def visitSum(self, node: AST.Sum):
'''
expr_out
i = 0
for (j = 0; j < n; j++)
expr_in = prog_in
expr_out = expr_out + expr_in
i++
1. for i in [0, C]:
2. expr_out[i] = expr_out[i] + shr(expr_in[i])
'''
var_idf = node.name
self.decls[var_idf] = Type.Int()
(prog_in, expr_in) = self.visit(node.expr)
start, end = node.start, node.end
expr_out = self.getTempVar()
type_out = node.type
var = IR.Var(var_idf)
var_iter = self.getTempIterator()
iters = self.getTempIterators(type_out.dim)
(scale_out, height_shr,
height_noshr) = self.getScaleForTreeSum(self.scales[expr_in.idf],
end - start)
intv_out = self.getIntervalForTreeSum(self.intvs[expr_in.idf],
end - start)
# Tree sum to sum output of each iteration
expr_in_idx = IRUtil.addIndex(expr_in, iters)
expr_out_idx = IRUtil.addIndex(expr_out, iters)
cmd1 = IR.Memset(expr_out, type_out.size())
cmd2 = IR.Assn(
expr_out_idx,
IRUtil.add(expr_out_idx, IRUtil.shr(expr_in_idx, height_shr)))
treeSum = IRUtil.loop(type_out.shape, iters, [cmd2])
# Final program to sum output of each iteration
prog_sum = [
cmd1,
IR.Assn(var, IR.Int(start)),
IR.For(var_iter, 0, IRUtil.lt(var_iter, IR.Int(end - start)),
prog_in.cmd_l + treeSum + [IR.Assn(var, IRUtil.inc(var))])
]
prog_out = IR.Prog(prog_sum)
self.decls[expr_out.idf] = type_out
self.scales[expr_out.idf] = scale_out
self.intvs[expr_out.idf] = intv_out
return (prog_out, expr_out)
def printSuffix(self, expr: IR.Expr):
self.out.printf('\n')
type = self.decls[expr.idf]
if Type.isInt(type):
self.out.printf('return ', indent=True)
self.print(expr)
self.out.printf(';\n')
elif Type.isTensor(type):
idfr = expr.idf
exponent = self.scales[expr.idf]
num = 2**exponent
if type.dim == 0:
self.out.printf('cout << ', indent=True)
self.out.printf('float(' + idfr + ')*' + str(num))
self.out.printf(' << endl;\n')
else:
iters = []
for i in range(type.dim):
s = chr(ord('i') + i)
tempVar = IR.Var(s)
iters.append(tempVar)
expr_1 = IRUtil.addIndex(expr, iters)
cmds = IRUtil.loop(type.shape, iters,
[IR.PrintAsFloat(expr_1, exponent)])
self.print(IR.Prog(cmds))
else:
assert False
self.out.decreaseIndent()
self.out.printf('}\n', indent=True)
self.out.close()
def visitMathExp(self, node: AST.Func):
# Tunable parameter
MIN = 0.1
(prog_in, expr_in) = self.visit(node.expr)
type_in = node.expr.type
scale_in = self.scales[expr_in.idf]
intv_in = self.intvs[expr_in.idf]
'''
1. y = ((int) (exp(((float)e) / shr1) * shr2))
'''
maxExp = np.exp(-MIN)
expr_out = self.getTempVar()
scale_out = self.getScale(maxExp)
intv_out = self.getInterval(scale_out, maxExp, maxExp)
shr1 = IR.Int(2**-scale_in)
shr2 = IR.Int(2**-scale_out)
expr_in_idx = IRUtil.addIndex(expr_in, [IRUtil.zero] * type_in.dim)
expr_out_idx = IRUtil.addIndex(expr_out, [IRUtil.zero] * type_in.dim)
cmd0 = IR.Comment('exp(' + expr_in.idf + ')')
cmd_assn = IR.Assn(
expr_out_idx,
IRUtil.castToInt(
IRUtil.mul(
IR.Exp(IRUtil.div(IRUtil.castToFloat(expr_in_idx), shr1)),
shr2)))
prog_exp = IR.Prog([cmd0, cmd_assn])
prog_out = IRUtil.concatPrograms(prog_in, prog_exp)
self.decls[expr_out.idf] = type_in
self.scales[expr_out.idf] = scale_out
self.intvs[expr_out.idf] = intv_out
return (prog_out, expr_out)
def visitIndex(self, node: AST.Index):
(prog_in, expr_in) = self.visit(node.expr)
(prog_idx, expr_idx) = self.visit(node.index)
prog_out = IRUtil.concatPrograms(prog_in, prog_idx)
expr_out = IRUtil.addIndex(expr_in, [expr_idx])
return (prog_out, expr_out)
def visitUop(self, node: AST.Uop):
(prog_in, expr_in) = self.visit(node.expr)
op = node.op
if op == SeeDotParser.ADD:
return (prog_in, expr_in)
assert op == SeeDotParser.SUB
type_out = node.type
# e : Int
if Type.isInt(type_out):
prog_out = prog_in
expr_out = IRUtil.negate(expr_in)
