nir,ac/llvm,radeonsi: replace nir_buffer_atomic_add_amd with ssbo atomic

Now that radeonsi support pass desc to ssbo atomic ops,
we can use ssbo atomic instead. aco does not implement
nir_buffer_atomic_add either.

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/23096>

src/amd/llvm/ac_nir_to_llvm.c

@@ -3746,29 +3746,6 @@ static bool visit_intrinsic(struct ac_nir_context *ctx, nir_intrinsic_instr *ins
       ac_build_atomic_rmw(&ctx->ac, LLVMAtomicRMWBinOpAdd, gds_base, store_val, "workgroup-one-as");
       break;
    }
-   case nir_intrinsic_buffer_atomic_add_amd: {
-      LLVMValueRef desc = get_src(ctx, instr->src[0]);
-      LLVMValueRef data = get_src(ctx, instr->src[1]);
-      unsigned base = nir_intrinsic_base(instr);
-      LLVMTypeRef return_type = LLVMTypeOf(data);
-      unsigned cache_flags =
-         ac_get_hw_cache_flags(ctx->ac.gfx_level,
-			       ac_get_mem_access_flags(instr) | ACCESS_TYPE_ATOMIC).value;
-
-      LLVMValueRef args[] = {
-         data, desc,
-         LLVMConstInt(ctx->ac.i32, base, false),
-         ctx->ac.i32_0, /* soffset */
-         LLVMConstInt(ctx->ac.i32, cache_flags, 0),
-      };
-
-      char name[64], type[8];
-      ac_build_type_name_for_intr(return_type, type, sizeof(type));
-      snprintf(name, sizeof(name), "llvm.amdgcn.raw.buffer.atomic.add.%s", type);
-
-      result = ac_build_intrinsic(&ctx->ac, name, return_type, args, 5, 0);
-      break;
-   }
    case nir_intrinsic_elect:
       result = LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ, visit_first_invocation(ctx),
                              ac_get_thread_id(&ctx->ac), "");

src/compiler/nir/nir_divergence_analysis.c

@@ -558,7 +558,6 @@ visit_intrinsic(nir_shader *shader, nir_intrinsic_instr *instr)
    case nir_intrinsic_load_packed_passthrough_primitive_amd:
    case nir_intrinsic_load_initial_edgeflags_amd:
    case nir_intrinsic_gds_atomic_add_amd:
-   case nir_intrinsic_buffer_atomic_add_amd:
    case nir_intrinsic_load_rt_arg_scratch_offset_amd:
    case nir_intrinsic_load_intersection_opaque_amd:
    case nir_intrinsic_load_vector_arg_amd:

src/compiler/nir/nir_intrinsics.py

@@ -1363,9 +1363,6 @@ store("global_amd", [1, 1], indices=[BASE, ACCESS, ALIGN_MUL, ALIGN_OFFSET, WRIT
 # Same as shared_atomic_add, but with GDS. src[] = {store_val, gds_addr, m0}
 intrinsic("gds_atomic_add_amd",  src_comp=[1, 1, 1], dest_comp=1, indices=[BASE])
 
-# src[] = { descriptor, add_value }
-intrinsic("buffer_atomic_add_amd", src_comp=[4, 1], dest_comp=1, indices=[BASE])
-
 # src[] = { sample_id, num_samples }
 intrinsic("load_sample_positions_amd", src_comp=[1, 1], dest_comp=2, flags=[CAN_ELIMINATE, CAN_REORDER])
 

src/gallium/drivers/radeonsi/si_nir_lower_abi.c

@@ -401,7 +401,8 @@ static bool lower_intrinsic(nir_builder *b, nir_instr *instr, struct lower_abi_s
       unsigned offset = si_query_pipestat_end_dw_offset(sel->screen, index) * 4;
 
       nir_ssa_def *count = intrin->src[0].ssa;
-      nir_buffer_atomic_add_amd(b, 32, buf, count, .base = offset);
+      nir_ssbo_atomic(b, 32, buf, nir_imm_int(b, offset), count,
+                      .atomic_op = nir_atomic_op_iadd);
       break;
    }
    case nir_intrinsic_atomic_add_gen_prim_count_amd:
@@ -414,7 +415,8 @@ static bool lower_intrinsic(nir_builder *b, nir_instr *instr, struct lower_abi_s
          offsetof(struct gfx10_sh_query_buffer_mem, stream[stream].emitted_primitives);
 
       nir_ssa_def *prim_count = intrin->src[0].ssa;
-      nir_buffer_atomic_add_amd(b, 32, buf, prim_count, .base = offset);
+      nir_ssbo_atomic(b, 32, buf, nir_imm_int(b, offset), prim_count,
+                      .atomic_op = nir_atomic_op_iadd);
       break;
    }
    case nir_intrinsic_load_ring_attr_amd: