nir, ac, aco: Add index src to load_buffer_amd/store_buffer_amd.

Also modify all existing uses to pass a zero to this new src.

Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Georg Lehmann <dadschoorse@gmail.com>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Alyssa Rosenzweig <alyssa@collabora.com> (nir)
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17551>

diff --git a/src/amd/common/ac_nir_lower_esgs_io_to_mem.c b/src/amd/common/ac_nir_lower_esgs_io_to_mem.c
index cad54d1..9a483b3 100644 (file)
--- a/src/amd/common/ac_nir_lower_esgs_io_to_mem.c
+++ b/src/amd/common/ac_nir_lower_esgs_io_to_mem.c
@@ -71,13 +71,15 @@ emit_split_buffer_load(nir_builder *b, nir_ssa_def *desc, nir_ssa_def *v_off, ni
       full_dwords++;
    }
 
+   nir_ssa_def *zero = nir_imm_int(b, 0);
+
    for (unsigned i = 0; i < full_dwords; ++i)
-      comps[i] = nir_load_buffer_amd(b, 1, 32, desc, v_off, s_off,
+      comps[i] = nir_load_buffer_amd(b, 1, 32, desc, v_off, s_off, zero,
                                      .base = component_stride * i, .memory_modes = nir_var_shader_in,
                                      .access = ACCESS_COHERENT);
 
    if (remaining_bytes)
-      comps[full_dwords] = nir_load_buffer_amd(b, 1, remaining_bytes * 8, desc, v_off, s_off,
+      comps[full_dwords] = nir_load_buffer_amd(b, 1, remaining_bytes * 8, desc, v_off, s_off, zero,
                                                .base = component_stride * full_dwords,
                                                .memory_modes = nir_var_shader_in,
                                                .access = ACCESS_COHERENT);
@@ -90,6 +92,8 @@ emit_split_buffer_store(nir_builder *b, nir_ssa_def *d, nir_ssa_def *desc, nir_s
                         unsigned component_stride, unsigned num_components, unsigned bit_size,
                         unsigned writemask, bool swizzled, bool slc)
 {
+   nir_ssa_def *zero = nir_imm_int(b, 0);
+
    while (writemask) {
       int start, count;
       u_bit_scan_consecutive_range(&writemask, &start, &count);
@@ -106,7 +110,7 @@ emit_split_buffer_store(nir_builder *b, nir_ssa_def *d, nir_ssa_def *desc, nir_s
             store_bytes = MIN2(store_bytes, 2);
 
          nir_ssa_def *store_val = nir_extract_bits(b, &d, 1, start_byte * 8u, 1, store_bytes * 8u);
-         nir_store_buffer_amd(b, store_val, desc, v_off, s_off, .is_swizzled = swizzled, .slc_amd = slc,
+         nir_store_buffer_amd(b, store_val, desc, v_off, s_off, zero, .is_swizzled = swizzled, .slc_amd = slc,
                               .base = start_byte, .memory_modes = nir_var_shader_out, .access = ACCESS_COHERENT);
 
          start_byte += store_bytes;

diff --git a/src/amd/common/ac_nir_lower_ngg.c b/src/amd/common/ac_nir_lower_ngg.c
index 6c79575..7c1ad54 100644 (file)
--- a/src/amd/common/ac_nir_lower_ngg.c
+++ b/src/amd/common/ac_nir_lower_ngg.c
@@ -1700,9 +1700,10 @@ ngg_build_streamout_vertex(nir_builder *b, nir_xfb_info *info,
       nir_ssa_def *out_data =
          nir_load_shared(b, count, 32, vtx_lds_addr, .base = offset);
 
