syrk schedule (missing calls to gemm after tiling) (#94)

src/common/codegen_helpers.py

@@ -52,9 +52,7 @@ def generate_stride_1_proc(proc):
     proc = rename(proc, proc.name() + "_stride_1")
     for arg in proc.args():
         if arg.is_tensor():
-            proc = proc.add_assertion(
-                f"stride({arg.name()}, {len(arg.shape()) - 1}) == 1"
-            )
+            proc = proc.add_assertion(f"stride({arg.name()}, {len(arg.shape()) - 1}) == 1")
     return proc
 
 
@@ -110,15 +108,14 @@ def export_perf_features(kernel_name, perf_features):
         json.dump(perf_features, f, sort_keys=True, indent=4, separators=(",", ": "))
 
 
-def variants_generator(blas_op, opt_precisions=("f32", "f64"), targets=(AVX2, Neon)):
+def variants_generator(blas_op, opt_precisions=("f32", "f64"), targets=(AVX2, Neon), stage_scalars=True):
     def generate(proc, loop_name, *args, globals=None, **kwargs):
         perf_features = {}
         for precision in ("f32", "f64"):
             proc_variant = specialize_precision(proc, precision)
 
-            proc_variant = stage_scalar_args(proc_variant)
-
             stride_any = generate_stride_any_proc(proc_variant)
+            stride_any = stage_scalar_args(stride_any)
             stride_any = bind_builtins_args(stride_any, stride_any.body(), precision)
             export_exo_proc(globals, stride_any)
 
@@ -127,9 +124,8 @@ def generate(proc, loop_name, *args, globals=None, **kwargs):
             algorithm = get_perf_features(stride_1)
 
             if precision in opt_precisions and C.Machine.mem_type in targets:
-                stride_1 = blas_op(
-                    stride_1, loop, precision, C.Machine, *args, **kwargs
-                )
+                stride_1 = blas_op(stride_1, loop, precision, C.Machine, *args, **kwargs)
+            stride_1 = stage_scalar_args(stride_1)
             stride_1 = bind_builtins_args(stride_1, stride_1.body(), precision)
             scheduled = get_perf_features(stride_1)
 

src/common/perf_features.py

@@ -63,9 +63,7 @@ def get_stmt_ops(cursor, syms):
             iter_sym = sp.Symbol(cursor.name(), integer=integers, positive=integers)
             new_syms[cursor.name()] = iter_sym
             stmts_ops = [get_stmt_ops(s, new_syms) for s in cursor.body()]
-            sums = [
-                sp.summation(s_ops, (iter_sym, lo, hi - one)) for s_ops in stmts_ops
-            ]
+            sums = [sp.summation(s_ops, (iter_sym, lo, hi - one)) for s_ops in stmts_ops]
             return sum(sums)
         else:
             return zero
@@ -74,14 +72,14 @@ def get_stmt_ops(cursor, syms):
     for arg in proc.args():
         if arg.type().is_indexable():
             is_size = arg.type() == ExoType.Size
-            syms[arg.name()] = sp.Symbol(
-                arg.name(), integer=integers, positive=is_size and integers
-            )
-
-    ops = get_stmt_ops(proc.body(), syms)
-    ops = sp.simplify(ops)
-    ops = sp.factor(ops)
-    return ops
+            syms[arg.name()] = sp.Symbol(arg.name(), integer=integers, positive=is_size and integers)
+    try:
+        ops = get_stmt_ops(proc.body(), syms)
+        ops = sp.simplify(ops)
+        ops = sp.factor(ops)
+        return ops
+    except:
+        return zero
 
 
 def count_flops(proc, upper=False):
@@ -91,9 +89,7 @@ def count_flops(proc, upper=False):
     def get_expr_ops(proc, expr):
         expr = proc.forward(expr)
         if isinstance(expr, (BuiltInFunctionCursor, BinaryOpCursor, UnaryMinusCursor)):
-            children_ops = sum(
-                get_expr_ops(proc, c) for c in get_numeric_children(proc, expr)
-            )
+            children_ops = sum(get_expr_ops(proc, c) for c in get_numeric_children(proc, expr))
             return one + children_ops
         return zero
 
