def k_loop_simple(self, code):
kc, mr, nr = self.block.kc, self.block.mr, self.block.nr
a, b, c = self.a, self.b, self.c
p_tA, p_tB = self.p_tA, self.p_tB
# Inner loop over k
for k in syn_iter(code, kc,
mode=CTR): # syn_range(code, 0, kc * 8, 8):
# Load the next values from tA and tB -- generating loops
for ai in range(mr):
a[ai].load(p_tA, ai * self.A_strides.row)
for bj in range(nr):
b[bj].load(p_tB, bj * self.B_strides.col)
# Update c -- generating loop
for ci in range(mr):
for cj in range(nr):
c[ci][cj].v = ppcvar.fmadd(a[ci], b[cj], c[ci][cj])
p_tA.v = p_tA + self.A_strides.col
p_tB.v = p_tB + self.B_strides.row
# /end for k
return
def _pack_b(self, code):
kc, nc, = self.block.kc, self.block.nc
vb, vB, vtB, vBi, bij, tbji, vN = (
self.vb, self.vB, self.vtB, self.vBi, self.bij, self.tbji, self.vN)
for i in syn_iter(code, kc):
vBi.v = i * vN
for j in syn_iter(code, nc):
bij.v = (vBi + j) * 8
tbji.v = (j * kc + i) * 8
vb.load(vB, bij)
vb.store(vtB, tbji)
return
def _pack_b(self, code):
kc, nc, = self.block.kc, self.block.nc
vb, vB, vtB, vBi, bij, tbji, vN = (
self.vb, self.vB, self.vtB, self.vBi, self.bij, self.tbji, self.vN)
for i in syn_iter(code, kc):
vBi.v = i * vN
for j in syn_iter(code, nc):
bij.v = (vBi + j) * 8
tbji.v = (j * kc + i) * 8
vb.load(vB, bij)
vb.store(vtB, tbji)
return
def k_loop_prefetch_simple(self, code):
kc, mr, nr = self.block.kc, self.block.mr, self.block.nr
a, b, c = self.a, self.b, self.c
a_pre, b_pre = self.a_pre, self.b_pre
p_tA, p_tB = self.p_tA, self.p_tB
# Load the next values from tA and tB
for ai in range(mr):
a[ai].load(p_tA, ai * self.A_strides.row)
for bj in range(nr):
b[bj].load(p_tB, bj * self.B_strides.col)
p_tA.v = p_tA + self.A_strides.col
p_tB.v = p_tB + self.B_strides.row
# Inner loop over k
for k in syn_iter(code, kc / 2, mode = CTR): # syn_range(code, 0, kc * 8, 8):
# Iteration 1 -- load [a,b]_pre, compute [a,b]
# Load the prefetch values from tA and tB
for ai in range(mr):
a_pre[ai].load(p_tA, ai * self.A_strides.row)
for bj in range(nr):
b_pre[bj].load(p_tB, bj * self.B_strides.col)
p_tA.v = p_tA + self.A_strides.col
p_tB.v = p_tB + self.B_strides.row
# Update c
for ci in range(mr):
for cj in range(nr):
c[ci][cj].v = ppcvar.fmadd(a[ci], b[cj], c[ci][cj])
# Iteration 2l -- oad [a,b], compute [a,b]_pre
# Load the prefetch values from tA and tB
for ai in range(mr):
a_pre[ai].load(p_tA, ai * self.A_strides.row)
for bj in range(nr):
b_pre[bj].load(p_tB, bj * self.B_strides.col)
p_tA.v = p_tA + self.A_strides.col
p_tB.v = p_tB + self.B_strides.row
# Update c
for ci in range(mr):
for cj in range(nr):
c[ci][cj].v = ppcvar.fmadd(a_pre[ci], b_pre[cj], c[ci][cj])
