[Mesa-dev] [PATCH v2 26/31] nvir/nir: implement variable indexing

[Karol Herbst]

we store those arrays in local memory and reserve some space for each of the
arrays. The arrays are stored in a packed format, because we know quite easily
the context of each index. We don't do that in TGSI so far.

This causes various issues to come up in the MemoryOpt pass, because ld/st with
indirects aren't guarenteed to be aligned to 0x10 anymore.

Signed-off-by: Karol Herbst <kherbst at redhat.com>
---
 .../drivers/nouveau/codegen/nv50_ir_from_nir.cpp   | 58 ++++++++++++++++++++++
 1 file changed, 58 insertions(+)

diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_from_nir.cpp b/src/gallium/drivers/nouveau/codegen/nv50_ir_from_nir.cpp
index 34e9a85e13..1ac5436153 100644
--- a/src/gallium/drivers/nouveau/codegen/nv50_ir_from_nir.cpp
+++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_from_nir.cpp
@@ -74,6 +74,7 @@ public:
    typedef decltype(nir_ssa_def().index) NirSSADefIdx;
    typedef decltype(nir_ssa_def().bit_size) NirSSADefBitSize;
    typedef std::unordered_map<NirSSADefIdx, LValues> NirDefMap;
+   typedef std::unordered_map<NirSSADefIdx, uint32_t> NirArrayLMemOffsets;
    typedef std::unordered_map<decltype(nir_block().index), BasicBlock*> NirBlockMap;
 
    Converter(Program *, nir_shader *, nv50_ir_prog_info *);
@@ -140,6 +141,7 @@ private:
 
    NirDefMap ssaDefs;
    NirDefMap regDefs;
+   NirArrayLMemOffsets regToLmemOffset;
    NirBlockMap blocks;
    unsigned int curLoopDepth;
 
@@ -1091,6 +1093,7 @@ bool Converter::assignSlots() {
 bool
 Converter::parseNIR()
 {
+   info->bin.tlsSpace = 0;
    info->io.clipDistances = nir->info.clip_distance_array_size;
    info->io.cullDistances = nir->info.cull_distance_array_size;
 
@@ -1175,6 +1178,15 @@ Converter::visit(nir_function *function)
       break;
    }
 
+   nir_foreach_register(reg, &function->impl->registers) {
+      if (reg->num_array_elems) {
+         uint32_t size = reg->num_components * reg->num_array_elems * (reg->bit_size / 8);
+         // reserve some lmem
+         regToLmemOffset[reg->index] = info->bin.tlsSpace;
+         info->bin.tlsSpace += size;
+      }
+   }
+
    nir_index_ssa_defs(function->impl);
    foreach_list_typed(nir_cf_node, node, node, &function->impl->body) {
       if (!visit(node))
@@ -1737,6 +1749,51 @@ Converter::visit(nir_alu_instr *insn)
     *   2. they basically just merge multiple values into one data type
     */
    CASE_OPFI(mov):
+      if (!insn->dest.dest.is_ssa && insn->dest.dest.reg.reg->num_array_elems) {
+         nir_reg_dest& reg = insn->dest.dest.reg;
+         auto goffset = regToLmemOffset[reg.reg->index];
+         auto comps = reg.reg->num_components;
+         auto size = reg.reg->bit_size / 8;
+         auto csize = comps * size;
+         auto aoffset = csize * reg.base_offset;
+         Value *indirect = nullptr;
+
+         if (reg.indirect)
+            indirect = mkOp2v(OP_MUL, TYPE_U32, getSSA(4, FILE_ADDRESS), getSrc(reg.indirect, 0), mkImm(csize));
+
+         for (auto i = 0u; i < comps; ++i) {
+            if (!((1u << i) & insn->dest.write_mask))
+               continue;
+
+            Symbol *sym = mkSymbol(FILE_MEMORY_LOCAL, 0, dType, goffset + aoffset + i * size);
+            mkStore(OP_STORE, dType, sym, indirect, getSrc(&insn->src[0], i));
+         }
+         break;
+      } else if (!insn->src[0].src.is_ssa && insn->src[0].src.reg.reg->num_array_elems) {
+         LValues &newDefs = convert(&insn->dest);
+         nir_reg_src& reg = insn->src[0].src.reg;
+         auto goffset = regToLmemOffset[reg.reg->index];
+         auto comps = reg.reg->num_components;
+         auto size = reg.reg->bit_size / 8;
+         auto csize = comps * size;
+         auto aoffset = csize * reg.base_offset;
+         Value *indirect = nullptr;
+
+         if (reg.indirect)
+            indirect = mkOp2v(OP_MUL, TYPE_U32, getSSA(4, FILE_ADDRESS), getSrc(reg.indirect, 0), mkImm(csize));
+
+         for (auto i = 0u; i < newDefs.size(); ++i) {
+            Symbol *sym = mkSymbol(FILE_MEMORY_LOCAL, 0, dType, goffset + aoffset + i * size);
+            mkLoad(dType, newDefs[i], sym, indirect);
+         }
+         break;
+      } else {
+         LValues &newDefs = convert(&insn->dest);
+         for (LValues::size_type c = 0u; c < newDefs.size(); ++c) {
+            mkMov(newDefs[c], getSrc(&insn->src[0], c), dType);
+         }
+      }
+      break;
    case nir_op_vec2:
    case nir_op_vec3:
    case nir_op_vec4: {
@@ -2120,6 +2177,7 @@ Converter::run()
       NIR_PASS(progress, nir, nir_opt_dead_cf);
    } while (progress);
 
+   NIR_PASS_V(nir, nir_lower_locals_to_regs);
    NIR_PASS_V(nir, nir_remove_dead_variables, nir_var_local);
    NIR_PASS_V(nir, nir_convert_from_ssa, true);
 
-- 
2.14.3