Merge pull request #130 from slothy-optimizer/dilithium-m7

Cortex-M7 Dilithium Part 1: Non-NTT functions

example.py

@@ -1551,6 +1551,174 @@ def core(self, slothy):
 #############################################################################################
 
 
+class pointwise_montgomery_dilithium(Example):
+    def __init__(self, var="", arch=Arch_Armv7M, target=Target_CortexM7, timeout=None):
+        name = "pointwise_montgomery_dilithium"
+        infile = name
+        funcname = "pqcrystals_dilithium_asm_pointwise_montgomery"
+
+        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 = ["r14", "r12"]
+        slothy.config.inputs_are_outputs = True
+        slothy.config.sw_pipelining.enabled = True
+
+        slothy.optimize_loop("1")
+
+class pointwise_acc_montgomery_dilithium(Example):
+    def __init__(self, var="", arch=Arch_Armv7M, target=Target_CortexM7, timeout=None):
+        name = "pointwise_acc_montgomery_dilithium"
+        infile = name
+        funcname = "pqcrystals_dilithium_asm_pointwise_acc_montgomery"
+
+        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 = ["r12"]
+        slothy.config.inputs_are_outputs = True
+        slothy.config.sw_pipelining.enabled = True
+
+        slothy.optimize_loop("1")
+
+
+class basemul_257_dilithium(Example):
+    def __init__(self, var="", arch=Arch_Armv7M, target=Target_CortexM7, timeout=None):
+        name = "basemul_257_dilithium"
+        infile = name
+        funcname = "__asm_point_mul_257_16"
+
+        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 = ["r12", "r14"]
+        slothy.config.inputs_are_outputs = True
+
+        slothy.config.sw_pipelining.enabled = True
+        slothy.optimize_loop("_point_mul_16_loop")
+
+class basemul_257_asymmetric_dilithium(Example):
+    def __init__(self, var="", arch=Arch_Armv7M, target=Target_CortexM7, timeout=None):
+        name = "basemul_257_asymmetric_dilithium"
+        infile = name
+        funcname = "__asm_asymmetric_mul_257_16"
+
+        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 = ["r14", "r12"]
+        slothy.config.inputs_are_outputs = True
+
+        slothy.config.sw_pipelining.enabled = True
+        slothy.config.unsafe_address_offset_fixup = False
+        slothy.optimize_loop("_asymmetric_mul_16_loop")
+
+
+
+class pointwise_769_dilithium(Example):
+    def __init__(self, var="", arch=Arch_Armv7M, target=Target_CortexM7, timeout=None):
+        name = "pointwise_769_dilithium"
+        infile = name
+        funcname = "small_pointmul_asm_769"
+
+        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.inputs_are_outputs = True
+        slothy.config.variable_size = True
+
+        r = slothy.config.reserved_regs
+        r.add("r3")
+        slothy.config.reserved_regs = r
+        slothy.config.sw_pipelining.enabled = True
+        slothy.config.constraints.stalls_first_attempt = 16
+        slothy.optimize_loop("_point_mul_16_loop", forced_loop_type=Arch_Armv7M.CmpLoop)
+
+
+class pointwise_769_asymmetric_dilithium(Example):
+    def __init__(self, var="", arch=Arch_Armv7M, target=Target_CortexM7, timeout=None):
+        name = "pointwise_769_asymmetric_dilithium"
+        infile = name
+        funcname = "small_asymmetric_mul_asm_769"
+
+        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 = ["r10"]
+        slothy.config.inputs_are_outputs = True
+
+        slothy.config.sw_pipelining.enabled = True
+        slothy.optimize_loop("_asymmetric_mul_16_loop")
+
+class reduce32_dilithium(Example):
+    def __init__(self, var="", arch=Arch_Armv7M, target=Target_CortexM7, timeout=None):
+        name = "reduce32_dilithium"
+        infile = name
+        funcname = "pqcrystals_dilithium_asm_reduce32"
+
+        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 = ["r10"]
+        slothy.config.inputs_are_outputs = True
+        slothy.config.constraints.stalls_first_attempt = 4
+        slothy.config.sw_pipelining.enabled = True
+        slothy.optimize_loop("1")
+
+class caddq_dilithium(Example):
+    def __init__(self, var="", arch=Arch_Armv7M, target=Target_CortexM7, timeout=None):
+        name = "caddq_dilithium"
+        infile = name
+        funcname = "pqcrystals_dilithium_asm_caddq"
+
+        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 = ["r10"]
+        slothy.config.inputs_are_outputs = True
+        slothy.config.sw_pipelining.enabled = True
+        slothy.optimize_loop("1")
 def main():
     examples = [ Example0(),
                  Example1(),
@@ -1706,6 +1874,15 @@ def main():
                  fft_floatingpoint_radix4(),
                  # Fixed point
                  fft_fixedpoint_radix4(),
+
+                 pointwise_montgomery_dilithium(),
+                 pointwise_acc_montgomery_dilithium(),
+                 basemul_257_dilithium(),
+                 basemul_257_asymmetric_dilithium(),
+                 pointwise_769_dilithium(),
+                 pointwise_769_asymmetric_dilithium(),
+                 reduce32_dilithium(),
+                 caddq_dilithium(),
                  ]
 
     all_example_names = [e.name for e in examples]

