Skip to content

Commit

Permalink
Merge remote-tracking branch 'origin/main' into armv7m
Browse files Browse the repository at this point in the history
  • Loading branch information
@mkannwischer
mkannwischer committed Jan 7, 2025
2 parents c31d0b3 + 0f6b99d commit 11516d5
Show file tree
Hide file tree
Showing 9 changed files with 3,225 additions and 2,951 deletions.
143 changes: 18 additions & 125 deletions example.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -1868,18 +1868,16 @@ def core(self, slothy):
slothy.config.constraints.stalls_first_attempt = 32

r = slothy.config.reserved_regs
r = r.union(f"s{i}" for i in range(30)) # reserve FPR
r = r.union(f"s{i}" for i in range(31)) # reserve FPR
slothy.config.reserved_regs = r

### TODO
# - Experiment with lower split factors
# - Try to get stable performance: It currently varies a lot with each run

slothy.config.unsafe_address_offset_fixup = False
slothy.config.constraints.stalls_first_attempt = 16
slothy.config.variable_size = True
slothy.config.split_heuristic = True
slothy.config.constraints.stalls_precision = 1
slothy.config.timeout = 360 # Not more than 2min per step
slothy.config.split_heuristic_factor = 1
slothy.config.visualize_expected_performance = False
Expand All @@ -1891,39 +1889,19 @@ def core(self, slothy):

slothy.config.outputs = ["r14"]

slothy.config.absorb_spills = True
slothy.config.unsafe_address_offset_fixup = False
slothy.fusion_loop("layer567_loop", ssa=True)
slothy.fusion_loop("layer567_loop", ssa=False)
slothy.config.unsafe_address_offset_fixup = True

slothy.config.outputs = ["r14"]
slothy.config.constraints.functional_only = True
slothy.config.unsafe_address_offset_fixup = False
slothy.config.constraints.allow_reordering = False
slothy.config.inputs_are_outputs = True
slothy.config.split_heuristic = False
slothy.config.constraints.stalls_first_attempt = 64
slothy.config.constraints.allow_spills = True
slothy.config.absorb_spills = True
slothy.config.constraints.spill_type = { 'spill_to_vreg': 26 }
slothy.config.constraints.minimize_spills = True
slothy.config.objective_lower_bound = 2 # <2 stalls doesn't seem possible
slothy.optimize_loop("layer567_loop")

slothy.config.timeout = 360
slothy.config.constraints.maximize_register_lifetimes = False
slothy.config.variable_size = True
slothy.config.split_heuristic_optimize_seam = 0
slothy.config.split_heuristic = True
slothy.config.split_heuristic_repeat = 1
slothy.config.split_heuristic_factor = 2.25
slothy.config.split_heuristic_stepsize = 0.25
slothy.config.constraints.allow_spills = False
slothy.config.constraints.minimize_spills = False
slothy.config.absorb_spills = False
slothy.config.constraints.stalls_precision = 1
# slothy.config.unsafe_address_offset_fixup = True
slothy.config.constraints.functional_only = False
slothy.config.constraints.allow_reordering = True
slothy.optimize_loop("layer567_loop")

slothy.config.split_heuristic_optimize_seam = 6
Expand Down Expand Up @@ -2113,7 +2091,6 @@ def core(self, slothy):
slothy.optimize(start="slothy_start", end="slothy_end")



class ntt_kyber(Example):
def __init__(self, var="", arch=Arch_Armv7M, target=Target_CortexM7, timeout=None):
name = f"ntt_kyber"
Expand All @@ -2137,122 +2114,35 @@ def core(self, slothy):

slothy.config.inputs_are_outputs = True
slothy.config.variable_size = True
slothy.config.constraints.stalls_first_attempt = 32

slothy.config.unsafe_address_offset_fixup = False

### TODO
# - Experiment with lower split factors
# - Try to get stable performance: It currently varies a lot with each run

