Add GEP RTL support for 1D and 2D address generation

[ShangkunLi]

Thanks for the Claude Code!

Overview

Implements the GepRTL functional unit for GetElementPtr (GEP) operations. 1D and 2D address generation are handled directly in RTL; decomposition of 3D/4D+ into basic 1D/2D with additional mul/shift is left to the compiler pass.

Changes

lib/opt_type.py — 4 new opcodes:

Opcode	Value	Semantics
OPT_GEP	88	result = base(in0) + index(in1)
OPT_GEP_CONST	89	result = base(const) + index(in0)
OPT_GEP_2D	90	result = base(in0) + index0(in1) * stride + index1(in2)
OPT_GEP_2D_CONST	91	result = base(const) + index0(in0) * stride + index1(in1)

lib/cmd_type.py — 1 new command:

• CMD_CONFIG_GEP_STRIDE (43): pre-configures the stride register in the GEP FU via recv_from_ctrl_mem before execution, following the same CMD-based configuration pattern used in LoopCounterRTL.

fu/single/GepRTL.py (new): implements all four GEP variants. The 2D stride is stored in a per-FU register latched on CMD_CONFIG_GEP_STRIDE. The const_queue is used for the base address in *_CONST variants.

fu/single/test/GepRTL_test.py (new): 20 test cases covering:

• 1D GEP with parametrized (base, index) combinations
• 1D GEP with const base, multiple sequential operations
• 2D GEP with CMD-configured stride
• 2D GEP with const base and CMD-configured stride
• Mixed 1D + 2D operations in sequence
• Predicate propagation (zero predicate on any input propagates to output)

[Jackcuii]

LGTM！ Thank you so much Shangkun~