+      nir_ssa_def *zero = nir_imm_int(b, 0);
       nir_store_buffer_amd(b, out_data, so_buffer[out->buffer],
                            vtx_buffer_offsets[out->buffer],
-                           nir_imm_int(b, 0),
+                           zero, zero,
                            .base = out->offset,
                            .slc_amd = true);
    }
@@ -3188,7 +3189,8 @@ ms_store_arrayed_output_intrin(nir_builder *b,
    } else if (out_mode == ms_out_mode_vram) {
       nir_ssa_def *ring = nir_load_ring_mesh_scratch_amd(b);
       nir_ssa_def *off = nir_load_ring_mesh_scratch_offset_amd(b);
-      nir_store_buffer_amd(b, store_val, ring, addr, off,
+      nir_ssa_def *zero = nir_imm_int(b, 0);
+      nir_store_buffer_amd(b, store_val, ring, addr, off, zero,
                            .base = const_off,
                            .write_mask = write_mask,
                            .memory_modes = nir_var_shader_out,
@@ -3242,7 +3244,8 @@ ms_load_arrayed_output(nir_builder *b,
    } else if (out_mode == ms_out_mode_vram) {
       nir_ssa_def *ring = nir_load_ring_mesh_scratch_amd(b);
       nir_ssa_def *off = nir_load_ring_mesh_scratch_offset_amd(b);
-      return nir_load_buffer_amd(b, num_components, load_bit_size, ring, addr, off,
+      nir_ssa_def *zero = nir_imm_int(b, 0);
+      return nir_load_buffer_amd(b, num_components, load_bit_size, ring, addr, off, zero,
                                  .base = const_off,
                                  .memory_modes = nir_var_shader_out,
                                  .access = ACCESS_COHERENT);

diff --git a/src/amd/common/ac_nir_lower_taskmesh_io_to_mem.c b/src/amd/common/ac_nir_lower_taskmesh_io_to_mem.c
index 17cba47..2808248 100644 (file)
--- a/src/amd/common/ac_nir_lower_taskmesh_io_to_mem.c
+++ b/src/amd/common/ac_nir_lower_taskmesh_io_to_mem.c
@@ -240,8 +240,9 @@ task_write_draw_ring(nir_builder *b,
    nir_ssa_def *ring = nir_load_ring_task_draw_amd(b);
    nir_ssa_def *scalar_off = nir_imul_imm(b, ptr, s->draw_entry_bytes);
    nir_ssa_def *vector_off = nir_imm_int(b, 0);
+   nir_ssa_def *zero = nir_imm_int(b, 0);
 
-   nir_store_buffer_amd(b, store_val, ring, vector_off, scalar_off,
+   nir_store_buffer_amd(b, store_val, ring, vector_off, scalar_off, zero,
                         .base = const_off, .memory_modes = nir_var_shader_out,
                         .access = ACCESS_COHERENT);
 }
@@ -305,8 +306,9 @@ lower_task_payload_store(nir_builder *b,
    nir_ssa_def *ring = nir_load_ring_task_payload_amd(b);
    nir_ssa_def *ptr = task_ring_entry_index(b, s);
    nir_ssa_def *ring_off = nir_imul_imm(b, ptr, s->payload_entry_bytes);
+   nir_ssa_def *zero = nir_imm_int(b, 0);
 
-   nir_store_buffer_amd(b, store_val, ring, addr, ring_off, .base = base,
+   nir_store_buffer_amd(b, store_val, ring, addr, ring_off, zero, .base = base,
                         .write_mask = write_mask,
                         .memory_modes = nir_var_mem_task_payload,
                         .access = ACCESS_COHERENT);
@@ -331,8 +333,9 @@ lower_taskmesh_payload_load(nir_builder *b,
    nir_ssa_def *addr = intrin->src[0].ssa;
    nir_ssa_def *ring = nir_load_ring_task_payload_amd(b);
    nir_ssa_def *ring_off = nir_imul_imm(b, ptr, s->payload_entry_bytes);
+   nir_ssa_def *zero = nir_imm_int(b, 0);
 
-   return nir_load_buffer_amd(b, num_components, bit_size, ring, addr, ring_off, .base = base,
+   return nir_load_buffer_amd(b, num_components, bit_size, ring, addr, ring_off, zero, .base = base,
                               .memory_modes = nir_var_mem_task_payload,
                               .access = ACCESS_COHERENT);
 }

diff --git a/src/amd/common/ac_nir_lower_tess_io_to_mem.c b/src/amd/common/ac_nir_lower_tess_io_to_mem.c
index 198ff52..0dfe8ab 100644 (file)
--- a/src/amd/common/ac_nir_lower_tess_io_to_mem.c
+++ b/src/amd/common/ac_nir_lower_tess_io_to_mem.c
@@ -442,7 +442,8 @@ lower_hs_output_store(nir_builder *b,
 
