSL/HLOperationLower.cpp#L2451C1-L2467C1)

The format for integer dot product is silightly different. You don't fetch all the  vector  indices up front you do them in stages
extract a[i] and b[i]
multiply a[i] and b[i]
extract a[i+1] and b[i+1]
Perform an IMad on  (a[i+1], b[i+1],a[i]*b[i])  // `%IMad = call i32 @dx.op.tertiary.i32(i32 48, i32 %4, i32 %5, i32 %3)`


For every vector size increase you daisy chain the  extract elements and IMad results.
</pre>
<img width="1px" height="1px" alt="" src="http://email.email.llvm.org/o/eJzEVk2P2zYQ_TX0ZWBBoj5sH3xIbBhdwEECZIumR0ocWSxoUh1Sdry_viAtK9pFkqanXqyRPF-aN-9Rwjl1MohbVr5nnLeCXqzRjHNW7hdi8J2l7fhwUVt52z53SAiCEHxHiNAL8o6le5a-OyhyHrwFaaHVVnhW7jqhW5DWQ09WDo13cEUwiDL4uaHvLfkxk-0bK9HBVfkOLoJuypxA0Gk4o_Gg0Zx8N1a6_y6BVSkrUvk1sX0iredJm3PG1yrnUBaQJAnjG1alsIRiyuRAsPJ9yso947toZ6Ndz57Xj-ffr5TPK5WvKlW_UCnY_GdVR5v_uINi3kH1qoP1f-wg2PmvdjPa-dTZiL0lQNF0YFvwHToMOF-V1hPY0kIRrmvAr55E4wE1RmhrbC3FgEZoHcJBGU_KONUk8yKfBHnI4GoHLaHGkK0hFB5BQC-cA98JD9Bq4T0aFzNdsPGWXFy3RmhBLiS38b_W0hlcZ68GRG0v-K0KBxVDzlaq9gb7L0_HxMfXIOwJXWh7__Tl6Qi2d2_ncLJWAmGLhKZBB9ZAjZ24KEvgEFk-7W_nfe_CPT8wfjhZWVvtE0snxg8vjB-e64vrPjWPtV8CK99rVTN--O34-RgvH3sk4ZU1R3tFSpq-jyit36RWvhvqpLFnxg9n1ZB1tvWMH_aKsPFfPndCIu3suVcaifFDrW0ocxbKMH7495o8P1ZVUe3K5bGqymyXr8I6vsUuH8f6IL4yHk9IrwTiFeJ_vFmiCe79V6WfSRlBt6kbuHaq6cIeSWsYX3loBiI0Xt-gExcEYSS0A_kOaaY9CDq8RlAb20aow9jydx_7nZV4t3__ICTE-B84PH0Qct74Z8RfBdlR_ULdh_rvGcjPJIzTwuPT3npW7uF_QZQXZbbLlkdeVKtd9gbQ55FAwofL96AMYDul1akLCEjVRkb4BP60wwRRi77p4MH7ka4AykgVuDP00JI1Hm4xJjidQRlwXpxwpMVDT4KOqTCsgFM93txdzoP2qte3n_m8TsOj9M3c-Ow8-IQUtUMYCMAHhgeE5pEPoZzdP2oz_m6qzDchMgAaBZ6XMR_L93ctDOo-qb5H8krQLVGT8hfrUCd68bKY2eXMzu8nwxy8mWhfkG6PsTv1ErQ3sMzhfeRCuRs0nVDmDtAb8XZxQrFpQjdo75KF3OZyk2_EArfZKt3keZpn2aLbyqLENsf1RtZrbEtcb7JNy3naruqmTtPVQm15yos0TzOeles8S7J1VTecC1nnxYqvM1akeBZKJ1pfzoFCC-XcgNt1XvFqoUWN2o1fMgavEP8cv2VoG2KW9XByrEi1ct59y-KV1_ETKJK73IM69-PhFPZ5Oo0mpVgMpLc_IWTIPF6WPdm_sAnUjP04xg-x338CAAD__3s55NM">