slothy.config.unsafe_address_offset_fixup = False
slothy.config.constraints.stalls_first_attempt = 16
slothy.config.variable_size = True
slothy.config.split_heuristic = True
slothy.config.constraints.stalls_precision = 1
slothy.config.timeout = 360 # Not more than 2min per step
slothy.config.split_heuristic_factor = 1
slothy.config.timeout = 360 # Not more than 6min per step
slothy.config.visualize_expected_performance = False
slothy.config.split_heuristic_factor = 6
slothy.config.split_heuristic_stepsize = 0.15
slothy.optimize_loop("1")
slothy.config.split_heuristic_stepsize = 0.1
slothy.optimize_loop("1", forced_loop_type=Arch_Armv7M.BranchLoop)
slothy.config.split_heuristic_optimize_seam = 6
slothy.optimize_loop("1")
slothy.optimize_loop("1", forced_loop_type=Arch_Armv7M.BranchLoop)

slothy.config.outputs = ["r14"]

slothy.config.absorb_spills = True
slothy.config.unsafe_address_offset_fixup = False
slothy.fusion_loop("2", ssa=True)

slothy.config.outputs = ["r14"]
slothy.config.constraints.functional_only = True
slothy.config.unsafe_address_offset_fixup = False
slothy.config.constraints.allow_reordering = False
slothy.config.inputs_are_outputs = True
slothy.config.split_heuristic = False
slothy.config.constraints.stalls_first_attempt = 64
slothy.config.constraints.allow_spills = True
slothy.config.absorb_spills = True
slothy.config.constraints.spill_type = { 'spill_to_vreg': 26 }
slothy.config.constraints.minimize_spills = True
slothy.config.objective_lower_bound = 2 # <2 stalls doesn't seem possible
slothy.optimize_loop("2")
slothy.fusion_loop("2", ssa=False, forced_loop_type=Arch_Armv7M.BranchLoop)

slothy.config.timeout = 360
slothy.config.constraints.maximize_register_lifetimes = False
slothy.config.variable_size = True
slothy.config.split_heuristic_optimize_seam = 0
slothy.config.split_heuristic = True
slothy.config.split_heuristic_repeat = 1
slothy.config.split_heuristic_factor = 4
slothy.config.split_heuristic_stepsize = 0.25
slothy.config.constraints.allow_spills = False
slothy.config.constraints.minimize_spills = False
slothy.config.absorb_spills = False
slothy.config.constraints.stalls_precision = 1
slothy.config.split_heuristic_stepsize = 0.1
slothy.config.unsafe_address_offset_fixup = True
slothy.config.constraints.functional_only = False
slothy.config.constraints.allow_reordering = True
slothy.optimize_loop("2")
slothy.optimize_loop("2", forced_loop_type=Arch_Armv7M.BranchLoop)

slothy.config.split_heuristic_optimize_seam = 6
slothy.optimize_loop("2")

class ntt_kyber_symbolic(Example):
def __init__(self, var="", arch=Arch_Armv7M, target=Target_CortexM7, timeout=None):
name = f"ntt_kyber_symbolic"
infile = name
funcname = "ntt_fast_symbolic"

if var != "":
name += f"_{var}"
infile += f"_{var}"
name += f"_{target_label_dict[target]}"

super().__init__(infile, name, rename=True, arch=arch, target=target, timeout=timeout, funcname=funcname)

def core(self, slothy):
slothy.config.outputs = ["s23"]
slothy.config.reserved_regs = ["r13", "s25", "s26", "s27", "s28", "s29", "s30", "s31"]
slothy.config.inputs_are_outputs = True
slothy.config.variable_size = True
slothy.config.constraints.stalls_first_attempt = 16
orig_functional_only = slothy.config.constraints.functional_only
orig_allow_reordering = slothy.config.constraints.allow_reordering