# e: Tensor(), or Tensor(..)
else:
expr_out = self.getTempVar()
iters = self.getTempIterators(type_out.dim)
scale_out = self.scales[expr_in.idf]
(m, M) = self.intvs[expr_in.idf]
intv_out = (-M, -m)
expr_in_idx = IRUtil.addIndex(expr_in, iters)
expr_out_idx = IRUtil.addIndex(expr_out, iters)
rhs = IRUtil.negate(expr_in_idx)
loop = IRUtil.loop(type_out.shape, iters,
[IR.Assn(expr_out_idx, rhs)])
prog_uop = IR.Prog(loop)
prog_out = IRUtil.concatPrograms(prog_in, prog_uop)
self.decls[expr_out.idf] = type_out
self.scales[expr_out.idf] = scale_out
self.intvs[expr_out.idf] = intv_out
return (prog_out, expr_out)
def visitSgn(self, node: AST.Func):
(prog_in, expr_in) = self.visit(node.expr)
expr_out = self.getTempVar()
type_in = node.expr.type
expr_in_idx = IRUtil.addIndex(expr_in, [IRUtil.zero] * type_in.dim)
cmd0 = IR.Comment('sgn(' + expr_in.idf + ')')
cmd1 = IR.Assn(expr_out,
IRUtil.cond_zero(expr_in_idx, IRUtil.one, IRUtil.zero))
prog_sgn = IR.Prog([cmd0, cmd1])
prog_out = IRUtil.concatPrograms(prog_in, prog_sgn)
self.decls[expr_out.idf] = Type.Int()
return (prog_out, expr_out)
def printSuffix(self, expr: IR.Expr):
self.out.printf('\n')
type = self.decls[expr.idf]
if Type.isInt(type):
self.out.printf('return ', indent=True)
self.print(expr)
self.out.printf(';\n')
elif Type.isTensor(type):
idfr = expr.idf
exponent = self.scales[expr.idf]
num = 2**exponent
if type.dim == 0:
self.out.printf('Serial.println(', indent=True)
self.out.printf('float(' + idfr + ')*' + str(num))
self.out.printf(', 6);\n')
else:
iters = []
for i in range(type.dim):
s = chr(ord('i') + i)
tempVar = IR.Var(s)
iters.append(tempVar)
expr_1 = IRUtil.addIndex(expr, iters)
cmds = IRUtil.loop(type.shape, iters,
[IR.PrintAsFloat(expr_1, exponent)])
self.print(IR.Prog(cmds))
else:
assert False
self.out.decreaseIndent()
self.out.printf('}\n', indent=True)
self.out.close()
with open(os.path.join(self.outputDir, "ram.usage"), "w") as f:
f.write("Estimate RAM usage :: %d bytes" % (self.maxRAMestimate))
def printSuffix(self, expr: IR.Expr):
self.out.printf('\n')
if config.vbwEnabled and forFixed():
bw = self.varsForBitwidth['X']
typ_str = "int%d_t" % bw
size = self.decls['X'].shape
sizestr = ''.join([("[%d]" % i) for i in size])
Xindexstr = ''
Xintstar = ''.join(["*" for i in size])
for i in range(len(size)):
Xindexstr += (("[i%d]" % (i - 1)) if i > 0 else "")
self.out.printf("for (int i%d = 0; i%d < %d; i%d ++ ){\n" %
(i, i, size[i], i),
indent=True)
self.out.increaseIndent()
for i in range(len(size) - 1, -1, -1):
self.out.decreaseIndent()
self.out.printf("}\n", indent=True)
self.out.printf("delete[] X%s;\n" % (Xindexstr), indent=True)
Xindexstr = Xindexstr[:-4] if len(Xindexstr) > 0 else Xindexstr
assert len(
size
) < 10, "Too simple logic for printing indices used, cannot handle 10+ Dim Tensors"
type = self.decls[expr.idf]
if Type.isInt(type):
self.out.printf('return ', indent=True)
self.print(expr)
self.out.printf(';\n')
elif Type.isTensor(type):
idfr = expr.idf
exponent = self.scales[expr.idf]
num = 2**exponent
if type.dim == 0:
self.out.printf('cout << ', indent=True)
self.out.printf('float(' + idfr + ')*' + str(num))
self.out.printf(' << endl;\n')
else:
iters = []
for i in range(type.dim):
s = chr(ord('i') + i)
tempVar = IR.Var(s)
iters.append(tempVar)
expr_1 = IRUtil.addIndex(expr, iters)
cmds = IRUtil.loop(type.shape, iters,
[IR.PrintAsFloat(expr_1, exponent)])