# /end for k
return
def k_loop_prefetch_simple(self, code):
kc, mr, nr = self.block.kc, self.block.mr, self.block.nr
a, b, c = self.a, self.b, self.c
a_pre, b_pre = self.a_pre, self.b_pre
p_tA, p_tB = self.p_tA, self.p_tB
# Load the next values from tA and tB
for ai in range(mr):
a[ai].load(p_tA, ai * self.A_strides.row)
for bj in range(nr):
b[bj].load(p_tB, bj * self.B_strides.col)
p_tA.v = p_tA + self.A_strides.col
p_tB.v = p_tB + self.B_strides.row
# Inner loop over k
for k in syn_iter(code, kc / 2, mode = CTR): # syn_range(code, 0, kc * 8, 8):
# Iteration 1 -- load [a,b]_pre, compute [a,b]
# Load the prefetch values from tA and tB
for ai in range(mr):
a_pre[ai].load(p_tA, ai * self.A_strides.row)
for bj in range(nr):
b_pre[bj].load(p_tB, bj * self.B_strides.col)
p_tA.v = p_tA + self.A_strides.col
p_tB.v = p_tB + self.B_strides.row
# Update c
for ci in range(mr):
for cj in range(nr):
c[ci][cj].v = ppcvar.fmadd(a[ci], b[cj], c[ci][cj])
# Iteration 2l -- oad [a,b], compute [a,b]_pre
# Load the prefetch values from tA and tB
for ai in range(mr):
a_pre[ai].load(p_tA, ai * self.A_strides.row)
for bj in range(nr):
b_pre[bj].load(p_tB, bj * self.B_strides.col)
p_tA.v = p_tA + self.A_strides.col
p_tB.v = p_tB + self.B_strides.row
# Update c
for ci in range(mr):
for cj in range(nr):
c[ci][cj].v = ppcvar.fmadd(a_pre[ci], b_pre[cj], c[ci][cj])
# /end for k
return
def synthesize(self, prgm, tB, M, K, N, kc, nc, mr = 1, nr = 1):
code = prgm.get_stream()
old_code = ppc.get_active_code()
ppc.set_active_code(code)
gepb = SynGEPB(self.gepb_mode)
packb = SynPackB()
gepb._init_constants(M, K, N, kc, nc, mr, nr, True)
packb._init_constants(prgm, tB, N)
gepb._init_vars()
# Reuse the C/C_aux registers for B. They are set in init pointers.
packb._init_vars2(gepb.p_C, gepb.c[0][0], gepb.r_tB_addr)
gepb._load_params()
packb._load_params(pvB = 7)
# kN = k * N * 8
# for j in range(0, N * 8, nc * 8):
for j in syn_iter(code, N, nc):
# # Pack B into tB -- tB1.transpose(B[k:k+kc, j:j+nc])
# pack_params.p1 = B_addr + kN + j # (k * N + j) * 8
packb.vN.v = N
packb._pack_b(code)
# proc.execute(cgepb, params = pm)
gepb._init_pointers()
gepb._gepb(code)
# pm.p3 += nc8
gepb.r_C_addr.v = gepb.r_C_addr + nc * 8
packb.vB.v = packb.vB + nc * 8
# /end for j
ppc.set_active_code(old_code)
return
def synthesize(self, prgm, tB, M, K, N, kc, nc, mr=1, nr=1):
code = prgm.get_stream()
old_code = ppc.get_active_code()
ppc.set_active_code(code)
gepb = SynGEPB(self.gepb_mode)
packb = SynPackB()
gepb._init_constants(M, K, N, kc, nc, mr, nr, True)
packb._init_constants(prgm, tB, N)
gepb._init_vars()
