r600/sfn: improve scheduling of tex sources

Let tex sources switch the channel to unused channels.

Signed-off-by: Gert Wollny <gert.wollny@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19300>

diff --git a/src/gallium/drivers/r600/sfn/sfn_instr.h b/src/gallium/drivers/r600/sfn/sfn_instr.h
index 720f9a3..dee1744 100644 (file)
--- a/src/gallium/drivers/r600/sfn/sfn_instr.h
+++ b/src/gallium/drivers/r600/sfn/sfn_instr.h
@@ -127,6 +127,8 @@ public:
    const InstrList& dependend_instr() { return m_dependend_instr;}
 
    virtual AluInstr *as_alu() { return nullptr;}
+   virtual uint8_t allowed_dest_chan_mask() const { return 0; }
+
 protected:
 
    const InstrList& required_instr() const {return m_required_instr; }

diff --git a/src/gallium/drivers/r600/sfn/sfn_instr_alu.cpp b/src/gallium/drivers/r600/sfn/sfn_instr_alu.cpp
index a55148e..80c0c1e 100644 (file)
--- a/src/gallium/drivers/r600/sfn/sfn_instr_alu.cpp
+++ b/src/gallium/drivers/r600/sfn/sfn_instr_alu.cpp
@@ -444,6 +444,18 @@ void AluInstr::set_sources(SrcValues src)
    }
 }
 
+uint8_t AluInstr::allowed_dest_chan_mask() const
+{
+   if (alu_slots() != 1) {
+      if (has_alu_flag(alu_is_cayman_trans)) {
+         return (1 << alu_slots()) - 1;
+      } else {
+         return 0;
+      }
+   }
+   return 0xf;
+}
+
 bool AluInstr::replace_dest(PRegister new_dest, AluInstr *move_instr)
 {
    if (m_dest->equal_to(*new_dest))

diff --git a/src/gallium/drivers/r600/sfn/sfn_instr_alu.h b/src/gallium/drivers/r600/sfn/sfn_instr_alu.h
index a2ff296..1c28270 100644 (file)
--- a/src/gallium/drivers/r600/sfn/sfn_instr_alu.h
+++ b/src/gallium/drivers/r600/sfn/sfn_instr_alu.h
@@ -160,6 +160,8 @@ public:
 
    AluInstr *as_alu() override { return this;}
 
+   uint8_t allowed_dest_chan_mask() const override;
+
 private:
    friend class AluGroup;
 

diff --git a/src/gallium/drivers/r600/sfn/sfn_instr_alugroup.cpp b/src/gallium/drivers/r600/sfn/sfn_instr_alugroup.cpp
index 35b7676..9f2cc87 100644 (file)
--- a/src/gallium/drivers/r600/sfn/sfn_instr_alugroup.cpp
+++ b/src/gallium/drivers/r600/sfn/sfn_instr_alugroup.cpp
@@ -25,6 +25,9 @@
  */
 
 #include "sfn_instr_alugroup.h"
+#include "sfn_instr_export.h"
+#include "sfn_instr_mem.h"
+#include "sfn_instr_tex.h"
 #include "sfn_debug.h"
 #include <algorithm>
 
@@ -158,25 +161,6 @@ int AluGroup::free_slots() const
    return free_mask;
 }
 
-class AluAllowSlotSwitch : public AluInstrVisitor {
-public:
-   using AluInstrVisitor::visit;
-
-   void visit(AluInstr *alu) {
-      if (alu->alu_slots() != 1) {
-         if (alu->has_alu_flag(alu_is_cayman_trans)) {
-            free_mask &= (1 << alu->alu_slots()) - 1;
-         } else {
-            yes = false;
-         }
-      }
-   }
-
-   bool yes{true};
-   uint8_t free_mask{0xf};
-
-};
-
 bool AluGroup::add_vec_instructions(AluInstr *instr)
 {
    if (!update_indirect_access(instr))
@@ -213,18 +197,18 @@ bool AluGroup::add_vec_instructions(AluInstr *instr)
    } else {
 
       auto dest = instr->dest();
-      if (dest && dest->pin() == pin_free) {
+      if (dest && (dest->pin() == pin_free || dest->pin() == pin_group)) {
 
-         AluAllowSlotSwitch swich_allowed;
+         int free_mask = 0xf;
          for (auto u : dest->uses()) {
-            u->accept(swich_allowed);
-            if (!swich_allowed.yes)
+            free_mask &= u->allowed_dest_chan_mask();
+            if (!free_mask)
                return false;
          }
 
          int free_chan = 0;
          while (free_chan < 4 && (m_slots[free_chan] ||
-                !(swich_allowed.free_mask & (1 << free_chan))))
+                !(free_mask & (1 << free_chan))))
             free_chan++;
 
          if (free_chan < 4) {
@@ -255,8 +239,12 @@ bool AluGroup::try_readport(AluInstr *instr, AluBankSwizzle cycle)
       m_has_lds_op |= instr->has_lds_access();
       sfn_log << SfnLog::schedule << "V: " << *instr << "\n";
       auto dest = instr->dest();
-      if (dest && dest->pin() == pin_free)
-         dest->set_pin(pin_chan);
+      if (dest) {
+         if (dest->pin() == pin_free)
+            dest->set_pin(pin_chan);
+         else if (dest->pin() == pin_group)
+            dest->set_pin(pin_chgr);
+      }
       instr->pin_sources_to_chan();
       return true;
    }

diff --git a/src/gallium/drivers/r600/sfn/sfn_instr_export.cpp b/src/gallium/drivers/r600/sfn/sfn_instr_export.cpp
index 5d4a3b7..dac0e10 100644 (file)
--- a/src/gallium/drivers/r600/sfn/sfn_instr_export.cpp
+++ b/src/gallium/drivers/r600/sfn/sfn_instr_export.cpp
@@ -147,6 +147,11 @@ ExportInstr::Pointer ExportInstr::from_string_impl(std::istream& is, ValueFactor
    return new ExportInstr( type, pos, value);
 }
 
+uint8_t ExportInstr::allowed_dest_chan_mask() const
+{
+   return value().free_chan_mask();
+}
+
 ScratchIOInstr::ScratchIOInstr(const RegisterVec4& value, PRegister addr,
                                int align, int align_offset, int writemask,
                                int array_size, bool is_read):

diff --git a/src/gallium/drivers/r600/sfn/sfn_instr_export.h b/src/gallium/drivers/r600/sfn/sfn_instr_export.h
index d19580a..d8ab4f0 100644 (file)
--- a/src/gallium/drivers/r600/sfn/sfn_instr_export.h
+++ b/src/gallium/drivers/r600/sfn/sfn_instr_export.h
@@ -43,6 +43,8 @@ public:
 
    const RegisterVec4& value() const {return m_value;};
    RegisterVec4& value() {return m_value;};
+
+
 private:
 
    RegisterVec4 m_value;
@@ -79,6 +81,8 @@ public:
    static Instr::Pointer from_string(std::istream& is, ValueFactory &vf);
    static Instr::Pointer last_from_string(std::istream& is, ValueFactory &vf);
 
+   uint8_t allowed_dest_chan_mask() const override;
+
 private:
    static ExportInstr::Pointer from_string_impl(std::istream& is, ValueFactory &vf);
 

diff --git a/src/gallium/drivers/r600/sfn/sfn_instr_tex.cpp b/src/gallium/drivers/r600/sfn/sfn_instr_tex.cpp
index bb480a8..f687392 100644 (file)
--- a/src/gallium/drivers/r600/sfn/sfn_instr_tex.cpp
+++ b/src/gallium/drivers/r600/sfn/sfn_instr_tex.cpp
@@ -376,6 +376,11 @@ bool TexInstr::replace_source(PRegister old_src, PVirtualValue new_src)
    return success;
 }
 
+uint8_t TexInstr::allowed_dest_chan_mask() const
+{
+   return m_src.free_chan_mask();
+}
+
 struct SamplerId {
    int id;
    bool indirect;

diff --git a/src/gallium/drivers/r600/sfn/sfn_instr_tex.h b/src/gallium/drivers/r600/sfn/sfn_instr_tex.h
index 0eddf63..4b40a13 100644 (file)
--- a/src/gallium/drivers/r600/sfn/sfn_instr_tex.h
+++ b/src/gallium/drivers/r600/sfn/sfn_instr_tex.h
@@ -152,6 +152,7 @@ public:
 
    bool replace_source(PRegister old_src, PVirtualValue new_src) override;
 
+   uint8_t allowed_dest_chan_mask() const override;
 private:
 
    bool do_ready() const override;

diff --git a/src/gallium/drivers/r600/sfn/sfn_virtualvalues.cpp b/src/gallium/drivers/r600/sfn/sfn_virtualvalues.cpp
index b41e705..3b40895 100644 (file)
--- a/src/gallium/drivers/r600/sfn/sfn_virtualvalues.cpp
+++ b/src/gallium/drivers/r600/sfn/sfn_virtualvalues.cpp
@@ -451,7 +451,7 @@ void RegisterVec4::print(std::ostream& os) const
 {
    os << (m_values[0]->value()->is_ssa() ? 'S' : 'R') << sel() << ".";
    for (int i = 0; i < 4; ++i)
-      os << VirtualValue::chanchar[m_swz[i]];
+      os << VirtualValue::chanchar[m_values[i]->value()->chan()];
 }
 
 bool operator == (const RegisterVec4& lhs, const RegisterVec4& rhs)

diff --git a/src/gallium/drivers/r600/sfn/sfn_virtualvalues.h b/src/gallium/drivers/r600/sfn/sfn_virtualvalues.h
index 52f817f..d71c0fd 100644 (file)
--- a/src/gallium/drivers/r600/sfn/sfn_virtualvalues.h
+++ b/src/gallium/drivers/r600/sfn/sfn_virtualvalues.h
@@ -297,6 +297,17 @@ public:
       }
    }
 
