def instantiate(self, arch_input_chn, psum_chn, arr_x, arr_y,
chn_per_word):
# PE static configuration (immutable)
self.arr_x = arr_x
self.arr_y = arr_y
self.chn_per_word = chn_per_word
self.arch_input_chn = arch_input_chn
self.psum_chn = psum_chn
self.ifmap = None
self.weights = None
self.bias = None
self.image_size = (0, 0)
self.filter_size = (0, 0)
self.ifmap_psum_done = True
self.pass_done = Reg(False)
# State Counters
self.curr_set = 0
self.curr_filter = 0
self.iteration = 0
self.fmap_idx = 0
def instantiate(self, depth=2, name=None):
self.data = [None] * depth
self.depth = depth
self.name = name
self.rd_ptr = Reg(0)
self.wr_ptr = Reg(0)
def instantiate(self, depth, peek_window, enq_window, deq_window):
self.data = [0] * depth
self.depth = depth
self.peek_window = peek_window
self.enq_window = enq_window
self.deq_window = deq_window
self.rd_ptr = Reg(0)
self.wr_ptr = Reg(0)
def instantiate(self, locx, locy, bias_chn, ofmap_in_chn, ofmap_out_chn): #ofmap_in
self.locx = locx
self.locy = locy
self.bias_chn = bias_chn
self.ofmap_in_chn = ofmap_in_chn
self.ofmap_out_chn = ofmap_out_chn
self.transform_done = Reg(False)
self.stat_type = 'aggregate'
self.raw_stats = {'post_tr_alu_comp' : 0, 'post_tr_rf_rd' : 0, 'post_tr_ifmap_rf_wr' : 0}
def instantiate(self, locx, locy, weight_in_chn, weight_out_chn):
self.locx = locx
self.locy = locy
self.weight_in_chn = weight_in_chn
self.weight_out_chn = weight_out_chn
self.transform_done = Reg(False)
self.stat_type = 'aggregate'
self.raw_stats = {
'pre_tr_weights_alu_comp': 0,
'pre_tr_weights_rf_rd': 0,
'pre_tr_weights_rf_wr': 0
}
def instantiate(self, locx, locy, ifmap_in_chn, ifmap_out_chn):
self.locx = locx
self.locy = locy
self.ifmap_in_chn = ifmap_in_chn
self.ifmap_out_chn = ifmap_out_chn
self.transform_done = Reg(False)
self.stat_type = 'aggregate'
self.raw_stats = {
'pre_tr_ifmap_alu_comp': 0,
'pre_tr_ifmap_rf_rd': 0,
'pre_tr_ifmap_rf_wr': 0
}
def instantiate(self, arch_output_chn, arr_x, arr_y, chn_per_word):
# PE static configuration (immutable)
self.arr_x = arr_x
self.arr_y = arr_y
self.chn_per_word = chn_per_word
self.arch_output_chn = arch_output_chn
self.ofmap = None
self.reference = None
self.image_size = (0, 0)
self.curr_set = 0
self.fmap_idx = 0
self.pass_done = Reg(False)
def instantiate(self, arch_output_chn, arr_y, block_size, num_nonzero):
# PE static configuration (immutable)
self.arr_y = arr_y
self.block_size = block_size
self.num_nonzero = num_nonzero
self.arch_output_chn = arch_output_chn
self.ofmap = None
self.reference = None
self.image_size = (0, 0)
self.curr_set = 0
self.fmap_idx = 0
self.pass_done = Reg(False)
class RegTB(Module):
def instantiate(self):
self.ra = Reg(0)
self.rb = Reg(10)
def tick(self):
print("ra val: %d" % self.ra.rd())
print("rb val: %d" % self.rb.rd())
self.ra.wr(self.ra.rd() + 1)
self.rb.wr(self.ra.rd() + self.rb.rd())
class FIFOTB(Module):
def instantiate(self, check_fifo=True):
self.check_fifo = check_fifo
self.counter = Reg(0)
self.fifo = FIFO(4)
def tick(self):
count = self.counter.rd()
self.counter.wr((count + 1) % 256)
if count % 4 < 2 and (self.fifo.not_full() or not self.check_fifo):
self.fifo.enq(count)
print("enq: %d" % count)
if count % 4 == 3 and (self.fifo.not_empty() or not self.check_fifo):
peek = self.fifo.peek()
self.fifo.deq()
print("deq: %d" % peek)
def instantiate(self, arch_input_chn, arr_y, block_size, num_nonzero,
pruner_name):
# PE static configuration (immutable)
#self.arr_x = arr_x
self.arr_y = arr_y
#self.chn_per_word = chn_per_word
self.block_size = block_size
self.num_nonzero = num_nonzero
self.convert_chn = Channel()
self.prune_chn = Channel()
self.arch_input_chn = arch_input_chn
# Although both InputSerializer and pruner will be pushing to arch_input_chn
# There is no conflict issue because all weights will be pushed by IS first
# then all inputs by pruner
self.converter = Converter(self.convert_chn, self.prune_chn, \
self.block_size, self.block_size)
# self.pruner = NaivePruner(self.prune_chn,self.arch_input_chn, \
# self.num_nonzero,True)
#user defined pruner for this layer, default to naive pruner
self.pruner = getattr(pruner, pruner_name)(self.prune_chn,self.arch_input_chn, \
self.num_nonzero, self.block_size, True)
self.ifmap = None
self.weights = None
self.bias = None
self.image_size = (0, 0)
self.filter_size = (0, 0)
self.ifmap_psum_done = True
self.pass_done = Reg(False)
# State Counters
self.curr_set = 0
self.curr_filter = 0
self.iteration = 0
self.fmap_idx = 0
self.curr_chn = 0
self.curr_x = 0 # run through first two dimensions of input
self.curr_y = 0
self.bias_set = 0
class InputSerializer(Module):
def instantiate(self, arch_input_chn, arr_x, arr_y, chn_per_word):
# PE static configuration (immutable)
self.arr_x = arr_x
self.arr_y = arr_y
self.chn_per_word = chn_per_word
self.arch_input_chn = arch_input_chn
self.ifmap = None
self.weights = None
self.bias = None
self.image_size = (0, 0)
self.filter_size = (0, 0)
self.ifmap_psum_done = True
self.pass_done = Reg(False)
# State Counters
self.curr_set = 0
self.curr_filter = 0
self.iteration = 0
self.fmap_idx = 0
def configure(self, ifmap, weights, bias, image_size, filter_size):
self.ifmap = ifmap
self.weights = weights
self.bias = bias
self.image_size = image_size
self.filter_size = filter_size
self.ifmap_psum_done = False
self.pass_done.wr(False)
def tick(self):
if self.pass_done.rd():
return
in_sets = self.arr_y // self.chn_per_word
out_sets = self.arr_x // self.chn_per_word
fmap_per_iteration = self.image_size[0] * self.image_size[1]
num_iteration = self.filter_size[0] * self.filter_size[1]
if not self.ifmap_psum_done:
if self.arch_input_chn.vacancy():
# print "input append"
x = self.fmap_idx % self.image_size[0]
y = self.fmap_idx // self.image_size[0]
if self.curr_set < in_sets:
cmin = self.curr_set * self.chn_per_word
cmax = cmin + self.chn_per_word
# Write ifmap to glb
data = np.array(
[self.ifmap[x, y, c] for c in range(cmin, cmax)])
else:
cmin = (self.curr_set - in_sets) * self.chn_per_word
cmax = cmin + self.chn_per_word
# Write bias to glb
data = np.array([self.bias[c] for c in range(cmin, cmax)])
self.arch_input_chn.push(data)
self.curr_set += 1
if self.curr_set == (in_sets + out_sets):
self.curr_set = 0
self.fmap_idx += 1
if self.fmap_idx == fmap_per_iteration:
self.fmap_idx = 0
self.ifmap_psum_done = True
# print "---- Wrote inputs and biases ----"
else:
f_x = self.iteration % self.filter_size[0]
f_y = self.iteration // self.filter_size[0]
# Push filters to PE columns. (PE is responsible for pop)
if self.arch_input_chn.vacancy(
) and self.iteration < num_iteration:
cmin = self.curr_set * self.chn_per_word
cmax = cmin + self.chn_per_word
data = np.array([self.weights[f_x, f_y, c, self.curr_filter] \
for c in range(cmin, cmax) ])
self.arch_input_chn.push(data)
self.curr_set += 1
if self.curr_set == in_sets:
self.curr_set = 0