# Step 1: find minimum number of stack spills in first loop
slothy.config.objective_lower_bound = 8
slothy.config.constraints.functional_only = True
slothy.config.constraints.allow_spills = True
slothy.config.constraints.minimize_spills = True
# TODO: it would be much better if we could allow re-ordering; but that seems out of reach for SLOTHY right now
slothy.config.constraints.allow_reordering = False
slothy.optimize_loop("1")
slothy.config.constraints.functional_only = orig_functional_only
slothy.config.constraints.allow_spills = False
slothy.config.constraints.allow_reordering = orig_allow_reordering
slothy.config.absorb_spills = False

# Step 2: optimize first loop
# TODO: use a small factor and larger repeat
slothy.config.sw_pipelining.halving_heuristic = True
slothy.config.split_heuristic = True
slothy.config.split_heuristic_factor = 4
slothy.config.split_heuristic_stepsize = 0.15
slothy.config.split_heuristic_repeat = 1
slothy.optimize_loop("1")

# Step 3: optimize second loop
# TODO: use a small factor and larger repeat
slothy.config.split_heuristic = True
slothy.config.sw_pipelining.enabled = True
slothy.config.sw_pipelining.halving_heuristic = True
slothy.config.split_heuristic_factor = 2
slothy.config.split_heuristic_repeat = 1
slothy.fusion_loop("2",ssa=False)
slothy.optimize_loop("2")


slothy.optimize_loop("2", forced_loop_type=Arch_Armv7M.BranchLoop)


class intt_kyber(Example):
Expand Down Expand Up @@ -2281,7 +2171,7 @@ def core(self, slothy):
slothy.config.split_heuristic_stepsize = 0.15
slothy.config.split_heuristic_repeat = 1
slothy.config.outputs = ["r14", "s8"]
slothy.optimize_loop("1")
slothy.optimize_loop("1", forced_loop_type=Arch_Armv7M.BranchLoop)

# Step 2: optimize the start of the second loop
slothy.config.split_heuristic = True
Expand All @@ -2299,9 +2189,10 @@ def core(self, slothy):
slothy.config.split_heuristic_stepsize = 0.2
slothy.config.outputs = ["r14", "s14"]
slothy.config.unsafe_address_offset_fixup = False
slothy.fusion_loop("2", ssa=False)
slothy.fusion_loop("2", ssa=False, forced_loop_type=Arch_Armv7M.BranchLoop)
slothy.config.unsafe_address_offset_fixup = True
slothy.optimize_loop("2")
slothy.optimize_loop("2", forced_loop_type=Arch_Armv7M.BranchLoop)


class basemul_16_32_kyber(Example):
def __init__(self, var="", arch=Arch_Armv7M, target=Target_CortexM7, timeout=None):
Expand Down Expand Up @@ -3049,6 +2940,8 @@ def main():
Keccak(var="adomnicai_m4"),
Keccak(var="adomnicai_m7"),

ntt_kyber(),
intt_kyber(),
basemul_16_32_kyber(),
basemul_acc_32_32_kyber(),
basemul_acc_32_16_kyber(),
Expand Down
8 changes: 4 additions & 4 deletions examples/naive/armv7m/intt_kyber.s
View file
Original file line number Diff line number Diff line change
Expand Up @@ -303,10 +303,10 @@ invntt_fast:
uadd16 tmp, poly0, poly1
usub16 twiddle2, poly0, poly1
str.w twiddle2, [poly, #offset]
str.w tmp, [poly], #strincr // increase 2*8*4 = 64 (2 * 8 loads of 4 bytes each)
str.w tmp, [poly], #strincr // @slothy:core // @slothy:before=cmp // increase 2*8*4 = 64 (2 * 8 loads of 4 bytes each)
//0,2,4,6: < 1.5q
vmov tmp, s8
cmp.w poly, tmp
cmp.w poly, tmp // @slothy:id=cmp
bne.w 1b

sub.w poly, #8*strincr
Expand Down Expand Up @@ -360,10 +360,10 @@ invntt_fast:
str.w poly1, [poly, #distance2/4]
str.w poly2, [poly, #2*distance2/4]
str.w poly3, [poly, #3*distance2/4]
str.w poly0, [poly], #4
str.w poly0, [poly], #4 // @slothy:core // @slothy:before=cmp

