Forum Discussion
New ContributorSIMD using DSP in Stratix 10
Unlike Xilinx DSP48E1 slices, Altera DSP blocks are primarily optimized for Multiply-Accumulate (MAC) operations — SIMD-style packed adders are not as directly exposed or as flexible in the same way Xilinx’s DSP48E slices are.
Can you see if the below code help?
Verilog RTL — 4×12-bit Parallel Adder (Using ALM Carry Chains)
module parallel_4x12b_adder (
input [11:0] a0, a1, a2, a3, // 4× 12-bit inputs
input [11:0] b0, b1, b2, b3, // 4× 12-bit inputs
output [12:0] sum0, sum1, sum2, sum3 // 4× 13-bit outputs to handle carry
);
assign sum0 = a0 + b0;
assign sum1 = a1 + b1;
assign sum2 = a2 + b2;
assign sum3 = a3 + b3;
endmodule
Explanation: Each assign statement uses the ALM carry chain. Quartus synthesizer will map this to the dedicated carry chains in the ALMs.
Outputs are 13-bit wide to handle potential overflow.
Very efficient — no LUT wasting.
If You Want to Pack into a 48-bit DSP-friendly Add (More Complex) If you really wanted to try packing them and using a DSP block (assuming 48-bit add support, e.g., Arria 10/Stratix 10), here’s a conceptual version:
verilog
module packed_4x12b_adder (
input [47:0] a_packed, // 4× 12-bit packed inputs
input [47:0] b_packed, // 4× 12-bit packed inputs
output [47:0] sum_packed // 4× 12-bit packed results (overflow risk!)
);
assign sum_packed = a_packed + b_packed;
endmodule
Notes: You'd need to align each 12-bit value properly in the 48-bit word.
Risk of overflow if sum exceeds 12 bits in each lane.
Post-add masking and saturation may be needed.
This is risky since Intel DSPs don’t natively split this into SIMD lanes like Xilinx. Quartus might split this into ALM logic anyway.
Recommendation ✔ Use the first version — Quartus will map those independent 12-bit additions onto ALMs using fast carry chains, highly efficient, no LUT waste.
✔ Avoid the packed 48-bit add unless you're certain the device and toolchain will optimize it safely into a DSP block.
Let me know if the above helps to some extent. If not, we may have to leave this for a future enhancement in Quartus.
Super ContributorAfter a detailed analysis, it’s clear that Xilinx offers a distinct advantage in supporting SIMD-style operations through its DSP slices (such as the DSP48E1), which allow for parallel processing of multiple narrow-width additions—for example, four 12-bit additions per clock cycle. This is made possible by internal ALU partitioning specifically optimized for such use cases.
In contrast, Intel (Altera) DSP blocks—such as those in the Stratix 10 and Arria 10 families—do not provide a native SIMD adder mode. Intel's DSP architecture is instead focused on:
- 18x19 or 27x27 multipliers
- Optional pre-adders
- Accumulators and chainout logic
- Dynamic add/subtract/negate features
Because of this, attempting to implement Xilinx-style 4-lane SIMD additions (e.g., 4 parallel 10–14-bit adders) on Intel FPGAs requires alternative approaches:
- Soft logic (LUTs): Quartus can infer these adders from RTL, but this approach uses more logic resources and may impact performance.
- DSP blocks using workaround modes, such as the plus36 mode (Templates 9 and 10), where multiplier outputs are combined using DSP-based post-adders. These methods can only approximate SIMD behavior when combined with dummy or fixed multipliers, not for standalone parallel additions.
Among the provided DSP templates:
- Template 9 allows combining two multipliers with additional operands using dynamic add/sub/negate. Although not a direct SIMD implementation, it can be adapted by setting multiplier inputs to constants (e.g., 1) and injecting other values as summands.
- Template 10 extends this with accumulation and preload support, enabling a pseudo-SIMD operation over multiple cycles.
When exploring the Quartus IP Catalog, there is no current DSP IP that directly implements SIMD-style packed addition (e.g., 4x 12-bit additions per clock) similar to Xilinx. The closest configuration available is the dual 18x19 with post-add/subtract, but this still does not offer true SIMD parallelism within a single DSP block.
Given these limitations, we believe this is a candidate for an enhancement request in a future release of Quartus. Adding native SIMD-style addition support would significantly improve resource efficiency for common signal processing applications and enable better competitiveness.
To better support this request, we would appreciate your input on the specific application use case, including the target algorithm or workload, expected performance improvements, and the resource constraints currently being faced. This information will help justify the feature request and demonstrate the practical impact such an enhancement would provide.
jjxichnThank you for the detailed and insightful comparison between Xilinx and Intel DSP architectures regarding SIMD-style operations.
As you said, since Altera is lacking the SIMD feature compared to Xilinx and I can use template 9 or template 10 as a workaround. This is good enough for help fix my current issue.
Thanks again for your deep analysis and reply in great details.
