Module: Mesa Branch: main Commit: 1fabf535faae9ec651a0574247e885f9cdcbdaad URL: http://cgit.freedesktop.org/mesa/mesa/commit/?id=1fabf535faae9ec651a0574247e885f9cdcbdaad
Author: Qiang Yu <[email protected]> Date: Fri Nov 17 10:46:33 2023 +0800 aco: handle GL_TEXTURE_RECTANGLE in tg4_integer_workarounds Ported from LLVM side lower_gather4_integer(). Reviewed-by: Georg Lehmann <[email protected]> Signed-off-by: Qiang Yu <[email protected]> Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26244> --- src/amd/compiler/aco_instruction_selection.cpp | 74 ++++++++++++++------------ 1 file changed, 39 insertions(+), 35 deletions(-) diff --git a/src/amd/compiler/aco_instruction_selection.cpp b/src/amd/compiler/aco_instruction_selection.cpp index 4eef843d7c5..bfb05e76cf8 100644 --- a/src/amd/compiler/aco_instruction_selection.cpp +++ b/src/amd/compiler/aco_instruction_selection.cpp @@ -9411,42 +9411,46 @@ visit_tex(isel_context* ctx, nir_tex_instr* instr) Temp tg4_compare_cube_wa64 = Temp(); if (tg4_integer_workarounds) { - Temp tg4_lod = bld.copy(bld.def(v1), Operand::zero()); - Temp size = bld.tmp(v2); - MIMG_instruction* tex = emit_mimg(bld, aco_opcode::image_get_resinfo, size, resource, - Operand(s4), std::vector<Temp>{tg4_lod}); - tex->dim = dim; - tex->dmask = 0x3; - tex->da = da; - emit_split_vector(ctx, size, size.size()); - Temp half_texel[2]; - for (unsigned i = 0; i < 2; i++) { - half_texel[i] = emit_extract_vector(ctx, size, i, v1); - half_texel[i] = bld.vop1(aco_opcode::v_cvt_f32_i32, bld.def(v1), half_texel[i]); - half_texel[i] = bld.vop1(aco_opcode::v_rcp_iflag_f32, bld.def(v1), half_texel[i]); - half_texel[i] = bld.vop2(aco_opcode::v_mul_f32, bld.def(v1), - Operand::c32(0xbf000000 /*-0.5*/), half_texel[i]); - } - - if (instr->sampler_dim == GLSL_SAMPLER_DIM_2D && !instr->is_array) { - /* In vulkan, whether the sampler uses unnormalized - * coordinates or not is a dynamic property of the - * sampler. Hence, to figure out whether or not we - * need to divide by the texture size, we need to test - * the sampler at runtime. This tests the bit set by - * radv_init_sampler(). - */ - unsigned bit_idx = ffs(S_008F30_FORCE_UNNORMALIZED(1)) - 1; - Temp dword0 = emit_extract_vector(ctx, sampler, 0, s1); - Temp not_needed = - bld.sopc(aco_opcode::s_bitcmp0_b32, bld.def(s1, scc), dword0, Operand::c32(bit_idx)); - - not_needed = bool_to_vector_condition(ctx, not_needed); - half_texel[0] = bld.vop2(aco_opcode::v_cndmask_b32, bld.def(v1), - Operand::c32(0xbf000000 /*-0.5*/), half_texel[0], not_needed); - half_texel[1] = bld.vop2(aco_opcode::v_cndmask_b32, bld.def(v1), - Operand::c32(0xbf000000 /*-0.5*/), half_texel[1], not_needed); + if (instr->sampler_dim == GLSL_SAMPLER_DIM_RECT) { + half_texel[0] = half_texel[1] = bld.copy(bld.def(v1), Operand::c32(0xbf000000 /*-0.5*/)); + } else { + Temp tg4_lod = bld.copy(bld.def(v1), Operand::zero()); + Temp size = bld.tmp(v2); + MIMG_instruction* tex = emit_mimg(bld, aco_opcode::image_get_resinfo, size, resource, + Operand(s4), std::vector<Temp>{tg4_lod}); + tex->dim = dim; + tex->dmask = 0x3; + tex->da = da; + emit_split_vector(ctx, size, size.size()); + + for (unsigned i = 0; i < 2; i++) { + half_texel[i] = emit_extract_vector(ctx, size, i, v1); + half_texel[i] = bld.vop1(aco_opcode::v_cvt_f32_i32, bld.def(v1), half_texel[i]); + half_texel[i] = bld.vop1(aco_opcode::v_rcp_iflag_f32, bld.def(v1), half_texel[i]); + half_texel[i] = bld.vop2(aco_opcode::v_mul_f32, bld.def(v1), + Operand::c32(0xbf000000 /*-0.5*/), half_texel[i]); + } + + if (instr->sampler_dim == GLSL_SAMPLER_DIM_2D && !instr->is_array) { + /* In vulkan, whether the sampler uses unnormalized + * coordinates or not is a dynamic property of the + * sampler. Hence, to figure out whether or not we + * need to divide by the texture size, we need to test + * the sampler at runtime. This tests the bit set by + * radv_init_sampler(). + */ + unsigned bit_idx = ffs(S_008F30_FORCE_UNNORMALIZED(1)) - 1; + Temp dword0 = emit_extract_vector(ctx, sampler, 0, s1); + Temp not_needed = + bld.sopc(aco_opcode::s_bitcmp0_b32, bld.def(s1, scc), dword0, Operand::c32(bit_idx)); + + not_needed = bool_to_vector_condition(ctx, not_needed); + half_texel[0] = bld.vop2(aco_opcode::v_cndmask_b32, bld.def(v1), + Operand::c32(0xbf000000 /*-0.5*/), half_texel[0], not_needed); + half_texel[1] = bld.vop2(aco_opcode::v_cndmask_b32, bld.def(v1), + Operand::c32(0xbf000000 /*-0.5*/), half_texel[1], not_needed); + } } Temp new_coords[2] = {bld.vop2(aco_opcode::v_add_f32, bld.def(v1), coords[0], half_texel[0]),
