AlexIndustrial/mesa
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

ac/nir: mark all input loads as reorderable and speculatable (for LICM)

Browse Source 
These are only memory loads.

We could do the same for LDS loads, which are not truly speculatable
in merged shaders (can't be moved before the barrier), but that's fine
because LICM only moves code out of loops, which can't have barriers.

Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35950>
This commit is contained in:
Marek Olšák
2025-03-18 16:51:24 -04:00
committed by Marge Bot
parent a7291074c8
commit 43f18f5615
3 changed files with 8 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/amd/common/nir/ac_nir_create_gs_copy_shader.c
+2 -1
View File
@@ -54,7 +54,8 @@ ac_nir_create_gs_copy_shader(const nir_shader *gs_nir, ac_nir_lower_legacy_gs_op
out->outputs[i][j] =
nir_load_buffer_amd(&b, 1, 32, gsvs_ring, vtx_offset, zero, zero,
.base = base,
.access = ACCESS_COHERENT | ACCESS_NON_TEMPORAL);
.access = ACCESS_COHERENT | ACCESS_NON_TEMPORAL |
ACCESS_CAN_REORDER | ACCESS_CAN_SPECULATE);
offset += 4;
}
}
src/amd/common/nir/ac_nir_lower_esgs_io_to_mem.c
+4 -2
View File
@@ -64,13 +64,15 @@ emit_split_buffer_load(nir_builder *b, unsigned num_components, unsigned bit_siz
for (unsigned i = 0; i < full_dwords; ++i)
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);
.access = ACCESS_COHERENT | ACCESS_CAN_REORDER |
ACCESS_CAN_SPECULATE);
if (remaining_bytes)
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);
.access = ACCESS_COHERENT | ACCESS_CAN_REORDER |
ACCESS_CAN_SPECULATE);
return nir_extract_bits(b, comps, full_dwords + !!remaining_bytes, 0, num_components, bit_size);
}
src/amd/common/nir/ac_nir_lower_tess_io_to_mem.c
+2 -1
View File
@@ -1499,7 +1499,8 @@ lower_tes_input_load(nir_builder *b,
nir_def *load = NULL;
AC_NIR_LOAD_IO(load, b, intrin->def.num_components, intrin->def.bit_size, io_sem.high_16bits,
nir_load_buffer_amd, offchip_ring, off, offchip_offset, zero, .access = ACCESS_COHERENT,
nir_load_buffer_amd, offchip_ring, off, offchip_offset, zero,
.access = ACCESS_COHERENT | ACCESS_CAN_REORDER | ACCESS_CAN_SPECULATE,
.memory_modes = nir_var_shader_in);
return load;