ac/nir: add ac_nir_lower_legacy_gs

For legacy (non-NGG) GS to lower outputs to memory stores and add
shader query when required.

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Signed-off-by: Qiang Yu <yuq825@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20158>

diff --git a/src/amd/common/ac_nir.c b/src/amd/common/ac_nir.c
index 8e738d5..e447b09 100644 (file)
--- a/src/amd/common/ac_nir.c
+++ b/src/amd/common/ac_nir.c
@@ -438,3 +438,231 @@ ac_nir_gs_shader_query(nir_builder *b,
    nir_pop_if(b, if_shader_query);
    return true;
 }
+
+typedef struct {
+   nir_ssa_def *outputs[64][4];
+   nir_ssa_def *outputs_16bit_lo[16][4];
+   nir_ssa_def *outputs_16bit_hi[16][4];
+
+   ac_nir_gs_output_info *info;
+
+   nir_ssa_def *vertex_count[4];
+   nir_ssa_def *primitive_count[4];
+} lower_legacy_gs_state;
+
+static bool
+lower_legacy_gs_store_output(nir_builder *b, nir_intrinsic_instr *intrin,
+                             lower_legacy_gs_state *s)
+{
+   /* Assume:
+    * - the shader used nir_lower_io_to_temporaries
+    * - 64-bit outputs are lowered
+    * - no indirect indexing is present
+    */
+   assert(nir_src_is_const(intrin->src[1]) && !nir_src_as_uint(intrin->src[1]));
+
+   b->cursor = nir_before_instr(&intrin->instr);
+
+   unsigned component = nir_intrinsic_component(intrin);
+   unsigned write_mask = nir_intrinsic_write_mask(intrin);
+   nir_io_semantics sem = nir_intrinsic_io_semantics(intrin);
+
+   nir_ssa_def **outputs;
+   if (sem.location < VARYING_SLOT_VAR0_16BIT) {
+      outputs = s->outputs[sem.location];
+   } else {
+      unsigned index = sem.location - VARYING_SLOT_VAR0_16BIT;
+      if (sem.high_16bits)
+         outputs = s->outputs_16bit_hi[index];
+      else
+         outputs = s->outputs_16bit_lo[index];
+   }
+
+   nir_ssa_def *store_val = intrin->src[0].ssa;
+   /* 64bit output has been lowered to 32bit */
+   assert(store_val->bit_size <= 32);
+
+   u_foreach_bit (i, write_mask) {
+      unsigned comp = component + i;
+      outputs[comp] = nir_channel(b, store_val, i);
+   }
+
+   nir_instr_remove(&intrin->instr);
+   return true;
+}
+
+static bool
+lower_legacy_gs_emit_vertex_with_counter(nir_builder *b, nir_intrinsic_instr *intrin,
+                                         lower_legacy_gs_state *s)
+{
+   b->cursor = nir_before_instr(&intrin->instr);
+
+   unsigned stream = nir_intrinsic_stream_id(intrin);
+   nir_ssa_def *vtxidx = intrin->src[0].ssa;
+
+   nir_ssa_def *gsvs_ring = nir_load_ring_gsvs_amd(b, .stream_id = stream);
+   nir_ssa_def *soffset = nir_load_ring_gs2vs_offset_amd(b);
+
+   unsigned offset = 0;
+   u_foreach_bit64 (i, b->shader->info.outputs_written) {
+      for (unsigned j = 0; j < 4; j++) {
+         nir_ssa_def *output = s->outputs[i][j];
+         /* Next vertex emit need a new value, reset all outputs. */
+         s->outputs[i][j] = NULL;
+
+         if (!(s->info->usage_mask[i] & (1 << j)) ||
+             ((s->info->streams[i] >> (j * 2)) & 0x3) != stream)
+            continue;
+
+         unsigned base = offset * b->shader->info.gs.vertices_out;
+         offset++;
+
+         /* no one set this output, skip the buffer store */
+         if (!output)
+            continue;
+
+         nir_ssa_def *voffset = nir_iadd_imm(b, vtxidx, base);
+         voffset = nir_ishl_imm(b, voffset, 2);
+
+         /* extend 8/16 bit to 32 bit, 64 bit has been lowered */
+         nir_ssa_def *data = nir_u2uN(b, output, 32);
+
+         nir_store_buffer_amd(b, data, gsvs_ring, voffset, soffset, nir_imm_int(b, 0),
+                              .is_swizzled = true, .slc_amd = true,
+                              .access = ACCESS_COHERENT,
+                              /* For ACO to not reorder this store around EmitVertex/EndPrimitve */
+                              .memory_modes = nir_var_shader_out);
+      }
+   }
+
+   u_foreach_bit (i, b->shader->info.outputs_written_16bit) {
+      for (unsigned j = 0; j < 4; j++) {
+         nir_ssa_def *output_lo = s->outputs_16bit_lo[i][j];
+         nir_ssa_def *output_hi = s->outputs_16bit_hi[i][j];
+         /* Next vertex emit need a new value, reset all outputs. */
+         s->outputs_16bit_lo[i][j] = NULL;
+         s->outputs_16bit_hi[i][j] = NULL;
+
+         bool has_lo_16bit = (s->info->usage_mask_16bit_lo[i] & (1 << j)) &&
+            ((s->info->streams_16bit_lo[i] >> (j * 2)) & 0x3) == stream;
+         bool has_hi_16bit = (s->info->usage_mask_16bit_hi[i] & (1 << j)) &&
+            ((s->info->streams_16bit_hi[i] >> (j * 2)) & 0x3) == stream;
+         if (!has_lo_16bit && !has_hi_16bit)
+            continue;
+
+         unsigned base = offset * b->shader->info.gs.vertices_out;
+         offset++;
+
+         bool has_lo_16bit_out = has_lo_16bit && output_lo;
+         bool has_hi_16bit_out = has_hi_16bit && output_hi;
+
+         /* no one set needed output, skip the buffer store */
+         if (!has_lo_16bit_out && !has_hi_16bit_out)
+            continue;
+
+         if (!has_lo_16bit_out)
+            output_lo = nir_ssa_undef(b, 1, 16);
+
+         if (!has_hi_16bit_out)
+            output_hi = nir_ssa_undef(b, 1, 16);
+
+         nir_ssa_def *voffset = nir_iadd_imm(b, vtxidx, base);
+         voffset = nir_ishl_imm(b, voffset, 2);
+
+         nir_store_buffer_amd(b, nir_pack_32_2x16_split(b, output_lo, output_hi),
+                              gsvs_ring, voffset, soffset, nir_imm_int(b, 0),
+                              .is_swizzled = true, .slc_amd = true,
+                              .access = ACCESS_COHERENT,
+                              /* For ACO to not reorder this store around EmitVertex/EndPrimitve */
+                              .memory_modes = nir_var_shader_out);
+      }
+   }
+
+   /* Keep this instruction to signal vertex emission. */
+
+   return true;
+}
+
+static bool
+lower_legacy_gs_set_vertex_and_primitive_count(nir_builder *b, nir_intrinsic_instr *intrin,
+                                               lower_legacy_gs_state *s)
+{
+   b->cursor = nir_before_instr(&intrin->instr);
+
+   unsigned stream = nir_intrinsic_stream_id(intrin);
+
+   s->vertex_count[stream] = intrin->src[0].ssa;
+   s->primitive_count[stream] = intrin->src[1].ssa;
+
+   nir_instr_remove(&intrin->instr);
+   return true;
+}
+
+static bool
+lower_legacy_gs_intrinsic(nir_builder *b, nir_instr *instr, void *state)
+{
+   lower_legacy_gs_state *s = (lower_legacy_gs_state *) state;
+
+   if (instr->type != nir_instr_type_intrinsic)
+      return false;
+
+   nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
+
+   if (intrin->intrinsic == nir_intrinsic_store_output)
+      return lower_legacy_gs_store_output(b, intrin, s);
+   else if (intrin->intrinsic == nir_intrinsic_emit_vertex_with_counter)
+      return lower_legacy_gs_emit_vertex_with_counter(b, intrin, s);
+   else if (intrin->intrinsic == nir_intrinsic_set_vertex_and_primitive_count)
+      return lower_legacy_gs_set_vertex_and_primitive_count(b, intrin, s);
+
+   return false;
+}
+
+void
+ac_nir_lower_legacy_gs(nir_shader *nir,
+                       bool has_gen_prim_query,
+                       bool has_pipeline_stats_query,
+                       ac_nir_gs_output_info *output_info)
+{
+   lower_legacy_gs_state s = {
+      .info = output_info,
+   };
+
+   unsigned num_vertices_per_primitive = 0;
+   switch (nir->info.gs.output_primitive) {
+   case SHADER_PRIM_POINTS:
+      num_vertices_per_primitive = 1;
+      break;
+   case SHADER_PRIM_LINE_STRIP:
+      num_vertices_per_primitive = 2;
+      break;
+   case SHADER_PRIM_TRIANGLE_STRIP:
+      num_vertices_per_primitive = 3;
+      break;
+   default:
+      unreachable("Invalid GS output primitive.");
+      break;
+   }
+
+   nir_shader_instructions_pass(nir, lower_legacy_gs_intrinsic,
+                                nir_metadata_block_index | nir_metadata_dominance, &s);
+
+   nir_function_impl *impl = nir_shader_get_entrypoint(nir);
+
+   nir_builder builder;
+   nir_builder *b = &builder;
+   nir_builder_init(b, impl);
+
+   b->cursor = nir_after_cf_list(&impl->body);
+
+   /* Emit shader query for mix use legacy/NGG GS */
+   bool progress = ac_nir_gs_shader_query(b,
+                                          has_gen_prim_query,
+                                          has_pipeline_stats_query,
+                                          num_vertices_per_primitive,
+                                          64,
+                                          s.vertex_count,
+                                          s.primitive_count);
+   if (progress)
+      nir_metadata_preserve(impl, nir_metadata_none);
+}