+   uint8_t free_chan_mask() const {
+      int mask = 0xf;
+      for (int i = 0; i < 4; ++i) {
+         int chan = m_values[i]->value()->chan();
+         if (chan <= 3) {
+            mask &= ~(1 << chan);
+         }
+      }
+      return mask;
+   }
+
    bool ready(int block_id, int index) const;
 private:
    int m_sel;

diff --git a/src/gallium/drivers/r600/sfn/tests/sfn_liverange_test.cpp b/src/gallium/drivers/r600/sfn/tests/sfn_liverange_test.cpp
index f12f600..1a9b5e1 100644 (file)
--- a/src/gallium/drivers/r600/sfn/tests/sfn_liverange_test.cpp
+++ b/src/gallium/drivers/r600/sfn/tests/sfn_liverange_test.cpp
@@ -173,7 +173,7 @@ TEST_F(LiveRangeTests, SimpleArrayAccess)
    auto s2x = vf.dest_from_string("S2.x");
    auto s2y = vf.dest_from_string("S2.y");
 
-   auto s3 = vf.dest_vec4_from_string("S3.xy01", dummy, pin_group);
+   auto s3 = vf.dest_vec4_from_string("S3.xy01", dummy, pin_chgr);
 
 
    LiveRangeMap expect = vf.prepare_live_range_map();

diff --git a/src/gallium/drivers/r600/sfn/tests/sfn_optimizer_test.cpp b/src/gallium/drivers/r600/sfn/tests/sfn_optimizer_test.cpp
index 487cb20..12f7288 100644 (file)
--- a/src/gallium/drivers/r600/sfn/tests/sfn_optimizer_test.cpp
+++ b/src/gallium/drivers/r600/sfn/tests/sfn_optimizer_test.cpp
@@ -277,6 +277,13 @@ TEST_F(TestShaderFromNir, fs_opt_tex_coord)
    check(sh, fs_opt_tex_coord_expect);
 }
 
+TEST_F(TestShaderFromNir, fs_shed_tex_coord)
+{
+   auto sh = from_string(fs_sched_tex_coord_init);
+   check(schedule(sh), fs_sched_tex_coord_expect);
+}
+
+
 void TestShaderFromNir::check(Shader *s, const char *expect_orig)
 {
    ostringstream test_str;

diff --git a/src/gallium/drivers/r600/sfn/tests/sfn_test_shaders.cpp b/src/gallium/drivers/r600/sfn/tests/sfn_test_shaders.cpp
index c1e312e..1f874ac 100644 (file)
--- a/src/gallium/drivers/r600/sfn/tests/sfn_test_shaders.cpp
+++ b/src/gallium/drivers/r600/sfn/tests/sfn_test_shaders.cpp
@@ -2409,6 +2409,89 @@ ALU ADD S5.z@group : S3.z@group S4.z@group {W}
 ALU ADD S5.w@group : S3.w@group S4.w@group {W}
 EXPORT_DONE PIXEL 0 S5.xyzw)";
 