       nir_ssa_def *hs_ring_tess_offchip = nir_load_ring_tess_offchip_amd(b);
       nir_ssa_def *offchip_offset = nir_load_ring_tess_offchip_offset_amd(b);
-      nir_store_buffer_amd(b, store_val, hs_ring_tess_offchip, vmem_off, offchip_offset,
+      nir_ssa_def *zero = nir_imm_int(b, 0);
+      nir_store_buffer_amd(b, store_val, hs_ring_tess_offchip, vmem_off, offchip_offset, zero,
                            .write_mask = write_mask, .memory_modes = nir_var_shader_out,
                            .access = ACCESS_COHERENT);
    }
@@ -587,6 +588,7 @@ hs_emit_write_tess_factors(nir_shader *shader,
                                                       .align_mul = 16u, .align_offset = st->tcs_tess_lvl_in_loc % 16u)
                                     : NULL;
 
+   nir_ssa_def *zero = nir_imm_int(b, 0);
    nir_ssa_def *rel_patch_id = nir_load_tess_rel_patch_id_amd(b);
    nir_ssa_def *tess_factors_base = nir_load_ring_tess_factors_offset_amd(b);
    nir_ssa_def *tess_factors_offset = nir_imul_imm(b, rel_patch_id, (inner_comps + outer_comps) * 4u);
@@ -596,7 +598,7 @@ hs_emit_write_tess_factors(nir_shader *shader,
       /* Store the dynamic HS control word. */
       nir_if *rel_patch_id_zero = nir_push_if(b, nir_ieq_imm(b, rel_patch_id, 0));
       nir_ssa_def *ctrlw = nir_imm_int(b, 0x80000000u);
-      nir_store_buffer_amd(b, ctrlw, tessfactor_ring, nir_imm_zero(b, 1, 32), tess_factors_base,
+      nir_store_buffer_amd(b, ctrlw, tessfactor_ring, zero, tess_factors_base, zero,
                            .access = ACCESS_COHERENT);
       tess_factors_const_offset += 4;
       nir_pop_if(b, rel_patch_id_zero);
@@ -606,17 +608,17 @@ hs_emit_write_tess_factors(nir_shader *shader,
    if (shader->info.tess._primitive_mode == TESS_PRIMITIVE_ISOLINES) {
       /* LINES reversal */
       nir_ssa_def *t = nir_vec2(b, nir_channel(b, tessfactors_outer, 1), nir_channel(b, tessfactors_outer, 0));
-      nir_store_buffer_amd(b, t, tessfactor_ring, tess_factors_offset, tess_factors_base,
+      nir_store_buffer_amd(b, t, tessfactor_ring, tess_factors_offset, tess_factors_base, zero,
                            .base = tess_factors_const_offset, .access = ACCESS_COHERENT);
    } else if (shader->info.tess._primitive_mode == TESS_PRIMITIVE_TRIANGLES) {
       nir_ssa_def *t = nir_vec4(b, nir_channel(b, tessfactors_outer, 0), nir_channel(b, tessfactors_outer, 1),
                                 nir_channel(b, tessfactors_outer, 2), nir_channel(b, tessfactors_inner, 0));
-      nir_store_buffer_amd(b, t, tessfactor_ring, tess_factors_offset, tess_factors_base,
+      nir_store_buffer_amd(b, t, tessfactor_ring, tess_factors_offset, tess_factors_base, zero,
                            .base = tess_factors_const_offset, .access = ACCESS_COHERENT);
    } else {
-      nir_store_buffer_amd(b, tessfactors_outer, tessfactor_ring, tess_factors_offset, tess_factors_base,
+      nir_store_buffer_amd(b, tessfactors_outer, tessfactor_ring, tess_factors_offset, tess_factors_base, zero,
                            .base = tess_factors_const_offset, .access = ACCESS_COHERENT);
-      nir_store_buffer_amd(b, tessfactors_inner, tessfactor_ring, tess_factors_offset, tess_factors_base,
+      nir_store_buffer_amd(b, tessfactors_inner, tessfactor_ring, tess_factors_offset, tess_factors_base, zero,
                            .base = tess_factors_const_offset + 4u * outer_comps, .access = ACCESS_COHERENT);
    }
 
