def c_aux_save_hand(self, code):
mr = self.block.mr
nc = self.block.nc
p_C = self.p_C
p_C_aux = self.p_C_aux
c_mem = self.a + self.b
if mr > 1: c_aux = self.c[0] + self.c[1]
else: c_aux = self.c[0]
for jj in syn_range(code, 0, nc * self.C_strides.col,
self.C_strides.col):
JJ = jj
# ~ 650 Mflops (one extra unroll does not affect result)
if mr >= 1:
c_mem[0].load(p_C[0], JJ)
c_aux[0].load(p_C_aux[0], JJ)
if mr >= 2:
c_mem[1].load(p_C[1], JJ)
c_aux[1].load(p_C_aux[1], JJ)
if mr >= 1:
c_mem[0].v = c_mem[0] + c_aux[0]
if mr >= 2:
c_mem[1].v = c_mem[1] + c_aux[1]
if mr >= 3:
c_mem[2].load(p_C[2], JJ)
c_aux[2].load(p_C_aux[2], JJ)
if mr >= 4:
c_mem[3].load(p_C[3], JJ)
c_aux[3].load(p_C_aux[3], JJ)
if mr >= 3:
c_mem[2].v = c_mem[2] + c_aux[2]
if mr >= 4:
c_mem[3].v = c_mem[3] + c_aux[3]
if mr >= 1: c_mem[0].store(p_C[0], JJ)
if mr >= 2: c_mem[1].store(p_C[1], JJ)
if mr >= 3: c_mem[2].store(p_C[2], JJ)
if mr >= 4: c_mem[3].store(p_C[3], JJ)
# /end for jj
return
def c_aux_save_hand(self, code):
mr = self.block.mr
nc = self.block.nc
p_C = self.p_C
p_C_aux = self.p_C_aux
c_mem = self.a + self.b
if mr > 1: c_aux = self.c[0] + self.c[1]
else: c_aux = self.c[0]
for jj in syn_range(code, 0, nc * self.C_strides.col, self.C_strides.col):
JJ = jj
# ~ 650 Mflops (one extra unroll does not affect result)
if mr >= 1:
c_mem[0].load(p_C[0], JJ)
c_aux[0].load(p_C_aux[0], JJ)
if mr >= 2:
c_mem[1].load(p_C[1], JJ)
c_aux[1].load(p_C_aux[1], JJ)
if mr >= 1:
c_mem[0].v = c_mem[0] + c_aux[0]
if mr >= 2:
c_mem[1].v = c_mem[1] + c_aux[1]
if mr >= 3:
c_mem[2].load(p_C[2], JJ)
c_aux[2].load(p_C_aux[2], JJ)
if mr >= 4:
c_mem[3].load(p_C[3], JJ)
c_aux[3].load(p_C_aux[3], JJ)
if mr >= 3:
c_mem[2].v = c_mem[2] + c_aux[2]
if mr >= 4:
c_mem[3].v = c_mem[3] + c_aux[3]
if mr >= 1: c_mem[0].store(p_C[0], JJ)
if mr >= 2: c_mem[1].store(p_C[1], JJ)
if mr >= 3: c_mem[2].store(p_C[2], JJ)
if mr >= 4: c_mem[3].store(p_C[3], JJ)
# /end for jj
return
def c_aux_save_simple(self, code):
# Copy C_aux to C
# ~620 MFlops
nc, mr, a, b, p_C, p_C_aux, C_col_stride = (
self.block.nc, self.block.mr, self.a, self.b, self.p_C, self.p_C_aux, self.C_strides.col)
for jj in syn_range(code, 0, nc * C_col_stride, C_col_stride):
for ci in range(mr):
a[ci].load(p_C[ci], jj)
b[ci].load(p_C_aux[ci], jj)
a[ci].v = a[ci] + b[ci]
a[ci].store(p_C[ci], jj)
# /end for ci
# /end for jj
return
def c_aux_save_simple(self, code):
# Copy C_aux to C
# ~620 MFlops
nc, mr, a, b, p_C, p_C_aux, C_col_stride = (
self.block.nc, self.block.mr, self.a, self.b, self.p_C, self.p_C_aux, self.C_strides.col)
for jj in syn_range(code, 0, nc * C_col_stride, C_col_stride):
for ci in range(mr):
a[ci].load(p_C[ci], jj)
b[ci].load(p_C_aux[ci], jj)
a[ci].v = a[ci] + b[ci]
a[ci].store(p_C[ci], jj)
# /end for ci
# /end for jj
return
def _gepb(self, code):
a, b, c = self.a, self.b, self.c
p_tA, p_tB = self.p_tA, self.p_tB
M = self.dims.M