self.curr_filter += 1
if self.curr_filter == self.arr_x:
self.curr_filter = 0
# print "---- Wrote weights iteration: %d ----" % self.iteration
self.iteration += 1
if self.iteration == num_iteration:
# print "---- Wrote all weights ----"
self.pass_done.wr(True)
class WindowFIFO(Module):
def instantiate(self, depth, peek_window, enq_window, deq_window):
self.data = [0] * depth
self.depth = depth
self.peek_window = peek_window
self.enq_window = enq_window
self.deq_window = deq_window
self.rd_ptr = Reg(0)
self.wr_ptr = Reg(0)
def peek(self):
if (self.wr_ptr.rd() - self.rd_ptr.rd()) % (2*self.depth) \
< self.peek_window:
raise FIFOError("Reading from empty FIFO")
peek_output = [0] * self.peek_window
for i in xrange(self.peek_window):
peek_output[i] = self.data[(self.rd_ptr.rd() + i) % self.depth]
return peek_output
def enq(self, x):
if (self.wr_ptr.rd() - self.rd_ptr.rd() + self.enq_window - 1) % \
(2*self.depth) >= self.depth:
raise FIFOError("Enqueueing into full FIFO")
for i in xrange(self.enq_window):
self.data[(self.wr_ptr.rd() + i) % self.depth] = x[i]
self.wr_ptr.wr((self.wr_ptr.rd() + self.enq_window) % (2 * self.depth))
def deq(self):
if (self.wr_ptr.rd() - self.rd_ptr.rd()) % (2*self.depth) \
< self.deq_window:
raise FIFOError("Dequeueing from empty FIFO")
self.rd_ptr.wr((self.rd_ptr.rd() + self.deq_window) % (2 * self.depth))
def not_full(self):
return not (
(self.wr_ptr.rd() - self.rd_ptr.rd() + self.enq_window - 1) %
(2 * self.depth) >= self.depth)
def valid(self):
return not ((self.wr_ptr.rd() - self.rd_ptr.rd()) %
(2 * self.depth) < self.peek_window)
def not_empty(self):
return not ((self.wr_ptr.rd() - self.rd_ptr.rd()) %
(2 * self.depth) < self.deq_window)
def clear(self):
self.rd_ptr.wr(self.wr_ptr.rd())
class OutputDeserializer(Module):
def instantiate(self, arch_output_chn, arr_x, arr_y, chn_per_word):
# PE static configuration (immutable)
self.arr_x = arr_x
self.arr_y = arr_y
self.chn_per_word = chn_per_word
self.arch_output_chn = arch_output_chn
self.ofmap = None
self.reference = None
self.image_size = (0, 0)
self.curr_set = 0
self.fmap_idx = 0
self.pass_done = Reg(False)
def configure(self, ofmap, reference, image_size, filter_size):
self.ofmap = ofmap
self.reference = reference
self.image_size = image_size
self.filter_size = filter_size
self.curr_set = 0
self.fmap_idx = 0
self.pass_done.wr(False)
def tick(self):
if self.pass_done.rd():
return
out_sets = self.arr_x // self.chn_per_word
#fmap_per_iteration = self.image_size[0]*self.image_size[1]
fmap_per_iteration = (self.image_size[0] - self.filter_size[0] + 1) * (
self.image_size[1] - self.filter_size[1] + 1)
if self.arch_output_chn.valid():
data = [e for e in self.arch_output_chn.pop()]
x = self.fmap_idx % self.image_size[0]
y = self.fmap_idx // self.image_size[0]
x = self.fmap_idx % (self.image_size[0] - self.filter_size[0] + 1)
y = self.fmap_idx // (self.image_size[0] - self.filter_size[0] + 1)
if self.curr_set < out_sets:
cmin = self.curr_set * self.chn_per_word
cmax = cmin + self.chn_per_word
for c in range(cmin, cmax):
self.ofmap[x, y, c] = data[c - cmin]
self.curr_set += 1
if self.curr_set == out_sets:
self.curr_set = 0
self.fmap_idx += 1
if self.fmap_idx == fmap_per_iteration:
self.fmap_idx = 0
self.pass_done.wr(True)
if np.all(self.ofmap == self.reference):
raise Finish("Success")
else:
print(self.ofmap)
print(self.reference)
print(self.ofmap - self.reference)
raise Finish("Validation Failed")
class FIFO(Module):
def instantiate(self, depth=2):
self.data = [0] * depth
self.depth = depth
self.rd_ptr = Reg(0)
self.wr_ptr = Reg(0)
def peek(self):
if self.wr_ptr.rd() == self.rd_ptr.rd():
raise FIFOError("Reading from empty FIFO")
return self.data[self.rd_ptr.rd() % self.depth]
def enq(self, x):
if (self.wr_ptr.rd() - self.rd_ptr.rd()) % (2 *
self.depth) == self.depth:
raise FIFOError("Enqueueing into full FIFO")
self.data[self.wr_ptr.rd() % self.depth] = x
self.wr_ptr.wr((self.wr_ptr.rd() + 1) % (2 * self.depth))
def deq(self):
if self.wr_ptr.rd() == self.rd_ptr.rd():
raise FIFOError("Dequeueing from empty FIFO")
self.rd_ptr.wr((self.rd_ptr.rd() + 1) % (2 * self.depth))
def not_full(self):
return not ((self.wr_ptr.rd() - self.rd_ptr.rd()) %
(2 * self.depth) == self.depth)
def not_empty(self):
return not (self.wr_ptr.rd() == self.rd_ptr.rd())
def reset(self):
self.rd_ptr.wr(self.wr_ptr.rd())
class PreTransformIFMap(Module):
def instantiate(self, locx, locy, ifmap_in_chn, ifmap_out_chn):
self.locx = locx
self.locy = locy
self.ifmap_in_chn = ifmap_in_chn
self.ifmap_out_chn = ifmap_out_chn
self.transform_done = Reg(False)
self.stat_type = 'aggregate'
self.raw_stats = {
'pre_tr_ifmap_alu_comp': 0,
'pre_tr_ifmap_rf_rd': 0,
'pre_tr_ifmap_rf_wr': 0
}
def configure(self):
self.iteration = 0
self.push_ctr = 0
self.V = np.zeros([4, 4]).astype(np.int64)
self.raw_stats['pre_tr_ifmap_rf_wr'] += 16 # write zeros into rf
self.transform_done.wr(False)
# Explanation of algorithm: transform ifmap D into V, performing Winograd transform v = B_T*D*M
# D = [D00 D01 D02 D03
# D10 D11 D12 D13
# D20 D21 D22 D23
# D30 D31 D32 D33]
#
# B_T = [1 0 -1 0
# 0 1 1 0
# 0 -1 1 0
# 0 1 0 -1]
#
# B = [ 1 0 0 0
# 0 1 -1 1
# -1 1 1 0
# 0 0 0 -1]
#
# Performing this transform yields a 4x4 output for a given 4x4 input:
#
# V = [v00 v01 v02 v03
# v10 v11 v12 v13
# v20 v21 v22 v23
# v30 v31 v32 v33]
#
# ... such that:
# v00 = (D00 - D02 - D20 + D22);
# v01 = (D01 + D02 - D21 - D22);
# v02 = (D02 - D01 + D21 - D22);
# v03 = (D01 - D03 - D21 + D23);
# v10 = (D10 - D12 + D20 - D22);
# v11 = (D11 + D12 + D21 + D22);
# v12 = (D12 - D11 - D21 + D22);
# v13 = (D11 - D13 + D21 - D23);
# v20 = (D12 - D10 + D20 - D22);
# v21 = (D21 - D12 - D11 + D22);
# v22 = (D11 - D12 - D21 + D22);
# v23 = (D13 - D11 + D21 - D23);
# v30 = (D10 - D12 - D30 + D32);
# v31 = (D11 + D12 - D31 - D32);
# v32 = (D12 - D11 + D31 - D32);
# v33 = (D11 - D13 - D31 + D33);
def tick(self):
if self.transform_done.rd():
return
if self.ifmap_in_chn.valid() and self.ifmap_out_chn.vacancy():
d = (self.ifmap_in_chn.pop())
#print ("pre transform ifmap pop - locx, locy, data: ",self.locx,self.locy,d)
if (self.iteration == 0): # get D_00
self.V[0][0] += d
self.raw_stats['pre_tr_ifmap_alu_comp'] += 1
self.raw_stats['pre_tr_ifmap_rf_rd'] += 1
self.raw_stats['pre_tr_ifmap_rf_wr'] += 1
self.iteration += 1
elif (self.iteration == 1): # get D_01
self.V[0][1] += d
self.V[0][2] -= d
self.V[0][3] += d
self.raw_stats['pre_tr_ifmap_alu_comp'] += 3
self.raw_stats['pre_tr_ifmap_rf_rd'] += 3
self.raw_stats['pre_tr_ifmap_rf_wr'] += 3
self.iteration += 1
elif (self.iteration == 2): # get D_02
self.V[0][0] -= d
self.V[0][1] += d
self.V[0][2] += d
self.raw_stats['pre_tr_ifmap_alu_comp'] += 3
self.raw_stats['pre_tr_ifmap_rf_rd'] += 3
self.raw_stats['pre_tr_ifmap_rf_wr'] += 3
self.iteration += 1
elif (self.iteration == 3): # get D_03
self.V[0][3] -= d