@@ -626,12 +628,12 @@ hs_emit_write_tess_factors(nir_shader *shader,
       nir_ssa_def *offchip_offset = nir_load_ring_tess_offchip_offset_amd(b);
 
       nir_ssa_def *vmem_off_outer = hs_per_patch_output_vmem_offset(b, st, NULL, st->tcs_tess_lvl_out_loc);
-      nir_store_buffer_amd(b, tessfactors_outer, hs_ring_tess_offchip, vmem_off_outer, offchip_offset,
+      nir_store_buffer_amd(b, tessfactors_outer, hs_ring_tess_offchip, vmem_off_outer, offchip_offset, zero,
                            .memory_modes = nir_var_shader_out, .access = ACCESS_COHERENT);
 
       if (inner_comps) {
          nir_ssa_def *vmem_off_inner = hs_per_patch_output_vmem_offset(b, st, NULL, st->tcs_tess_lvl_in_loc);
-         nir_store_buffer_amd(b, tessfactors_inner, hs_ring_tess_offchip, vmem_off_inner, offchip_offset,
+         nir_store_buffer_amd(b, tessfactors_inner, hs_ring_tess_offchip, vmem_off_inner, offchip_offset, zero,
                               .memory_modes = nir_var_shader_out, .access = ACCESS_COHERENT);
       }
    }
@@ -655,8 +657,11 @@ lower_tes_input_load(nir_builder *b,
                     ? hs_per_vertex_output_vmem_offset(b, st, intrin)
                     : hs_per_patch_output_vmem_offset(b, st, intrin, 0);
 
+   nir_ssa_def *zero = nir_imm_int(b, 0);
+
    return nir_load_buffer_amd(b, intrin->dest.ssa.num_components,
-                              intrin->dest.ssa.bit_size, offchip_ring, off, offchip_offset,
+                              intrin->dest.ssa.bit_size, offchip_ring,
+                              off, offchip_offset, zero,
                               .access = ACCESS_COHERENT);
 }
 

diff --git a/src/amd/compiler/aco_instruction_selection.cpp b/src/amd/compiler/aco_instruction_selection.cpp
index 4993529..8df3b7d 100644 (file)
--- a/src/amd/compiler/aco_instruction_selection.cpp
+++ b/src/amd/compiler/aco_instruction_selection.cpp
@@ -3991,6 +3991,7 @@ struct LoadEmitInfo {
    unsigned num_components;
    unsigned component_size;
    Temp resource = Temp(0, s1); /* buffer resource or base 64-bit address */
+   Temp idx = Temp(0, v1); /* buffer index */
    unsigned component_stride = 0;
    unsigned const_offset = 0;
    unsigned align_mul = 0;
@@ -4417,6 +4418,14 @@ mubuf_load_callback(Builder& bld, const LoadEmitInfo& info, Temp offset, unsigne
       soffset = Operand(info.soffset);
    }
 
+   bool offen = !vaddr.isUndefined();
+   bool idxen = info.idx.id();
+
+   if (offen && idxen)
+      vaddr = bld.pseudo(aco_opcode::p_create_vector, bld.def(v2), info.idx, vaddr);
+   else if (idxen)
+      vaddr = Operand(info.idx);
+
    unsigned bytes_size = 0;
    aco_opcode op;
    if (bytes_needed == 1 || align_ % 2) {
@@ -4442,7 +4451,8 @@ mubuf_load_callback(Builder& bld, const LoadEmitInfo& info, Temp offset, unsigne
    mubuf->operands[0] = Operand(info.resource);
    mubuf->operands[1] = vaddr;
    mubuf->operands[2] = soffset;
-   mubuf->offen = (offset.type() == RegType::vgpr);
+   mubuf->offen = offen;
+   mubuf->idxen = idxen;
    mubuf->glc = info.glc;
    mubuf->dlc =
       info.glc && (bld.program->gfx_level == GFX10 || bld.program->gfx_level == GFX10_3);
@@ -5078,7 +5088,7 @@ resolve_excess_vmem_const_offset(Builder& bld, Temp& voffset, unsigned const_off
 }
 
 void
-emit_single_mubuf_store(isel_context* ctx, Temp descriptor, Temp voffset, Temp soffset, Temp vdata,
+emit_single_mubuf_store(isel_context* ctx, Temp descriptor, Temp voffset, Temp soffset, Temp idx, Temp vdata,
                         unsigned const_offset, memory_sync_info sync, bool glc, bool slc,
                         bool swizzled)
 {
@@ -5090,20 +5100,30 @@ emit_single_mubuf_store(isel_context* ctx, Temp descriptor, Temp voffset, Temp s
    aco_opcode op = get_buffer_store_op(vdata.bytes());
    const_offset = resolve_excess_vmem_const_offset(bld, voffset, const_offset);
 
