nvc0: fix bindless multisampled images on Maxwell+

NVC0_CB_AUX_BINDLESS_INFO isn't written to on Maxwell+ and it's too small
anyway.

With these changes, TXQ is used to determine the number of samples and
the coordinate adjustment information looked up in a small array in the
driver constant buffer.

v2: rework to use TXQ and a small array instead of a larger array with an
    entry for each texture
v3: get rid of the small array and calculate the adjustments in the shader

Signed-off-by: Rhys Perry <[email protected]>
Fixes: c2ae9b40527 ('nvc0: implement multisampled images on Maxwell+')
Reviewed-by: Ilia Mirkin <[email protected]>

3 files changed, 45 insertions, 5 deletions

diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_gm107.cpp b/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_gm107.cpp
index c7436e2e297..49a5f3b01f2 100644
--- a/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_gm107.cpp
+++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_gm107.cpp
@@ -320,11 +320,11 @@ GM107LoweringPass::handleSUQ(TexInstruction *suq)
 
       if (mask & 0x1)
          bld.mkOp2(OP_SHR, TYPE_U32, suq->getDef(0), suq->getDef(0),
-                   loadSuInfo32(ind, slot, NVC0_SU_INFO_MS(0), suq->tex.bindless));
+                   loadMsAdjInfo32(suq->tex.target, 0, slot, ind, suq->tex.bindless));
       if (mask & 0x2) {
          int d = util_bitcount(mask & 0x1);
          bld.mkOp2(OP_SHR, TYPE_U32, suq->getDef(d), suq->getDef(d),
-                   loadSuInfo32(ind, slot, NVC0_SU_INFO_MS(1), suq->tex.bindless));
+                   loadMsAdjInfo32(suq->tex.target, 1, slot, ind, suq->tex.bindless));
       }
    }
 
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 176e0cf608d..0cbf4d071f2 100644
--- a/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.cpp
+++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.cpp
@@ -1732,6 +1732,45 @@ NVC0LoweringPass::loadSuInfo32(Value *ptr, int slot, uint32_t off, bool bindless
                         prog->driver->io.suInfoBase);
 }
 
+inline Value *
+NVC0LoweringPass::loadMsAdjInfo32(TexInstruction::Target target, uint32_t index, int slot, Value *ind, bool bindless)
+{
+   if (!bindless || targ->getChipset() < NVISA_GM107_CHIPSET)
+      return loadSuInfo32(ind, slot, NVC0_SU_INFO_MS(index), bindless);
+
+   assert(bindless);
+
+   Value *samples = bld.getSSA();
+   // this shouldn't be lowered because it's being inserted before the current instruction
+   TexInstruction *tex = new_TexInstruction(func, OP_TXQ);
+   tex->tex.target = target;
+   tex->tex.query = TXQ_TYPE;
+   tex->tex.mask = 0x4;
+   tex->tex.r = 0xff;
+   tex->tex.s = 0x1f;
+   tex->tex.rIndirectSrc = 0;
+   tex->setDef(0, samples);
+   tex->setSrc(0, ind);
+   tex->setSrc(1, bld.loadImm(NULL, 0));
+   bld.insert(tex);
+
+   // doesn't work with sample counts other than 1/2/4/8 but they aren't supported
+   switch (index) {
+   case 0: {
+      Value *tmp = bld.mkOp2v(OP_ADD, TYPE_U32, bld.getSSA(), samples, bld.mkImm(2));
+      return bld.mkOp2v(OP_SHR, TYPE_U32, bld.getSSA(), tmp, bld.mkImm(2));
+   }
+   case 1: {
+      Value *tmp = bld.mkCmp(OP_SET, CC_GT, TYPE_U32, bld.getSSA(), TYPE_U32, samples, bld.mkImm(2))->getDef(0);
+      return bld.mkOp2v(OP_AND, TYPE_U32, bld.getSSA(), tmp, bld.mkImm(1));
+   }
+   default: {
+      assert(false);
+      return NULL;
+   }
+   }
+}
+
 static inline uint16_t getSuClampSubOp(const TexInstruction *su, int c)
 {
    switch (su->tex.target.getEnum()) {
@@ -1817,8 +1856,8 @@ NVC0LoweringPass::adjustCoordinatesMS(TexInstruction *tex)
    Value *tx = bld.getSSA(), *ty = bld.getSSA(), *ts = bld.getSSA();
    Value *ind = tex->getIndirectR();
 
-   Value *ms_x = loadSuInfo32(ind, slot, NVC0_SU_INFO_MS(0), tex->tex.bindless);
-   Value *ms_y = loadSuInfo32(ind, slot, NVC0_SU_INFO_MS(1), tex->tex.bindless);
+   Value *ms_x = loadMsAdjInfo32(tex->tex.target, 0, slot, ind, tex->tex.bindless);
+   Value *ms_y = loadMsAdjInfo32(tex->tex.target, 1, slot, ind, tex->tex.bindless);
 
    bld.mkOp2(OP_SHL, TYPE_U32, tx, x, ms_x);
    bld.mkOp2(OP_SHL, TYPE_U32, ty, y, ms_y);
diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.h b/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.h
index 5dbb3e4f009..4136b1ecfeb 100644
--- a/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.h
+++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.h
@@ -148,7 +148,7 @@ protected:
    void handlePIXLD(Instruction *);
 
    void checkPredicate(Instruction *);
-   Value *loadSuInfo32(Value *ptr, int slot, uint32_t off, bool bindless);
+   Value *loadMsAdjInfo32(TexInstruction::Target targ, uint32_t index, int slot, Value *ind, bool bindless);
 
    virtual bool visit(Instruction *);
 
@@ -161,6 +161,7 @@ private:
    Value *loadResInfo32(Value *ptr, uint32_t off, uint16_t base);
    Value *loadResInfo64(Value *ptr, uint32_t off, uint16_t base);
    Value *loadResLength32(Value *ptr, uint32_t off, uint16_t base);
+   Value *loadSuInfo32(Value *ptr, int slot, uint32_t off, bool bindless);
    Value *loadBufInfo64(Value *ptr, uint32_t off);
    Value *loadBufLength32(Value *ptr, uint32_t off);
    Value *loadUboInfo64(Value *ptr, uint32_t off);