self.raw_stats['pre_tr_ifmap_alu_comp'] += 1
self.raw_stats['pre_tr_ifmap_rf_rd'] += 1
self.raw_stats['pre_tr_ifmap_rf_wr'] += 1
self.iteration += 1
elif (self.iteration == 4): # get D_10
self.V[1][0] += d
self.V[2][0] -= d
self.V[3][0] += d
self.raw_stats['pre_tr_ifmap_alu_comp'] += 3
self.raw_stats['pre_tr_ifmap_rf_rd'] += 3
self.raw_stats['pre_tr_ifmap_rf_wr'] += 3
self.iteration += 1
elif (self.iteration == 5): # get D_11
self.V[1][1] += d
self.V[1][2] -= d
self.V[1][3] += d
self.V[2][1] -= d
self.V[2][2] += d
self.V[2][3] -= d
self.V[3][1] += d
self.V[3][2] -= d
self.V[3][3] += d
self.raw_stats['pre_tr_ifmap_alu_comp'] += 9
self.raw_stats['pre_tr_ifmap_rf_rd'] += 9
self.raw_stats['pre_tr_ifmap_rf_wr'] += 9
self.iteration += 1
elif (self.iteration == 6): # get D_12
self.V[1][0] -= d
self.V[1][1] += d
self.V[1][2] += d
self.V[2][0] += d
self.V[2][1] -= d
self.V[2][2] -= d
self.V[3][0] -= d
self.V[3][1] += d
self.V[3][2] += d
self.raw_stats['pre_tr_ifmap_alu_comp'] += 9
self.raw_stats['pre_tr_ifmap_rf_rd'] += 9
self.raw_stats['pre_tr_ifmap_rf_wr'] += 9
self.iteration += 1
elif (self.iteration == 7): # get D_13
self.V[1][3] -= d
self.V[2][3] += d
self.V[3][3] -= d
self.raw_stats['pre_tr_ifmap_alu_comp'] += 3
self.raw_stats['pre_tr_ifmap_rf_rd'] += 3
self.raw_stats['pre_tr_ifmap_rf_wr'] += 3
self.iteration += 1
elif (self.iteration == 8): # get D_20
self.V[0][0] -= d
self.V[1][0] += d
self.V[2][0] += d
self.raw_stats['pre_tr_ifmap_alu_comp'] += 3
self.raw_stats['pre_tr_ifmap_rf_rd'] += 3
self.raw_stats['pre_tr_ifmap_rf_wr'] += 3
self.iteration += 1
elif (self.iteration == 9): # get D_21
self.V[0][1] -= d
self.V[0][2] += d
self.V[0][3] -= d
self.V[1][1] += d
self.V[1][2] -= d
self.V[1][3] += d
self.V[2][1] += d
self.V[2][2] -= d
self.V[2][3] += d
self.raw_stats['pre_tr_ifmap_alu_comp'] += 9
self.raw_stats['pre_tr_ifmap_rf_rd'] += 9
self.raw_stats['pre_tr_ifmap_rf_wr'] += 9
self.iteration += 1
elif (self.iteration == 10
): # get D_22 & start pushing transformed data out
self.V[0][0] += d
self.V[0][1] -= d
self.V[0][2] -= d
self.V[1][0] -= d
self.V[1][1] += d
self.V[1][2] += d
self.V[2][0] -= d
self.V[2][1] += d
self.V[2][2] += d
self.raw_stats['pre_tr_ifmap_alu_comp'] += 9
self.raw_stats['pre_tr_ifmap_rf_rd'] += 9
self.raw_stats['pre_tr_ifmap_rf_wr'] += 9
self.ifmap_out_chn.push(
self.V[self.push_ctr // 4][self.push_ctr % 4]) # push v00
self.raw_stats[
'pre_tr_ifmap_rf_rd'] -= 1 # push v00 immediately w/o writing to rf
#print ("pre transform ifmap - locx, locy, iteration, transformed ifmap: ", \
# self.locx, self.locy, self.iteration, self.V[self.push_ctr // 4][self.push_ctr % 4])
self.push_ctr += 1
self.iteration += 1
elif (self.iteration == 11): # get D_23
self.V[0][3] += d
self.V[1][3] -= d
self.V[2][3] -= d
self.raw_stats['pre_tr_ifmap_alu_comp'] += 3
self.raw_stats['pre_tr_ifmap_rf_rd'] += 3
self.raw_stats['pre_tr_ifmap_rf_wr'] += 3
self.ifmap_out_chn.push(
self.V[self.push_ctr // 4][self.push_ctr % 4]) # push v01
self.raw_stats['pre_tr_ifmap_rf_rd'] += 1 # read v01 from rf
#print ("pre transform ifmap - locx, locy, iteration, transformed ifmap: ", \
# self.locx, self.locy, self.iteration, self.V[self.push_ctr // 4][self.push_ctr % 4])
self.push_ctr += 1
self.iteration += 1
elif (self.iteration == 12): # get D_30
self.V[3][0] -= d
self.raw_stats['pre_tr_ifmap_alu_comp'] += 1
self.raw_stats['pre_tr_ifmap_rf_rd'] += 1
self.raw_stats['pre_tr_ifmap_rf_wr'] += 1
self.ifmap_out_chn.push(
self.V[self.push_ctr // 4][self.push_ctr % 4]) # push v02
self.raw_stats['pre_tr_ifmap_rf_rd'] += 1 # read v02 from rf
#print ("pre transform ifmap - locx, locy, iteration, transformed ifmap: ", \
# self.locx, self.locy, self.iteration, self.V[self.push_ctr // 4][self.push_ctr % 4])
self.push_ctr += 1
self.iteration += 1
elif (self.iteration == 13): # get D_31
self.V[3][1] -= d
self.V[3][2] += d
self.V[3][3] -= d
self.raw_stats['pre_tr_ifmap_alu_comp'] += 3
self.raw_stats['pre_tr_ifmap_rf_rd'] += 3
self.raw_stats['pre_tr_ifmap_rf_wr'] += 3
self.ifmap_out_chn.push(
self.V[self.push_ctr // 4][self.push_ctr % 4]) # push v03
self.raw_stats['pre_tr_ifmap_rf_rd'] += 1 # read v03 from rf
#print ("pre transform ifmap - locx, locy, iteration, transformed ifmap: ", \
# self.locx, self.locy, self.iteration, self.V[self.push_ctr // 4][self.push_ctr % 4])
self.push_ctr += 1
self.iteration += 1
elif (self.iteration == 14): # get D_32
self.V[3][0] += d
self.V[3][1] -= d
self.V[3][2] -= d
self.raw_stats['pre_tr_ifmap_alu_comp'] += 3
self.raw_stats['pre_tr_ifmap_rf_rd'] += 3
self.raw_stats['pre_tr_ifmap_rf_wr'] += 3
self.ifmap_out_chn.push(
self.V[self.push_ctr // 4][self.push_ctr % 4]) # push v10
self.raw_stats['pre_tr_ifmap_rf_wr'] += 1 # read v10 from rf
#print ("pre transform ifmap - locx, locy, iteration, transformed ifmap: ", \
# self.locx, self.locy, self.iteration, self.V[self.push_ctr // 4][self.push_ctr % 4])
self.push_ctr += 1
self.iteration += 1
elif (self.iteration == 15): # get D_33
self.V[3][3] += d
self.raw_stats['pre_tr_ifmap_alu_comp'] += 1
self.raw_stats['pre_tr_ifmap_rf_rd'] += 1
self.raw_stats['pre_tr_ifmap_rf_wr'] += 1
self.ifmap_out_chn.push(
self.V[self.push_ctr // 4][self.push_ctr % 4]) # push v11
self.raw_stats['pre_tr_ifmap_rf_wr'] += 1 # read v11 from rf
#print ("pre transform ifmap - locx, locy, iteration, transformed ifmap: ", \
# self.locx, self.locy, self.iteration, self.V[self.push_ctr // 4][self.push_ctr % 4])
self.push_ctr += 1
self.iteration += 1
elif self.iteration == 16 and self.ifmap_out_chn.vacancy(
): # done computing transform, push remaining V's sequentially
self.ifmap_out_chn.push(self.V[self.push_ctr // 4][self.push_ctr %
4])
self.raw_stats['pre_tr_ifmap_rf_rd'] += 1 # read vXX from rf
#print ("pre transform ifmap - locx, locy, iteration, transformed ifmap: ", \
# self.locx, self.locy, self.iteration, self.V[self.push_ctr // 4][self.push_ctr % 4])
self.push_ctr += 1
if self.push_ctr == 16: # all 16 transformed ifmap values have been pushed
self.transform_done.wr(True)
class OutputDeserializer(Module):
def instantiate(self, arch_output_chn, arr_x, arr_y, chn_per_word):
# PE static configuration (immutable)
self.arr_x = arr_x
self.arr_y = arr_y
self.chn_per_word = chn_per_word
self.arch_output_chn = arch_output_chn
self.ofmap = None
self.reference = None
self.image_size = (0, 0)
self.curr_set = 0
self.fmap_idx = 0
self.pass_done = Reg(False)
def configure(self, ofmap, reference, image_size):
self.ofmap = ofmap
self.reference = reference
self.image_size = image_size
self.curr_set = 0
self.fmap_idx = 0
self.pass_done.wr(False)
def tick(self):
if self.pass_done.rd():
return
# How many psums packets we expect to receive
out_sets = self.arr_x // self.chn_per_word
fmap_per_iteration = self.image_size[0] * self.image_size[1]
if self.arch_output_chn.valid():
data = [e for e in self.arch_output_chn.pop()]
