Review Src register re-use in matmul

[rtawfik01]

Source register re-use for the llk_unpack_AB_matmul.h and llk_math_matmul.h kernels was a necessity for wormhole_b0, due to the unpacker bandwidth being too low to saturate the compute engine. For blackhole, that is no longer the case, we should review the matmul kernel, and measure performance without re-using Source register as found here:

    for (uint t = 0; t < t_dim; t++) {

        std::uint32_t offset_address_a =tile_size_a*(tile_index_a + (reuse_a ? (t*kt_dim) : (0)));
        std::uint32_t offset_address_b = tile_size_b*(tile_index_b + (reuse_a ? (0       ) : (t)));
        std::uint32_t address_a = base_address_a + offset_address_a;
        std::uint32_t address_b = base_address_b + offset_address_b;

        // Wait for free context
        wait_for_next_context(2);

        semaphore_post(semaphore::UNPACK_SYNC);  // Trisc::SEMPOST for context acquire

        // Program unpacker 1 base address
        if (0 == unp_cfg_context) {
            cfg[THCON_SEC0_REG3_Base_address_ADDR32] = address_b;
            cfg[THCON_SEC1_REG3_Base_address_ADDR32] = address_a;
        } else {
            cfg[THCON_SEC0_REG3_Base_cntx1_address_ADDR32] = address_b;
            cfg[THCON_SEC1_REG3_Base_cntx1_address_ADDR32] = address_a;
        }

        if (reuse_a) {
            #if SKIP_UNP == 1
                TTI_NOP;
            #else
                if (partial_face) {
                    // Do face by face unpacking
                    TTI_UNPACR(SrcB, 0b00010001, 0, 0, 0, 1 /*Set OvrdThreadId*/, 0 /*Set Dvalid*/, p_unpacr::RAREFYB_DISABLE, 0, 0 /* Set ContextIdInc */, 0, 0, 1);
                    TTI_UNPACR(SrcB, 0b00010001, 0, 0, 0, 1 /*Set OvrdThreadId*/, 1 /*Set Dvalid*/, p_unpacr::RAREFYB_DISABLE, 0, 0 /* Set ContextIdInc */, 0, 0, 1);
                    TTI_SETADCZW(p_setadc::UNP_B, 0, 0, 0, 0, 0b0101); // Set ch0_z=0, ch1_z=0
                } else {
                    TTI_UNPACR(SrcB, 0, 0, 0, 0, 1 /*Set OvrdThreadId*/, 1 /*Set Dvalid*/, p_unpacr::RAREFYB_DISABLE, 0, 0 /* Set ContextIdInc */, 0, 0, 1);
                }
            #endif
        } else {
            #if SKIP_UNP == 1
                TTI_NOP;
            #else
                TTI_UNPACR(SrcA, 0, 0, 0, 0, 1 /*Set OvrdThreadId*/, 1 /*Set Dvalid*/, p_unpacr::RAREFYB_DISABLE, 0, 0 /* Set ContextIdInc */, 0, 0, 1);
            #endif
        }


        TT_MOP(0, (reuse_a ? ct_dim : rt_dim) - 1, unp_cfg_context == 0 ? 0 : 0xff); // Run the MOP

        // T6::SEMGET for context release
        t6_semaphore_get(semaphore::UNPACK_SYNC);

        // Switch unpacker config context
        switch_config_context(unp_cfg_context);
    }

If the matmul kernels performance is similar or greater without re-use, then the re-use flags and functionality should be removed.

@ttmtrajkovic @rdjogoTT fyi.