diff --git a/src/amd/common/ac_nir.h b/src/amd/common/ac_nir.h
index e32b694..3f5752a 100644 (file)
--- a/src/amd/common/ac_nir.h
+++ b/src/amd/common/ac_nir.h
@@ -187,6 +187,16 @@ ac_nir_lower_global_access(nir_shader *shader);
 
 bool ac_nir_lower_resinfo(nir_shader *nir, enum amd_gfx_level gfx_level);
 
+typedef struct ac_nir_gs_output_info {
+   const uint8_t *streams;
+   const uint8_t *streams_16bit_lo;
+   const uint8_t *streams_16bit_hi;
+
+   const uint8_t *usage_mask;
+   const uint8_t *usage_mask_16bit_lo;
+   const uint8_t *usage_mask_16bit_hi;
+} ac_nir_gs_output_info;
+
 nir_shader *
 ac_nir_create_gs_copy_shader(const nir_shader *gs_nir,
                              bool disable_streamout,
@@ -207,6 +217,12 @@ ac_nir_gs_shader_query(nir_builder *b,
                        nir_ssa_def *vertex_count[4],
                        nir_ssa_def *primitive_count[4]);
 
+void
+ac_nir_lower_legacy_gs(nir_shader *nir,
+                       bool has_gen_prim_query,
+                       bool has_pipeline_stats_query,
+                       ac_nir_gs_output_info *output_info);
+
 #ifdef __cplusplus
 }
 #endif