aboutsummaryrefslogtreecommitdiffstats
path: root/riscV/OpWeights.ml
diff options
context:
space:
mode:
Diffstat (limited to 'riscV/OpWeights.ml')
-rw-r--r--riscV/OpWeights.ml335
1 files changed, 163 insertions, 172 deletions
diff --git a/riscV/OpWeights.ml b/riscV/OpWeights.ml
index 35ae81e6..0a1d9ad4 100644
--- a/riscV/OpWeights.ml
+++ b/riscV/OpWeights.ml
@@ -1,177 +1,168 @@
-open Op;;
-open PrepassSchedulingOracleDeps;;
-
-module Rocket =
- struct
- (* Attempt at modeling the Rocket core *)
-
- let resource_bounds = [| 1 |];;
- let nr_non_pipelined_units = 1;; (* divider *)
-
- let latency_of_op (op : operation) (nargs : int) =
- match op with
- | Omul | Omulhs | Omulhu
- | Omull | Omullhs | Omullhu -> 4
-
- | Onegf -> 1 (*r [rd = - r1] *)
- | Oabsf (*r [rd = abs(r1)] *)
- | Oaddf (*r [rd = r1 + r2] *)
- | Osubf (*r [rd = r1 - r2] *)
- | Omulf -> 6 (*r [rd = r1 * r2] *)
- | Onegfs -> 1 (*r [rd = - r1] *)
- | Oabsfs (*r [rd = abs(r1)] *)
- | Oaddfs (*r [rd = r1 + r2] *)
- | Osubfs (*r [rd = r1 - r2] *)
- | Omulfs -> 4 (*r [rd = r1 * r2] *)
- | Osingleoffloat (*r [rd] is [r1] truncated to single-precision float *)
- | Ofloatofsingle -> 4 (*r [rd] is [r1] extended to double-precision float *)
- (*c Conversions between int and float: *)
- | Ointoffloat (*r [rd = signed_int_of_float64(r1)] *)
- | Ointuoffloat (*r [rd = unsigned_int_of_float64(r1)] *)
- | Ofloatofint (*r [rd = float64_of_signed_int(r1)] *)
- | Ofloatofintu -> 6 (*r [rd = float64_of_unsigned_int(r1)] *)
- | Ointofsingle (*r [rd = signed_int_of_float32(r1)] *)
- | Ointuofsingle (*r [rd = unsigned_int_of_float32(r1)] *)
- | Osingleofint (*r [rd = float32_of_signed_int(r1)] *)
- | Osingleofintu -> 4 (*r [rd = float32_of_unsigned_int(r1)] *)
- | Olongoffloat (*r [rd = signed_long_of_float64(r1)] *)
- | Olonguoffloat (*r [rd = unsigned_long_of_float64(r1)] *)
- | Ofloatoflong (*r [rd = float64_of_signed_long(r1)] *)
- | Ofloatoflongu -> 6 (*r [rd = float64_of_unsigned_long(r1)] *)
- | Olongofsingle (*r [rd = signed_long_of_float32(r1)] *)
- | Olonguofsingle (*r [rd = unsigned_long_of_float32(r1)] *)
- | Osingleoflong (*r [rd = float32_of_signed_long(r1)] *)
- | Osingleoflongu -> 4 (*r [rd = float32_of_unsigned_int(r1)] *)
-
- | Odiv | Odivu | Odivl | Odivlu -> 16
- | Odivfs -> 35
- | Odivf -> 50
-
- | Ocmp cond ->
- (match cond with
- | Ccomp _
- | Ccompu _
- | Ccompimm _
- | Ccompuimm _
- | Ccompl _
- | Ccomplu _
- | Ccomplimm _
- | Ccompluimm _
- | CEbeqw _
- | CEbnew _
- | CEbequw _
- | CEbneuw _
- | CEbltw _
- | CEbltuw _
- | CEbgew _
- | CEbgeuw _
- | CEbeql _
- | CEbnel _
- | CEbequl _
- | CEbneul _
- | CEbltl _
- | CEbltul _
- | CEbgel _
- | CEbgeul _ -> 1
- | Ccompf _
- | Cnotcompf _ -> 6
- | Ccompfs _
- | Cnotcompfs _ -> 4)
- | _ -> 1;;
-
- let resources_of_op (op : operation) (nargs : int) = resource_bounds;;
-
- let non_pipelined_resources_of_op (op : operation) (nargs : int) =
- match op with
- | Odiv | Odivu -> [| 29 |]
- | Odivfs -> [| 20 |]
