doc: Add ADR for storing and processing shares (#933)

adr/001-storage-and-processing-of-iris-shares.md

@@ -0,0 +1,23 @@
+# ADR-001: Storage and processing of iris shares
+
+## Context
+- We have iris code shares that, in their raw form, are 12800 bytes (u8) each and stored in a PostgreSQL database.
+- For secret sharing purposes, these iris codes need to be represented internally as u16.
+- The cuBLAS library (used for efficient GPU-accelerated comparisons) requires the data to be in i8 format.
+- Performing repeated on-the-fly conversions from u16 to i8 can degrade performance significantly due to large data volumes.
+
+## Decision
+Store and load iris code shares in a pre-processed i8 format (split into two limbs from each u16), so that runtime conversions are minimized or eliminated.
+
+## Rationale
+- **Performance**: The conversion from u16 to i8 (subtracting 128 for cuBLAS alignment) can be expensive if done repeatedly for large datasets.
+- **cuBLAS Requirements**: The library operates natively on i8 data. Having the data already in the correct format avoids overhead.
+- **Simplicity in Usage**: Once data is in i8 format, usage in GPU kernels is straightforward.
+
+## Consequences
+- **Pros**:
+    - Streamlined GPU-based operations (no extra conversion step in the processing pipeline).
+- **Cons**:
+    - Requires additional pre-processing upon ingestion.
+
+By pre-processing iris code shares into i8 and storing them in that format, we reduce runtime overhead and align with the cuBLAS requirements, resulting in more efficient iris-code comparisons.