# Reuse the C/C_aux registers for B. They are set in init pointers.
packb._init_vars2(gepb.p_C, gepb.c[0][0], gepb.r_tB_addr)
gepb._load_params()
packb._load_params(pvB=7)
# kN = k * N * 8
# for j in range(0, N * 8, nc * 8):
for j in syn_iter(code, N, nc):
# # Pack B into tB -- tB1.transpose(B[k:k+kc, j:j+nc])
# pack_params.p1 = B_addr + kN + j # (k * N + j) * 8
packb.vN.v = N
packb._pack_b(code)
# proc.execute(cgepb, params = pm)
gepb._init_pointers()
gepb._gepb(code)
# pm.p3 += nc8
gepb.r_C_addr.v = gepb.r_C_addr + nc * 8
packb.vB.v = packb.vB + nc * 8
# /end for j
ppc.set_active_code(old_code)
return
def k_loop_simple(self, code):
kc, mr, nr = self.block.kc, self.block.mr, self.block.nr
a, b, c = self.a, self.b, self.c
p_tA, p_tB = self.p_tA, self.p_tB
# Inner loop over k
for k in syn_iter(code, kc, mode = CTR): # syn_range(code, 0, kc * 8, 8):
# Load the next values from tA and tB -- generating loops
for ai in range(mr):
a[ai].load(p_tA, ai * self.A_strides.row)
for bj in range(nr):
b[bj].load(p_tB, bj * self.B_strides.col)
# Update c -- generating loop
for ci in range(mr):
for cj in range(nr):
c[ci][cj].v = ppcvar.fmadd(a[ci], b[cj], c[ci][cj])
p_tA.v = p_tA + self.A_strides.col
p_tB.v = p_tB + self.B_strides.row
# /end for k
return
def k_loop_prefetch(self, code):
A_row_stride = self.A_strides.row
A_col_stride = self.A_strides.col
B_row_stride = self.B_strides.row
B_col_stride = self.B_strides.col
kc, mr, nr = self.block.kc, self.block.mr, self.block.nr
a, b, c = self.a, self.b, self.c
a_pre, b_pre = self.a_pre, self.b_pre
p_tA, p_tB = self.p_tA, self.p_tB
# Increment p_tA, p_tB
a[0].load(p_tA, 0 * A_row_stride)
b[0].load(p_tB, 0 * B_col_stride)
b[1].load(p_tB, 1 * B_col_stride)
a[1].load(p_tA, 1 * A_row_stride)
b[2].load(p_tB, 2 * B_col_stride)
b[3].load(p_tB, 3 * B_col_stride)
a[2].load(p_tA, 2 * A_row_stride)
a[3].load(p_tA, 3 * A_row_stride)
p_tA.v = p_tA + A_col_stride
p_tB.v = p_tB + B_row_stride
for k in syn_iter(code, kc / 2 , mode = CTR): # syn_range(code, 0, kc * 8, 8):
# self.reg_loop_simple(a, b, c, mr, nr, p_tA, p_tB, A_row_stride, B_row_stride)
a_pre[0].load(p_tA, 0 * A_row_stride)
b_pre[1].load(p_tB, 1 * B_col_stride)
b_pre[0].load(p_tB, 0 * B_col_stride)
a_pre[1].load(p_tA, 1 * A_row_stride)
b_pre[2].load(p_tB, 2 * B_col_stride)
a_pre[2].load(p_tA, 2 * A_row_stride)
b_pre[3].load(p_tB, 3 * B_col_stride)
a_pre[3].load(p_tA, 3 * A_row_stride)
p_tA.v = p_tA + A_col_stride
p_tB.v = p_tB + B_row_stride
c[0][0].v = ppcvar.fmadd(a[0], b[0], c[0][0]);
