aco,nir: add dpp16_shift_amd intrinsic

Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24650>

src/amd/compiler/aco_instruction_selection.cpp

@@ -8594,6 +8594,20 @@ visit_intrinsic(isel_context* ctx, nir_intrinsic_instr* instr)
       set_wqm(ctx);
       break;
    }
+   case nir_intrinsic_dpp16_shift_amd: {
+      Temp src = as_vgpr(ctx, get_ssa_temp(ctx, instr->src[0].ssa));
+      Temp dst = get_ssa_temp(ctx, &instr->def);
+      int delta = nir_intrinsic_base(instr);
+      assert(delta >= -15 && delta <= 15 && delta != 0);
+      assert(instr->def.bit_size != 1 && instr->def.bit_size < 64);
+      assert(ctx->options->gfx_level >= GFX8);
+
+      uint16_t dpp_ctrl = delta < 0 ? dpp_row_sr(-delta) : dpp_row_sl(delta);
+      bld.vop1_dpp(aco_opcode::v_mov_b32, Definition(dst), src, dpp_ctrl);
+
+      set_wqm(ctx);
+      break;
+   }
    case nir_intrinsic_quad_broadcast:
    case nir_intrinsic_quad_swap_horizontal:
    case nir_intrinsic_quad_swap_vertical:

src/amd/compiler/aco_instruction_selection_setup.cpp

@@ -484,6 +484,7 @@ init_context(isel_context* ctx, nir_shader* shader)
                case nir_intrinsic_write_invocation_amd:
                case nir_intrinsic_mbcnt_amd:
                case nir_intrinsic_lane_permute_16_amd:
+               case nir_intrinsic_dpp16_shift_amd:
                case nir_intrinsic_load_instance_id:
                case nir_intrinsic_ssbo_atomic:
                case nir_intrinsic_ssbo_atomic_swap:

src/compiler/nir/nir_divergence_analysis.c

@@ -705,6 +705,7 @@ visit_intrinsic(nir_intrinsic_instr *instr, struct divergence_state *state)
    case nir_intrinsic_write_invocation_amd:
    case nir_intrinsic_mbcnt_amd:
    case nir_intrinsic_lane_permute_16_amd:
+   case nir_intrinsic_dpp16_shift_amd:
    case nir_intrinsic_elect:
    case nir_intrinsic_elect_any_ir3:
    case nir_intrinsic_load_tlb_color_brcm:

src/compiler/nir/nir_intrinsics.py

@@ -545,6 +545,11 @@ intrinsic("write_invocation_amd", src_comp=[0, 0, 1], dest_comp=0, bit_sizes=src
 intrinsic("mbcnt_amd", src_comp=[1, 1], dest_comp=1, bit_sizes=[32], flags=[CAN_ELIMINATE])
 # Compiled to v_permlane16_b32. src = [ value, lanesel_lo, lanesel_hi ]
 intrinsic("lane_permute_16_amd", src_comp=[1, 1, 1], dest_comp=1, bit_sizes=[32], flags=[CAN_ELIMINATE])
+# subgroup shuffle up/down with cluster size 16.
+# base in [-15, -1]: DPP_ROW_SR
+# base in [  1, 15]: DPP_ROW_SL, otherwise invalid.
+# Returns zero for invocations that try to read out of bounds
+intrinsic("dpp16_shift_amd", src_comp=[0], dest_comp=0, bit_sizes=src0, indices=[BASE], flags=[CAN_ELIMINATE])
 
 # Basic Geometry Shader intrinsics.
 #