www.ti.com | Shared-Processing (SP) Mode |
Each sub-frame is independent of each other. There are three types of sub-frames. The first sub-frame starts the trellis from the zero state. The last sub-frame ends the trellis from a known state. The remaining middle subframes do not start or end from a known state.
The EDMA3 transfers ACNT*BCNT number of bytes in A-Sync Mode and ACNT*BCNT*CCNT number of bytes in AB-Sync Mode. The total number of bytes for both modes should be a multiple of 8. Also, the starting address of the first sub-frame that the EDMA3 will transfer needs to be memory-mapped.
In the shared processing mode:
∙Prolog length must be multiples of 8
∙Starting address for reading extrinsic RAM must be:
RAM base address + middle and last subframes prolog length
∙CRC is turned off
∙SNR is turned off
∙Prolog reduction is turned off
∙Extrinsic scaling is turned off
The turbo decoding of the full frame is performed in several steps as described below:
∙The EDMA3 sends the input buffers for one sub-frame (the MAP0 inputs are described in Figure 19).
∙The TCP2 performs the MAP0 for the current sub-frame.
∙The EDMA3 reads the MAP output (extrinsic) of the current sub-frame and writes it into the DSP memory.
The steps for the MAP0 process are repeated for all the other sub-frames.
Once all the sub-frames MAP0 have been computed, the full MAP0 extrinsic (= apriori 1) is then available. This allows the DSP to interleave the extrinsic output 1 to prepare the next MAP (= MAP1). Once this interleaving is done, the same process is applied, in MAP1 configuration:
∙The EDMA3 sends the input buffers for one sub-frame (the MAP0 inputs are described in Figure 19).
∙The TCP2 performs the MAP1 for the current sub-frame
∙The EDMA3 reads the MAP output (extrinsic) of the current sub-frame and writes it into the DSP memory.
The steps for the MAP1 process are repeated for all the other sub-frames.
Once all the sub-frames MAP1 have been computed, the full extrinsic (=apriori 2) is then available. This allows the DSP to de-interleave the extrinsic output 2 to prepare the next MAP (=MAP0). Once this de-interleaving is done, the same process is applied, in MAP0 configuration. Steps 1-4 are then repeated for all iterations. The DSP is in charge of any stopping criteria algorithm implementation and computing the final hard decisions. Figure 19 shows a description of the TCP2 processing unit functional block diagram in shared processing mode.
SPRUGK1–March 2009 | TMS320C6457 Turbo-Decoder Coprocessor 2 | 21 |