[Beignet] [PATCH 4/5] refine code starting from header in typedwrite
Guo, Yejun
yejun.guo at intel.com
Fri Dec 9 10:07:56 UTC 2016
With this refine, the virtual reg and physical reg will be logically
1:1 mapping, and it helps the later instruction sends
Signed-off-by: Guo, Yejun <yejun.guo at intel.com>
---
backend/src/backend/gen_insn_selection.cpp | 145 ++++++++++++++++-------------
1 file changed, 78 insertions(+), 67 deletions(-)
diff --git a/backend/src/backend/gen_insn_selection.cpp b/backend/src/backend/gen_insn_selection.cpp
index 1ebbbe6..ec0897c 100644
--- a/backend/src/backend/gen_insn_selection.cpp
+++ b/backend/src/backend/gen_insn_selection.cpp
@@ -6809,86 +6809,97 @@ extern bool OCL_DEBUGINFO; // first defined by calling BVAR in program.cpp
{
INLINE bool emitOne(Selection::Opaque &sel, const ir::TypedWriteInstruction &insn, bool &markChildren) const
{
- using namespace ir;
- const uint32_t simdWidth = sel.ctx.getSimdWidth();
- GenRegister msgs[9]; // (header + U + V + R + LOD + 4)
- const uint32_t msgNum = (8 / (simdWidth / 8)) + 1;
- const uint32_t dim = insn.getSrcNum() - 4;
-
- if (simdWidth == 16) {
- for(uint32_t i = 0; i < msgNum; i++)
- msgs[i] = sel.selReg(sel.reg(FAMILY_DWORD), TYPE_U32);
- } else {
- uint32_t valueID = 0;
- uint32_t msgID = 0;
- msgs[msgID++] = sel.selReg(sel.reg(FAMILY_DWORD), TYPE_U32);
- for(; msgID < 1 + dim; msgID++, valueID++)
- msgs[msgID] = sel.selReg(insn.getSrc(msgID - 1), insn.getCoordType());
-
- // fake v.
- if (dim < 2)
- msgs[msgID++] = sel.selReg(sel.reg(FAMILY_DWORD), TYPE_U32);
- // fake w.
- if (dim < 3)
- msgs[msgID++] = sel.selReg(sel.reg(FAMILY_DWORD), TYPE_U32);
- // LOD.
- msgs[msgID++] = sel.selReg(sel.reg(FAMILY_DWORD), TYPE_U32);
- for(; valueID < insn.getSrcNum(); msgID++, valueID++)
- msgs[msgID] = sel.selReg(insn.getSrc(valueID), insn.getSrcType());
- }
-
+ const GenRegister header = GenRegister::ud8grf(sel.reg(ir::FAMILY_REG));
sel.push();
sel.curr.predicate = GEN_PREDICATE_NONE;
sel.curr.noMask = 1;
- sel.MOV(msgs[0], GenRegister::immud(0));
+ sel.MOV(header, GenRegister::immud(0));
sel.curr.execWidth = 1;
-
- GenRegister channelEn = sel.getOffsetReg(msgs[0], 0, 7*4);
+ GenRegister channelEn = sel.getOffsetReg(header, 0, 7*4);
// Enable all channels.
sel.MOV(channelEn, GenRegister::immud(0xffff));
- sel.curr.execWidth = 8;
- // Set zero LOD.
- if (simdWidth == 8)
- sel.MOV(msgs[4], GenRegister::immud(0));
- else
- sel.MOV(GenRegister::Qn(msgs[2], 0), GenRegister::immud(0));
sel.pop();
+ const uint32_t simdWidth = sel.ctx.getSimdWidth();
+ if (simdWidth == 16)
+ emitWithSimd16(sel, insn, markChildren, header);
+ else if (simdWidth == 8)
+ emitWithSimd8(sel, insn, markChildren, header);
+ else
+ assert(!"not supported");
+ return true;
+ }
+
+ INLINE bool emitWithSimd16(Selection::Opaque &sel, const ir::TypedWriteInstruction &insn, bool &markChildren, const GenRegister& header) const
+ {
+ using namespace ir;
+
+ GenRegister msgs[9]; // (header + U + V + W + LOD + 4)
+ msgs[0] = header;
+ for (uint32_t i = 1; i < 9; ++i) {
+ //SIMD16 will be split into two SIMD8,
+ //each virtual reg in msgs requires one physical reg with 8 DWORDs (32 bytes),
+ //so, declare with FAMILY_WORD, and the allocated size will be sizeof(WORD)*SIMD16 = 32 bytes
+ msgs[i] = sel.selReg(sel.reg(FAMILY_WORD), TYPE_U32);
+ }
+
+ const uint32_t dims = insn.getSrcNum() - 4;
uint32_t bti = insn.getImageIndex();
- if (simdWidth == 8)
- sel.TYPED_WRITE(msgs, msgNum, bti, dim == 3);
- else {
- sel.push();
- sel.curr.execWidth = 8;
- for( uint32_t quarter = 0; quarter < 2; quarter++)
- {
- #define QUARTER_MOV0(msgs, msgid, src) \
- sel.MOV(GenRegister::Qn(GenRegister::retype(msgs[msgid/2], GEN_TYPE_UD), msgid % 2), \
- GenRegister::Qn(src, quarter))
-
- #define QUARTER_MOV1(msgs, msgid, src) \
- sel.MOV(GenRegister::Qn(GenRegister::retype(msgs[msgid/2], src.type), msgid % 2), \
- GenRegister::Qn(src, quarter))
- sel.curr.quarterControl = (quarter == 0) ? GEN_COMPRESSION_Q1 : GEN_COMPRESSION_Q2;
- // Set U,V,W
- QUARTER_MOV0(msgs, 1, sel.selReg(insn.getSrc(0), insn.getCoordType()));
- if (dim > 1)
- QUARTER_MOV0(msgs, 2, sel.selReg(insn.getSrc(1), insn.getCoordType()));
- if (dim > 2)
- QUARTER_MOV0(msgs, 3, sel.selReg(insn.getSrc(2), insn.getCoordType()));
- // Set R, G, B, A
- QUARTER_MOV1(msgs, 5, sel.selReg(insn.getSrc(dim), insn.getSrcType()));
- QUARTER_MOV1(msgs, 6, sel.selReg(insn.getSrc(dim + 1), insn.getSrcType()));
- QUARTER_MOV1(msgs, 7, sel.selReg(insn.getSrc(dim + 2), insn.getSrcType()));
- QUARTER_MOV1(msgs, 8, sel.selReg(insn.getSrc(dim + 3), insn.getSrcType()));
- sel.TYPED_WRITE(msgs, msgNum, bti, dim == 3);
- #undef QUARTER_MOV0
- #undef QUARTER_MOV1
+
+ sel.push();
+ sel.curr.execWidth = 8;
+ for (uint32_t i = 0; i < 2; ++i) { //SIMD16 split to two SIMD8
+ sel.curr.quarterControl = (i == 0) ? GEN_COMPRESSION_Q1 : GEN_COMPRESSION_Q2;
+ uint32_t msgid = 1;
+ for (uint32_t dim = 0; dim < dims; ++dim) { //the coords
+ GenRegister coord = sel.selReg(insn.getSrc(dim), insn.getCoordType());
+ sel.MOV(GenRegister::retype(msgs[msgid++], coord.type), GenRegister::Qn(coord, i));
}
- sel.pop();
+
+ while (msgid < 5) //fill fake coords
+ sel.MOV(msgs[msgid++], GenRegister::immud(0));
+
+ for (uint32_t j = 0; j < 4; ++j) { //the data
+ GenRegister data = sel.selReg(insn.getSrc(j + dims), insn.getSrcType());
+ sel.MOV(GenRegister::retype(msgs[msgid++], data.type), GenRegister::Qn(data, i));
+ }
+
+ sel.TYPED_WRITE(msgs, 9, bti, dims == 3);
}
+ sel.pop();
return true;
}
+
+ INLINE bool emitWithSimd8(Selection::Opaque &sel, const ir::TypedWriteInstruction &insn, bool &markChildren, const GenRegister& header) const
+ {
+ using namespace ir;
+ GenRegister msgs[9]; // (header + U + V + W + LOD + 4)
+ msgs[0] = header;
+
+ const uint32_t dims = insn.getSrcNum() - 4;
+ uint32_t bti = insn.getImageIndex();
+ uint32_t msgid = 1;
+
+ for (uint32_t dim = 0; dim < dims; ++dim) { //the coords
+ GenRegister coord = sel.selReg(insn.getSrc(dim), insn.getCoordType());
+ msgs[msgid++] = coord;
+ }
+
+ while (msgid < 5) { //fill fake coords
+ GenRegister fake = sel.selReg(sel.reg(FAMILY_DWORD), TYPE_U32);
+ sel.MOV(fake, GenRegister::immud(0));
+ msgs[msgid++] = fake;
+ }
+
+ for (uint32_t j = 0; j < 4; ++j) { //the data
+ GenRegister data = sel.selReg(insn.getSrc(j + dims), insn.getSrcType());
+ msgs[msgid++] = data;
+ }
+
+ sel.TYPED_WRITE(msgs, 9, bti, dims == 3);
+ return true;
+ }
+
DECL_CTOR(TypedWriteInstruction, 1, 1);
};
--
1.9.1
More information about the Beignet
mailing list