self.print(IR.Prog(cmds))
else:
assert False
self.out.decreaseIndent()
self.out.printf('}\n', indent=True)
def isInt(a):
try:
int(a)
return True
except:
return False
if forFixed():
if (int(self.printSwitch) if isInt(self.printSwitch) else -2) > -1:
self.out.printf("const int switches = %d;\n" %
(int(self.printSwitch)),
indent=True)
self.out.printf(
'void seedotFixedSwitch(int i, MYINT **X_temp, int& res) {\n',
indent=True)
self.out.increaseIndent()
self.out.printf('switch(i) {\n', indent=True)
self.out.increaseIndent()
for i in range(int(self.printSwitch)):
self.out.printf(
'case %d: res = seedotFixed%d(X_temp); return;\n' %
(i, i + 1),
indent=True)
self.out.printf('default: res = -1; return;\n', indent=True)
self.out.decreaseIndent()
self.out.printf('}\n', indent=True)
self.out.decreaseIndent()
self.out.printf('}\n', indent=True)
if debugCompiler():
print("Closing File after outputting cpp code: ID " + self.idStr)
self.out.close()
def visitReshape(self, node: AST.Reshape):
(prog_in, expr_in) = self.visit(node.expr)
'''
reshape(A, n, h, w)
cmd1: t1 = t2 = 0;
loop2: for n in 0:N:
for h in 0:H:
for w in 0:W:
cmd3: B[n][h][w] = A[t1][t2][t3]
cmd4: t3++;
cmd5: if (t3 == WW)
t3 = 0;
t2++;
if (t2 == HH)
t2 = 0;
t1++;
'''
type_in = node.expr.type
type_out = node.type
# Compute scaling factors
scale_out = self.scales[expr_in.idf]
intv_out = self.intvs[expr_in.idf]
# Declare variables
expr_out = self.getTempVar()
iters_in = self.getTempIterators(type_in.dim)
iters_out = self.getTempVars(type_out.dim)
# Initialize to 0
cmd1 = [IR.Assn(var, IRUtil.zero) for var in iters_out]
# Incrementing the first index
first_iter = iters_out[0]
cmd4 = IRUtil.incCmd(first_iter)
# Incrementing other indices using a loop
cmd5 = [cmd4]
for i in range(1, type_out.dim):
curr_iter = iters_out[i]
curr_size = IR.Int(type_out.shape[i])
cmd5 = [
IRUtil.incCmd(curr_iter),
IR.If(IRUtil.eq(curr_iter, curr_size),
[IRUtil.initVarToZero(curr_iter)] + cmd5)
]
# Outer loop
loopShape = []
loopIters = []
for order in node.order:
order = order - 1
loopShape.append(type_in.shape[order])
loopIters.append(iters_in[order])
loop2 = IRUtil.loop(loopShape, loopIters, [
IR.Assn(IRUtil.addIndex(expr_out, iters_out),
IRUtil.addIndex(expr_in, iters_in))
] + cmd5)
# Finalize
comment = IR.Comment("reshape(" + expr_in.idf + ", " +
', '.join(str(e) for e in type_out.shape) + ")")
prog_reshape = IR.Prog([comment] + cmd1 + loop2)
prog_out = IRUtil.concatPrograms(prog_in, prog_reshape)
# Update context
self.decls[expr_out.idf] = type_out
self.scales[expr_out.idf] = scale_out
self.intvs[expr_out.idf] = intv_out
# Update declarations
self.decls.update(dict((var.idf, Type.Int()) for var in iters_out))
return (prog_out, expr_out)
def visitCond(self, node: AST.Cond):
(prog_in_cond, expr_in_cond) = self.visit(node.expr)
(prog_in_A, expr_in_A) = self.visit(node.trueBlock)
(prog_in_B, expr_in_B) = self.visit(node.falseBlock)
type_in_cond = node.expr.type
type_in_A = node.trueBlock.type
if Type.isInt(type_in_cond):
expr_in_cond_idx = expr_in_cond
else:
expr_in_cond_idx = IRUtil.addIndex(expr_in_cond, [IRUtil.zero] *
type_in_cond.dim)
# e2,e3 : Int
if Type.isInt(type_in_A):
# TODO: Update the scale and intv of expr_out based on in_A and
# in_B
prog_out = IRUtil.concatPrograms(prog_in_cond, prog_in_A,
prog_in_B)
expr_out = IRUtil.cond_zero(expr_in_cond_idx, expr_in_A, expr_in_B)
