r600/sfn: Add support for geometry shader
authorGert Wollny <gert.wollny@collabora.com>
Fri, 27 Dec 2019 16:49:27 +0000 (17:49 +0100)
committerMarge Bot <eric+marge@anholt.net>
Mon, 10 Feb 2020 19:09:08 +0000 (19:09 +0000)
v2: fix compilation with gcc-6

Signed-off-by: Gert Wollny <gert.wollny@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3225>

src/gallium/drivers/r600/meson.build
src/gallium/drivers/r600/sfn/sfn_nir.cpp
src/gallium/drivers/r600/sfn/sfn_shader_base.cpp
src/gallium/drivers/r600/sfn/sfn_shader_geometry.cpp [new file with mode: 0644]
src/gallium/drivers/r600/sfn/sfn_shader_geometry.h [new file with mode: 0644]
src/gallium/drivers/r600/sfn/sfn_shader_vertex.cpp
src/gallium/drivers/r600/sfn/sfn_shader_vertex.h

index bfc17dae15393e6f13cd8ba8cd50991455b074ad..3a4cf25a0981f1d8a5ab65788a85b147a7ad24d6 100644 (file)
@@ -148,6 +148,8 @@ files_r600 = files(
   'sfn/sfn_shader_base.h',
   'sfn/sfn_shader_fragment.cpp',
   'sfn/sfn_shader_fragment.h',
+  'sfn/sfn_shader_geometry.cpp',
+  'sfn/sfn_shader_geometry.h',
   'sfn/sfn_shader_vertex.cpp',
   'sfn/sfn_shader_vertex.h',
   'sfn/sfn_shaderio.cpp',
index 7eb025f46a45b3e7a90e60066041da4c63ae8548..845bb6c3b368574ede1499f49f14f496e36ae049 100644 (file)
@@ -34,6 +34,7 @@
 
 #include "sfn_shader_vertex.h"
 #include "sfn_shader_fragment.h"
+#include "sfn_shader_geometry.h"
 #include "sfn_nir_lower_fs_out_to_vector.h"
 #include "sfn_ir_to_assembly.h"
 
@@ -59,7 +60,8 @@ bool ShaderFromNir::lower(const nir_shader *shader, r600_pipe_shader *pipe_shade
    switch (shader->info.stage) {
    case MESA_SHADER_VERTEX:
       if (key.vs.as_es) {
-         sfn_log << SfnLog::trans << "VS; next type GS not yet supported\n";
+         sfn_log << SfnLog::trans << "Start VS for GS\n";
+         impl.reset(new VertexShaderFromNirForGS(pipe_shader, *sel, key, gs_shader));
       } else if (key.vs.as_ls) {
          sfn_log << "VS: next type TCS and TES not yet supported\n";
          return false;
@@ -68,6 +70,10 @@ bool ShaderFromNir::lower(const nir_shader *shader, r600_pipe_shader *pipe_shade
          impl.reset(new VertexShaderFromNirForFS(pipe_shader, *sel, key));
       }
       break;
+   case MESA_SHADER_GEOMETRY:
+      sfn_log << SfnLog::trans << "Start GS\n";
+      impl.reset(new GeometryShaderFromNir(pipe_shader, *sel, key));
+      break;
    case MESA_SHADER_FRAGMENT:
       sfn_log << SfnLog::trans << "Start FS\n";
       impl.reset(new FragmentShaderFromNir(*shader, pipe_shader->shader, *sel, key));
index d1f3f2feba03490457b449379b0535e406ee06f6..6da9a8035c22b9e887b7bd6b879d5936c55b0ea5 100644 (file)
@@ -29,6 +29,7 @@
 #include "sfn_shader_vertex.h"
 
