refactor: slim PLENA compiler frontend and codegen

asm_templates/__init__.py

@@ -18,7 +18,7 @@
 from .elementwise_add_asm import elementwise_add_asm
 from .embedding_asm import embedding_asm
 from .ffn_asm import ffn_asm, ffn_intermediate_asm, ffn_up_silu_asm
-from .flash_attn_asm import flash_attn_asm
+from .flashattn import flash_attn_asm
 from .gelu_asm import gelu_asm
 from .im2col_asm import im2col_asm
 from .im2col_asm_no_shift import im2col_asm_no_shift

asm_templates/tests/test_vram_sub_projection.py

@@ -3,7 +3,7 @@
 Verifies the extracted free function produces the expected ISA output for
 looped/unrolled and transposed/non-transposed variants, and asserts
 byte-identical parity with the delegating
-``TileCompiler._vram_sub_projection_asm_impl`` method.
+``IsaCompiler._vram_sub_projection_asm_impl`` method.
 """
 
 import sys
@@ -95,10 +95,10 @@ def test_unrolled_no_loops(self):
         self.assertIn("M_MM_WO", asm)
 
     def test_output_byte_identical_to_method(self):
-        """The free function must produce byte-identical output to TileCompiler's method."""
-        from compiler.aten.plena_compiler import TileCompiler
+        """The free function must produce byte-identical output to IsaCompiler's method."""
+        from compiler.aten.plena import IsaCompiler
 
-        tc = TileCompiler(mlen=64, blen=4, unroll_loops=False)
+        compiler = IsaCompiler(mlen=64, blen=4, unroll_loops=False)
 
         method_kwargs = dict(
             header_lines=["; header"],
@@ -113,7 +113,7 @@ def test_output_byte_identical_to_method(self):
             caller_name="test",
         )
 
-        method_out = tc._vram_sub_projection_asm_impl(**method_kwargs)
+        method_out = compiler._vram_sub_projection_asm_impl(**method_kwargs)
         free_out = vram_sub_projection_asm_impl(
             mlen=64,
             blen=4,

asm_templates/vram_sub_projection_asm.py

@@ -1,7 +1,7 @@
 """Pure emitter for VRAM sub-projection ISA.
 
-Shared implementation kernel used by ``TileCompiler.vram_sub_projection_asm``
-and ``TileCompiler.vram_sub_projection_T_asm``. The caller resolves all
+Shared implementation kernel used by ``IsaCompiler.vram_sub_projection_asm``
+and ``IsaCompiler.vram_sub_projection_T_asm``. The caller resolves all
 instance-dependent state (register allocator, tile layouts, MRAM addresses,
 ``unroll_loops`` default) and passes it in as plain parameters, so this
 emitter can be unit-tested in isolation.

aten/__init__.py

@@ -1,6 +1,6 @@
 """compiler.aten — ATen-style PLENA compiler path.
 
-PlenaCompiler DSL + op backend registry. Pairs with compiler.generator
+PlenaCompiler program builder + op backend registry. Pairs with compiler.generator
 for the two-path compiler (template vs. aten).
 """
 
@@ -9,30 +9,22 @@
 PLENA_PKG_DIR = Path(__file__).parent
 NATIVE_OPS_YAML = PLENA_PKG_DIR / "native_ops.yaml"
 
-from compiler.aten.plena_compiler import (  # noqa: E402, F401
+from compiler.aten.isa_builder import (  # noqa: E402, F401
+    Comment,
+    Instr,
+    IsaBuilder,
+    Register,
+    addr,
+    fp,
+    gp,
+)
+from compiler.aten.plena import (  # noqa: E402, F401
+    FPVar,
+    InputVar,
+    IsaCompiler,
+    MemoryStateMixin,
     PlenaCompiler,
-    TileCompiler,
-    DeveloperCompiler,
-    RegisterAllocator,
     TensorVar,
-    InputVar,
     VRAMMatrixVar,
-    FPVar,
-    TensorKind,
-    tensor_kind,
-    Tensor,
-    TensorInfo,
-    TileLayout,
-    MemoryBlock,
-    VirtualMemoryManager,
-    MRAMAllocator,
-    VRAMAllocator,
-    FPRAMAllocator,
-    SubMatrixInfo,
-    MatrixBlockLayout,
-    VRAMSubMatrixInfo,
-    VRAMMatrixBlockLayout,
-    MemoryObjectInfo,
-    FPRAMObjectLayout,
 )
 from compiler.aten.ops.registry import OpRegistry, Backend  # noqa: E402, F401