# Calculate the end coords of where these ending psums must go...
x = self.fmap_idx % self.image_size[0]
y = self.fmap_idx // self.image_size[0]
if self.curr_set < out_sets:
cmin = self.curr_set * self.chn_per_word
cmax = cmin + self.chn_per_word
for c in range(cmin, cmax):
self.ofmap[x, y, c] = data[c - cmin]
self.curr_set += 1
# After recieving all the elements for pixel 0, do pixel 1, etc...
if self.curr_set == out_sets:
self.curr_set = 0
self.fmap_idx += 1
if self.fmap_idx == fmap_per_iteration:
self.fmap_idx = 0
self.pass_done.wr(True)
if np.all(self.ofmap == self.reference):
raise Finish("Success")
else:
print(self.ofmap)
print(self.reference)
print(self.ofmap - self.reference)
raise Finish("Validation Failed")
class OutputDeserializer(Module):
def instantiate(self, arch_output_chn, arr_x, arr_y, chn_per_word,
finish_signal_chn):
# PE static configuration (immutable)
self.arr_x = arr_x
self.arr_y = arr_y
self.chn_per_word = chn_per_word
self.arch_output_chn = arch_output_chn
self.finish_signal_chn = finish_signal_chn
self.ofmap = None
self.ofmap_transformed = None
self.reference = None
self.image_size = (0, 0)
self.curr_set = 0
self.fmap_idx = 0
self.pass_done = Reg(False)
def configure(self, ofmap, reference, image_size, bias):
# self.ofmap = np.zeros((2, 2, self.arr_x, 4)).astype(np.int64) # 2x2x8x4
self.ofmap = np.zeros((image_size[0], image_size[1],
self.arr_x)).astype(np.int64) # 4x4x8
self.reference = reference
self.num_tiles = 4
self.curr_tile = 0
self.image_size = image_size
self.bias = bias # TODO
self.curr_set = 0
self.fmap_idx = 0
self.curr_chn = 0
self.A_T = np.array([[1, 1, 1, 0], [0, 1, -1, -1]])
self.A = self.A_T.transpose()
self.pass_done.wr(False)
def tick(self):
if self.pass_done.rd():
# partly parallelized to be on chip:
# x_idx = (self.curr_tile // 2)*2
# y_idx = (self.curr_tile % 2)*2
# self.ofmap_transformed[x_idx:x_idx+2, y_idx:y_idx+2, self.curr_chn] += np.dot(self.A_T, np.dot(self.ofmap[:,:,self.curr_chn, self.curr_tile],self.A))
# self.curr_tile += 1
# if self.curr_tile == 4:
# self.curr_tile = 0
# self.ofmap_transformed[:,:,self.curr_chn] += self.bias[self.curr_chn] # add bias
# self.curr_chn += 1
# if self.curr_chn == 8:
# print ("reference shape: ", self.reference.shape)
# print ("ofmap shape: ", self.ofmap.shape)
# FOR LOOPS USED B/C NOT COUNTING OFF CHIP PROCESSING IN PERFORMANCE STATISTICS (will unroll loops in on chip processing)
# for k in range(8):
# self.ofmap_transformed[:,:,k] += self.bias[k] # add bias
# for t in range(self.num_tiles):
# x_idx = (t // 2)*2
# y_idx = (t % 2)*2
# self.ofmap_transformed[x_idx:x_idx+2,y_idx:y_idx+2,k] += np.dot(self.A_T,np.dot(self.ofmap[:,:,k,t],self.A))
# self.finish_signal_chn.push(True)
if np.all(self.ofmap == self.reference):
raise Finish("Success")
else:
print("ofmap: ")
print(self.ofmap)
print("reference: ")
print(self.reference)
print("difference: ")
print(self.ofmap - self.reference)
raise Finish("Validation Failed")
else:
#print ("output deser curr_tile, fmap_idx: ", self.curr_tile, self.fmap_idx)
out_sets = self.arr_x // self.chn_per_word # 2
fmap_per_iteration = 4 # ofmap size, parametrize .. TODO
if self.arch_output_chn.valid():
data = [e for e in self.arch_output_chn.pop()]
x_idx = (self.curr_tile // 2) * 2
y_idx = (self.curr_tile % 2) * 2
x = (self.fmap_idx % 2) + x_idx
y = self.fmap_idx // 2 + y_idx
# self.ofmap_transformed[x_idx:x_idx+2,y_idx:y_idx+2,k] += np.dot(self.A_T,np.dot(self.ofmap[:,:,k,t],self.A))
if self.curr_set < out_sets:
cmin = self.curr_set * self.chn_per_word
cmax = cmin + self.chn_per_word
for c in range(cmin, cmax):
self.ofmap[x, y, c] = data[c - cmin]
self.curr_set += 1
if self.curr_set == out_sets:
self.curr_set = 0
#self.fmap_idx += 1
self.curr_tile += 1
if self.curr_tile == 4:
self.fmap_idx += 1
self.curr_tile = 0
if self.fmap_idx == fmap_per_iteration:
self.fmap_idx = 0
self.curr_tile = 0
# self.ofmap = self.ofmap//(128*128)
self.pass_done.wr(True)
def instantiate(self, check_fifo=True):
self.check_fifo = check_fifo
self.counter = Reg(0)
self.fifo = FIFO(4)
class OutputDeserializer(Module):
def instantiate(self, arch_output_chn, psum_chn, arr_x, arr_y,
chn_per_word):
# PE static configuration (immutable)
self.arr_x = arr_x
self.arr_y = arr_y
self.chn_per_word = chn_per_word
self.arch_output_chn = arch_output_chn
self.psum_chn = psum_chn
self.ofmap = None
self.reference = None
self.image_size = (0, 0)
self.curr_set = 0
self.fmap_idx = 0
self.pass_done = Reg(False)
def configure(self, ofmap, reference, image_size, curr_pass):
if (curr_pass == 0): # so that ofmap doesnt get rewritten with zeros
self.ofmap = ofmap
self.reference = reference
self.image_size = image_size
self.curr_set = 0
self.fmap_idx = 0
self.curr_pass = curr_pass
self.num_passes = 4
self.pass_done.wr(False)
def tick(self):
if self.pass_done.rd():
return
out_sets = self.arr_x // self.chn_per_word
fmap_per_iteration = self.image_size[0] * self.image_size[1]
if self.arch_output_chn.valid() and (self.psum_chn.vacancy()
or self.curr_pass % 2 == 1):
data = [e for e in self.arch_output_chn.pop()]
if ((self.curr_pass %
2) == 0): # push ofmap psum to serializer on pass 0 and 2
self.psum_chn.push(data)
x = self.fmap_idx % self.image_size[0]
y = self.fmap_idx // self.image_size[0]
if self.curr_set < out_sets:
channel_offset = 0
if (self.curr_pass > 1):
channel_offset = 8
cmin = self.curr_set * self.chn_per_word + channel_offset
cmax = cmin + self.chn_per_word
for c in range(cmin, cmax):
self.ofmap[x, y, c] = data[c - cmin]
self.curr_set += 1
if self.curr_set == out_sets:
self.curr_set = 0
self.fmap_idx += 1
if self.fmap_idx == fmap_per_iteration:
self.fmap_idx = 0
if (self.curr_pass == self.num_passes - 1):
self.pass_done.wr(True)
if np.all(self.ofmap == self.reference):
raise Finish("Success")
else:
print(self.ofmap)
print(self.reference)
print(self.ofmap - self.reference)
raise Finish("Validation Failed")
def instantiate(self):
self.ra = Reg(0)
self.rb = Reg(10)
class InputSerializer(Module):
def instantiate(self, arch_input_chn, arr_y, block_size, num_nonzero,
pruner_name):
# PE static configuration (immutable)
#self.arr_x = arr_x
self.arr_y = arr_y
#self.chn_per_word = chn_per_word
self.block_size = block_size
self.num_nonzero = num_nonzero
self.convert_chn = Channel()
self.prune_chn = Channel()
self.arch_input_chn = arch_input_chn
# Although both InputSerializer and pruner will be pushing to arch_input_chn
# There is no conflict issue because all weights will be pushed by IS first
# then all inputs by pruner
self.converter = Converter(self.convert_chn, self.prune_chn, \
self.block_size, self.block_size)
# self.pruner = NaivePruner(self.prune_chn,self.arch_input_chn, \
# self.num_nonzero,True)
#user defined pruner for this layer, default to naive pruner