@@ -140,9 +136,7 @@ def get_bytes_traffic(p, c):
     def get_expr_loads(proc, expr):
         expr = proc.forward(expr)
         if isinstance(expr, (BuiltInFunctionCursor, BinaryOpCursor, UnaryMinusCursor)):
-            return sum(
-                get_expr_loads(proc, c) for c in get_numeric_children(proc, expr)
-            )
+            return sum(get_expr_loads(proc, c) for c in get_numeric_children(proc, expr))
         elif isinstance(expr, ReadCursor):
             return get_bytes_traffic(proc, expr)
         return zero

src/common/stdlib.py

@@ -566,8 +566,7 @@ def get_depth(proc, scope):
     def push_scope_in(proc, scope, depth, size=None):
         scope = proc.forward(scope)
         allocs = list(filter_cursors(is_alloc)(proc, scope.body()))
-        proc = apply(attempt(parallelize_and_lift_alloc))(proc, allocs)
-
+        proc = apply(parallelize_and_lift_alloc)(proc, allocs)
         if const_allocs and size:
             proc = apply(lambda p, a: resize_dim(p, a, 0, size, 0))(proc, allocs)
 
@@ -670,8 +669,9 @@ def get_window(cursor):
         for loop in my_loops:
             lo_rng = eval_rng(loop.lo(), env)
             hi_rng = eval_rng(loop.hi(), env)
-            env[loop.name()] = (lo_rng[0], hi_rng[1])
-        window = tuple(eval_rng(idx, env) for idx in cursor.idx())
+            env[loop.name()] = (lo_rng[0], f"{hi_rng[1]} - 1")
+        extend_hi = lambda rng: (rng[0], f"{rng[1]} + 1") if rng[0] != rng[1] else rng
+        window = tuple(extend_hi(eval_rng(idx, env)) for idx in cursor.idx())
         return window
 
     def window_to_str(window):
@@ -1014,11 +1014,18 @@ def bound_loop_by_if(proc, loop):
     if not isinstance(if_c.orelse(), InvalidCursor):
         raise BLAS_SchedulingError(err)
 
-    if not isinstance(if_c.cond().lhs(), ReadCursor) or if_c.cond().lhs().name() != loop.name() or if_c.cond().op() != "<":
+    cond = if_c.cond()
+    cond_lhs = cond.lhs()
+
+    if is_read(proc, cond_lhs, loop.name()) and cond.op() == "<":
+        cut_point = FormattedExprStr("_ + _", loop.lo(), cond.rhs())
+    elif is_add(proc, cond_lhs) and is_read(proc, cond_lhs.lhs(), loop.name()) and cond.op() == "<":
+        cut_point = FormattedExprStr("_ + (_ - _)", loop.lo(), cond.rhs(), cond_lhs.rhs())
+    else:
         raise BLAS_SchedulingError(err)
 
     if_c = loop.body()[0]
-    proc = cut_loop(proc, loop, FormattedExprStr("_ + _", loop.lo(), if_c.cond().rhs()))
+    proc = cut_loop(proc, loop, cut_point)
     loop1 = proc.forward(loop)
     loop2 = loop1.next()
     proc = eliminate_dead_code(proc, loop1.body()[0])
@@ -1333,9 +1340,7 @@ def add_loop_nest(loop):
     for src_dim, _ in reversed(list(enumerate(alloc.shape()))):
         for dst_dim, size in reversed(divisions[src_dim][1:]):
             proc = divide_dim(proc, alloc, src_dim, size)
-            proc = apply(divide_loop_)(
-                proc, loop_nest[src_dim], size, tail="cut"
-            )  # TODO: This should be enabled but it slows down compilation
+            proc = apply(divide_loop_)(proc, loop_nest[src_dim], size, tail="cut")
             perm.append(dst_dim)
         perm.append(divisions[src_dim][0][0])
 
@@ -1358,3 +1363,9 @@ def add_loop_nest(loop):
         load = load.as_block().expand(0, diff)
 
     return proc, pack_mem_cursors(alloc, load, block, store)
+
+
+def inline_calls(proc, block=InvalidCursor(), subproc=None):
+    calls = filter_cursors(is_call)(proc, nlr_stmts(proc, block), subproc)
+    proc = simplify(apply(inline_proc_and_wins)(proc, calls))
+    return proc

src/level3/gemm.py

@@ -127,8 +127,7 @@ def schedule(main_gemm, i_loop, precision, machine, m_r, n_r_fac, M_tile, N_tile
     gemm_tiled = apply(repeate_n(reorder_loops))(gemm_tiled, gemm_tiled.find_loop("ki", many=True), n=2)
     gemm_tiled = replace_all_stmts(gemm_tiled, [gemm_macro])
 
-    macro_calls = filter_cursors(is_call)(gemm_tiled, nlr_stmts(gemm_tiled))
-    gemm_tiled = simplify(apply(inline_proc_and_wins)(gemm_tiled, macro_calls))
+    gemm_tiled = inline_calls(gemm_tiled, subproc=gemm_macro[1])
 
     gemm_tiled = apply(hoist_from_loop)(gemm_tiled, gemm_tiled.find_loop("jo", many=True))
     gemm_tiled = squash_buffers(gemm_tiled, gemm_tiled.find("packed_A : _", many=True))