c[0][1].v = ppcvar.fmadd(a[0], b[1], c[0][1]);
c[1][0].v = ppcvar.fmadd(a[1], b[0], c[1][0])
c[1][1].v = ppcvar.fmadd(a[1], b[1], c[1][1])
c[1][2].v = ppcvar.fmadd(a[1], b[2], c[1][2])
c[0][2].v = ppcvar.fmadd(a[0], b[2], c[0][2]);
c[2][0].v = ppcvar.fmadd(a[2], b[0], c[2][0])
c[2][1].v = ppcvar.fmadd(a[2], b[1], c[2][1])
c[2][2].v = ppcvar.fmadd(a[2], b[2], c[2][2])
c[2][3].v = ppcvar.fmadd(a[2], b[3], c[2][3])
c[0][3].v = ppcvar.fmadd(a[0], b[3], c[0][3]);
c[1][3].v = ppcvar.fmadd(a[1], b[3], c[1][3])
c[3][0].v = ppcvar.fmadd(a[3], b[0], c[3][0])
c[3][1].v = ppcvar.fmadd(a[3], b[1], c[3][1])
c[3][2].v = ppcvar.fmadd(a[3], b[2], c[3][2])
c[3][3].v = ppcvar.fmadd(a[3], b[3], c[3][3])
a[0].load(p_tA, 0 * A_row_stride)
b[1].load(p_tB, 1 * B_col_stride)
b[0].load(p_tB, 0 * B_col_stride)
a[1].load(p_tA, 1 * A_row_stride)
b[2].load(p_tB, 2 * B_col_stride)
b[3].load(p_tB, 3 * B_col_stride)
a[2].load(p_tA, 2 * A_row_stride)
a[3].load(p_tA, 3 * A_row_stride)
p_tA.v = p_tA + A_col_stride
p_tB.v = p_tB + B_row_stride
c[0][0].v = ppcvar.fmadd(a_pre[0], b_pre[0], c[0][0]);
c[0][1].v = ppcvar.fmadd(a_pre[0], b_pre[1], c[0][1]);
c[1][0].v = ppcvar.fmadd(a_pre[1], b_pre[0], c[1][0])
c[1][1].v = ppcvar.fmadd(a_pre[1], b_pre[1], c[1][1])
c[1][2].v = ppcvar.fmadd(a_pre[1], b_pre[2], c[1][2])
c[0][2].v = ppcvar.fmadd(a_pre[0], b_pre[2], c[0][2]);
c[2][0].v = ppcvar.fmadd(a_pre[2], b_pre[0], c[2][0])
c[2][1].v = ppcvar.fmadd(a_pre[2], b_pre[1], c[2][1])
c[2][2].v = ppcvar.fmadd(a_pre[2], b_pre[2], c[2][2])
c[2][3].v = ppcvar.fmadd(a_pre[2], b_pre[3], c[2][3])
c[0][3].v = ppcvar.fmadd(a_pre[0], b_pre[3], c[0][3]);
c[1][3].v = ppcvar.fmadd(a_pre[1], b_pre[3], c[1][3])
c[3][0].v = ppcvar.fmadd(a_pre[3], b_pre[0], c[3][0])
c[3][1].v = ppcvar.fmadd(a_pre[3], b_pre[1], c[3][1])
c[3][2].v = ppcvar.fmadd(a_pre[3], b_pre[2], c[3][2])
c[3][3].v = ppcvar.fmadd(a_pre[3], b_pre[3], c[3][3])
# /end for k
return
def k_loop_prefetch(self, code):
A_row_stride = self.A_strides.row
A_col_stride = self.A_strides.col
B_row_stride = self.B_strides.row
B_col_stride = self.B_strides.col
kc, mr, nr = self.block.kc, self.block.mr, self.block.nr
a, b, c = self.a, self.b, self.c
a_pre, b_pre = self.a_pre, self.b_pre
p_tA, p_tB = self.p_tA, self.p_tB
# Increment p_tA, p_tB
a[0].load(p_tA, 0 * A_row_stride)
b[0].load(p_tB, 0 * B_col_stride)
b[1].load(p_tB, 1 * B_col_stride)
a[1].load(p_tA, 1 * A_row_stride)
b[2].load(p_tB, 2 * B_col_stride)
b[3].load(p_tB, 3 * B_col_stride)
a[2].load(p_tA, 2 * A_row_stride)
a[3].load(p_tA, 3 * A_row_stride)