self.pruner = getattr(pruner, pruner_name)(self.prune_chn,self.arch_input_chn, \
self.num_nonzero, self.block_size, True)
self.ifmap = None
self.weights = None
self.bias = None
self.image_size = (0, 0)
self.filter_size = (0, 0)
self.ifmap_psum_done = True
self.pass_done = Reg(False)
# State Counters
self.curr_set = 0
self.curr_filter = 0
self.iteration = 0
self.fmap_idx = 0
self.curr_chn = 0
self.curr_x = 0 # run through first two dimensions of input
self.curr_y = 0
self.bias_set = 0
#self.send_bias = False
def configure(self, ifmap, weights, bias, in_chn, out_chn, image_size,
filter_size):
self.ifmap = ifmap
self.weights = weights
self.bias = bias
self.in_chn = in_chn
self.out_chn = out_chn
self.image_size = image_size
self.filter_size = filter_size
self.ifmap_psum_done = False
self.weights_done = False
self.pass_done.wr(False)
# State Counters
self.curr_set = 0
self.curr_filter = 0
self.iteration = 0
self.curr_chn = 0
self.curr_x = 0 # run through first two dimensions of input
self.curr_y = 0
self.bias_set = 0
#self.send_bias = False
def tick(self):
if self.pass_done.rd():
return
if self.ifmap_psum_done:
if self.convert_chn.vacancy():
data = np.zeros(self.block_size)
self.convert_chn.push(data)
return
in_sets = self.in_chn // self.block_size
out_sets = self.out_chn // self.block_size
num_iteration = self.filter_size[0] * self.filter_size[1]
# read and hold all weights at the beginning for ease of implementation
if not self.weights_done:
f_x = self.iteration // self.filter_size[0]
f_y = self.iteration % self.filter_size[0]
# Push filters to PE columns. (PE is responsible for pop)
if self.arch_input_chn.vacancy(
) and self.iteration < num_iteration:
cmin = self.curr_filter * self.block_size
cmax = cmin + self.block_size
data = np.array([self.weights[f_x, f_y, self.curr_chn, c] \
for c in range(cmin, cmax) ])
#print("{},{},{},{}-{}".format(f_x,f_y,self.curr_chn,cmin,cmax))
#print(data)
self.arch_input_chn.push(
data) # Gives groups of four along num_filters axis
self.curr_filter += 1
if (self.curr_filter == out_sets
): # Loop through blocks of filters
self.curr_filter = 0
self.curr_chn += 1
if (self.curr_chn == self.in_chn): # Loop through channels
self.curr_chn = 0
self.iteration += 1
if (self.iteration == num_iteration
): # Loop through 2D filter support
self.iteration = 0
#print("Weights done")
self.weights_done = True
elif self.arch_input_chn.vacancy() and self.bias_set < out_sets:
cmin = self.bias_set * self.block_size
cmax = cmin + self.block_size
data = np.array([self.bias[c] for c in range(cmin, cmax)])
#print("bias (input serializer):")
#print(data)
self.arch_input_chn.push(data)
self.bias_set += 1
elif not self.ifmap_psum_done:
if self.convert_chn.vacancy():
cmin = self.curr_set * self.block_size
cmax = cmin + self.block_size
#xmin = x
#xmax = x+self.arr_x
# Write ifmap to glb
#data = np.array([ self.ifmap[x, self.curr_y, self.curr_chn] for x in range(xmin, xmax) ])
data = np.array([
self.ifmap[self.curr_x, self.curr_y, c]
for c in range(cmin, cmax)
])
#print("{},{},{}-{}".format(self.curr_x, self.curr_y, cmin, cmax))
#print(data)
self.curr_set += 1
if (self.curr_set == in_sets):
self.curr_set = 0
self.curr_y += 1
if (self.curr_y == self.image_size[1]):
self.curr_y = 0
self.curr_x += 1
self.convert_chn.push(data)
if (self.curr_x == self.image_size[0]):
self.curr_x = 0
self.ifmap_psum_done = True
class PreTransformWeights(Module):
def instantiate(self, locx, locy, weight_in_chn, weight_out_chn):
self.locx = locx
self.locy = locy
self.weight_in_chn = weight_in_chn
self.weight_out_chn = weight_out_chn
self.transform_done = Reg(False)
self.stat_type = 'aggregate'
self.raw_stats = {
'pre_tr_weights_alu_comp': 0,
'pre_tr_weights_rf_rd': 0,
'pre_tr_weights_rf_wr': 0
}
def configure(self):
self.iteration = 0
self.push_ctr = 0
self.U = np.zeros([4, 4]).astype(np.int64)
self.transform_done.wr(False)
# Explanation of algorithm: transform filter weights G into U, performing Winograd transform U = H*G*H_T
# G = [G00 G01 G02
# G10 G11 G12
# G20 G21 G22]
#
# H = [1 0 0
# 0.5 0.5 0.5
# 0.5 -0.5 0.5
# 0 0 1 ]
#
# H_T = [1 0.5 0.5 0
# 0 0.5 -0.5 0
# 0 0.5 0.5 1]
#
# Performing this transform yields a 4x4 output for a given 4x4 input:
#
# U = [u00 u01 u02 u03
# u10 u11 u12 u13
# u20 u21 u22 u23
# u30 u31 u32 u33]
# ... such that:
# u00 = (G00)<<7;
# u01 = (G00 + G01 + G02)<<6;
# u02 = (G00 - G01 + G02)<<6;
# u03 = (G02)<<7;
# u10 = (G00 + G10 + G20)<<6;
# u11 = (G00 + G01 + G02 + G10 + G11 + G12 + G20 + G21 + G22)<<5;
# u12 = (G00 - G01 + G02 + G10 - G11 + G12 + G20 - G21 + G22)<<5;
# u13 = (G02 + G12 + G22)<<6;
# u20 = (G00 - G10 + G20)<<6;
# u21 = (G00 + G01 + G02 - G10 - G11 - G12 + G20 + G21 + G22)<<5;
# u22 = (G00 - G01 + G02 - G10 + G11 - G12 + G20 - G21 + G22)<<5;
# u23 = (G02 - G12 + G22)<<6;
# u30 = (G20)<<7;
# u31 = (G20 + G21 + G22)<<6;
# u32 = (G20 - G21 + G22)<<6;
# u33 = (G22)<<7;
def tick(self):
if self.transform_done.rd():
return
if self.weight_in_chn.valid() and self.weight_out_chn.vacancy():
g = (self.weight_in_chn.pop())
#print("pre tr weight: locx, locy, receive weight: ", self.locx, self.locy, g)
if (self.iteration == 0): # get G_00
self.U[0][0] += g
self.U[0][1] += g
self.U[0][2] += g
self.U[1][0] += g
self.U[1][1] += g
self.U[1][2] += g
self.U[2][0] += g
self.U[2][1] += g
self.U[2][2] += g
self.U[0][0] = self.U[0][0] * 128 # left shift by 7
self.raw_stats[
'pre_tr_weights_alu_comp'] += 10 #9 adds, 1 shift
self.raw_stats['pre_tr_weights_rf_rd'] += 9
self.raw_stats['pre_tr_weights_rf_wr'] += 9
self.weight_out_chn.push(
self.U[self.push_ctr // 4][self.push_ctr % 4]) # send U00
self.raw_stats[
'pre_tr_weights_rf_wr'] -= 1 # u00 sent immediately, not written back to rf
#print ("pre transform weights - locx, locy, iteration, transformed weight: ", \
# self.locx, self.locy, self.iteration, self.U[self.push_ctr // 4][self.push_ctr % 4])
self.push_ctr += 1
self.iteration += 1
elif (self.iteration == 1): # get G_01
self.U[0][1] += g
self.U[0][2] -= g
self.U[1][1] += g
self.U[1][2] -= g
self.U[2][1] += g
self.U[2][2] -= g
self.raw_stats['pre_tr_weights_alu_comp'] += 6
self.raw_stats['pre_tr_weights_rf_rd'] += 6
self.raw_stats['pre_tr_weights_rf_wr'] += 6
self.iteration += 1
elif (self.iteration == 2): # get G_02
self.U[0][1] += g
self.U[0][2] += g
self.U[0][3] += g
self.U[1][1] += g
self.U[1][2] += g
self.U[1][3] += g
self.U[2][1] += g
self.U[2][2] += g
self.U[2][3] += g
self.U[0][1] = self.U[0][1] * 64 # left shift by 6
self.U[0][2] = self.U[0][2] * 64 # left shift by 6
self.U[0][3] = self.U[0][3] * 128 # left shift by 7
self.raw_stats[
'pre_tr_weights_alu_comp'] += 12 #9 adds/subt, 3 shift
self.raw_stats['pre_tr_weights_rf_rd'] += 9
self.raw_stats['pre_tr_weights_rf_wr'] += 9
self.weight_out_chn.push(
self.U[self.push_ctr // 4][self.push_ctr % 4]) # send U01
self.raw_stats[
