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 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 reg_loop_simple(self, a, b, c, mr, nr, p_tA, p_tB, A_row_stride, B_col_stride):
# Load the next values from tA and tB -- generating loops
for ai in range(mr):
a[ai].load(p_tA, ai * A_row_stride)
for bj in range(nr):
b[bj].load(p_tB, bj * B_col_stride)
# 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])
return
def reg_loop_simple(self, a, b, c, mr, nr, p_tA, p_tB, A_row_stride, B_col_stride):
# Load the next values from tA and tB -- generating loops
for ai in range(mr):
a[ai].load(p_tA, ai * A_row_stride)
for bj in range(nr):
b[bj].load(p_tB, bj * B_col_stride)
# 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])
return
def TestExternalStop():
prgm = synppc.Program()
code = prgm.get_stream()
prgm.add(code)
ppc.set_active_code(code)
# Data
data = array.array('d', range(5*5))
# Constants - read only
n_rows = vars.SignedWord(5)
n_cols = vars.SignedWord(5)
addr = vars.SignedWord(data.buffer_info()[0])
dbl_size = vars.SignedWord(synppc.WORD_SIZE * 2)
row_bytes = vars.SignedWord(synppc.WORD_SIZE * 5 * 2)
# Variables - read/write
sum = vars.DoubleFloat(0.0)
x = vars.DoubleFloat(0.0)
offset = vars.SignedWord(0)
# Iterators
i_iter = syn_iter(code, 0, mode = INC)
i_iter.set_external_stop(n_rows.reg)
j_ctr = syn_iter(code, 0, mode = CTR)
j_ctr.set_external_stop(n_cols.reg)
for i in i_iter:
offset.v = vars.SignedWord.cast(i) * row_bytes
# Note that j_cnt is unreadable since it's in the ctr register
for j_cnt in j_ctr:
# Load the next vaule in the matrix
ppc.lfdx(x, addr, offset)
sum.v = vars.fmadd(x, x, sum) # sum += x*x
offset.v = offset + dbl_size
# code.add(ppc.Illegal())
util.return_var(sum)
proc = synppc.Processor()
r = proc.execute(prgm, mode = 'fp')
# print 'Test external stop: ', r
assert(r == 4900.0)
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 compute(i,j): c[i][j].v = ppcvar.fmadd(a[i], b[j], c[i][j]); load_a(0)
def reg_loop_4x4(self, a, b, c, mr, nr, p_tA, p_tB, A_row_stride, B_col_stride):
a[0].load(p_tA, 0 * A_row_stride)
b[0].load(p_tB, 0 * B_col_stride)
c[0][0].v = ppcvar.fmadd(a[0], b[0], c[0][0]);
b[1].load(p_tB, 1 * B_col_stride)
a[1].load(p_tA, 1 * A_row_stride)
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])
a[2].load(p_tA, 2 * A_row_stride)
b[2].load(p_tB, 2 * B_col_stride)
c[1][1].v = ppcvar.fmadd(a[1], b[1], c[1][1])
c[2][0].v = ppcvar.fmadd(a[2], b[0], c[2][0])
c[0][2].v = ppcvar.fmadd(a[0], b[2], c[0][2]);
c[1][2].v = ppcvar.fmadd(a[1], b[2], c[1][2])
c[2][1].v = ppcvar.fmadd(a[2], b[1], c[2][1])
a[3].load(p_tA, 3 * A_row_stride)
b[3].load(p_tB, 3 * B_col_stride)
c[2][2].v = ppcvar.fmadd(a[2], b[2], c[2][2])
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])
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[2][3].v = ppcvar.fmadd(a[2], b[3], c[2][3])
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 compute(i,j): c[i][j].v = ppcvar.fmadd(a[i], b[j], c[i][j]); load_a(0)
def reg_loop_4x4(self, a, b, c, mr, nr, p_tA, p_tB, A_row_stride, B_col_stride):
a[0].load(p_tA, 0 * A_row_stride)
b[0].load(p_tB, 0 * B_col_stride)
c[0][0].v = ppcvar.fmadd(a[0], b[0], c[0][0]);
b[1].load(p_tB, 1 * B_col_stride)
a[1].load(p_tA, 1 * A_row_stride)
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])
a[2].load(p_tA, 2 * A_row_stride)
b[2].load(p_tB, 2 * B_col_stride)
c[1][1].v = ppcvar.fmadd(a[1], b[1], c[1][1])
c[2][0].v = ppcvar.fmadd(a[2], b[0], c[2][0])
c[0][2].v = ppcvar.fmadd(a[0], b[2], c[0][2]);
c[1][2].v = ppcvar.fmadd(a[1], b[2], c[1][2])
c[2][1].v = ppcvar.fmadd(a[2], b[1], c[2][1])
a[3].load(p_tA, 3 * A_row_stride)
b[3].load(p_tB, 3 * B_col_stride)
c[2][2].v = ppcvar.fmadd(a[2], b[2], c[2][2])
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])
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[2][3].v = ppcvar.fmadd(a[2], b[3], c[2][3])
return
