Author: Richard Plangger <planri...@gmail.com>
Branch: ppc-vsx-support
Changeset: r85489:89ec178c8b17
Date: 2016-07-01 15:37 +0200
http://bitbucket.org/pypy/pypy/changeset/89ec178c8b17/
Log: finished accum reduce function for f64/i64 + and * (ppc)
diff --git a/rpython/jit/backend/ppc/codebuilder.py
b/rpython/jit/backend/ppc/codebuilder.py
--- a/rpython/jit/backend/ppc/codebuilder.py
+++ b/rpython/jit/backend/ppc/codebuilder.py
@@ -612,7 +612,7 @@
# mul
xvmuldp = XX3(60, XO9=112)
xvmulsp = XX3(60, XO9=80)
- xsmuldp = XX3(60, XO9=46)
+ xsmuldp = XX3(60, XO9=48)
# div
xvdivdp = XX3(60, XO9=102)
xvdivsp = XX3(60, XO9=88)
diff --git a/rpython/jit/backend/ppc/vector_ext.py
b/rpython/jit/backend/ppc/vector_ext.py
--- a/rpython/jit/backend/ppc/vector_ext.py
+++ b/rpython/jit/backend/ppc/vector_ext.py
@@ -333,38 +333,40 @@
if not scalar_loc.is_reg():
scalar_loc = regalloc.force_allocate_reg(scalar_arg)
assert scalar_arg is not None
- if accum_info.accum_operation == '+':
- self._accum_reduce_sum(scalar_arg, vector_loc, scalar_loc)
- elif accum_info.accum_operation == '*':
- self._accum_reduce_mul(scalar_arg, vector_loc, scalar_loc)
- else:
- not_implemented("accum operator %s not implemented" %
- (accum_info.accum_operation))
+ op = accum_info.accum_operation
+ self._accum_reduce(op, scalar_arg, vector_loc, scalar_loc)
accum_info = accum_info.next()
- def _accum_reduce_mul(self, arg, accumloc, targetloc):
- notimplemented("[ppc reduce mul]")
- #scratchloc = X86_64_XMM_SCRATCH_REG
- #self.mov(accumloc, scratchloc)
- ## swap the two elements
- #self.mc.SHUFPD_xxi(scratchloc.value, scratchloc.value, 0x01)
- #self.mc.MULSD(accumloc, scratchloc)
- #if accumloc is not targetloc:
- # self.mov(accumloc, targetloc)
-
- def _accum_reduce_sum(self, arg, accumloc, targetloc):
+ def _accum_reduce(self, op, arg, accumloc, targetloc):
# Currently the accumulator can ONLY be the biggest
# 64 bit float/int
tgt = targetloc.value
acc = accumloc.value
if arg.type == FLOAT:
# r = (r[0]+r[1],r[0]+r[1])
- self.mc.xvmr(tgt, acc, acc)
if IS_BIG_ENDIAN:
self.mc.xxspltd(tgt, acc, acc, 0b00)
else:
- self.mc.xxspltd(tgt, acc, acc, 0b01)
- self.mc.xsadddp(tgt, tgt, acc)
+ self.mc.xxspltd(tgt, acc, acc, 0b10)
+ if op == '+':
+ self.mc.xsadddp(tgt, tgt, acc)
+ elif op == '*':
+ self.mc.xsmuldp(tgt, tgt, acc)
+ else:
+ not_implemented("sum not implemented")
+ return
+ else:
+ assert arg.type == INT
+ self.mc.load_imm(r.SCRATCH2, PARAM_SAVE_AREA_OFFSET)
+ self.mc.stvx(acc, r.SCRATCH2.value, r.SP.value)
+ self.mc.load(tgt, r.SP.value, PARAM_SAVE_AREA_OFFSET)
+ self.mc.load(r.SCRATCH.value, r.SP.value, PARAM_SAVE_AREA_OFFSET+8)
+ if op == '+':
+ self.mc.add(tgt, tgt, acc)
+ elif op == '*':
+ self.mc.mul(tgt, tgt, acc)
+ else:
+ not_implemented("sum not implemented")
return
not_implemented("reduce sum for %s not impl." % arg)
@@ -514,59 +516,37 @@
else:
notimplemented("[expand int size not impl]")
- #def genop_vec_pack_i(self, op, arglocs, regalloc):
- # resultloc, sourceloc, residxloc, srcidxloc, countloc, sizeloc =
arglocs
- # assert isinstance(resultloc, RegLoc)
- # assert isinstance(sourceloc, RegLoc)
- # size = sizeloc.value
- # srcidx = srcidxloc.value
- # residx = residxloc.value
- # count = countloc.value
- # # for small data type conversion this can be quite costy
- # # NOTE there might be some combinations that can be handled
- # # more efficiently! e.g.