mr, nr, nc = self.block.mr, self.block.nr, self.block.nc
# For each row in C
for ii in syn_range(code, 0, M, mr):
# Reset p_tB
p_tB.v = self.r_tB_addr
# For each column in C
for jj in syn_range(code, 0, nc, nr):
# Set p_tA to the current row in tA
p_tA.v = self.r_tA_addr + ii * self.A_strides.row # * mr
# Set p_tB to the current col in tB
p_tB.v = self.r_tB_addr + jj * self.B_strides.col
# Zero the c register block
ppc.fsubx(c[0][0], c[0][0], c[0][0])
for ci in range(mr):
for cj in range(nr):
c[ci][cj].copy_register(c[0][0])
# self.k_loop_simple(code)
if self.mode == gepb_simple:
self.k_loop_simple(code)
elif self.mode == gepb_prefetch:
self.k_loop_prefetch_simple(code)
elif self.mode == gepb_prefetch_hand:
self.k_loop_prefetch(code)
else:
raise Exception("Unknown inner loop mode: %s" % (str(self.mode)))
# Save c_current to c_aux
# (this is OK performance-wise)
for ci in range(mr):
for cj in range(nr):
c[ci][cj].store(self.p_C_aux[ci], cj * self.C_strides.col)
# Increment the sub-matrix in C_aux
for ci in range(mr):
self.p_C_aux[ci].v = self.p_C_aux[ci] + self.C_strides.col * nr
# /end for jj
# Reset p_C_aux
for ci in range(mr):
self.p_C_aux[ci].v = self.r_C_aux_addr + ci * (nc * self.C_strides.col)
# Copy C_aux to C
# self.c_aux_save_simple(code)
self.c_aux_save_hand(code)
# Increment p_C
for ci in range(mr):
self.p_C[ci].v = self.p_C[ci] + self.vC_row_stride * mr
# /end for ii
return
def _gepb(self, code):
a, b, c = self.a, self.b, self.c
p_tA, p_tB = self.p_tA, self.p_tB
M = self.dims.M
mr, nr, nc = self.block.mr, self.block.nr, self.block.nc
# For each row in C
for ii in syn_range(code, 0, M, mr):
# Reset p_tB
p_tB.v = self.r_tB_addr
# For each column in C
for jj in syn_range(code, 0, nc, nr):
# Set p_tA to the current row in tA
p_tA.v = self.r_tA_addr + ii * self.A_strides.row # * mr
# Set p_tB to the current col in tB
p_tB.v = self.r_tB_addr + jj * self.B_strides.col
# Zero the c register block
ppc.fsubx(c[0][0], c[0][0], c[0][0])
for ci in range(mr):
for cj in range(nr):
c[ci][cj].copy_register(c[0][0])
# self.k_loop_simple(code)
if self.mode == gepb_simple:
self.k_loop_simple(code)
elif self.mode == gepb_prefetch:
self.k_loop_prefetch_simple(code)
elif self.mode == gepb_prefetch_hand:
self.k_loop_prefetch(code)
else:
raise Exception("Unknown inner loop mode: %s" % (str(self.mode)))
# Save c_current to c_aux
# (this is OK performance-wise)
for ci in range(mr):
for cj in range(nr):
c[ci][cj].store(self.p_C_aux[ci], cj * self.C_strides.col)
# Increment the sub-matrix in C_aux
for ci in range(mr):
self.p_C_aux[ci].v = self.p_C_aux[ci] + self.C_strides.col * nr
# /end for jj
# Reset p_C_aux
for ci in range(mr):
self.p_C_aux[ci].v = self.r_C_aux_addr + ci * (nc * self.C_strides.col)
# Copy C_aux to C
# self.c_aux_save_simple(code)
self.c_aux_save_hand(code)
# Increment p_C
for ci in range(mr):
self.p_C[ci].v = self.p_C[ci] + self.vC_row_stride * mr
# /end for ii
return
