aco: Implement bvh64_intersect_ray_amd intrinsic.

Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10818>

diff --git a/src/amd/compiler/aco_assembler.cpp b/src/amd/compiler/aco_assembler.cpp
index 19b63d1..ce42f59 100644 (file)
--- a/src/amd/compiler/aco_assembler.cpp
+++ b/src/amd/compiler/aco_assembler.cpp
@@ -427,7 +427,8 @@ void emit_instruction(asm_context& ctx, std::vector<uint32_t>& out, Instruction*
       MIMG_instruction& mimg = instr->mimg();
       uint32_t encoding = (0b111100 << 26);
       encoding |= mimg.slc ? 1 << 25 : 0;
-      encoding |= opcode << 18;
+      encoding |= (opcode & 0x7f) << 18;
+      encoding |= (opcode >> 7) & 1;
       encoding |= mimg.lwe ? 1 << 17 : 0;
       encoding |= mimg.tfe ? 1 << 16 : 0;
       encoding |= mimg.glc ? 1 << 13 : 0;

diff --git a/src/amd/compiler/aco_instruction_selection.cpp b/src/amd/compiler/aco_instruction_selection.cpp
index 0794ca3..496dafb 100644 (file)
--- a/src/amd/compiler/aco_instruction_selection.cpp
+++ b/src/amd/compiler/aco_instruction_selection.cpp
@@ -5598,6 +5598,39 @@ static MIMG_instruction *emit_mimg(Builder& bld, aco_opcode op,
    return res;
 }
 
+void visit_bvh64_intersect_ray_amd(isel_context *ctx, nir_intrinsic_instr *instr)
+{
+   Builder bld(ctx->program, ctx->block);
+   Temp dst = get_ssa_temp(ctx, &instr->dest.ssa);
+   Temp resource = get_ssa_temp(ctx, instr->src[0].ssa);
+   Temp node = get_ssa_temp(ctx, instr->src[1].ssa);
+   Temp tmax = get_ssa_temp(ctx, instr->src[2].ssa);
+   Temp origin = get_ssa_temp(ctx, instr->src[3].ssa);
+   Temp dir = get_ssa_temp(ctx, instr->src[4].ssa);
+   Temp inv_dir = get_ssa_temp(ctx, instr->src[5].ssa);
+
+   std::vector<Temp> args;
+   args.push_back(emit_extract_vector(ctx, node, 0, v1));
+   args.push_back(emit_extract_vector(ctx, node, 1, v1));
+   args.push_back(as_vgpr(ctx, tmax));
+   args.push_back(emit_extract_vector(ctx, origin, 0, v1));
+   args.push_back(emit_extract_vector(ctx, origin, 1, v1));
+   args.push_back(emit_extract_vector(ctx, origin, 2, v1));
+   args.push_back(emit_extract_vector(ctx, dir, 0, v1));
+   args.push_back(emit_extract_vector(ctx, dir, 1, v1));
+   args.push_back(emit_extract_vector(ctx, dir, 2, v1));
+   args.push_back(emit_extract_vector(ctx, inv_dir, 0, v1));
+   args.push_back(emit_extract_vector(ctx, inv_dir, 1, v1));
+   args.push_back(emit_extract_vector(ctx, inv_dir, 2, v1));
+
+   MIMG_instruction *mimg = emit_mimg(bld, aco_opcode::image_bvh64_intersect_ray,
+                                      Definition(dst), resource, Operand(s4), args);
+   mimg->dim = ac_image_1d;
+   mimg->dmask = 0xf;
+   mimg->unrm = true;
+   mimg->r128 = true;
+}
+
 /* Adjust the sample index according to FMASK.
  *
  * For uncompressed MSAA surfaces, FMASK should return 0x76543210,
@@ -8515,6 +8548,9 @@ void visit_intrinsic(isel_context *ctx, nir_intrinsic_instr *instr)
    case nir_intrinsic_load_sbt_amd:
       visit_load_sbt_amd(ctx, instr);
       break;
+   case nir_intrinsic_bvh64_intersect_ray_amd:
+      visit_bvh64_intersect_ray_amd(ctx, instr);
+      break;
    default:
       isel_err(&instr->instr, "Unimplemented intrinsic instr");
       abort();

diff --git a/src/amd/compiler/aco_instruction_selection_setup.cpp b/src/amd/compiler/aco_instruction_selection_setup.cpp
index 1e544a9..9c3dde6 100644 (file)
--- a/src/amd/compiler/aco_instruction_selection_setup.cpp
+++ b/src/amd/compiler/aco_instruction_selection_setup.cpp
@@ -805,6 +805,7 @@ void init_context(isel_context *ctx, nir_shader *shader)
                   case nir_intrinsic_load_packed_passthrough_primitive_amd:
                   case nir_intrinsic_gds_atomic_add_amd:
                   case nir_intrinsic_load_sbt_amd:
+                  case nir_intrinsic_bvh64_intersect_ray_amd:
                      type = RegType::vgpr;
                      break;
                   case nir_intrinsic_shuffle:

diff --git a/src/amd/compiler/aco_opcodes.py b/src/amd/compiler/aco_opcodes.py
index abe852c..a28f1d5 100644 (file)
--- a/src/amd/compiler/aco_opcodes.py
+++ b/src/amd/compiler/aco_opcodes.py
@@ -1509,6 +1509,7 @@ IMAGE_GATHER4 = {
 for (code, name) in IMAGE_GATHER4:
    opcode(name, code, code, code, Format.MIMG, InstrClass.VMem)
 
+opcode("image_bvh64_intersect_ray", -1, -1, 231, Format.MIMG, InstrClass.VMem)
 
 FLAT = {
    #GFX7, GFX8_9, GFX10