ggml : reduce the values of ggml_op enum

[ggerganov]

Here is another idea to consider for reducing the amount of unary ops and code duplication in ggml.c:

The following ops:

    enum ggml_op {
        ...
        GGML_OP_ABS,
        GGML_OP_SGN,
        GGML_OP_NEG,
        GGML_OP_STEP,
        GGML_OP_TANH,
        GGML_OP_ELU,
        GGML_OP_RELU,
        GGML_OP_GELU,
        GGML_OP_GELU_QUICK,
        GGML_OP_SILU,
        ...
    };

Can all become:

    enum ggml_unary_op {
        GGML_UNARY_OP_ABS,
        GGML_UNARY_OP_SGN,
        GGML_UNARY_OP_NEG,
        GGML_UNARY_OP_STEP,
        GGML_UNARY_OP_TANH,
        GGML_UNARY_OP_ELU,
        GGML_UNARY_OP_RELU,
        GGML_UNARY_OP_GELU,
        GGML_UNARY_OP_GELU_QUICK,
        GGML_UNARY_OP_SILU,    
    };

    GGML_API struct ggml_tensor * ggml_unary(
            struct ggml_context * ctx,
             struct ggml_tensor * a,
             enum ggml_unary_op   op);

---

Original task (completed via #313)

Certain ggml operators can be combined into a singe operator with extra arguments.

For example:

• GGML_OP_CONV_1D_S1_PH
• GGML_OP_CONV_1D_S2_PH

can all become:

• GGML_OP_CONV_1D

with a signature of:

    GGML_API struct ggml_tensor * ggml_conv_1d(
            struct ggml_context * ctx,
            struct ggml_tensor  * a,
            struct ggml_tensor  * b
                            int   stride,
                            int   padding);

Internally, we can have custom implementations of ggml_compute_forward_conv_1d() for specific values of the parameters and GGML_ASSERT for everything else that is not supported.

This will allow to extend the functionality of these operator more easily in the future

[ggerganov]

bump (description updated)

[rlanday]

@ggerganov, it appears you did this already in 5b2b2dc. I recommend closing this issue.

[rlanday]

Oh never mind, I am looking at issues that are already completed.