p_tA.v = p_tA + A_col_stride
p_tB.v = p_tB + B_row_stride
for k in syn_iter(code, kc / 2 , mode = CTR): # syn_range(code, 0, kc * 8, 8):
# self.reg_loop_simple(a, b, c, mr, nr, p_tA, p_tB, A_row_stride, B_row_stride)
a_pre[0].load(p_tA, 0 * A_row_stride)
b_pre[1].load(p_tB, 1 * B_col_stride)
b_pre[0].load(p_tB, 0 * B_col_stride)
a_pre[1].load(p_tA, 1 * A_row_stride)
b_pre[2].load(p_tB, 2 * B_col_stride)
a_pre[2].load(p_tA, 2 * A_row_stride)
b_pre[3].load(p_tB, 3 * B_col_stride)
a_pre[3].load(p_tA, 3 * A_row_stride)
p_tA.v = p_tA + A_col_stride
p_tB.v = p_tB + B_row_stride
c[0][0].v = ppcvar.fmadd(a[0], b[0], c[0][0]);
c[0][1].v = ppcvar.fmadd(a[0], b[1], c[0][1]);
c[1][0].v = ppcvar.fmadd(a[1], b[0], c[1][0])
c[1][1].v = ppcvar.fmadd(a[1], b[1], c[1][1])
c[1][2].v = ppcvar.fmadd(a[1], b[2], c[1][2])
c[0][2].v = ppcvar.fmadd(a[0], b[2], c[0][2]);
c[2][0].v = ppcvar.fmadd(a[2], b[0], c[2][0])
c[2][1].v = ppcvar.fmadd(a[2], b[1], c[2][1])
c[2][2].v = ppcvar.fmadd(a[2], b[2], c[2][2])
c[2][3].v = ppcvar.fmadd(a[2], b[3], c[2][3])
c[0][3].v = ppcvar.fmadd(a[0], b[3], c[0][3]);
c[1][3].v = ppcvar.fmadd(a[1], b[3], c[1][3])
c[3][0].v = ppcvar.fmadd(a[3], b[0], c[3][0])
c[3][1].v = ppcvar.fmadd(a[3], b[1], c[3][1])
c[3][2].v = ppcvar.fmadd(a[3], b[2], c[3][2])
c[3][3].v = ppcvar.fmadd(a[3], b[3], c[3][3])
a[0].load(p_tA, 0 * A_row_stride)
b[1].load(p_tB, 1 * B_col_stride)
b[0].load(p_tB, 0 * B_col_stride)
a[1].load(p_tA, 1 * A_row_stride)
b[2].load(p_tB, 2 * B_col_stride)
b[3].load(p_tB, 3 * B_col_stride)
a[2].load(p_tA, 2 * A_row_stride)
a[3].load(p_tA, 3 * A_row_stride)
p_tA.v = p_tA + A_col_stride
p_tB.v = p_tB + B_row_stride
c[0][0].v = ppcvar.fmadd(a_pre[0], b_pre[0], c[0][0]);
c[0][1].v = ppcvar.fmadd(a_pre[0], b_pre[1], c[0][1]);
c[1][0].v = ppcvar.fmadd(a_pre[1], b_pre[0], c[1][0])
c[1][1].v = ppcvar.fmadd(a_pre[1], b_pre[1], c[1][1])
c[1][2].v = ppcvar.fmadd(a_pre[1], b_pre[2], c[1][2])
c[0][2].v = ppcvar.fmadd(a_pre[0], b_pre[2], c[0][2]);
c[2][0].v = ppcvar.fmadd(a_pre[2], b_pre[0], c[2][0])
c[2][1].v = ppcvar.fmadd(a_pre[2], b_pre[1], c[2][1])
c[2][2].v = ppcvar.fmadd(a_pre[2], b_pre[2], c[2][2])
c[2][3].v = ppcvar.fmadd(a_pre[2], b_pre[3], c[2][3])
c[0][3].v = ppcvar.fmadd(a_pre[0], b_pre[3], c[0][3]);
c[1][3].v = ppcvar.fmadd(a_pre[1], b_pre[3], c[1][3])
c[3][0].v = ppcvar.fmadd(a_pre[3], b_pre[0], c[3][0])
c[3][1].v = ppcvar.fmadd(a_pre[3], b_pre[1], c[3][1])
c[3][2].v = ppcvar.fmadd(a_pre[3], b_pre[2], c[3][2])
c[3][3].v = ppcvar.fmadd(a_pre[3], b_pre[3], c[3][3])
# /end for k
return