Module: Mesa Branch: main Commit: 909895ae2a377758812e3ac3098e97e21f952393 URL: http://cgit.freedesktop.org/mesa/mesa/commit/?id=909895ae2a377758812e3ac3098e97e21f952393
Author: Qiang Yu <yuq...@gmail.com> Date: Wed Nov 1 17:29:11 2023 +0800 radeonsi: only vectorize nir ops that aco support To fix si_compute_blit created nir code compilation with ACO. Two 16bit vector ops are used in it: con 16x2 %11 = u2u16 %10.xy con 16x2 %25 = f2f16 %22.xy which is not supported by ACO yet. PS. now ACO supports vec2 f2f16. Reviewed-by: Marek Olšák <marek.ol...@amd.com> Signed-off-by: Qiang Yu <yuq...@gmail.com> Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25990> --- src/gallium/drivers/radeonsi/si_shader_nir.c | 27 ++++++++++++++++++--------- 1 file changed, 18 insertions(+), 9 deletions(-) diff --git a/src/gallium/drivers/radeonsi/si_shader_nir.c b/src/gallium/drivers/radeonsi/si_shader_nir.c index c85f7b6af8b..7f8219cf8db 100644 --- a/src/gallium/drivers/radeonsi/si_shader_nir.c +++ b/src/gallium/drivers/radeonsi/si_shader_nir.c @@ -8,15 +8,17 @@ #include "nir_xfb_info.h" #include "si_pipe.h" #include "ac_nir.h" +#include "aco_interface.h" bool si_alu_to_scalar_packed_math_filter(const nir_instr *instr, const void *data) { if (instr->type == nir_instr_type_alu) { nir_alu_instr *alu = nir_instr_as_alu(instr); + bool use_aco = (bool)data; - if (alu->def.bit_size == 16 && - alu->def.num_components == 2) + if (alu->def.bit_size == 16 && alu->def.num_components == 2 && + (!use_aco || aco_nir_op_supports_packed_math_16bit(alu))) return false; } @@ -29,7 +31,14 @@ static uint8_t si_vectorize_callback(const nir_instr *instr, const void *data) return 0; nir_alu_instr *alu = nir_instr_as_alu(instr); - if (alu->def.bit_size == 16) { + if (alu->def.bit_size != 16) + return 1; + + bool use_aco = (bool)data; + + if (use_aco) { + return aco_nir_op_supports_packed_math_16bit(alu) ? 2 : 1; + } else { switch (alu->op) { case nir_op_unpack_32_2x16_split_x: case nir_op_unpack_32_2x16_split_y: @@ -38,8 +47,6 @@ static uint8_t si_vectorize_callback(const nir_instr *instr, const void *data) return 2; } } - - return 1; } static unsigned si_lower_bit_size_callback(const nir_instr *instr, void *data) @@ -73,7 +80,7 @@ void si_nir_opts(struct si_screen *sscreen, struct nir_shader *nir, bool first) NIR_PASS(progress, nir, nir_lower_vars_to_ssa); NIR_PASS(progress, nir, nir_lower_alu_to_scalar, - nir->options->lower_to_scalar_filter, NULL); + nir->options->lower_to_scalar_filter, (void *)sscreen->use_aco); NIR_PASS(progress, nir, nir_lower_phis_to_scalar, false); if (first) { @@ -97,7 +104,7 @@ void si_nir_opts(struct si_screen *sscreen, struct nir_shader *nir, bool first) if (lower_alu_to_scalar) { NIR_PASS_V(nir, nir_lower_alu_to_scalar, - nir->options->lower_to_scalar_filter, NULL); + nir->options->lower_to_scalar_filter, (void *)sscreen->use_aco); } if (lower_phis_to_scalar) NIR_PASS_V(nir, nir_lower_phis_to_scalar, false); @@ -139,8 +146,10 @@ void si_nir_opts(struct si_screen *sscreen, struct nir_shader *nir, bool first) if (nir->info.stage == MESA_SHADER_FRAGMENT) NIR_PASS_V(nir, nir_opt_move_discards_to_top); - if (sscreen->info.has_packed_math_16bit) - NIR_PASS(progress, nir, nir_opt_vectorize, si_vectorize_callback, NULL); + if (sscreen->info.has_packed_math_16bit) { + NIR_PASS(progress, nir, nir_opt_vectorize, si_vectorize_callback, + (void *)sscreen->use_aco); + } } while (progress); NIR_PASS_V(nir, nir_lower_var_copies);