'pre_tr_weights_rf_wr'] -= 1 # u01 sent immediately, not written back to rf
#print ("pre transform weights - locx, locy, iteration, transformed weight: ", \
# self.locx, self.locy, self.iteration, self.U[self.push_ctr // 4][self.push_ctr % 4])
self.push_ctr += 1
self.iteration += 1
elif (self.iteration == 3): # get G_10
self.U[1][0] += g
self.U[1][1] += g
self.U[1][2] += g
self.U[2][0] -= g
self.U[2][1] -= g
self.U[2][2] -= g
self.raw_stats['pre_tr_weights_alu_comp'] += 6
self.raw_stats['pre_tr_weights_rf_rd'] += 6
self.raw_stats['pre_tr_weights_rf_wr'] += 6
self.weight_out_chn.push(
self.U[self.push_ctr // 4][self.push_ctr % 4]) # send U02
self.raw_stats['pre_tr_weights_rf_rd'] += 1 # read u02 from rf
#print ("pre transform weights - locx, locy, iteration, transformed weight: ", \
# self.locx, self.locy, self.iteration, self.U[self.push_ctr // 4][self.push_ctr % 4])
self.push_ctr += 1
self.iteration += 1
elif (self.iteration == 4): # get G_11
self.U[1][1] += g
self.U[1][2] -= g
self.U[2][1] -= g
self.U[2][2] += g
self.raw_stats['pre_tr_weights_alu_comp'] += 4
self.raw_stats['pre_tr_weights_rf_rd'] += 4
self.raw_stats['pre_tr_weights_rf_wr'] += 4
self.weight_out_chn.push(
self.U[self.push_ctr // 4][self.push_ctr % 4]) # send U03
self.raw_stats['pre_tr_weights_rf_rd'] += 1 # read u03 from rf
#print ("pre transform weights - locx, locy, iteration, transformed weight: ", \
# self.locx, self.locy, self.iteration, self.U[self.push_ctr // 4][self.push_ctr % 4])
self.push_ctr += 1
self.iteration += 1
elif (self.iteration == 5): # get G_12
self.U[1][1] += g
self.U[1][2] += g
self.U[1][3] += g
self.U[2][1] -= g
self.U[2][2] -= g
self.U[2][3] -= g
self.raw_stats['pre_tr_weights_alu_comp'] += 6
self.raw_stats['pre_tr_weights_rf_rd'] += 6
self.raw_stats['pre_tr_weights_rf_wr'] += 6
# no new completed weights this round
self.iteration += 1
elif (self.iteration == 6): # get G_20
self.U[1][0] += g
self.U[1][1] += g
self.U[1][2] += g
self.U[2][0] += g
self.U[2][1] += g
self.U[2][2] += g
self.U[3][0] += g
self.U[3][1] += g
self.U[3][2] += g
self.U[1][0] = self.U[1][0] * 64 # left shift 6
self.U[2][0] = self.U[2][0] * 64 # left shift 6
self.U[3][0] = self.U[3][0] * 128 # left shift 7
self.raw_stats[
'pre_tr_weights_alu_comp'] += 12 # 9 add, 3 shift ops
self.raw_stats['pre_tr_weights_rf_rd'] += 9
self.raw_stats['pre_tr_weights_rf_wr'] += 9
self.weight_out_chn.push(
self.U[self.push_ctr // 4][self.push_ctr % 4]) # send U10
self.raw_stats[
'pre_tr_weights_rf_wr'] -= 1 # send u10 immediately, w/o writing to rf
#print ("pre transform weights - locx, locy, iteration, transformed weight: ", \
# self.locx, self.locy, self.iteration, self.U[self.push_ctr // 4][self.push_ctr % 4])
self.push_ctr += 1
self.iteration += 1
elif (self.iteration == 7): # get G_21
self.U[1][1] += g
self.U[1][2] -= g
self.U[2][1] += g
self.U[2][2] -= g
self.U[3][1] += g
self.U[3][2] -= g
self.raw_stats[
'pre_tr_weights_alu_comp'] += 6 # 9 add, 3 shift ops
self.raw_stats['pre_tr_weights_rf_rd'] += 6
self.raw_stats['pre_tr_weights_rf_wr'] += 6
# no new completed weights this round
self.iteration += 1
elif (self.iteration == 8): # get G_22
#print ("iteration 8")
self.U[1][1] += g
self.U[1][2] += g
self.U[1][3] += g
self.U[2][1] += g
self.U[2][2] += g
self.U[2][3] += g
self.U[3][1] += g
self.U[3][2] += g
self.U[3][3] += g
self.raw_stats['pre_tr_weights_alu_comp'] += 9 # 9 add
self.raw_stats['pre_tr_weights_rf_rd'] += 9
self.raw_stats['pre_tr_weights_rf_wr'] += 9
self.iteration += 1
elif (self.iteration == 9 and self.weight_out_chn.vacancy()):
#print ("iteration 9")
self.U[1][1] = self.U[1][1] * 32 # left shift by 5
self.U[1][2] = self.U[1][2] * 32 # left shift by 5
self.U[1][3] = self.U[1][3] * 64
self.U[2][1] = self.U[2][1] * 32
self.U[2][2] = self.U[2][2] * 32
self.U[2][3] = self.U[2][3] * 64
self.U[3][1] = self.U[3][1] * 64
self.U[3][2] = self.U[3][2] * 64
self.U[3][3] = self.U[3][3] * 128
self.raw_stats['pre_tr_weights_alu_comp'] += 9 # 9 shift ops
self.raw_stats['pre_tr_weights_rf_rd'] += 9
self.raw_stats['pre_tr_weights_rf_wr'] += 9
self.weight_out_chn.push(self.U[self.push_ctr // 4][self.push_ctr %
4]) # send U11
self.raw_stats[
'pre_tr_weights_rf_wr'] -= 1 # send u11 immediately w/o writing to rf
# print ("pre transform weights - locx, locy, iteration, transformed weight: ", \
# self.locx, self.locy, self.iteration, self.U[self.push_ctr // 4][self.push_ctr % 4])
self.push_ctr += 1
self.iteration += 1
elif self.iteration == 10 and self.weight_out_chn.vacancy(
): # finish pushing transformed weights
self.weight_out_chn.push(self.U[self.push_ctr // 4][self.push_ctr %
4])
self.raw_stats['pre_tr_weights_rf_rd'] += 1 # read uXX from rf
#print ("pre transform weights - locx, locy, iteration, transformed weight: ", \
# self.locx, self.locy, self.iteration, self.U[self.push_ctr // 4][self.push_ctr % 4])
self.push_ctr += 1
if self.push_ctr == 16: # all 16 transformed weight values have been pushed
self.transform_done.wr(True)
class PostTransform(Module):
def instantiate(self, locx, locy, bias_chn, ofmap_in_chn, ofmap_out_chn): #ofmap_in
self.locx = locx
self.locy = locy
self.bias_chn = bias_chn
self.ofmap_in_chn = ofmap_in_chn
self.ofmap_out_chn = ofmap_out_chn
self.transform_done = Reg(False)
self.stat_type = 'aggregate'
self.raw_stats = {'post_tr_alu_comp' : 0, 'post_tr_rf_rd' : 0, 'post_tr_ifmap_rf_wr' : 0}
def configure(self):
self.bias = 0
self.iteration = 0
self.y00 = None
self.y01 = None
self.y10 = None
self.y11 = None
self.transform_done.wr(False)
self.bias_read = False
# Explanation of algorithm: transform ofmap M into y, performing inverse Winograd transform y = A_T*M*A
# M = [M00 M01 M02 M03
# M10 M11 M12 M13
# M20 M21 M22 M23
# M30 M31 M32 M33]
#
# A_T = [1 1 1 0
# 0 1 -1 -1]
#
# A = [1 0
# 1 1
# 1 -1
# 0 -1]
#
# Performing this transform yields a 2x2 output for a given 4x4 input:
#
# y = [y00 y01
# y10 y11]
# ... such that:
# y00 = M00+M01+M02+M10+M11+M12+M20+M21+M22
# y01 = M01-M02-M03+M11-M12-M13+M21-M22-M23
# y10 = M10+M11+M12-M20-M21-M22-M30-M31-M32
# y11 = M11-M12-M13-M21+M22+M23-M31+M32+M33
def tick(self):
if self.transform_done.rd():
return
if self.bias_chn.valid(): # should only ever be valid once
self.bias = self.bias_chn.pop()
self.bias_read = True
self.y00 = self.bias
self.y01 = self.bias
self.y10 = self.bias
self.y11 = self.bias
self.raw_stats['post_tr_alu_comp'] += 4
self.raw_stats['post_tr_ifmap_rf_wr'] += 4
elif self.ofmap_in_chn.valid() and self.ofmap_out_chn.vacancy():
m = (self.ofmap_in_chn.pop())//(128) # right shift by 7 bits
self.raw_stats['post_tr_alu_comp'] += 1
#print("post tr -- iteration ", self.iteration)
if (self.iteration == 0): # get M_00
self.y00 += m
self.raw_stats['post_tr_alu_comp'] += 1
self.raw_stats['post_tr_rf_rd'] += 1
self.raw_stats['post_tr_ifmap_rf_wr'] += 1
