nvc0/ir: initial implementation of nve4 scheduling hints

5 files changed, 141 insertions, 4 deletions

diff --git a/src/gallium/drivers/nv50/codegen/nv50_ir.h b/src/gallium/drivers/nv50/codegen/nv50_ir.h
index da9042066ad..e544d071b52 100644
--- a/src/gallium/drivers/nv50/codegen/nv50_ir.h
+++ b/src/gallium/drivers/nv50/codegen/nv50_ir.h
@@ -681,6 +681,8 @@ public:
 
    uint8_t subOp; // quadop, 1 for mul-high, etc.
 
+   uint8_t sched; // scheduling data (NOTE: maybe move to separate storage)
+
    unsigned encSize    : 4; // encoding size in bytes
    unsigned saturate   : 1; // to [0.0f, 1.0f]
    unsigned join       : 1; // converge control flow (use OP_JOIN until end)
diff --git a/src/gallium/drivers/nv50/codegen/nv50_ir_print.cpp b/src/gallium/drivers/nv50/codegen/nv50_ir_print.cpp
index 4652bb95f69..9d92b7bc24b 100644
--- a/src/gallium/drivers/nv50/codegen/nv50_ir_print.cpp
+++ b/src/gallium/drivers/nv50/codegen/nv50_ir_print.cpp
@@ -53,6 +53,26 @@ static const char *colour[8] =
 #endif
 };
 
+static const char *OpClassStr[OPCLASS_OTHER + 1] =
+{
+   "MOVE",
+   "LOAD",
+   "STORE",
+   "ARITH",
+   "SHIFT",
+   "SFU",
+   "LOGIC",
+   "COMPARE",
+   "CONVERT",
+   "ATOMIC",
+   "TEXTURE",
+   "SURFACE",
+   "FLOW",
+   "(INVALID)",
+   "PSEUDO",
+   "OTHER"
+};
+
 const char *operationStr[OP_LAST + 1] =
 {
    "nop",
diff --git a/src/gallium/drivers/nv50/codegen/nv50_ir_target.cpp b/src/gallium/drivers/nv50/codegen/nv50_ir_target.cpp
index e3eae69554c..f718912fb39 100644
--- a/src/gallium/drivers/nv50/codegen/nv50_ir_target.cpp
+++ b/src/gallium/drivers/nv50/codegen/nv50_ir_target.cpp
@@ -52,6 +52,65 @@ const uint8_t Target::operationSrcNr[OP_LAST + 1] =
    0
 };
 