 #include "sfn_shader_fragment.h"
+#include "sfn_shader_geometry.h"
 #include "sfn_liverange.h"
 #include "sfn_ir_to_assembly.h"
 #include "sfn_nir.h"
diff --git a/src/gallium/drivers/r600/sfn/sfn_shader_geometry.cpp b/src/gallium/drivers/r600/sfn/sfn_shader_geometry.cpp
new file mode 100644 (file)
index 0000000..931b5cd
--- /dev/null
@@ -0,0 +1,297 @@
+/* -*- mesa-c++  -*-
+ *
+ * Copyright (c) 2018 Collabora LTD
+ *
+ * Author: Gert Wollny <gert.wollny@collabora.com>
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * on the rights to use, copy, modify, merge, publish, distribute, sub
+ * license, and/or sell copies of the Software, and to permit persons to whom
+ * the Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
+ * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
+ * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
+ * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
+ * USE OR OTHER DEALINGS IN THE SOFTWARE.
+ */
+
+#include "tgsi/tgsi_from_mesa.h"
+#include "sfn_shader_geometry.h"
+#include "sfn_instruction_misc.h"
+#include "sfn_instruction_fetch.h"
+
+namespace r600 {
+
+GeometryShaderFromNir::GeometryShaderFromNir(r600_pipe_shader *sh,
+                                             r600_pipe_shader_selector &sel,
+                                             const r600_shader_key &key):
+   ShaderFromNirProcessor (PIPE_SHADER_GEOMETRY, sel, sh->shader,
+                           sh->scratch_space_needed),
+   m_pipe_shader(sh),
+   m_so_info(&sel.so),
+   m_first_vertex_emitted(false),
+   m_offset(0),
+   m_next_input_ring_offset(0),
+   m_key(key),
+   m_num_clip_dist(0),
+   m_cur_ring_output(0),
+   m_gs_tri_strip_adj_fix(false)
+{
+   sh_info().atomic_base = key.gs.first_atomic_counter;
+}
+
+bool GeometryShaderFromNir::do_emit_load_deref(UNUSED const nir_variable *in_var, UNUSED nir_intrinsic_instr* instr)
+{
+   return false;
+}
+
+bool GeometryShaderFromNir::do_emit_store_deref(const nir_variable *out_var, nir_intrinsic_instr* instr)
+{
+   uint32_t write_mask =  (1 << instr->num_components) - 1;
+   GPRVector::Swizzle swz = swizzle_from_mask(instr->num_components);
+   std::unique_ptr<GPRVector> vec(vec_from_nir_with_fetch_constant(instr->src[1], write_mask, swz));
+
+   GPRVector out_value  = *vec;
+
+   sh_info().output[out_var->data.driver_location].write_mask =
+         (1 << instr->num_components) - 1;
+
+   auto ir = new MemRingOutIntruction(cf_mem_ring, mem_write_ind, out_value,
+                                      4 * out_var->data.driver_location,
+                                      4, m_export_base);
+   emit_instruction(ir);
+
+   return true;
+}
+
+bool GeometryShaderFromNir::scan_sysvalue_access(UNUSED nir_instr *instr)
+{
+   return true;
+}
+
+bool GeometryShaderFromNir::do_process_inputs(nir_variable *input)
+{
+
+   if (input->data.location == VARYING_SLOT_POS ||
+       input->data.location == VARYING_SLOT_PSIZ ||
+       input->data.location == VARYING_SLOT_CLIP_DIST0 ||
+       input->data.location == VARYING_SLOT_CLIP_DIST1 ||
+       (input->data.location >= VARYING_SLOT_VAR0 &&
+       input->data.location <= VARYING_SLOT_VAR31) ||
+       (input->data.location >= VARYING_SLOT_TEX0 &&
+       input->data.location <= VARYING_SLOT_TEX7)) {
+
+      r600_shader_io& io = sh_info().input[input->data.driver_location];
+      tgsi_get_gl_varying_semantic(static_cast<gl_varying_slot>( input->data.location),
+                                   true, &io.name, &io.sid);
+      io.ring_offset = 16 * input->data.driver_location;
+      ++sh_info().ninput;
+      m_next_input_ring_offset += 16;
+      return true;
+   }
+
+   return false;
+}
+
+bool GeometryShaderFromNir::do_process_outputs(nir_variable *output)
+{
+   if (output->data.location == VARYING_SLOT_COL0 ||
+       output->data.location == VARYING_SLOT_COL1 ||
+       (output->data.location >= VARYING_SLOT_VAR0 &&
+       output->data.location <= VARYING_SLOT_VAR31) ||
+       (output->data.location >= VARYING_SLOT_TEX0 &&
+       output->data.location <= VARYING_SLOT_TEX7) ||
+       output->data.location == VARYING_SLOT_BFC0 ||
+       output->data.location == VARYING_SLOT_BFC1 ||
+       output->data.location == VARYING_SLOT_CLIP_VERTEX ||
+       output->data.location == VARYING_SLOT_CLIP_DIST0 ||
+       output->data.location == VARYING_SLOT_CLIP_DIST1 ||
+       output->data.location == VARYING_SLOT_PRIMITIVE_ID ||