-   Operand voffset_op = voffset.id() ? Operand(as_vgpr(ctx, voffset)) : Operand(v1);
+   bool offen = voffset.id();
+   bool idxen = idx.id();
+
    Operand soffset_op = soffset.id() ? Operand(soffset) : Operand::zero();
    glc &= ctx->program->gfx_level < GFX11;
+
+   Operand vaddr_op(v1);
+   if (offen && idxen)
+      vaddr_op = bld.pseudo(aco_opcode::p_create_vector, bld.def(v2), idx, voffset);
+   else if (offen)
+      vaddr_op = Operand(voffset);
+   else if (idxen)
+      vaddr_op = Operand(idx);
+
    Builder::Result r =
-      bld.mubuf(op, Operand(descriptor), voffset_op, soffset_op, Operand(vdata), const_offset,
-                /* offen */ !voffset_op.isUndefined(), /* swizzled */ swizzled,
-                /* idxen*/ false, /* addr64 */ false, /* disable_wqm */ false,
-                /* glc */ glc, /* dlc*/ false, /* slc */ slc);
+      bld.mubuf(op, Operand(descriptor), vaddr_op, soffset_op, Operand(vdata), const_offset,
+                offen, swizzled, idxen, /* addr64 */ false, /* disable_wqm */ false, glc,
+                /* dlc*/ false, slc);
 
    r.instr->mubuf().sync = sync;
 }
 
 void
-store_vmem_mubuf(isel_context* ctx, Temp src, Temp descriptor, Temp voffset, Temp soffset,
+store_vmem_mubuf(isel_context* ctx, Temp src, Temp descriptor, Temp voffset, Temp soffset, Temp idx,
                  unsigned base_const_offset, unsigned elem_size_bytes, unsigned write_mask,
                  bool swizzled, memory_sync_info sync, bool glc, bool slc)
 {
@@ -5120,13 +5140,13 @@ store_vmem_mubuf(isel_context* ctx, Temp src, Temp descriptor, Temp voffset, Tem
 
    for (unsigned i = 0; i < write_count; i++) {
       unsigned const_offset = offsets[i] + base_const_offset;
-      emit_single_mubuf_store(ctx, descriptor, voffset, soffset, write_datas[i], const_offset, sync,
+      emit_single_mubuf_store(ctx, descriptor, voffset, soffset, idx, write_datas[i], const_offset, sync,
                               glc, slc, swizzled);
    }
 }
 
 void
-load_vmem_mubuf(isel_context* ctx, Temp dst, Temp descriptor, Temp voffset, Temp soffset,
+load_vmem_mubuf(isel_context* ctx, Temp dst, Temp descriptor, Temp voffset, Temp soffset, Temp idx,
                 unsigned base_const_offset, unsigned elem_size_bytes, unsigned num_components,
                 unsigned swizzle_element_size, bool glc, bool slc, memory_sync_info sync)
 {
@@ -5136,6 +5156,7 @@ load_vmem_mubuf(isel_context* ctx, Temp dst, Temp descriptor, Temp voffset, Temp
    Builder bld(ctx->program, ctx->block);
 
    LoadEmitInfo info = {Operand(voffset), dst, num_components, elem_size_bytes, descriptor};
+   info.idx = idx;
    info.component_stride = swizzle_element_size;
    info.glc = glc;
    info.slc = slc;
@@ -7175,10 +7196,13 @@ visit_load_buffer(isel_context* ctx, nir_intrinsic_instr* intrin)
 {
    Builder bld(ctx->program, ctx->block);
 
+   bool idxen = !nir_src_is_const(intrin->src[3]) || nir_src_as_uint(intrin->src[3]);
+
    Temp dst = get_ssa_temp(ctx, &intrin->dest.ssa);
    Temp descriptor = bld.as_uniform(get_ssa_temp(ctx, intrin->src[0].ssa));
    Temp v_offset = as_vgpr(ctx, get_ssa_temp(ctx, intrin->src[1].ssa));
    Temp s_offset = bld.as_uniform(get_ssa_temp(ctx, intrin->src[2].ssa));
+   Temp idx = idxen ? as_vgpr(ctx, get_ssa_temp(ctx, intrin->src[3].ssa)) : Temp();
 
    bool swizzled = nir_intrinsic_is_swizzled(intrin);
    bool slc = nir_intrinsic_slc_amd(intrin);
@@ -7192,7 +7216,7 @@ visit_load_buffer(isel_context* ctx, nir_intrinsic_instr* intrin)
    nir_variable_mode mem_mode = nir_intrinsic_memory_modes(intrin);
    memory_sync_info sync(aco_storage_mode_from_nir_mem_mode(mem_mode));
 