- # # v2 = pack(v0,v1,4,4)
- # si = srcidx
- # ri = residx
- # k = count
- # while k > 0:
- # if size == 8:
- # if resultloc.is_xmm and sourceloc.is_xmm: # both xmm
- # self.mc.PEXTRQ_rxi(X86_64_SCRATCH_REG.value,
sourceloc.value, si)
- # self.mc.PINSRQ_xri(resultloc.value,
X86_64_SCRATCH_REG.value, ri)
- # elif resultloc.is_xmm: # xmm <- reg
- # self.mc.PINSRQ_xri(resultloc.value, sourceloc.value, ri)
- # else: # reg <- xmm
- # self.mc.PEXTRQ_rxi(resultloc.value, sourceloc.value, si)
- # elif size == 4:
- # if resultloc.is_xmm and sourceloc.is_xmm:
- # self.mc.PEXTRD_rxi(X86_64_SCRATCH_REG.value,
sourceloc.value, si)
- # self.mc.PINSRD_xri(resultloc.value,
X86_64_SCRATCH_REG.value, ri)
- # elif resultloc.is_xmm:
- # self.mc.PINSRD_xri(resultloc.value, sourceloc.value, ri)
- # else:
- # self.mc.PEXTRD_rxi(resultloc.value, sourceloc.value, si)
- # elif size == 2:
- # if resultloc.is_xmm and sourceloc.is_xmm:
- # self.mc.PEXTRW_rxi(X86_64_SCRATCH_REG.value,
sourceloc.value, si)
- # self.mc.PINSRW_xri(resultloc.value,
X86_64_SCRATCH_REG.value, ri)
- # elif resultloc.is_xmm:
- # self.mc.PINSRW_xri(resultloc.value, sourceloc.value, ri)
- # else:
- # self.mc.PEXTRW_rxi(resultloc.value, sourceloc.value, si)
- # elif size == 1:
- # if resultloc.is_xmm and sourceloc.is_xmm:
- # self.mc.PEXTRB_rxi(X86_64_SCRATCH_REG.value,
sourceloc.value, si)
- # self.mc.PINSRB_xri(resultloc.value,
X86_64_SCRATCH_REG.value, ri)
- # elif resultloc.is_xmm:
- # self.mc.PINSRB_xri(resultloc.value, sourceloc.value, ri)
- # else:
- # self.mc.PEXTRB_rxi(resultloc.value, sourceloc.value, si)
- # si += 1
- # ri += 1
- # k -= 1
+ def emit_vec_pack_i(self, op, arglocs, regalloc):
+ resultloc, vloc, sourceloc, residxloc, srcidxloc, countloc = arglocs
+ srcidx = srcidxloc.value
+ residx = residxloc.value
+ count = countloc.value
+ # for small data type conversion this can be quite costy
+ # NOTE there might be some combinations that can be handled
+ # more efficiently! e.g.