+       output->data.location == VARYING_SLOT_POS ||
+       output->data.location == VARYING_SLOT_PSIZ ||
+       output->data.location == VARYING_SLOT_LAYER ||
+       output->data.location == VARYING_SLOT_VIEWPORT ||
+       output->data.location == VARYING_SLOT_FOGC) {
+      r600_shader_io& io = sh_info().output[output->data.driver_location];
+
+      tgsi_get_gl_varying_semantic(static_cast<gl_varying_slot>( output->data.location),
+                                   true, &io.name, &io.sid);
+      evaluate_spi_sid(io);
+      ++sh_info().noutput;
+
+      if (output->data.location == VARYING_SLOT_CLIP_DIST0 ||
+          output->data.location == VARYING_SLOT_CLIP_DIST1) {
+         m_num_clip_dist += 4;
+      }
+      return true;
+   }
+   return false;
+}
+
+
+bool GeometryShaderFromNir::allocate_reserved_registers()
+{
+   const int sel[6] = {0, 0 ,0, 1, 1, 1};
+   const int chan[6] = {0, 1 ,3, 0, 1, 2};
+
+   increment_reserved_registers();
+   increment_reserved_registers();
+
+   /* Reserve registers used by the shaders (should check how many
+    * components are actually used */
+   for (int i = 0; i < 6; ++i) {
+      auto reg = new GPRValue(sel[i], chan[i]);
+      reg->set_as_input();
+      m_per_vertex_offsets[i].reset(reg);
+      inject_register(sel[i], chan[i], m_per_vertex_offsets[i], false);
+   }
+   auto reg = new GPRValue(0, 2);
+   reg->set_as_input();
+   m_primitive_id.reset(reg);
+   inject_register(0, 2, m_primitive_id, false);
+
+   reg = new GPRValue(1, 3);
+   reg->set_as_input();
+   m_invocation_id.reset(reg);
+   inject_register(1, 3, m_invocation_id, false);
+
+   m_export_base = get_temp_register();
+   emit_instruction(new AluInstruction(op1_mov, m_export_base, Value::zero, {alu_write, alu_last_instr}));
+
+   sh_info().ring_item_sizes[0] = m_next_input_ring_offset;
+
+   if (m_key.gs.tri_strip_adj_fix)
+      emit_adj_fix();
+
+   return true;
+}
+
+void GeometryShaderFromNir::emit_adj_fix()
+{
+   PValue adjhelp0(new  GPRValue(m_export_base->sel(), 1));
+   emit_instruction(op2_and_int, adjhelp0, {m_primitive_id, Value::one_i}, {alu_write, alu_last_instr});
+
+   int help2 = allocate_temp_register();
+   int reg_indices[6];
+   int reg_chanels[6] = {0, 1, 2, 3, 2, 3};
+
+   int rotate_indices[6] = {4, 5, 0, 1, 2, 3};
+
+   reg_indices[0] = reg_indices[1] = reg_indices[2] = reg_indices[3] = help2;
+   reg_indices[4] = reg_indices[5] = m_export_base->sel();
+
+   std::array<PValue, 6> adjhelp;
+
+   AluInstruction *ir = nullptr;
+   for (int i = 0; i < 6; i++) {
+      adjhelp[i].reset(new GPRValue(reg_indices[i], reg_chanels[i]));
+      ir = new AluInstruction(op3_cnde_int, adjhelp[i],
+                             {adjhelp0, m_per_vertex_offsets[i],
+                              m_per_vertex_offsets[rotate_indices[i]]},
+                             {alu_write});
+      if (i == 3)
+         ir->set_flag(alu_last_instr);
+      emit_instruction(ir);
+   }
+   ir->set_flag(alu_last_instr);
+
+   for (int i = 0; i < 6; i++)
+      m_per_vertex_offsets[i] = adjhelp[i];
+}
+
+bool GeometryShaderFromNir::emit_deref_instruction_override(nir_deref_instr* instr)
+{
+   if (instr->deref_type  == nir_deref_type_array) {
+      auto var = get_deref_location(instr->parent);
+      ArrayDeref ad = {var, &instr->arr.index};
+      assert(instr->dest.is_ssa);
+      m_in_array_deref[instr->dest.ssa.index] = ad;
+
+      /* Problem: nir_intrinsice_load_deref tries to lookup the
+       * variable, and will not find it, need to override that too */
+      return true;
+   }
+   return false;
+}
+
+bool GeometryShaderFromNir::emit_intrinsic_instruction_override(nir_intrinsic_instr* instr)
+{
+   switch (instr->intrinsic) {
+   case nir_intrinsic_load_deref: {
+      auto& src = instr->src[0];
+      assert(src.is_ssa);
+      auto array = m_in_array_deref.find(src.ssa->index);
+      if (array != m_in_array_deref.end())
+         return emit_load_from_array(instr, array->second);
+   } break;
+   case nir_intrinsic_emit_vertex:
+      return emit_vertex(instr, false);
+   case nir_intrinsic_end_primitive:
+      return emit_vertex(instr, true);
+   case nir_intrinsic_load_primitive_id:
+      return load_preloaded_value(instr->dest, 0, m_primitive_id);
+   case nir_intrinsic_load_invocation_id:
+      return load_preloaded_value(instr->dest, 0, m_invocation_id);
+   default:
+      ;
+   }
+   return false;
+}
+
+bool GeometryShaderFromNir::emit_vertex(nir_intrinsic_instr* instr, bool cut)
+{
+   int stream = nir_intrinsic_stream_id(instr);
+   assert(stream < 4);
+
+   emit_instruction(new EmitVertex(stream, cut));
+
+   if (!cut)
+      emit_instruction(new AluInstruction(op2_add_int, m_export_base, m_export_base,
+                                          PValue(new LiteralValue(sh_info().noutput)),
+                                          {alu_write, alu_last_instr}));
+
+   return true;
+}
+
+bool GeometryShaderFromNir::emit_load_from_array(nir_intrinsic_instr* instr,
+                                                 const ArrayDeref& array_deref)
+{
+   auto dest = vec_from_nir(instr->dest, instr->num_components);
+
+   const nir_load_const_instr* literal_index = nullptr;
+
+   if (array_deref.index->is_ssa)
+      literal_index = get_literal_constant(array_deref.index->ssa->index);
+
+   if (!literal_index) {
+      sfn_log << SfnLog::err << "GS: Indirect input addressing not (yet) supported\n";
+      return false;
+   }
+   assert(literal_index->value[0].u32 < 6);
+   PValue addr = m_per_vertex_offsets[literal_index->value[0].u32];
+
+   auto fetch = new FetchInstruction(vc_fetch, no_index_offset, dest, addr,
+                                     16 * array_deref.var->data.driver_location,
+                                     R600_GS_RING_CONST_BUFFER, PValue(), bim_none, true);
+   emit_instruction(fetch);
+   return true;
+}
+
+void GeometryShaderFromNir::do_finalize()
+{
+   if (m_num_clip_dist) {
+      sh_info().cc_dist_mask = (1 << m_num_clip_dist) - 1;
+      sh_info().clip_dist_write = (1 << m_num_clip_dist) - 1;
+   }
+}
+
+}
diff --git a/src/gallium/drivers/r600/sfn/sfn_shader_geometry.h b/src/gallium/drivers/r600/sfn/sfn_shader_geometry.h
new file mode 100644 (file)
index 0000000..9f7f69f
--- /dev/null
@@ -0,0 +1,80 @@
+/* -*- mesa-c++  -*-
+ *
+ * Copyright (c) 2019 Collabora LTD
+ *
+ * Author: Gert Wollny <gert.wollny@collabora.com>
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * on the rights to use, copy, modify, merge, publish, distribute, sub
+ * license, and/or sell copies of the Software, and to permit persons to whom
+ * the Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
+ * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
+ * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
+ * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
+ * USE OR OTHER DEALINGS IN THE SOFTWARE.
+ */
+
+
+#ifndef SFN_GEOMETRYSHADERFROMNIR_H
+#define SFN_GEOMETRYSHADERFROMNIR_H
+
+#include "sfn_shader_base.h"
+
+namespace r600 {
+
+class GeometryShaderFromNir : public ShaderFromNirProcessor
+{
+public:
+   GeometryShaderFromNir(r600_pipe_shader *sh, r600_pipe_shader_selector& sel, const r600_shader_key& key);
+   bool do_emit_load_deref(const nir_variable *in_var, nir_intrinsic_instr* instr) override;
+   bool do_emit_store_deref(const nir_variable *out_var, nir_intrinsic_instr* instr) override;
+   bool scan_sysvalue_access(nir_instr *instr) override;
+
+private:
+   struct ArrayDeref {
+      const nir_variable *var;
+      const nir_src *index;
+   };
+
+   bool do_process_inputs(nir_variable *input) override;
+   bool allocate_reserved_registers() override;
+   bool do_process_outputs(nir_variable *output) override;
+   bool emit_deref_instruction_override(nir_deref_instr* instr) override;
+   bool emit_intrinsic_instruction_override(nir_intrinsic_instr* instr) override;
+   bool emit_load_from_array(nir_intrinsic_instr* instr, const ArrayDeref& array_deref);
+   bool emit_vertex(nir_intrinsic_instr* instr, bool cut);
+   void emit_adj_fix();
+
+   void do_finalize() override;
+
+   r600_pipe_shader *m_pipe_shader;
+   const pipe_stream_output_info *m_so_info;
+
+   std::array<PValue, 6> m_per_vertex_offsets;
+   PValue m_primitive_id;
+   PValue m_invocation_id;
+   PValue m_export_base;
+   bool m_first_vertex_emitted;
+
+   std::map<unsigned, ArrayDeref> m_in_array_deref;
+   int  m_offset;
+   int  m_next_input_ring_offset;
+   r600_shader_key m_key;
+   int m_num_clip_dist;
+   unsigned m_cur_ring_output;
+   bool m_gs_tri_strip_adj_fix;
+};
+
+}
+
+#endif // SFN_GEOMETRYSHADERFROMNIR_H
index e67cd4638f3a61be96b233c653239eab12676c25..86182d927e95667363921f75e5a5c2d904fcf26d 100644 (file)
@@ -489,4 +489,66 @@ void VertexShaderFromNirForFS::finalize_exports()
 
