nvc0/ir: fix lowering of RSQ/RCP/SQRT/MOD to work with F64

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>

diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir.h b/src/gallium/drivers/nouveau/codegen/nv50_ir.h
index 0ff5e5dc3b2f2d20f039aa02e0ca714b46e8c852..529dcb9bdc23ccecab5b1da1e4b66b7b22c5f883 100644 (file)
--- a/src/gallium/drivers/nouveau/codegen/nv50_ir.h
+++ b/src/gallium/drivers/nouveau/codegen/nv50_ir.h
@@ -175,6 +175,7 @@ enum operation
 #define NV50_IR_SUBOP_MOV_FINAL    1
 #define NV50_IR_SUBOP_EXTBF_REV    1
 #define NV50_IR_SUBOP_BFIND_SAMT   1
+#define NV50_IR_SUBOP_RCPRSQ_64H   1
 #define NV50_IR_SUBOP_PERMT_F4E    1
 #define NV50_IR_SUBOP_PERMT_B4E    2
 #define NV50_IR_SUBOP_PERMT_RC8    3

diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_emit_gk110.cpp b/src/gallium/drivers/nouveau/codegen/nv50_ir_emit_gk110.cpp
index 204d911ae1745007fa0a57fe5f9d6b92985a370a..674be692f8b9476b8625d64ba7ad0e5d1302fd89 100644 (file)
--- a/src/gallium/drivers/nouveau/codegen/nv50_ir_emit_gk110.cpp
+++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_emit_gk110.cpp
@@ -1771,10 +1771,10 @@ CodeEmitterGK110::emitInstruction(Instruction *insn)
       emitCVT(insn);
       break;
    case OP_RSQ:
-      emitSFnOp(insn, 5);
+      emitSFnOp(insn, 5 + 2 * insn->subOp);
       break;
    case OP_RCP:
-      emitSFnOp(insn, 4);
+      emitSFnOp(insn, 4 + 2 * insn->subOp);
       break;
    case OP_LG2:
       emitSFnOp(insn, 3);

diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_emit_gm107.cpp b/src/gallium/drivers/nouveau/codegen/nv50_ir_emit_gm107.cpp
index 3e1da7ec084d94c9d0ff597cd389650d802f5356..ee0487f15d6c788f2a1a71261b1a359fd3b892bf 100644 (file)
--- a/src/gallium/drivers/nouveau/codegen/nv50_ir_emit_gm107.cpp
+++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_emit_gm107.cpp
@@ -1265,8 +1265,8 @@ CodeEmitterGM107::emitMUFU()
    case OP_SIN: mufu = 1; break;
    case OP_EX2: mufu = 2; break;
    case OP_LG2: mufu = 3; break;
-   case OP_RCP: mufu = 4; break;
-   case OP_RSQ: mufu = 5; break;
+   case OP_RCP: mufu = 4 + 2 * insn->subOp; break;
+   case OP_RSQ: mufu = 5 + 2 * insn->subOp; break;
    default:
       assert(!"invalid mufu");
       break;

diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_emit_nvc0.cpp b/src/gallium/drivers/nouveau/codegen/nv50_ir_emit_nvc0.cpp
index e38a3b806b6b8e49b1a3b8d7539237489feb5dd4..1a4f6e09b75968cde753ab27c2b65534e2d26c2a 100644 (file)
--- a/src/gallium/drivers/nouveau/codegen/nv50_ir_emit_nvc0.cpp
+++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_emit_nvc0.cpp
@@ -2365,10 +2365,10 @@ CodeEmitterNVC0::emitInstruction(Instruction *insn)
       emitCVT(insn);
       break;
    case OP_RSQ:
-      emitSFnOp(insn, 5);
+      emitSFnOp(insn, 5 + 2 * insn->subOp);
       break;
    case OP_RCP:
-      emitSFnOp(insn, 4);
+      emitSFnOp(insn, 4 + 2 * insn->subOp);
       break;
    case OP_LG2:
       emitSFnOp(insn, 3);

diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.cpp b/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.cpp
index 5dfb77745a6080a72ff58acda02446e06dc8abe9..8ac3b26589334789816810c3dc66d0de600af6ac 100644 (file)
--- a/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.cpp
+++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.cpp
@@ -70,7 +70,30 @@ NVC0LegalizeSSA::handleDIV(Instruction *i)
 void
 NVC0LegalizeSSA::handleRCPRSQ(Instruction *i)
 {
-   // TODO
+   assert(i->dType == TYPE_F64);
+   // There are instructions that will compute the high 32 bits of the 64-bit
+   // float. We will just stick 0 in the bottom 32 bits.
+
+   bld.setPosition(i, false);
+
+   // 1. Take the source and it up.
+   Value *src[2], *dst[2], *def = i->getDef(0);
+   bld.mkSplit(src, 4, i->getSrc(0));
+
+   // 2. We don't care about the low 32 bits of the destination. Stick a 0 in.
+   dst[0] = bld.loadImm(NULL, 0);
+   dst[1] = bld.getSSA();
+
+   // 3. The new version of the instruction takes the high 32 bits of the
+   // source and outputs the high 32 bits of the destination.
+   i->setSrc(0, src[1]);
+   i->setDef(0, dst[1]);
+   i->setType(TYPE_F32);
+   i->subOp = NV50_IR_SUBOP_RCPRSQ_64H;
+
+   // 4. Recombine the two dst pieces back into the original destination.
+   bld.setPosition(i, true);
+   bld.mkOp2(OP_MERGE, TYPE_U64, def, dst[0], dst[1]);
 }
 
 bool
@@ -1520,7 +1543,7 @@ NVC0LoweringPass::handleDIV(Instruction *i)
    if (!isFloatType(i->dType))
       return true;
    bld.setPosition(i, false);
-   Instruction *rcp = bld.mkOp1(OP_RCP, i->dType, bld.getSSA(), i->getSrc(1));
+   Instruction *rcp = bld.mkOp1(OP_RCP, i->dType, bld.getSSA(typeSizeof(i->dType)), i->getSrc(1));
    i->op = OP_MUL;
    i->setSrc(1, rcp->getDef(0));
    return true;
@@ -1529,13 +1552,13 @@ NVC0LoweringPass::handleDIV(Instruction *i)
 bool
 NVC0LoweringPass::handleMOD(Instruction *i)
 {
-   if (i->dType != TYPE_F32)
+   if (!isFloatType(i->dType))
       return true;
-   LValue *value = bld.getScratch();
-   bld.mkOp1(OP_RCP, TYPE_F32, value, i->getSrc(1));
-   bld.mkOp2(OP_MUL, TYPE_F32, value, i->getSrc(0), value);
-   bld.mkOp1(OP_TRUNC, TYPE_F32, value, value);
-   bld.mkOp2(OP_MUL, TYPE_F32, value, i->getSrc(1), value);
+   LValue *value = bld.getScratch(typeSizeof(i->dType));
+   bld.mkOp1(OP_RCP, i->dType, value, i->getSrc(1));
+   bld.mkOp2(OP_MUL, i->dType, value, i->getSrc(0), value);
+   bld.mkOp1(OP_TRUNC, i->dType, value, value);
+   bld.mkOp2(OP_MUL, i->dType, value, i->getSrc(1), value);
    i->op = OP_SUB;
    i->setSrc(1, value);
    return true;
@@ -1544,8 +1567,8 @@ NVC0LoweringPass::handleMOD(Instruction *i)
 bool
 NVC0LoweringPass::handleSQRT(Instruction *i)
 {
-   Instruction *rsq = bld.mkOp1(OP_RSQ, TYPE_F32,
-                                bld.getSSA(), i->getSrc(0));
+   Instruction *rsq = bld.mkOp1(OP_RSQ, i->dType,
+                                bld.getSSA(typeSizeof(i->dType)), i->getSrc(0));
    i->op = OP_MUL;
    i->setSrc(1, rsq->getDef(0));