+const OpClass Target::operationClass[OP_LAST + 1] =
+{
+   // NOP; PHI; UNION, SPLIT, MERGE, CONSTRAINT
+   OPCLASS_OTHER,
+   OPCLASS_PSEUDO,
+   OPCLASS_PSEUDO, OPCLASS_PSEUDO, OPCLASS_PSEUDO, OPCLASS_PSEUDO,
+   // MOV; LOAD; STORE
+   OPCLASS_MOVE,
+   OPCLASS_LOAD,
+   OPCLASS_STORE,
+   // ADD, SUB, MUL; DIV, MOD; MAD, FMA, SAD
+   OPCLASS_ARITH, OPCLASS_ARITH, OPCLASS_ARITH,
+   OPCLASS_ARITH, OPCLASS_ARITH,
+   OPCLASS_ARITH, OPCLASS_ARITH, OPCLASS_ARITH,
+   // ABS, NEG; NOT, AND, OR, XOR; SHL, SHR
+   OPCLASS_CONVERT, OPCLASS_CONVERT,
+   OPCLASS_LOGIC, OPCLASS_LOGIC, OPCLASS_LOGIC, OPCLASS_LOGIC,
+   OPCLASS_SHIFT, OPCLASS_SHIFT,
+   // MAX, MIN
+   OPCLASS_COMPARE, OPCLASS_COMPARE,
+   // SAT, CEIL, FLOOR, TRUNC; CVT
+   OPCLASS_CONVERT, OPCLASS_CONVERT, OPCLASS_CONVERT, OPCLASS_CONVERT,
+   OPCLASS_CONVERT,
+   // SET(AND,OR,XOR); SELP, SLCT
+   OPCLASS_COMPARE, OPCLASS_COMPARE, OPCLASS_COMPARE, OPCLASS_COMPARE,
+   OPCLASS_COMPARE, OPCLASS_COMPARE,
+   // RCP, RSQ, LG2, SIN, COS; EX2, EXP, LOG, PRESIN, PREEX2; SQRT, POW
+   OPCLASS_SFU, OPCLASS_SFU, OPCLASS_SFU, OPCLASS_SFU, OPCLASS_SFU,
+   OPCLASS_SFU, OPCLASS_SFU, OPCLASS_SFU, OPCLASS_SFU, OPCLASS_SFU,
+   OPCLASS_SFU, OPCLASS_SFU,
+   // BRA, CALL, RET; CONT, BREAK, PRE(RET,CONT,BREAK); BRKPT, JOINAT, JOIN
+   OPCLASS_FLOW, OPCLASS_FLOW, OPCLASS_FLOW,
+   OPCLASS_FLOW, OPCLASS_FLOW, OPCLASS_FLOW, OPCLASS_FLOW, OPCLASS_FLOW,
+   OPCLASS_FLOW, OPCLASS_FLOW, OPCLASS_FLOW,
+   // DISCARD, EXIT
+   OPCLASS_FLOW, OPCLASS_FLOW,
+   // MEMBAR
+   OPCLASS_OTHER,
+   // VFETCH, PFETCH, EXPORT
+   OPCLASS_LOAD, OPCLASS_OTHER, OPCLASS_STORE,
+   // LINTERP, PINTERP
+   OPCLASS_SFU, OPCLASS_SFU,
+   // EMIT, RESTART
+   OPCLASS_OTHER, OPCLASS_OTHER,
+   // TEX, TXB, TXL, TXF; TXQ, TXD, TXG, TEXCSAA
+   OPCLASS_TEXTURE, OPCLASS_TEXTURE, OPCLASS_TEXTURE, OPCLASS_TEXTURE,
+   OPCLASS_TEXTURE, OPCLASS_TEXTURE, OPCLASS_TEXTURE, OPCLASS_TEXTURE,
+   // SULD, SUST
+   OPCLASS_SURFACE, OPCLASS_SURFACE,
+   // DFDX, DFDY, RDSV, WRSV; PIXLD, QUADOP, QUADON, QUADPOP
+   OPCLASS_OTHER, OPCLASS_OTHER, OPCLASS_OTHER, OPCLASS_OTHER,
+   OPCLASS_OTHER, OPCLASS_OTHER, OPCLASS_OTHER, OPCLASS_OTHER,
+   // POPCNT, INSBF, EXTBF
+   OPCLASS_OTHER, OPCLASS_OTHER, OPCLASS_OTHER,
+   // TEXBAR
+   OPCLASS_OTHER,
+   OPCLASS_PSEUDO // LAST
+};
+
 
 extern Target *getTargetNVC0(unsigned int chipset);
 extern Target *getTargetNV50(unsigned int chipset);
@@ -104,6 +163,11 @@ CodeEmitter::printBinary() const
    INFO("\n");
 }
 
+static inline uint32_t sizeToBundlesNVE4(uint32_t size)
+{
+   return (size + 55) / 56;
+}
+
 void
 CodeEmitter::prepareEmission(Program *prog)
 {
@@ -112,6 +176,23 @@ CodeEmitter::prepareEmission(Program *prog)
       Function *func = reinterpret_cast<Function *>(fi.get());
       func->binPos = prog->binSize;
       prepareEmission(func);
+
+      // adjust sizes & positions for schedulding info:
+      if (prog->getTarget()->hasSWSched) {
+         BasicBlock *bb = NULL;
+         for (int i = 0; i < func->bbCount; ++i) {
+            bb = func->bbArray[i];
+            const uint32_t oldPos = bb->binPos;
+            const uint32_t oldEnd = bb->binPos + bb->binSize;
+            uint32_t adjPos = oldPos + sizeToBundlesNVE4(oldPos) * 8;
+            uint32_t adjEnd = oldEnd + sizeToBundlesNVE4(oldEnd) * 8;
+            bb->binPos = adjPos;
+            bb->binSize = adjEnd - adjPos;
+         }
+         if (bb)
+            func->binSize = bb->binPos + bb->binSize;
+      }
+
       prog->binSize += func->binSize;
    }
 }
diff --git a/src/gallium/drivers/nv50/codegen/nv50_ir_target.h b/src/gallium/drivers/nv50/codegen/nv50_ir_target.h
index 88996ebbde3..c60ee0216f7 100644
--- a/src/gallium/drivers/nv50/codegen/nv50_ir_target.h
+++ b/src/gallium/drivers/nv50/codegen/nv50_ir_target.h
@@ -93,9 +93,31 @@ protected:
    RelocInfo *relocInfo;
 };
 