 }
 
+VertexShaderFromNirForGS::VertexShaderFromNirForGS(r600_pipe_shader *sh,
+                                                   r600_pipe_shader_selector& sel,
+                                                   const r600_shader_key &key,
+                                                   const r600_shader *gs_shader):
+   VertexShaderFromNir(sh, sel, key),
+   m_gs_shader(gs_shader)
+{
+   sh->shader.vs_as_es = true;
+}
+
+bool VertexShaderFromNirForGS::do_emit_store_deref(const nir_variable *out_var, nir_intrinsic_instr* instr)
+{
+
+   assert(m_gs_shader);
+
+   int ring_offset = -1;
+   const r600_shader_io& out_io = sh_info().output[out_var->data.driver_location];
+
+   sfn_log << SfnLog::io << "check output " << out_var->data.driver_location
+           << " name=" << out_io.name<< " sid=" << out_io.sid << "\n";
+   for (unsigned k = 0; k < m_gs_shader->ninput; ++k) {
+      auto& in_io = m_gs_shader->input[k];
+      sfn_log << SfnLog::io << "  against  " <<  k << " name=" << in_io.name<< " sid=" << in_io.sid << "\n";
+
+      if (in_io.name == out_io.name &&
+          in_io.sid == out_io.sid) {
+         ring_offset = in_io.ring_offset;
+         break;
+      }
+   }
+
+   if (out_var->data.location == VARYING_SLOT_VIEWPORT)
+      return true;
+
+   if (ring_offset == -1) {
+      sfn_log << SfnLog::err << "VS defines output at "
+              << out_var->data.driver_location << "name=" << out_io.name
+              << " sid=" << out_io.sid << " that is not consumed as GS input\n";
+      return true;
+   }
+
+   uint32_t write_mask =  (1 << instr->num_components) - 1;
+
+   std::unique_ptr<GPRVector> value(vec_from_nir_with_fetch_constant(instr->src[1], write_mask,
+                                    swizzle_from_mask(instr->num_components)));
+
+   auto ir = new MemRingOutIntruction(cf_mem_ring, mem_write, *value,
+                                      ring_offset >> 2, 4, PValue());
+   emit_export_instruction(ir);
+
+   sh_info().output[out_var->data.driver_location].write_mask |= write_mask;
+   if (out_var->data.location == VARYING_SLOT_CLIP_DIST0 ||
+       out_var->data.location == VARYING_SLOT_CLIP_DIST1)
+      m_num_clip_dist += 4;
+
+   return true;
+}
+
+void VertexShaderFromNirForGS::finalize_exports()
+{
+}
+
 }
index e7be40f7ceac616c6204dff61fb00f28d334617f..4b7fb75c7d73933cf75a5fe6ed4b861c7aeef63f 100644 (file)
@@ -81,6 +81,16 @@ private:
    void finalize_exports() override;
 };
 
+class VertexShaderFromNirForGS :  public VertexShaderFromNir {
+public:
+   VertexShaderFromNirForGS(r600_pipe_shader *sh, r600_pipe_shader_selector &sel,
+                            const r600_shader_key &key, const r600_shader *gs_shader);
+   bool do_emit_store_deref(const nir_variable *out_var, nir_intrinsic_instr* instr) override;
+   void finalize_exports() override;
+
+   const r600_shader *m_gs_shader;
+};
+
 }
 
 #endif