Add Keccak

example.py

@@ -58,7 +58,7 @@ class Example():
 
     def __init__(self, infile, name=None, funcname=None, suffix="opt",
                  rename=False, outfile="", arch=Arch_Armv81M, target=Target_CortexM55r1,
-                 timeout=None, **kwargs):
+                 timeout=None, outfile_full=False, **kwargs):
         if name is None:
             name = infile
 
@@ -67,17 +67,23 @@ def __init__(self, infile, name=None, funcname=None, suffix="opt",
         self.funcname = funcname
         self.infile = infile
         self.suffix = suffix
-        if outfile == "":
-            self.outfile = f"{infile}_{self.suffix}_{target_label_dict[self.target]}"
+        if outfile_full is True:
+            self.outfile = outfile
         else:
-            self.outfile = f"{outfile}_{self.suffix}_{target_label_dict[self.target]}"
+            if outfile == "":
+                self.outfile = f"{infile}_{self.suffix}_{target_label_dict[self.target]}"
+            else:
+                self.outfile = f"{outfile}_{self.suffix}_{target_label_dict[self.target]}"
         if funcname is None:
             self.funcname = self.infile
         subfolder = ""
         if self.arch == AArch64_Neon:
             subfolder = "aarch64/"
         self.infile_full = f"examples/naive/{subfolder}{self.infile}.s"
-        self.outfile_full = f"examples/opt/{subfolder}{self.outfile}.s"
+        if outfile_full is False:
+            self.outfile_full = f"examples/opt/{subfolder}{self.outfile}.s"
+        else:
+            self.outfile_full = self.outfile
         self.name = name
         self.rename = rename
         self.timeout = timeout
@@ -155,9 +161,12 @@ def run(self, debug=False, log_model=False, log_model_dir="models", dry_run=Fals
 
         self.core(slothy, *self.extra_args)
 
-        if self.rename:
-            slothy.rename_function(
-                self.funcname, f"{self.funcname}_{self.suffix}_{target_label_dict[self.target]}")
+        if self.rename is not False:
+            if self.rename is True:
+                slothy.rename_function(
+                    self.funcname, f"{self.funcname}_{self.suffix}_{target_label_dict[self.target]}")
+            elif isinstance(self.rename, str):
+                slothy.rename_function(self.funcname, self.rename)
 
         if dry_run is False:
             slothy.write_source_to_file(self.outfile_full)
@@ -1088,7 +1097,7 @@ def core(self, slothy):
         slothy.config.constraints.stalls_first_attempt = 110
         slothy.optimize_loop("layer123_start")
 
-        
+
 
 
 class ntt_dilithium_123(Example):
@@ -1211,7 +1220,7 @@ def core(self, slothy):
         slothy.optimize_loop("layer5678_start")
 
         slothy.config = conf.copy()
-        
+
         if self.timeout is not None:
             slothy.config.timeout = self.timeout // 12
 
@@ -1228,7 +1237,7 @@ def core(self, slothy):
         slothy.config.split_heuristic_stepsize = 0.1
         slothy.config.constraints.stalls_first_attempt = 14
         slothy.optimize_loop("layer1234_start")
-            
+
 
 class ntt_dilithium_1234(Example):
     def __init__(self, var="", arch=AArch64_Neon, target=Target_CortexA72):
@@ -1355,6 +1364,133 @@ def core(self, slothy):
         slothy.config.sw_pipelining.optimize_postamble = False
         slothy.optimize_loop("flt_radix4_fft_loop_start")
 
