[Feature] pto.vmi.vcvt(fp16->bf16) is not supported which takes bad workaround solution

[sundyCoder]

Summary

pto.vmi.vcvt(fp16->bf16) is not supported, which takes bad workaround solutions.

Cast path (still required by VMI/ptoas): si8 → si32 → f32 → f16 → f32 → bf16
(no direct f16→bf16; sitofp needs i32)

CCE example:

    __VEC_SCOPE__
    {
        vector_u8 vKvData;
        vector_f16 vCastFp16Res;
        vector_bf16 vOffset, vScale, vRes;

        vector_bool kvTypeMaskAll = pset_b8(PAT_ALL);
        vector_bool qTypeMaskLower64 = pset_b16(PAT_VL64);
        vector_bool qTypeMaskAll = pset_b16(PAT_ALL);
        vector_bool qTypeMaskHigher64;
        pxor(qTypeMaskHigher64, qTypeMaskLower64, qTypeMaskAll, qTypeMaskAll);

        // Load offset and scale (128 bf16 elements each)
        vlds(vOffset, offsetPtr, 0, UNPK_B16);
        vlds(vScale, scalePtr, 0, UNPK_B16);

        for (uint16_t i = 0; i < loopCnt; ++i) {
            // LoadAlign<UNPK_B8>: load 256 int8, unpack
            vlds(vKvData, srcPtr, 0, UNPK_B8);

            // Cast<half, int8, PART_EVEN>: int8→f16
            vcvt(vCastFp16Res, vKvData, kvTypeMaskAll, PART_EVEN, MODE_ZEROING);

            // Cast<bf16, half>: f16→bf16
            vcvt(vRes, vCastFp16Res, qTypeMaskAll, ROUND_R, MODE_ZEROING);

            // Add offset (lower 128 mask)
            vadd(vRes, vRes, vOffset, qTypeMaskAll, MODE_ZEROING);

            // Mul scale (lower 128 mask)
            vmul(vRes, vRes, vScale, qTypeMaskAll, MODE_ZEROING);

            // StoreAlign lower 64 elements
            vsts(vRes, dstPtr, i * VF_LEN_B16, NORM_B16, qTypeMaskLower64);

            // StoreAlign higher 64 elements (to second output buffer)
            // Note: ASC uses DATA_BLOCK_COPY with blockStride/repeatStride for interleaved store.
            // In CCE, we use a second store at offset + 64.
            vsts(vRes, dstPtr, i * VF_LEN_B16 + 64, NORM_B16, qTypeMaskHigher64);
        }
    }

ptoas 0.59 VMI solution:

    mask128 = pto.vmi.create_mask(VF_LEN, size=VF_LEN)
    v_offset = pto.vmi.vload(off_ub, 0, size=VF_LEN)
    v_scale = pto.vmi.vload(sc_ub, 0, size=VF_LEN)

    # One loop like CCE: 128 bf16 outputs per iteration (was 2×64).
    for i in range(LOOP_CNT):
        # Contiguous first-128 of each 256-byte block (VMI has no PART_EVEN 256→128).
        v_kv = pto.vmi.vload(src_ub, i * 256, size=VF_LEN)
        # CCE: u8→f16→bf16. VMI: sitofp needs i32; f16→bf16 needs f32 bridge.
        v_i32 = pto.vmi.vcvt(v_kv, to_dtype=pto.si32)
        v_f32 = pto.vmi.vcvt(v_i32, to_dtype=pto.f32)
        v_f16 = pto.vmi.vcvt(v_f32, to_dtype=pto.f16, rounding="R", saturate="SAT")
        v_f32 = pto.vmi.vcvt(v_f16, to_dtype=pto.f32)
        v_res = pto.vmi.vcvt(v_f32, to_dtype=pto.bf16, rounding="R", saturate="SAT")
        v_res = pto.vmi.vadd(v_res, v_offset, mask128)
        v_res = pto.vmi.vmul(v_res, v_scale, mask128)
        pto.vmi.vstore(v_res, dst_ub, i * VF_LEN, mask128)

Motivation / use case

Cast path (still required by VMI/ptoas): si8 → si32 → f32 → f16 → f32 → bf16
(no direct f16→bf16; sitofp needs i32)

Proposed API / behavior

No response

Alternatives considered

No response

Additional context

No response