+ # v2 = pack(v0,v1,4,4)
+ res = resultloc.value
+ vector = vloc.value
+ src = sourceloc.value
+ size = op.bytesize
+ if size == 8:
+ if resultloc.is_vector_reg() and sourceloc.is_vector_reg(): # both
vector
+ notimplemented("[ppc/vec_pack_i]")
+ elif resultloc.is_vector_reg(): # vector <- reg
+ self.mc.load_imm(r.SCRATCH, PARAM_SAVE_AREA_OFFSET)
+ self.mc.stvx(vector, r.SCRATCH2.value, r.SP.value)
+ self.mc.store(src, r.SP.value, PARAM_SAVE_AREA_OFFSET+8*residx)
+ self.mc.lvx(res, r.SCRATCH2.value, r.SP.value)
+ else:
+ notimplemented("[ppc/vec_pack_i]")
+ elif size == 4:
+ notimplemented("[ppc/vec_pack_i]")
+ elif size == 2:
+ notimplemented("[ppc/vec_pack_i]")
+ elif size == 1:
+ notimplemented("[ppc/vec_pack_i]")
- #genop_vec_unpack_i = genop_vec_pack_i
+ # TODO emit_vec_unpack_i = emit_vec_pack_i
def emit_vec_pack_f(self, op, arglocs, resultloc):
resloc, vloc, srcloc, residxloc, srcidxloc, countloc = arglocs
@@ -755,11 +735,12 @@
count = op.getarg(3)
assert isinstance(index, ConstInt)
assert isinstance(count, ConstInt)
- srcloc = self.ensure_vector_reg(arg)
+ vloc = self.ensure_vector_reg(op.getarg(0))
+ srcloc = self.ensure_reg(arg)
resloc = self.force_allocate_vector_reg(op)
residx = index.value # where to put it in result?
srcidx = 0
- return [resloc, srcloc, imm(residx), imm(srcidx), imm(count.value)]
+ return [resloc, vloc, srcloc, imm(residx), imm(srcidx),
imm(count.value)]
def prepare_vec_pack_f(self, op):
# new_res = vec_pack_i(res, src, index, count)
@@ -769,7 +750,7 @@
count = op.getarg(3)
assert isinstance(index, ConstInt)
assert isinstance(count, ConstInt)
- assert not arg.is_vector()
+ assert not arg.is_vector_reg()
srcloc = self.ensure_reg(arg)
vloc = self.ensure_vector_reg(op.getarg(0))
resloc = self.force_allocate_vector_reg(op)
diff --git a/rpython/jit/metainterp/test/test_vector.py
b/rpython/jit/metainterp/test/test_vector.py
--- a/rpython/jit/metainterp/test/test_vector.py
+++ b/rpython/jit/metainterp/test/test_vector.py
@@ -379,23 +379,30 @@
res = self.meta_interp(f, [count])
assert res == f(count) == breaks
- def test_sum(self):
+ @py.test.mark.parametrize('type,func,cast',
+ [(rffi.DOUBLE, lambda a,b: a+b, float),
+ (rffi.DOUBLE, lambda a,b: a*b, float),
+ (lltype.Signed, lambda a,b: a+b, int),
+ (lltype.Signed, lambda a,b: a*b, int),
+ ])
+ def test_reduce(self, type, func, cast):
+ func = always_inline(func)
myjitdriver = JitDriver(greens = [], reds = 'auto', vectorize=True)
- T = lltype.Array(rffi.DOUBLE, hints={'nolength': True})
+ T = lltype.Array(type, hints={'nolength': True})
def f(d):
va = lltype.malloc(T, d, flavor='raw', zero=True)
for j in range(d):
- va[j] = float(j)
+ va[j] = cast(j+1)
i = 0
accum = 0
while i < d:
myjitdriver.jit_merge_point()
- accum += va[i]
+ accum = func(accum,va[i])
i += 1
lltype.free(va, flavor='raw')
return accum
res = self.meta_interp(f, [60])
- assert res == f(60) == sum(range(60))
+ assert isclose(res, f(60))
def test_constant_expand(self):
myjitdriver = JitDriver(greens = [], reds = 'auto', vectorize=True)
_______________________________________________
pypy-commit mailing list
pypy-commit@python.org
https://mail.python.org/mailman/listinfo/pypy-commit