self.iteration += 1
elif (self.iteration == 1): # get M_01
self.y00 += m
self.y01 += m
self.raw_stats['post_tr_alu_comp'] += 2
self.raw_stats['post_tr_rf_rd'] += 2
self.raw_stats['post_tr_ifmap_rf_wr'] += 2
self.iteration += 1
elif (self.iteration == 2): # get M_02
self.y00 += m
self.y01 -= m
self.raw_stats['post_tr_alu_comp'] += 2
self.raw_stats['post_tr_rf_rd'] += 2
self.raw_stats['post_tr_ifmap_rf_wr'] += 2
self.iteration += 1
elif (self.iteration == 3): # get M_03
self.y01 -= m
self.raw_stats['post_tr_alu_comp'] += 1
self.raw_stats['post_tr_rf_rd'] += 1
self.raw_stats['post_tr_ifmap_rf_wr'] += 1
self.iteration += 1
elif (self.iteration == 4): # get M_10
self.y00 += m
self.y10 += m
self.raw_stats['post_tr_alu_comp'] += 2
self.raw_stats['post_tr_rf_rd'] += 2
self.raw_stats['post_tr_ifmap_rf_wr'] += 2
self.iteration += 1
elif (self.iteration == 5): # get M_11
self.y00 += m
self.y01 += m
self.y10 += m
self.y11 += m
self.raw_stats['post_tr_alu_comp'] += 4
self.raw_stats['post_tr_rf_rd'] += 4
self.raw_stats['post_tr_ifmap_rf_wr'] += 4
self.iteration += 1
elif (self.iteration == 6): # get M_12
self.y00 += m
self.y01 -= m
self.y10 += m
self.y11 -= m
self.raw_stats['post_tr_alu_comp'] += 4
self.raw_stats['post_tr_rf_rd'] += 4
self.raw_stats['post_tr_ifmap_rf_wr'] += 4
self.iteration += 1
elif (self.iteration == 7): # get M_13
self.y01 -= m
self.y11 -= m
self.raw_stats['post_tr_alu_comp'] += 2
self.raw_stats['post_tr_rf_rd'] += 2
self.raw_stats['post_tr_ifmap_rf_wr'] += 2
self.iteration += 1
elif (self.iteration == 8): # get M_20
self.y00 += m
self.y10 -= m
self.raw_stats['post_tr_alu_comp'] += 2
self.raw_stats['post_tr_rf_rd'] += 2
self.raw_stats['post_tr_ifmap_rf_wr'] += 2
self.iteration += 1
elif (self.iteration == 9): # get M_21
self.y00 += m
self.y01 += m
self.y10 -= m
self.y11 -= m
self.raw_stats['post_tr_alu_comp'] += 4
self.raw_stats['post_tr_rf_rd'] += 4
self.raw_stats['post_tr_ifmap_rf_wr'] += 4
self.iteration += 1
elif (self.iteration == 10 and self.bias_read == True): # get M_22
self.y00 += m
self.y01 -= m
self.y10 -= m
self.y11 += m
self.raw_stats['post_tr_alu_comp'] += 4
self.raw_stats['post_tr_rf_rd'] += 4
self.raw_stats['post_tr_ifmap_rf_wr'] += 4
self.iteration += 1
#print("post tr pushing y00: ", self.y00, self.bias)
self.ofmap_out_chn.push(self.y00) # y00 done
self.raw_stats['post_tr_ifmap_rf_wr'] -= 1 # send y00 immediately w/o writing to rf
elif (self.iteration == 11): # get M_23
self.y01 -= m
self.y11 += m
self.raw_stats['post_tr_alu_comp'] += 2
self.raw_stats['post_tr_rf_rd'] += 2
self.raw_stats['post_tr_ifmap_rf_wr'] += 2
self.iteration += 1
#print("post tr pushing y01: ", self.y01, self.bias)
self.ofmap_out_chn.push(self.y01) # y01 done
self.raw_stats['post_tr_ifmap_rf_wr'] -= 1 # send y01 immediately w/o writing to rf
elif (self.iteration == 12): # get M_30
self.y10 -= m
self.raw_stats['post_tr_alu_comp'] += 1
self.raw_stats['post_tr_rf_rd'] += 1
self.raw_stats['post_tr_ifmap_rf_wr'] += 1
self.iteration += 1
elif (self.iteration == 13): # get M_31
self.y10 -= m
self.y11 -= m
self.raw_stats['post_tr_alu_comp'] += 2
self.raw_stats['post_tr_rf_rd'] += 2
self.raw_stats['post_tr_ifmap_rf_wr'] += 2
self.iteration += 1
elif (self.iteration == 14): # get M_32
self.y10 -= m
self.y11 += m
self.raw_stats['post_tr_alu_comp'] += 2
self.raw_stats['post_tr_rf_rd'] += 2
self.raw_stats['post_tr_ifmap_rf_wr'] += 2
self.iteration += 1
#print("post tr pushing y10: ", self.y10, self.bias)
self.ofmap_out_chn.push(self.y10) # y10 done
self.raw_stats['post_tr_ifmap_rf_wr'] -= 1 # send y10 immediately w/o writing to rf
elif (self.iteration == 15): # get M_33
self.y11 += m
self.raw_stats['post_tr_alu_comp'] += 1
self.raw_stats['post_tr_rf_rd'] += 1
self.raw_stats['post_tr_ifmap_rf_wr'] += 1
self.iteration += 1
#print("post tr pushing y11: ", self.y11, self.bias)
self.ofmap_out_chn.push(self.y11) # y11 done
self.raw_stats['post_tr_ifmap_rf_wr'] -= 1 # send y11 immediately w/o writing to rf
#self.iteration += 1
if self.iteration == 16:
self.transform_done.wr(True)
class Channel(Module):
def instantiate(self, depth=2):
self.data = [None] * depth
self.depth = depth
self.rd_ptr = Reg(0)
self.wr_ptr = Reg(0)
def peek(self, idx=0):
if not self.valid(idx):
raise ChannelError("Reading from empty channel")
return self.data[(self.rd_ptr.rd() + idx) % self.depth]
def push(self, x):
if not self.vacancy():
raise ChannelError("Enqueueing into full channel")
self.data[self.wr_ptr.rd() % self.depth] = x
self.wr_ptr.wr((self.wr_ptr.rd() + 1) % (2 * self.depth))
def free(self, count=1):
if not self.valid(count - 1):
raise ChannelError("Dequeueing from empty channel")
self.rd_ptr.wr((self.rd_ptr.rd() + count) % (2 * self.depth))
def pop(self):
self.free(1)
return self.peek(0)
def valid(self, idx=0):
return ((self.wr_ptr.rd() - self.rd_ptr.rd()) % (2 * self.depth)) > idx
def vacancy(self, idx=0):
return ((self.rd_ptr.rd() + self.depth - self.wr_ptr.rd()) %
(2 * self.depth)) > idx
def clear(self):
# Use with care since it conflicts with enq and deq
self.rd_ptr.wr(self.wr_ptr.rd())
class InputSerializer(Module):
def instantiate(self, arch_input_chn, arr_x, arr_y, chn_per_word):
# PE static configuration (immutable)
self.arr_x = arr_x
self.arr_y = arr_y
self.chn_per_word = chn_per_word
self.arch_input_chn = arch_input_chn
self.ifmap = None
self.weights = None
self.bias = None
self.image_size = (0, 0)
self.filter_size = (0, 0)
self.ifmap_psum_done = True
self.pass_done = Reg(False)
# State Counters
self.curr_set = 0
self.curr_filter = 0
self.iteration = 0
self.fmap_idx = 0
self.bias_idx = 0
self.weight_idx = 0
def configure(self, ifmap, weights, bias, image_size, filter_size):
self.ifmap = ifmap
self.weights = weights
self.bias = bias
self.image_size = image_size
self.filter_size = filter_size
self.bias_wr_done = False
self.fmap_wr_done = False
self.weight_wr_done = False
self.pass_done.wr(False)
self.send_ifmap = True # used to interleave sending weights and ifmaps to chip
self.bias_sets = 2
def tick(self):
if self.pass_done.rd():
return
# in_sets = self.arr_y//self.chn_per_word
# out_sets = self.arr_x//self.chn_per_word
fmap_per_iteration = self.image_size[0] * self.image_size[1]
num_iteration = self.filter_size[0] * self.filter_size[1]
weights_per_filter = self.filter_size[0] * self.filter_size[1]
if self.arch_input_chn.vacancy() and not self.pass_done.rd():
if not self.bias_wr_done:
kmin = self.bias_idx * self.chn_per_word
kmax = kmin + self.chn_per_word
data = np.array([self.bias[k] for k in range(kmin, kmax)])
self.bias_idx += 1
#print ("input ser kmin,kmax,bias: ",kmin,kmax,data)
elif (not self.fmap_wr_done) and self.send_ifmap: # send ifmap
# send 4 elements of ifmap
x = self.fmap_idx % self.image_size[0]
y = self.fmap_idx // self.image_size[0]