- | Odivl | Odivlu | Odivf -> [| 50 |]
- | _ -> [| -1 |];;
-
- let resources_of_cond (cond : condition) (nargs : int) = resource_bounds;;
-
- let latency_of_load trap chunk (addr : addressing) (nargs : int) = 3;;
- let latency_of_call _ _ = 6;;
-
- let resources_of_load trap chunk addressing nargs = resource_bounds;;
-
- let resources_of_store chunk addressing nargs = resource_bounds;;
-
- let resources_of_call _ _ = resource_bounds;;
- let resources_of_builtin _ = resource_bounds;;
- end;;
-
-module SweRV_EH1 =
- struct
- (* Attempt at modeling SweRV EH1
- [| issues ; LSU ; multiplier |] *)
- let resource_bounds = [| 2 ; 1; 1 |];;
- let nr_non_pipelined_units = 1;; (* divider *)
-
- let latency_of_op (op : operation) (nargs : int) =
- match op with
- | Omul | Omulhs | Omulhu
- | Omull | Omullhs | Omullhu -> 3
- | Odiv | Odivu | Odivl | Odivlu -> 16
- | _ -> 1;;
-
- let resources_of_op (op : operation) (nargs : int) =
- match op with
- | Omul | Omulhs | Omulhu
- | Omull | Omullhs | Omullhu -> [| 1 ; 0 ; 1 |]
- | Odiv | Odivu | Odivl | Odivlu -> [| 0 ; 0; 0 |]
- | _ -> [| 1; 0; 0 |];;
-
- let non_pipelined_resources_of_op (op : operation) (nargs : int) =
- match op with
- | Odiv | Odivu -> [| 29 |]
- | Odivfs -> [| 20 |]
- | Odivl | Odivlu | Odivf -> [| 50 |]
- | _ -> [| -1 |];;
-
- let resources_of_cond (cond : condition) (nargs : int) = [| 1; 0; 0 |];;
-
- let latency_of_load trap chunk (addr : addressing) (nargs : int) = 3;;
- let latency_of_call _ _ = 6;;
-
- let resources_of_load trap chunk addressing nargs = [| 1; 1; 0 |];;
-
- let resources_of_store chunk addressing nargs = [| 1; 1; 0 |];;
-
- let resources_of_call _ _ = resource_bounds;;
- let resources_of_builtin _ = resource_bounds;;
- end;;
+open Op
+open PrepassSchedulingOracleDeps
+
+module Rocket = struct
+ (* Attempt at modeling the Rocket core *)
+
+ let resource_bounds = [| 1 |]
+
+ let nr_non_pipelined_units = 1
+
+ (* divider *)
+
+ let latency_of_op (op : operation) (nargs : int) =
+ match op with
+ | Omul | Omulhs | Omulhu | Omull | Omullhs | Omullhu -> 4
+ | Onegf -> 1 (*r [rd = - r1] *)
+ | Oabsf (*r [rd = abs(r1)] *)
+ | Oaddf (*r [rd = r1 + r2] *)
+ | Osubf (*r [rd = r1 - r2] *)
+ | Omulf ->
+ 6 (*r [rd = r1 * r2] *)
+ | Onegfs -> 1 (*r [rd = - r1] *)
+ | Oabsfs (*r [rd = abs(r1)] *)
+ | Oaddfs (*r [rd = r1 + r2] *)
+ | Osubfs (*r [rd = r1 - r2] *)
+ | Omulfs ->
+ 4 (*r [rd = r1 * r2] *)
+ | Osingleoffloat (*r [rd] is [r1] truncated to single-precision float *)
+ | Ofloatofsingle (*r [rd] is [r1] extended to double-precision float *)
+ (*c Conversions between int and float: *)
+ | Ofloatconst _ | Osingleconst _
+ | Ointoffloat (*r [rd = signed_int_of_float64(r1)] *)
+ | Ointuoffloat (*r [rd = unsigned_int_of_float64(r1)] *)
+ | Ofloatofint (*r [rd = float64_of_signed_int(r1)] *)
+ | Ofloatofintu (*r [rd = float64_of_unsigned_int(r1)] *)
+ | Ointofsingle (*r [rd = signed_int_of_float32(r1)] *)
+ | Ointuofsingle (*r [rd = unsigned_int_of_float32(r1)] *)