# e2,e3 : Tensor(), or Tensor(..)
else:
expr_out = self.getTempVar()
iters = self.getTempIterators(type_in_A.dim)
scale_in_A, scale_in_B = self.scales[expr_in_A.idf], self.scales[
expr_in_B.idf]
intv_in_A, intv_in_B = self.intvs[expr_in_A.idf], self.intvs[
expr_in_B.idf]
m_2, M_2 = intv_in_A
m_3, M_3 = intv_in_B
if scale_in_A >= scale_in_B:
shr_n_2, shr_n_3 = 0, scale_in_A - scale_in_B
else:
shr_n_2, shr_n_3 = scale_in_B - scale_in_A, 0
scale_out = max(scale_in_A, scale_in_B)
intv_out = (min(m_2 >> shr_n_2,
m_3 >> shr_n_3), max(M_2 >> shr_n_2,
M_3 >> shr_n_3))
# prog_assn
expr_in_A_idx = IRUtil.addIndex(expr_in_A, iters)
expr_in_B_idx = IRUtil.addIndex(expr_in_B, iters)
expr_out_idx = IRUtil.addIndex(expr_out, iters)
rhs = IRUtil.cond_zero(expr_in_cond_idx,
IRUtil.shr(expr_in_A_idx, shr_n_2),
IRUtil.shr(expr_in_B_idx, shr_n_3))
cmdl_assn = IRUtil.loop(type_in_A.shape, iters,
[IR.Assn(expr_out_idx, rhs)])
prog_cond = IR.Prog(cmdl_assn)
prog_out = IRUtil.concatPrograms(prog_in_cond, prog_in_A,
prog_in_B, prog_cond)
self.decls[expr_out.idf] = type_in_A
self.scales[expr_out.idf] = scale_out
self.intvs[expr_out.idf] = intv_out
return (prog_out, expr_out)
def visitTableExp(self, node: AST.Func):
(prog_in, expr_in) = self.visit(node.expr)
# TODO: use MAX_VAL_EXP
type_in = node.expr.type
scale_in = self.scales[expr_in.idf]
intv_in = self.intvs[expr_in.idf]
[m, M] = self.expRange
[m_scale,
M_scale] = [int(np.ldexp(m, -scale_in)),
int(np.ldexp(M, -scale_in))]
max = int(np.ldexp(M - m, -scale_in))
shl = self.getShl(max)
input = self.getTempVar()
[i, j] = self.getTempVars(2)
expr_out = self.getTempVar()
'''
1. if ((-x) < min) {
2. i = 0;
3. j = 0;
4. }
5. else {
6. y = ((-x) - min) << shl
7. i = (y >> shrI) & (2^b-1)
8. j = (y >> shrJ) & (2^b-1)
9. }
10. ans = T[i] * U[j]
'''
mask = IR.Int(2**self.expB - 1)
shrI = Common.wordLength - self.expB
shrJ = Common.wordLength - self.expB * 2
table = self.getExpTable(scale_in)
scale1 = self.getScale(1)
scale2 = self.getScale(abs(np.exp(-m)))
[shr1, shr2] = self.getShrForMul(scale1, scale2)
expr_1_elt = IRUtil.addIndex(expr_in, [IRUtil.zero] * type_in.dim)
expr_2_elt = IRUtil.addIndex(expr_out, [IRUtil.zero] * type_in.dim)
cond = IRUtil.lt(IRUtil.negate(expr_1_elt), IR.Int(m_scale))
cmd2 = IR.Assn(i, IR.Int(0))
cmd3 = IR.Assn(j, IR.Int(0))
cmd6 = IR.Assn(
input,
IRUtil.shl(IRUtil.sub(IRUtil.negate(expr_1_elt), IR.Int(m_scale)),
shl))
cmd7 = IR.Assn(i, IRUtil.bitAnd(IRUtil.shrUint(input, shrI), mask))
cmd8 = IR.Assn(j, IRUtil.bitAnd(IRUtil.shrUint(input, shrJ), mask))
cmd1 = IR.If(cond, [cmd2, cmd3], [cmd6, cmd7, cmd8])
cmd10 = IR.Assn(
expr_2_elt,
IRUtil.mul(IRUtil.shrUint(IRUtil.addIndex(table[0], [i]), shr1),
IRUtil.shrUint(IRUtil.addIndex(table[1], [j]), shr2)))
scale_out = self.getScaleForExp(scale1, shr1, scale2, shr2)
intv_out = self.getIntervalForExp(scale_out, [-m_scale, -M_scale])
cmd0 = IR.Comment('exp(' + expr_in.idf + ')')
prog_exp = IR.Prog([cmd0, cmd1, cmd10])
prog_out = IRUtil.concatPrograms(prog_in, prog_exp)
self.decls[expr_out.idf] = type_in
self.scales[expr_out.idf] = scale_out
self.intvs[expr_out.idf] = intv_out
self.decls.update(dict((var.idf, Type.Int()) for var in [input, i, j]))
return (prog_out, expr_out)