+const char *fs_sched_tex_coord_init =
+R"(FS
+CHIPCLASS EVERGREEN
+PROP MAX_COLOR_EXPORTS:1
+PROP COLOR_EXPORTS:1
+PROP COLOR_EXPORT_MASK:15
+INPUT LOC:0 NAME:5 INTERP:2 SID:9 SPI_SID:10
+OUTPUT LOC:0 NAME:1 MASK:15
+REGISTERS R0.x@fully R0.y@fully
+SHADER
+ALU_GROUP_BEGIN
+  ALU INTERP_XY S1.x@chan : R0.y@fully Param0.x {W} VEC_210
+  ALU INTERP_XY S1.y@chan : R0.x@fully Param0.y {W} VEC_210
+  ALU INTERP_XY __.z@chan : R0.y@fully Param0.z {} VEC_210
+  ALU INTERP_XY __.w@chan : R0.x@fully Param0.w {L} VEC_210
+ALU_GROUP_END
+ALU_GROUP_BEGIN
+  ALU INTERP_ZW __.x@chan : R0.y@fully Param0.x {} VEC_210
+  ALU INTERP_ZW __.y@chan : R0.x@fully Param0.y {} VEC_210
+  ALU INTERP_ZW S1.z@chan : R0.y@fully Param0.z {W} VEC_210
+  ALU INTERP_ZW S1.w@chan : R0.x@fully Param0.w {WL} VEC_210
+ALU_GROUP_END
+ALU ADD S2.x@group : S1.x@chan S1.z@chan {W}
+ALU ADD S2.y@group : S1.y@chan S1.w@chan {WL}
+ALU MUL_IEEE S3.x@group : S1.x@chan S1.z@chan {W}
+ALU MUL_IEEE S3.y@group : S1.y@chan S1.w@chan {WL}
+
+TEX SAMPLE S4.xyzw : S2.xy__ RID:18 SID:0 NNNN
+TEX SAMPLE S5.xyzw : S3.xy__ RID:18 SID:0 NNNN
+ALU ADD S6.x@group : S5.x@group S4.x@group {W}
+ALU ADD S6.y@group : S5.y@group S4.y@group {W}
+ALU ADD S6.z@group : S5.z@group S4.z@group {W}
+ALU ADD S6.w@group : S5.w@group S4.w@group {W}
+EXPORT_DONE PIXEL 0 S5.xyzw)";
+
+
+const char *fs_sched_tex_coord_expect =
+R"(FS
+CHIPCLASS EVERGREEN
+PROP MAX_COLOR_EXPORTS:1
+PROP COLOR_EXPORTS:1
+PROP COLOR_EXPORT_MASK:15
+INPUT LOC:0 NAME:5 INTERP:2 SID:9 SPI_SID:10
+OUTPUT LOC:0 NAME:1 MASK:15
+REGISTERS R0.x@fully R0.y@fully
+SHADER
+BLOCK_START
+ALU_GROUP_BEGIN
+  ALU INTERP_XY S1.x@chan : R0.y@fully Param0.x {W} VEC_210
+  ALU INTERP_XY S1.y@chan : R0.x@fully Param0.y {W} VEC_210
+  ALU INTERP_XY __.z@chan : R0.y@fully Param0.z {} VEC_210
+  ALU INTERP_XY __.w@chan : R0.x@fully Param0.w {L} VEC_210
+ALU_GROUP_END
+ALU_GROUP_BEGIN
+  ALU INTERP_ZW __.x@chan : R0.y@fully Param0.x {} VEC_210
+  ALU INTERP_ZW __.y@chan : R0.x@fully Param0.y {} VEC_210
+  ALU INTERP_ZW S1.z@chan : R0.y@fully Param0.z {W} VEC_210
+  ALU INTERP_ZW S1.w@chan : R0.x@fully Param0.w {WL} VEC_210
+ALU_GROUP_END
+ALU_GROUP_BEGIN
+  ALU ADD S2.x@group : S1.x@chan S1.z@chan {W}
+  ALU ADD S2.y@group : S1.y@chan S1.w@chan {W}
+  ALU MUL_IEEE S3.z@chgr : S1.x@chan S1.z@chan {W}
+  ALU MUL_IEEE S3.w@chgr : S1.y@chan S1.w@chan {WL}
+ALU_GROUP_END
+BLOCK_END
+BLOCK_START
+TEX SAMPLE S4.xyzw : S2.xy__ RID:18 SID:0 NNNN
+TEX SAMPLE S5.xyzw : S3.zw__ RID:18 SID:0 NNNN
+BLOCK_END
+BLOCK_START
+ALU_GROUP_BEGIN
+ALU ADD S6.x@group : S5.x@group S4.x@group {W}
+ALU ADD S6.y@group : S5.y@group S4.y@group {W}
+ALU ADD S6.z@group : S5.z@group S4.z@group {W}
+ALU ADD S6.w@group : S5.w@group S4.w@group {WL}
+ALU_GROUP_END
+BLOCK_END
+BLOCK_START
+EXPORT_DONE PIXEL 0 S5.xyzw
+BLOCK_END)";
+
+
 const char *fs_with_loop_multislot_reuse =
 R"(FS
 CHIPCLASS CAYMAN

diff --git a/src/gallium/drivers/r600/sfn/tests/sfn_test_shaders.h b/src/gallium/drivers/r600/sfn/tests/sfn_test_shaders.h
index bcd224e..e56506b 100644 (file)
--- a/src/gallium/drivers/r600/sfn/tests/sfn_test_shaders.h
+++ b/src/gallium/drivers/r600/sfn/tests/sfn_test_shaders.h
@@ -98,6 +98,9 @@ extern const char *vtx_for_tcs_sched;
 extern const char *fs_opt_tex_coord_init;
 extern const char *fs_opt_tex_coord_expect;
 
+extern const char *fs_sched_tex_coord_init;
+extern const char *fs_sched_tex_coord_expect;
+
 class TestShader : public ::testing::Test {
 
    void SetUp() override;