examples/naive/armv7m/basemul_257_asymmetric_dilithium.s

@@ -0,0 +1,45 @@
+// 2
+.macro barrett_32 a, Qbar, Q, tmp
+    smmulr \tmp, \a, \Qbar
+    mls \a, \tmp, \Q, \a
+.endm
+
+.syntax unified
+.cpu cortex-m4
+
+.align 2
+.global __asm_asymmetric_mul_257_16
+.type __asm_asymmetric_mul_257_16, %function
+__asm_asymmetric_mul_257_16:
+    push.w {r4-r11, lr}
+
+    .equ width, 4
+
+    add.w r12, r0, #256*width
+    _asymmetric_mul_16_loop:
+
+    ldr.w r7, [r1, #width]
+    ldr.w r4, [r1], #2*width
+    ldr.w r8, [r2, #width]
+    ldr.w r5, [r2], #2*width
+    ldr.w r9, [r3, #width]
+    ldr.w r6, [r3], #2*width
+
+    smuad r10, r4, r6
+    smuadx r11, r4, r5
+
+    str.w r11, [r0, #width]
+    str.w r10, [r0], #2*width // @slothy:core=True
+
+    smuad r10, r7, r9
+    smuadx r11, r7, r8
+
+    str.w r11, [r0, #width]
+    str.w r10, [r0], #2*width // @slothy:core=True
+
+    cmp.w r0, r12
+    bne.w _asymmetric_mul_16_loop
+
+    pop.w {r4-r11, pc}
+
+.size __asm_asymmetric_mul_257_16, .-__asm_asymmetric_mul_257_16

examples/naive/armv7m/basemul_257_dilithium.s

@@ -0,0 +1,61 @@
+// 2
+.macro barrett_32 a, Qbar, Q, tmp
+    smmulr \tmp, \a, \Qbar
+    mls \a, \tmp, \Q, \a
+.endm
+
+.syntax unified
+.cpu cortex-m4
+
+.align 2
+.global __asm_point_mul_257_16
+.type __asm_point_mul_257_16, %function
+__asm_point_mul_257_16:
+    push.w {r4-r11, lr}
+
+    ldr.w r14, [sp, #36]
+
+    .equ width, 4
+
+    add.w r12, r14, #64*width
+    _point_mul_16_loop:
+
+    ldr.w r7, [r1, #2*width]
+    ldr.w r8, [r1, #3*width]
+    ldr.w r9, [r14, #1*width]
+    ldr.w r5, [r1, #1*width]
+    ldr.w r4, [r1], #4*width
+    ldr.w r6, [r14], #2*width
+
+    smultb r10, r4, r6
+    barrett_32 r10, r2, r3, r11
+    pkhbt r4, r4, r10, lsl #16
+
+    neg r6, r6
+
+    smultb r10, r5, r6
+    barrett_32 r10, r2, r3, r11
+    pkhbt r5, r5, r10, lsl #16
+
+    str.w r5, [r0, #1*width]
+    str.w r4, [r0], #2*width
+
+    smultb r10, r7, r9
+    barrett_32 r10, r2, r3, r11
+    pkhbt r7, r7, r10, lsl #16
+
+    neg r9, r9
+
+    smultb r10, r8, r9
+    barrett_32 r10, r2, r3, r11
+    pkhbt r8, r8, r10, lsl #16
+
+    str.w r8, [r0, #1*width]
+    str.w r7, [r0], #2*width
+
+    cmp.w r14, r12
+    bne.w _point_mul_16_loop
+
+    pop.w {r4-r11, pc}
+
+.size __asm_point_mul_257_16, .-__asm_point_mul_257_16

examples/naive/armv7m/caddq_dilithium.s

@@ -0,0 +1,52 @@
+.syntax unified
+.thumb
+
+.macro caddq a, tmp, q
+    and     \tmp, \q, \a, asr #31
+    add     \a, \a, \tmp
+.endm
+
+// void asm_caddq(int32_t a[N]);
+.global pqcrystals_dilithium_asm_caddq
+.type pqcrystals_dilithium_asm_caddq, %function
+.align 2
+pqcrystals_dilithium_asm_caddq:
+    push {r4-r11, r14}
+
+    movw r12,#:lower16:8380417
+    movt r12,#:upper16:8380417
+
+    movw r10, #32
+    1:
+        ldr.w r1, [r0]
+        ldr.w r2, [r0, #1*4]
+        ldr.w r3, [r0, #2*4]
+        ldr.w r4, [r0, #3*4]
+        ldr.w r5, [r0, #4*4]
+        ldr.w r6, [r0, #5*4]
+        ldr.w r7, [r0, #6*4]
+        ldr.w r8, [r0, #7*4]
+
+        caddq r1, r9, r12
+        caddq r2, r9, r12
+        caddq r3, r9, r12
+        caddq r4, r9, r12
+        caddq r5, r9, r12
+        caddq r6, r9, r12
+        caddq r7, r9, r12
+        caddq r8, r9, r12
+
+        str.w r2, [r0, #1*4]
+        str.w r3, [r0, #2*4]
+        str.w r4, [r0, #3*4]
+        str.w r5, [r0, #4*4]
+        str.w r6, [r0, #5*4]
+        str.w r7, [r0, #6*4]
+        str.w r8, [r0, #7*4]
+        str r1, [r0], #8*4
+        subs r10, #1
+        bne.w 1b
+
+    pop {r4-r11, pc}
+
+.size pqcrystals_dilithium_asm_caddq, .-pqcrystals_dilithium_asm_caddq