cmin = self.curr_set * self.chn_per_word # 0
cmax = cmin + self.chn_per_word # 4
data = np.array(
[self.ifmap[x, y, c] for c in range(cmin, cmax)])
self.fmap_idx += 1
#print ("input ser x,y,cmin,cmax,ifmaps: ",x,y,cmin,cmax,data)
self.send_ifmap = False
else: # send weight
# send 4 elements of weights (twice in succession)
x = self.weight_idx % self.filter_size[0]
y = self.weight_idx // self.filter_size[1]
cmin = 0
cmax = cmin + self.chn_per_word
data = np.array([
self.weights[x, y, c, self.curr_filter]
for c in range(cmin, cmax)
])
self.curr_filter += 1
if (not self.fmap_wr_done):
self.send_ifmap = True
#print ("input ser x,y,cmin,cmax,curr_filter,weights: ",x,y,cmin,cmax,self.curr_filter,data)
self.arch_input_chn.push(data)
if self.fmap_idx == fmap_per_iteration:
self.fmap_wr_done = True
self.fmap_idx = 0
if self.curr_filter == self.arr_x:
self.weight_idx += 1
self.curr_filter = 0
if self.weight_idx == weights_per_filter:
self.weight_wr_done = True
self.pass_done.wr(True)
if self.bias_idx == self.bias_sets: #2
self.bias_wr_done = True
def instantiate(self, depth=2):
self.data = [None] * depth
self.depth = depth
self.rd_ptr = Reg(0)
self.wr_ptr = Reg(0)
class InputSerializer(Module):
def instantiate(self, arch_input_chn, arr_x, arr_y, chn_per_word):
# PE static configuration (immutable)
self.arr_x = arr_x
self.arr_y = arr_y
self.chn_per_word = chn_per_word
self.arch_input_chn = arch_input_chn
self.ifmap = None
self.weights = None
self.image_size = (0, 0)
self.filter_size = (0, 0)
self.ifmap_psum_done = True
self.pass_done = Reg(False)
# State Counters
self.curr_set = 0
self.curr_filter = 0
self.iteration = 0
self.fmap_idx = 0
self.bias_idx = 0
def configure(self, ifmap, weights, biases, image_size, filter_size):
#self.ifmap = ifmap
#self.weights = weights
self.biases = biases
self.image_size = image_size
self.filter_size = image_size
self.num_tiles = 4
self.send_ifmap = True
self.fmap_idx = 0
self.fmap_tile = 0
self.weight_idx = 0
self.bias_sets = 2
self.fmap_wr_done = False
self.weight_wr_done = False
self.bias_wr_done = False
self.pass_done.wr(False)
# pad the ifmaps
ifmap_padded = np.pad(ifmap, 1, 'constant')
ifmap_padded = ifmap_padded[:, :, 1:5]
# Winograd transforms
B_T = np.array([[1, 0, -1, 0], [0, 1, 1, 0], [0, -1, 1, 0],
[0, 1, 0, -1]])
B = B_T.transpose()
G = np.array([[1, 0, 0], [0.5, 0.5, 0.5], [0.5, -0.5, 0.5], [0, 0, 1]])
G_T = G.transpose()
A_T = np.array([[1, 1, 1, 0], [0, 1, -1, -1]])
A = A_T.transpose()
C = 4 # num channels
K = 8 # num filters
T = 4 # num tiles
U = np.zeros([4, 4, C, K]) # 4,4,4,8
V = np.zeros([4, 4, C, T]) # 4,4,4
# FOR LOOPS USED B/C NOT COUNTING OFF CHIP PROCESSING IN PERFORMANCE STATISTICS (will unroll loops in on chip processing)
for t in range(T):
for k in range(K): # filter
for c in range(C): # channel
g = weights[:, :, c, k] # 3x3 filter
U[:, :, c, k] = np.dot(G, np.dot(g, G_T)) # 4x4
for c in range(C): # channel
x_idx = (t // 2) * 2
y_idx = (t % 2) * 2
d = ifmap_padded[x_idx:x_idx + 4, y_idx:y_idx + 4,
c] # 4x4 ifmap tile
V[:, :, c, t] = np.dot(B_T, np.dot(d, B))
# Convert to integers for on chip processing, LOSE ACCURACY -> bit shift
U = 128 * U
# left shift by 7 bits to avoid precision loss when convert float to int
#V = 128*V;
self.weights = U.astype(np.int64) # transformed weights
self.ifmap = V.astype(np.int64) # transformed ifmap
#print ("U ser: ", self.weights)
#print ("V ser: ", self.ifmap)
def tick(self):
if self.pass_done.rd():
return
in_sets = self.arr_y // self.chn_per_word # 1
out_sets = self.arr_x // self.chn_per_word # 2
fmap_per_iteration = self.image_size[0] * self.image_size[1]
weights_per_filter = self.filter_size[0] * self.filter_size[1]
if self.arch_input_chn.vacancy() and not self.pass_done.rd():
if not self.bias_wr_done:
kmin = self.bias_idx * self.chn_per_word
kmax = kmin + self.chn_per_word
data = np.array([self.biases[k] for k in range(kmin, kmax)])
self.bias_idx += 1
#print ("input ser kmin,kmax,biases: ",kmin,kmax,data)
elif not self.fmap_wr_done: # send ifmap
# send 4 elements of ifmap
x = self.fmap_idx % self.image_size[0]
y = self.fmap_idx // self.image_size[0]
cmin = self.curr_set * self.chn_per_word # 0
cmax = cmin + self.chn_per_word # 4
data = np.array([
self.ifmap[x, y, c, self.fmap_tile]
for c in range(cmin, cmax)
])
self.fmap_tile += 1
#print ("input ser x,y,cmin,cmax,ifmaps: ",x,y,cmin,cmax,data)
else: # send weight
# send 4 elements of weights (twice in succession)
x = self.weight_idx % self.filter_size[0]
y = self.weight_idx // self.filter_size[1]
cmin = 0
cmax = cmin + self.chn_per_word
data = np.array([
self.weights[x, y, c, self.curr_filter]
for c in range(cmin, cmax)
])
self.curr_filter += 1
#print ("input ser x,y,cmin,cmax,curr_filter,weights: ",x,y,cmin,cmax,self.curr_filter,data)
self.arch_input_chn.push(data)
if self.fmap_tile == self.num_tiles:
self.fmap_tile = 0
self.fmap_idx += 1
if self.fmap_idx == fmap_per_iteration:
self.fmap_wr_done = True
self.fmap_idx = 0
if self.curr_filter == self.arr_x:
self.weight_idx += 1
self.curr_filter = 0
if self.weight_idx == weights_per_filter:
self.weight_wr_done = True
self.pass_done.wr(True)
if self.bias_idx == self.bias_sets: #2
self.bias_wr_done = True
class OutputDeserializer(Module):
def instantiate(self, arch_output_chn, arr_y, block_size, num_nonzero):
# PE static configuration (immutable)
self.arr_y = arr_y
self.block_size = block_size
self.num_nonzero = num_nonzero
self.arch_output_chn = arch_output_chn
self.ofmap = None
self.reference = None
self.image_size = (0, 0)
self.curr_set = 0
self.fmap_idx = 0
self.pass_done = Reg(False)
def configure(self, ofmap, reference, image_size, out_chn):
self.ofmap = ofmap
self.reference = reference
self.out_chn = out_chn
self.image_size = image_size
self.curr_set = 0
self.fmap_idx = 0
self.pass_done.wr(False)
def tick(self):
if self.pass_done.rd():
return
out_sets = self.out_chn // self.block_size
fmap_per_iteration = self.image_size[0] * self.image_size[1]
if self.arch_output_chn.valid():
rcvd = self.arch_output_chn.pop()
loc_tag = [e[0] for e in rcvd]
data = [e[1] for e in rcvd]
#print(loc_tag)
x = loc_tag[0] // self.image_size[1]
y = loc_tag[0] % self.image_size[1]
#x = self.fmap_idx % self.image_size[0]
#y = self.fmap_idx // self.image_size[0]
self.fmap_idx = x + y * self.image_size[0]
#print("{},{} received (output deserializer)".format(x,y))
#print(data)
if self.curr_set < out_sets:
cmin = self.curr_set * self.block_size
cmax = cmin + self.block_size
for c in range(cmin, cmax):
assert (self.ofmap[x, y, c] == 0
) # should never replace an existing value
self.ofmap[x, y, c] = data[c - cmin]
self.curr_set += 1
if self.curr_set == out_sets:
self.curr_set = 0
self.fmap_idx += 1
if self.fmap_idx == fmap_per_iteration:
self.fmap_idx = 0
self.pass_done.wr(True)
raise Finish("Done processing")