Function ggml_cl_mul_mat_use_f16

ggml-opencl.cpp:1823–1841

}

static bool ggml_cl_mul_mat_use_f16(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * /* dst */) {
    // If device doesn't support FP16
    if (!fp16_support) {
        return false;
    }

    size_t src0_sz = ggml_nbytes(src0);
    size_t src1_sz = ggml_nbytes(src1);

    // mul_mat_q: src0 is converted to fp32 on device
    size_t mul_mat_q_transfer = src0_sz + src1_sz;

    // mul_mat_f16: src1 is converted to fp16 on cpu
    size_t mul_mat_f16_transfer = src0_sz + sizeof(ggml_fp16_t) * ggml_nelements(src1);

    // choose the smaller one to transfer to the device
    // TODO: this is not always the best choice due to the overhead of converting to fp16
    return mul_mat_f16_transfer < mul_mat_q_transfer;
}

void ggml_cl_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize) {
    GGML_ASSERT(ggml_cl_can_mul_mat(src0, src1, dst));