+ | Osingleofint (*r [rd = float32_of_signed_int(r1)] *)
+ | Osingleofintu (*r [rd = float32_of_unsigned_int(r1)] *)
+ | Olongoffloat (*r [rd = signed_long_of_float64(r1)] *)
+ | Olonguoffloat (*r [rd = unsigned_long_of_float64(r1)] *)
+ | Ofloatoflong (*r [rd = float64_of_signed_long(r1)] *)
+ | Ofloatoflongu (*r [rd = float64_of_unsigned_long(r1)] *)
+ | Olongofsingle (*r [rd = signed_long_of_float32(r1)] *)
+ | Olonguofsingle (*r [rd = unsigned_long_of_float32(r1)] *)
+ | Osingleoflong (*r [rd = float32_of_signed_long(r1)] *)
+ | Osingleoflongu ->
+ 2 (*r [rd = float32_of_unsigned_int(r1)] *)
+ | OEfeqd | OEfltd | OEfeqs | OEflts | OEfles | OEfled | Obits_of_single
+ | Obits_of_float | Osingle_of_bits | Ofloat_of_bits ->
+ 2
+ | OEloadli _ -> 2
+ | Odiv | Odivu | Odivl | Odivlu -> 16
+ | Odivfs -> 35
+ | Odivf -> 50
+ | Ocmp cond -> (
+ match cond with
+ | Ccomp _ | Ccompu _ | Ccompimm _ | Ccompuimm _ | Ccompl _ | Ccomplu _
+ | Ccomplimm _ | Ccompluimm _ | CEbeqw _ | CEbnew _ | CEbequw _
+ | CEbneuw _ | CEbltw _ | CEbltuw _ | CEbgew _ | CEbgeuw _ | CEbeql _
+ | CEbnel _ | CEbequl _ | CEbneul _ | CEbltl _ | CEbltul _ | CEbgel _
+ | CEbgeul _ ->
+ 1
+ | Ccompf _ | Cnotcompf _ -> 2
+ | Ccompfs _ | Cnotcompfs _ -> 2)
+ | OEmayundef _ -> 0
+ | _ -> 1
+
+ let resources_of_op (op : operation) (nargs : int) = resource_bounds
+
+ let non_pipelined_resources_of_op (op : operation) (nargs : int) =
+ match op with
+ | Odiv | Odivu -> [| 29 |]
+ | Odivfs -> [| 20 |]
+ | Odivl | Odivlu | Odivf -> [| 50 |]
+ | _ -> [| -1 |]
+
+ let resources_of_cond (cond : condition) (nargs : int) = resource_bounds
+
+ let latency_of_load trap chunk (addr : addressing) (nargs : int) = 3
+
+ let latency_of_call _ _ = 6
+
+ let resources_of_load trap chunk addressing nargs = resource_bounds
+
+ let resources_of_store chunk addressing nargs = resource_bounds
+
+ let resources_of_call _ _ = resource_bounds
+
+ let resources_of_builtin _ = resource_bounds
+end
+
+module SweRV_EH1 = struct
+ (* Attempt at modeling SweRV EH1
+ [| issues ; LSU ; multiplier |] *)
+ let resource_bounds = [| 2; 1; 1 |]
+
+ let nr_non_pipelined_units = 1
+
+ (* divider *)
+
+ let latency_of_op (op : operation) (nargs : int) =
+ match op with
+ | Omul | Omulhs | Omulhu | Omull | Omullhs | Omullhu -> 3
+ | Odiv | Odivu | Odivl | Odivlu -> 16
+ | _ -> 1
+
+ let resources_of_op (op : operation) (nargs : int) =
+ match op with
+ | Omul | Omulhs | Omulhu | Omull | Omullhs | Omullhu -> [| 1; 0; 1 |]
+ | Odiv | Odivu | Odivl | Odivlu -> [| 0; 0; 0 |]
+ | _ -> [| 1; 0; 0 |]
+
+ let non_pipelined_resources_of_op (op : operation) (nargs : int) =
+ match op with
+ | Odiv | Odivu -> [| 29 |]
+ | Odivfs -> [| 20 |]
+ | Odivl | Odivlu | Odivf -> [| 50 |]
+ | _ -> [| -1 |]
+
+ let resources_of_cond (cond : condition) (nargs : int) = [| 1; 0; 0 |]
+
+ let latency_of_load trap chunk (addr : addressing) (nargs : int) = 3
+
+ let latency_of_call _ _ = 6
+
+ let resources_of_load trap chunk addressing nargs = [| 1; 1; 0 |]
+
+ let resources_of_store chunk addressing nargs = [| 1; 1; 0 |]
+
+ let resources_of_call _ _ = resource_bounds
+
+ let resources_of_builtin _ = resource_bounds
+end
let get_opweights () : opweights =
match !Clflags.option_mtune with
| "rocket" | "" ->
- {
- pipelined_resource_bounds = Rocket.resource_bounds;
- nr_non_pipelined_units = Rocket.nr_non_pipelined_units;
- latency_of_op = Rocket.latency_of_op;
- resources_of_op = Rocket.resources_of_op;
- non_pipelined_resources_of_op = Rocket.non_pipelined_resources_of_op;
- latency_of_load = Rocket.latency_of_load;
- resources_of_load = Rocket.resources_of_load;
- resources_of_store = Rocket.resources_of_store;
- resources_of_cond = Rocket.resources_of_cond;
- latency_of_call = Rocket.latency_of_call;
- resources_of_call = Rocket.resources_of_call;
- resources_of_builtin = Rocket.resources_of_builtin
- }
+ {
+ pipelined_resource_bounds = Rocket.resource_bounds;
+ nr_non_pipelined_units = Rocket.nr_non_pipelined_units;
+ latency_of_op = Rocket.latency_of_op;
+ resources_of_op = Rocket.resources_of_op;
+ non_pipelined_resources_of_op = Rocket.non_pipelined_resources_of_op;
+ latency_of_load = Rocket.latency_of_load;
+ resources_of_load = Rocket.resources_of_load;
+ resources_of_store = Rocket.resources_of_store;
+ resources_of_cond = Rocket.resources_of_cond;
+ latency_of_call = Rocket.latency_of_call;
+ resources_of_call = Rocket.resources_of_call;
+ resources_of_builtin = Rocket.resources_of_builtin;
+ }
| "SweRV_EH1" | "EH1" ->
- {
- pipelined_resource_bounds = SweRV_EH1.resource_bounds;
- nr_non_pipelined_units = SweRV_EH1.nr_non_pipelined_units;
- latency_of_op = SweRV_EH1.latency_of_op;
- resources_of_op = SweRV_EH1.resources_of_op;
- non_pipelined_resources_of_op = SweRV_EH1.non_pipelined_resources_of_op;
- latency_of_load = SweRV_EH1.latency_of_load;
- resources_of_load = SweRV_EH1.resources_of_load;
- resources_of_store = SweRV_EH1.resources_of_store;
- resources_of_cond = SweRV_EH1.resources_of_cond;
- latency_of_call = SweRV_EH1.latency_of_call;
- resources_of_call = SweRV_EH1.resources_of_call;
- resources_of_builtin = SweRV_EH1.resources_of_builtin
- }
- | xxx -> failwith (Printf.sprintf "unknown -mtune: %s" xxx);;
+ {
+ pipelined_resource_bounds = SweRV_EH1.resource_bounds;
+ nr_non_pipelined_units = SweRV_EH1.nr_non_pipelined_units;
+ latency_of_op = SweRV_EH1.latency_of_op;
+ resources_of_op = SweRV_EH1.resources_of_op;
+ non_pipelined_resources_of_op = SweRV_EH1.non_pipelined_resources_of_op;
+ latency_of_load = SweRV_EH1.latency_of_load;
+ resources_of_load = SweRV_EH1.resources_of_load;
+ resources_of_store = SweRV_EH1.resources_of_store;
+ resources_of_cond = SweRV_EH1.resources_of_cond;
+ latency_of_call = SweRV_EH1.latency_of_call;
+ resources_of_call = SweRV_EH1.resources_of_call;
+ resources_of_builtin = SweRV_EH1.resources_of_builtin;
+ }
+ | xxx -> failwith (Printf.sprintf "unknown -mtune: %s" xxx)