On 28.09.2016 02:01, Ilia Mirkin wrote:> On Tue, Sep 27, 2016 at 7:25 PM, Tobias Klausmann > <tobias.johannes.klausmann at mni.thm.de> wrote: >> Split the source immediate value into two new values and create OP_MOV >> instructions the two newly created values. >> >> Signed-off-by: Tobias Klausmann <tobias.johannes.klausmann at mni.thm.de> >> --- >> .../drivers/nouveau/codegen/nv50_ir_peephole.cpp | 23 ++++++++++++++++++++++ >> 1 file changed, 23 insertions(+) >> >> diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp b/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp >> index 74a5a85..fdddd71 100644 >> --- a/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp >> +++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp >> @@ -920,6 +920,29 @@ ConstantFolding::opnd(Instruction *i, ImmediateValue &imm0, int s) >> Instruction *newi = i; >> >> switch (i->op) { >> + case OP_SPLIT: { >> + uint16_t shift = 0; >> + DataType type = TYPE_NONE; >> + bld.setPosition(i, false); >> + if (i->sType == TYPE_U64 || i->sType == TYPE_S64) { >> + shift = 32; >> + type = (i->sType == TYPE_U64) ? TYPE_U32 : TYPE_S32; >> + } >> + if (i->sType == TYPE_U32 || i->sType == TYPE_S32) { >> + shift = 16; >> + type = (i->sType == TYPE_U32) ? TYPE_U16 : TYPE_S16; >> + } >> + if (i->sType == TYPE_U16 || i->sType == TYPE_S16) { >> + shift = 8; >> + type = (i->sType == TYPE_U16) ? TYPE_U8 : TYPE_S8; >> + } > shift = typeSizeOf(i->dType); > >> + if (type != TYPE_NONE) { >> + bld.mkMov(i->getDef(0), bld.mkImm(imm0.reg.data.u64 >> shift), type); >> + bld.mkMov(i->getDef(1), bld.mkImm(imm0.reg.data.u64), type); > u64 val = ...u64; > for (d = 0; i->defExists(d); ++d) { > bld.mkMov(i->getDef(d), bld.mkImm(val & ((1 << shift) - 1)); > val >>= shift; > } > > I think this will account for every case, and with a lot less > special-casing. What do you think?Well with this you'd not set the new type right: bld.mkMov(def, val, >>type<<), where you always would use TYPE_U32. Not sure if that is what we want... other than that that, shorten it like this would be nice!> >> + delete_Instruction(prog, i); >> + } >> + } >> + break; >> case OP_MUL: >> if (i->dType == TYPE_F32) >> tryCollapseChainedMULs(i, s, imm0); >> -- >> 2.10.0 >> >> _______________________________________________ >> Nouveau mailing list >> Nouveau at lists.freedesktop.org >> https://lists.freedesktop.org/mailman/listinfo/nouveau
On Fri, Sep 30, 2016 at 3:21 PM, Tobias Klausmann <tobias.johannes.klausmann at mni.thm.de> wrote:> > > On 28.09.2016 02:01, Ilia Mirkin wrote: >> >> On Tue, Sep 27, 2016 at 7:25 PM, Tobias Klausmann >> <tobias.johannes.klausmann at mni.thm.de> wrote: >>> >>> Split the source immediate value into two new values and create OP_MOV >>> instructions the two newly created values. >>> >>> Signed-off-by: Tobias Klausmann <tobias.johannes.klausmann at mni.thm.de> >>> --- >>> .../drivers/nouveau/codegen/nv50_ir_peephole.cpp | 23 >>> ++++++++++++++++++++++ >>> 1 file changed, 23 insertions(+) >>> >>> diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp >>> b/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp >>> index 74a5a85..fdddd71 100644 >>> --- a/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp >>> +++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp >>> @@ -920,6 +920,29 @@ ConstantFolding::opnd(Instruction *i, ImmediateValue >>> &imm0, int s) >>> Instruction *newi = i; >>> >>> switch (i->op) { >>> + case OP_SPLIT: { >>> + uint16_t shift = 0; >>> + DataType type = TYPE_NONE; >>> + bld.setPosition(i, false); >>> + if (i->sType == TYPE_U64 || i->sType == TYPE_S64) { >>> + shift = 32; >>> + type = (i->sType == TYPE_U64) ? TYPE_U32 : TYPE_S32; >>> + } >>> + if (i->sType == TYPE_U32 || i->sType == TYPE_S32) { >>> + shift = 16; >>> + type = (i->sType == TYPE_U32) ? TYPE_U16 : TYPE_S16; >>> + } >>> + if (i->sType == TYPE_U16 || i->sType == TYPE_S16) { >>> + shift = 8; >>> + type = (i->sType == TYPE_U16) ? TYPE_U8 : TYPE_S8; >>> + } >> >> shift = typeSizeOf(i->dType); >> >>> + if (type != TYPE_NONE) { >>> + bld.mkMov(i->getDef(0), bld.mkImm(imm0.reg.data.u64 >> shift), >>> type); >>> + bld.mkMov(i->getDef(1), bld.mkImm(imm0.reg.data.u64), type); >> >> u64 val = ...u64; >> for (d = 0; i->defExists(d); ++d) { >> bld.mkMov(i->getDef(d), bld.mkImm(val & ((1 << shift) - 1)); >> val >>= shift; >> } >> >> I think this will account for every case, and with a lot less >> special-casing. What do you think? > > > Well with this you'd not set the new type right: bld.mkMov(def, val, >>>type<<), where you always would use TYPE_U32. Not sure if that is what we > want... other than that that, shorten it like this would be nice!pah-shaw :( typeOfSize(shift / 2, isFloatType(i->dType), isSignedType(i->dType)) How's that :p> > >> >>> + delete_Instruction(prog, i); >>> + } >>> + } >>> + break; >>> case OP_MUL: >>> if (i->dType == TYPE_F32) >>> tryCollapseChainedMULs(i, s, imm0); >>> -- >>> 2.10.0 >>> >>> _______________________________________________ >>> Nouveau mailing list >>> Nouveau at lists.freedesktop.org >>> https://lists.freedesktop.org/mailman/listinfo/nouveau > >
Tobias Klausmann
2016-Sep-30 21:50 UTC
[Nouveau] [PATCH v2] nv50/ir: constant fold OP_SPLIT
Split the source immediate value into two new values and create OP_MOV
instructions the two newly created values.
V2: get rid of special cases
Signed-off-by: Tobias Klausmann <tobias.johannes.klausmann at mni.thm.de>
---
src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp | 16 ++++++++++++++++
1 file changed, 16 insertions(+)
diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp
b/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp
index 9875738..d56b057 100644
--- a/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp
+++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp
@@ -932,6 +932,22 @@ ConstantFolding::opnd(Instruction *i, ImmediateValue
&imm0, int s)
Instruction *newi = i;
switch (i->op) {
+ case OP_SPLIT: {
+ uint8_t size = typeSizeof(i->dType);
+ DataType type = typeOfSize(size / 2, isFloatType(i->dType),
+ isSignedType(i->dType));
+ if (likely(type != TYPE_NONE)) {
+ uint64_t val = imm0.reg.data.u64;
+ uint16_t shift = size * 8;
+ bld.setPosition(i, false);
+ for (int8_t d = 0; i->defExists(d); ++d) {
+ bld.mkMov(i->getDef(d), bld.mkImm(val & ((1 << shift)
- 1)), type);
+ val >>= shift;
+ }
+ delete_Instruction(prog, i);
+ }
+ }
+ break;
case OP_MUL:
if (i->dType == TYPE_F32)
tryCollapseChainedMULs(i, s, imm0);
--
2.10.0