-   load_vmem_mubuf(ctx, dst, descriptor, v_offset, s_offset, const_offset, elem_size_bytes,
+   load_vmem_mubuf(ctx, dst, descriptor, v_offset, s_offset, idx, const_offset, elem_size_bytes,
                    num_components, swizzle_element_size, glc, slc, sync);
 }
 
@@ -7201,10 +7225,13 @@ visit_store_buffer(isel_context* ctx, nir_intrinsic_instr* intrin)
 {
    Builder bld(ctx->program, ctx->block);
 
+   bool idxen = !nir_src_is_const(intrin->src[4]) || nir_src_as_uint(intrin->src[4]);
+
    Temp store_src = get_ssa_temp(ctx, intrin->src[0].ssa);
    Temp descriptor = bld.as_uniform(get_ssa_temp(ctx, intrin->src[1].ssa));
    Temp v_offset = as_vgpr(ctx, get_ssa_temp(ctx, intrin->src[2].ssa));
    Temp s_offset = bld.as_uniform(get_ssa_temp(ctx, intrin->src[3].ssa));
+   Temp idx = idxen ? as_vgpr(ctx, get_ssa_temp(ctx, intrin->src[4].ssa)) : Temp();
 
    bool swizzled = nir_intrinsic_is_swizzled(intrin);
    bool glc = nir_intrinsic_access(intrin) & ACCESS_COHERENT;
@@ -7217,7 +7244,7 @@ visit_store_buffer(isel_context* ctx, nir_intrinsic_instr* intrin)
    nir_variable_mode mem_mode = nir_intrinsic_memory_modes(intrin);
    memory_sync_info sync(aco_storage_mode_from_nir_mem_mode(mem_mode));
 
-   store_vmem_mubuf(ctx, store_src, descriptor, v_offset, s_offset, const_offset, elem_size_bytes,
+   store_vmem_mubuf(ctx, store_src, descriptor, v_offset, s_offset, idx, const_offset, elem_size_bytes,
                     write_mask, swizzled, sync, glc, slc);
 }
 
@@ -11942,7 +11969,7 @@ select_gs_copy_shader(Program* program, struct nir_shader* gs_shader, ac_shader_
 
             Temp val = bld.tmp(v1);
             unsigned const_offset = offset * program->info.gs.vertices_out * 16 * 4;
-            load_vmem_mubuf(&ctx, val, gsvs_ring, vtx_offset, Temp(), const_offset, 4, 1, 0u, true,
+            load_vmem_mubuf(&ctx, val, gsvs_ring, vtx_offset, Temp(), Temp(), const_offset, 4, 1, 0, true,
                             true, memory_sync_info());
 
             ctx.outputs.mask[i] |= 1 << j;

diff --git a/src/amd/llvm/ac_nir_to_llvm.c b/src/amd/llvm/ac_nir_to_llvm.c
index 9014513..7da1ff6 100644 (file)
--- a/src/amd/llvm/ac_nir_to_llvm.c
+++ b/src/amd/llvm/ac_nir_to_llvm.c
@@ -4184,9 +4184,12 @@ static bool visit_intrinsic(struct ac_nir_context *ctx, nir_intrinsic_instr *ins
       /* Currently ignored. */
       break;
    case nir_intrinsic_load_buffer_amd: {
+      bool idxen = !nir_src_is_const(instr->src[3]) || nir_src_as_uint(instr->src[3]);
+
       LLVMValueRef descriptor = get_src(ctx, instr->src[0]);
       LLVMValueRef addr_voffset = get_src(ctx, instr->src[1]);
       LLVMValueRef addr_soffset = get_src(ctx, instr->src[2]);
+      LLVMValueRef vidx = idxen ? get_src(ctx, instr->src[3]) : NULL;
       unsigned num_components = instr->dest.ssa.num_components;
       unsigned const_offset = nir_intrinsic_base(instr);
       bool swizzled = nir_intrinsic_is_swizzled(instr);
@@ -4218,16 +4221,20 @@ static bool visit_intrinsic(struct ac_nir_context *ctx, nir_intrinsic_instr *ins
 