+
+enum OpClass
+{
+   OPCLASS_MOVE          = 0,
+   OPCLASS_LOAD          = 1,
+   OPCLASS_STORE         = 2,
+   OPCLASS_ARITH         = 3,
+   OPCLASS_SHIFT         = 4,
+   OPCLASS_SFU           = 5,
+   OPCLASS_LOGIC         = 6,
+   OPCLASS_COMPARE       = 7,
+   OPCLASS_CONVERT       = 8,
+   OPCLASS_ATOMIC        = 9,
+   OPCLASS_TEXTURE       = 10,
+   OPCLASS_SURFACE       = 11,
+   OPCLASS_FLOW          = 12,
+   OPCLASS_PSEUDO        = 14,
+   OPCLASS_OTHER         = 15
+};
+
 class Target
 {
 public:
+   Target(bool j, bool s) : joinAnterior(j), hasSWSched(s) { }
+
    static Target *create(uint32_t chipset);
    static void destroy(Target *);
 
@@ -153,6 +175,9 @@ public:
    virtual bool mayPredicate(const Instruction *,
                              const Value *) const = 0;
 
+   // whether @insn can be issued together with @next (order matters)
+   virtual bool canDualIssue(const Instruction *insn,
+                             const Instruction *next) const { return false; }
    virtual int getLatency(const Instruction *) const { return 1; }
    virtual int getThroughput(const Instruction *) const { return 1; }
 
@@ -162,9 +187,20 @@ public:
    virtual uint32_t getSVAddress(DataFile, const Symbol *) const = 0;
 
 public:
-   bool joinAnterior; // true if join is executed before the op
+   const bool joinAnterior; // true if join is executed before the op
+   const bool hasSWSched;   // true if code should provide scheduling data
 
    static const uint8_t operationSrcNr[OP_LAST + 1];
+   static const OpClass operationClass[OP_LAST + 1];
+
+   static inline uint8_t getOpSrcNr(operation op)
+   {
+      return operationSrcNr[op];
+   }
+   static inline OpClass getOpClass(operation op)
+   {
+      return operationClass[op];
+   }
 
 protected:
    uint32_t chipset;
diff --git a/src/gallium/drivers/nv50/codegen/nv50_ir_target_nv50.cpp b/src/gallium/drivers/nv50/codegen/nv50_ir_target_nv50.cpp
index a64f7f72255..5e541e514cb 100644
--- a/src/gallium/drivers/nv50/codegen/nv50_ir_target_nv50.cpp
+++ b/src/gallium/drivers/nv50/codegen/nv50_ir_target_nv50.cpp
@@ -29,7 +29,7 @@ Target *getTargetNV50(unsigned int chipset)
    return new TargetNV50(chipset);
 }
 
-TargetNV50::TargetNV50(unsigned int card)
+TargetNV50::TargetNV50(unsigned int card) : Target(true, false)
 {
    chipset = card;
 
@@ -132,8 +132,6 @@ void TargetNV50::initOpInfo()
       OP_CALL, OP_PREBREAK, OP_PRERET, OP_QUADON, OP_QUADPOP, OP_JOINAT
    };
 
-   joinAnterior = true;
-
    for (i = 0; i < DATA_FILE_COUNT; ++i)
       nativeFileMap[i] = (DataFile)i;
    nativeFileMap[FILE_PREDICATE] = FILE_FLAGS;