+class neon_keccak_x1_no_symbolic(Example):
+    def __init__(self, var="", arch=AArch64_Neon, target=Target_CortexA55):
+        name = "keccak_f1600_x1_scalar_slothy_no_symbolic"
+        infile = "keccak_f1600_x1_scalar_slothy"
+        outfile = "keccak_f1600_x1_scalar_no_symbolic"
+
+        super().__init__(infile, name, outfile=outfile, rename=True, arch=arch, target=target)
+
+    def core(self, slothy):
+        slothy.config.reserved_regs = ["x18", "sp"]
+
+        slothy.config.inputs_are_outputs = True
+        slothy.config.variable_size = True
+        slothy.config.visualize_expected_performance = False
+        slothy.config.timeout = 10800
+
+        slothy.config.selfcheck_failure_logfile = "selfcheck_fail.log"
+
+        slothy.config.outputs = ["flags"]
+        slothy.config.constraints.stalls_first_attempt = 64
+        slothy.config.constraints.minimize_spills = True
+        slothy.config.constraints.allow_reordering = True
+        slothy.config.constraints.allow_spills = True
+        slothy.config.constraints.minimize_spills = True
+        slothy.config.visualize_expected_performance = True
+        slothy.optimize(start="loop", end="end_loop")
+
+        slothy.config.outputs = ["hint_STACK_OFFSET_COUNT"]
+        slothy.optimize(start="initial_round_start", end="initial_round_end")
+
+class neon_keccak_x1_scalar_opt(Example):
+    def __init__(self, var="", arch=AArch64_Neon, target=Target_CortexA55):
+        name = "keccak_f1600_x1_scalar_opt"
+        infile = "keccak_f1600_x1_scalar_pre_opt"
+        outfile = "keccak_f1600_x1_scalar"
+
+        super().__init__(infile, name, outfile=outfile, rename=True, arch=arch, target=target)
+
+    def core(self, slothy):
+        slothy.config.reserved_regs = ["x18", "sp"]
+
+        slothy.config.inputs_are_outputs = True
+        slothy.config.variable_size = True
+        slothy.config.timeout = 10800
+
+        slothy.config.selfcheck_failure_logfile = "selfcheck_fail.log"
+
+        slothy.config.outputs = ["flags"]
+        slothy.config.constraints.stalls_first_attempt = 32
+        slothy.config.visualize_expected_performance = True
+        slothy.config.split_heuristic = True
+        slothy.config.split_heuristic_factor = 1.5
+        slothy.config.split_heuristic_stepsize = 0.3
+        slothy.config.split_heuristic_repeat = 1
+        slothy.config.split_heuristic_optimize_seam = 5
+
+        slothy.optimize(start="loop", end="end_loop")
+
+        slothy.config.outputs = ["hint_STACK_OFFSET_COUNT"]
+        slothy.optimize(start="initial_round_start", end="initial_round_end")
+
+class neon_keccak_x4_hybrid_no_symbolic(Example):
+    def __init__(self, var="v84a", arch=AArch64_Neon, target=Target_CortexA55):
+        name = f"keccak_f1600_x4_{var}_hybrid_slothy_no_symbolic"
+        infile = f"keccak_f1600_x4_{var}_hybrid_slothy_symbolic"
+        outfile = f"examples/naive/aarch64/keccak_f1600_x4_{var}_hybrid_slothy_clean.s"
+
+        super().__init__(infile, name, outfile=outfile, rename=f"keccak_f1600_x4_{var}_hybrid_no_symbolic", arch=arch, target=target)
+
+    def core(self, slothy):
+        slothy.config.reserved_regs = ["x18", "sp"]
+
+        slothy.config.inputs_are_outputs = True
+        slothy.config.variable_size = True
+        slothy.config.visualize_expected_performance = False
+        slothy.config.timeout = 10800
+
+        slothy.config.selfcheck_failure_logfile = "selfcheck_fail.log"
+
+        slothy.config.outputs = ["flags"]
+        slothy.config.constraints.stalls_first_attempt = 64
+        slothy.config.ignore_objective = True
+        slothy.config.constraints.functional_only = True
+        slothy.config.constraints.allow_reordering = False
+        slothy.config.constraints.allow_spills = True
+        slothy.config.visualize_expected_performance = True
+
+        slothy.optimize(start="loop", end="loop_end")
+        slothy.config.outputs = ["hint_STACK_OFFSET_COUNT"]
+        slothy.optimize(start="initial", end="loop")
+
+class neon_keccak_x4_hybrid_interleave(Example):
+    def __init__(self, var="v84a", arch=AArch64_Neon, target=Target_CortexA55):
+        name = f"keccak_f1600_x4_{var}_hybrid_slothy_interleave"
+        infile = f"keccak_f1600_x4_{var}_hybrid_slothy_clean"
+        outfile = f"examples/naive/aarch64/keccak_f1600_x4_{var}_hybrid_slothy_interleaved.s"
+
+        super().__init__(infile, name, outfile=outfile, rename=f"keccak_f1600_x4_{var}_hybrid_slothy_interleaved",
+                         arch=arch, target=target, outfile_full=True)
+
+    def core(self, slothy):
+        slothy.config.reserved_regs = ["x18", "sp"]
+
+        slothy.config.inputs_are_outputs = True
+        slothy.config.variable_size = True
+        slothy.config.visualize_expected_performance = False
+        slothy.config.timeout = 10800
+
+        slothy.config.selfcheck_failure_logfile = "selfcheck_fail.log"
+
+        slothy.config.outputs = ["flags", "hint_STACK_OFFSET_COUNT"]
+        slothy.config.constraints.stalls_first_attempt = 64
+        slothy.config.ignore_objective = True
+        slothy.config.constraints.functional_only = True
+        slothy.config.constraints.allow_reordering = False
+        slothy.config.constraints.allow_spills = True
+        slothy.config.visualize_expected_performance = True
+
+        slothy.config.split_heuristic = True
+        slothy.config.split_heuristic_repeat = 0
+        slothy.config.split_heuristic_preprocess_naive_interleaving = True
+        slothy.config.split_heuristic_preprocess_naive_interleaving_strategy = "alternate"
+        slothy.config.split_heuristic_estimate_performance = False
+        slothy.config.absorb_spills = False
+
+        slothy.optimize(start="loop", end="loop_end")
+
 #############################################################################################
 
 
@@ -1497,6 +1633,15 @@ def main():
                  fft_floatingpoint_radix4(),
                  # Fixed point
                  fft_fixedpoint_radix4(),
+                 # Keccak
+                 neon_keccak_x1_no_symbolic(),
+                 neon_keccak_x1_scalar_opt(),
+                 neon_keccak_x4_hybrid_no_symbolic(var="v84a"),
+                 neon_keccak_x4_hybrid_interleave(var="v84a"),
+                 neon_keccak_x4_hybrid_no_symbolic(var="v8a"),
+                 neon_keccak_x4_hybrid_interleave(var="v8a"),
+                 neon_keccak_x4_hybrid_no_symbolic(var="v8a_v84a"),
+                 neon_keccak_x4_hybrid_interleave(var="v8a_v84a"),
                  ]
 
     all_example_names = [e.name for e in examples]