       LLVMValueRef voffset = LLVMBuildAdd(ctx->ac.builder, addr_voffset,
                                           LLVMConstInt(ctx->ac.i32, const_offset, 0), "");
-      result = ac_build_buffer_load(&ctx->ac, descriptor, num_components, NULL, voffset,
+      result = ac_build_buffer_load(&ctx->ac, descriptor, num_components, vidx, voffset,
                                     addr_soffset, channel_type, cache_policy, reorder, false);
+
       result = ac_to_integer(&ctx->ac, ac_trim_vector(&ctx->ac, result, num_components));
       break;
    }
    case nir_intrinsic_store_buffer_amd: {
+      bool idxen = !nir_src_is_const(instr->src[4]) || nir_src_as_uint(instr->src[4]);
+
       LLVMValueRef store_data = get_src(ctx, instr->src[0]);
       LLVMValueRef descriptor = get_src(ctx, instr->src[1]);
       LLVMValueRef addr_voffset = get_src(ctx, instr->src[2]);
       LLVMValueRef addr_soffset = get_src(ctx, instr->src[3]);
+      LLVMValueRef vidx = idxen ? get_src(ctx, instr->src[4]) : NULL;
       unsigned const_offset = nir_intrinsic_base(instr);
       bool swizzled = nir_intrinsic_is_swizzled(instr);
       bool coherent = nir_intrinsic_access(instr) & ACCESS_COHERENT;
@@ -4251,7 +4258,7 @@ static bool visit_intrinsic(struct ac_nir_context *ctx, nir_intrinsic_instr *ins
             LLVMConstInt(ctx->ac.i32, const_offset + start * 4, 0), "");
 
          LLVMValueRef data = extract_vector_range(&ctx->ac, store_data, start, count);
-         ac_build_buffer_store_dword(&ctx->ac, descriptor, data, NULL, voffset, addr_soffset,
+         ac_build_buffer_store_dword(&ctx->ac, descriptor, data, vidx, voffset, addr_soffset,
                                      cache_policy);
       }
       break;

diff --git a/src/compiler/nir/nir_intrinsics.py b/src/compiler/nir/nir_intrinsics.py
index 3452f43..8f29e8a 100644 (file)
--- a/src/compiler/nir/nir_intrinsics.py
+++ b/src/compiler/nir/nir_intrinsics.py
@@ -1290,10 +1290,12 @@ store("tf_r600", [])
 
 # AMD GCN/RDNA specific intrinsics
 
-# src[] = { descriptor, base address, scalar offset }
-intrinsic("load_buffer_amd", src_comp=[4, 1, 1], dest_comp=0, indices=[BASE, IS_SWIZZLED, SLC_AMD, MEMORY_MODES, ACCESS], flags=[CAN_ELIMINATE])
-# src[] = { store value, descriptor, base address, scalar offset }
-intrinsic("store_buffer_amd", src_comp=[0, 4, 1, 1], indices=[BASE, WRITE_MASK, IS_SWIZZLED, SLC_AMD, MEMORY_MODES, ACCESS])
+# src[] = { descriptor, vector byte offset, scalar byte offset, index offset }
+# The index offset is multiplied by the stride in the descriptor. The vertex/scalar byte offsets
+# are in bytes.
+intrinsic("load_buffer_amd", src_comp=[4, 1, 1, 1], dest_comp=0, indices=[BASE, IS_SWIZZLED, SLC_AMD, MEMORY_MODES, ACCESS], flags=[CAN_ELIMINATE])
+# src[] = { store value, descriptor, vector byte offset, scalar byte offset, index offset }
+intrinsic("store_buffer_amd", src_comp=[0, 4, 1, 1, 1], indices=[BASE, WRITE_MASK, IS_SWIZZLED, SLC_AMD, MEMORY_MODES, ACCESS])
 
 # src[] = { address, unsigned 32-bit offset }.
 load("global_amd", [1, 1], indices=[BASE, ACCESS, ALIGN_MUL, ALIGN_OFFSET], flags=[CAN_ELIMINATE])