vmov tmp, s14
cmp.w poly, tmp
cmp.w poly, tmp // @slothy:id=cmp
bne.w 2b

vpop.w {s16-s23}
Expand Down
8 changes: 4 additions & 4 deletions examples/naive/armv7m/ntt_769_dilithium.s
View file
Original file line number Diff line number Diff line change
Expand Up @@ -268,22 +268,22 @@ small_ntt_asm_769:
vmov s13, tmp
vmov twiddle_ptr, s25
layer567_loop:
vmov s23, poly // @slothy:is_spill
vmov s23, poly
load poly, poly0, poly1, poly2, poly3, #0, #distance2/4, #2*distance2/4, #3*distance2/4
load poly, poly4, poly5, poly6, poly7, #distance2, #5*distance2/4, #6*distance2/4, #7*distance2/4

movw qa, #24608
_3_layer_double_CT_16_plant poly0, poly1, poly2, poly3, poly4, poly5, poly6, poly7, twiddle1, twiddle2, twiddle_ptr, q, qa, tmp

vmov poly, s23 // @slothy:is_restore
vmov poly, s23
store poly, poly4, poly5, poly6, poly7, #distance2, #5*distance2/4, #6*distance2/4, #7*distance2/4
str.w poly1, [poly, #distance2/4]
str.w poly2, [poly, #2*distance2/4]
str.w poly3, [poly, #3*distance2/4]
str.w poly0, [poly], #strincr2 // @slothy:core
str.w poly0, [poly], #strincr2 // @slothy:core // @slothy:before=cmp

vmov tmp, s13
cmp.w poly, tmp
cmp.w poly, tmp // @slothy:id=cmp
bne.w layer567_loop
vpop.w {s16-s27}
pop {r4-r11, pc}
Expand Down
15 changes: 7 additions & 8 deletions examples/naive/armv7m/ntt_kyber.s
View file
Original file line number Diff line number Diff line change
Expand Up @@ -267,10 +267,10 @@ ntt_fast:
uadd16 tmp, poly0, poly1
usub16 twiddle1, poly0, poly1
str.w twiddle1, [poly, #offset]
str.w tmp, [poly], #4
str.w tmp, [poly], #4 // @slothy:core // @slothy:before=cmp

vmov tmp, s24
cmp.w poly, tmp
cmp.w poly, tmp // @slothy:id=cmp
bne.w 1b

sub.w poly, #8*strincr
Expand All @@ -283,23 +283,22 @@ ntt_fast:
add.w tmp, poly, #strincr2*16
vmov s13, tmp
2:
vmov s23, poly // @slothy:is_spill
vmov s23, poly
load poly, poly0, poly1, poly2, poly3, #0, #distance2/4, #2*distance2/4, #3*distance2/4
load poly, poly4, poly5, poly6, poly7, #distance2, #5*distance2/4, #6*distance2/4, #7*distance2/4

movw qa, #26632
_3_layer_double_CT_16_plant poly0, poly1, poly2, poly3, poly4, poly5, poly6, poly7, twiddle1, twiddle2, twiddle_ptr, q, qa, tmp

vmov poly, s23 // @slothy:is_restore
vmov poly, s23
store poly, poly4, poly5, poly6, poly7, #distance2, #5*distance2/4, #6*distance2/4, #7*distance2/4
str.w poly1, [poly, #distance2/4]
str.w poly2, [poly, #2*distance2/4]
str.w poly3, [poly, #3*distance2/4]
str.w poly0, [poly], #strincr2

vmov tmp, s13
str.w poly0, [poly], #strincr2 // @slothy:core // @slothy:before=cmp

cmp.w poly, tmp
vmov tmp, s13
cmp.w poly, tmp // @slothy:id=cmp
bne.w 2b
vpop.w {s16-s26}
pop {r4-r11, pc}
Expand Down
Loading

0 comments on commit 11516d5

Please sign in to comment.