From 8fd2f2c276ccf3dec96365c7d76d5f4186a3d2ba Mon Sep 17 00:00:00 2001 From: Jason Ekstrand Date: Fri, 26 Jul 2019 16:03:08 -0500 Subject: [PATCH] intel/fs: Implement quad_swap_horizontal with a swizzle on gen7 This fixes dEQP-VK.subgroups.quad.compute.subgroupquadswaphorizontal_* on all gen7 platforms. Cc: mesa-stable@lists.freedesktop.org Reviewed-by: Matt Turner --- src/intel/compiler/brw_fs_nir.cpp | 27 ++++++++++++++++++++------- 1 file changed, 20 insertions(+), 7 deletions(-) diff --git a/src/intel/compiler/brw_fs_nir.cpp b/src/intel/compiler/brw_fs_nir.cpp index 2451fbf0349..ee3634b7b99 100644 --- a/src/intel/compiler/brw_fs_nir.cpp +++ b/src/intel/compiler/brw_fs_nir.cpp @@ -5114,16 +5114,29 @@ fs_visitor::nir_emit_intrinsic(const fs_builder &bld, nir_intrinsic_instr *instr case nir_intrinsic_quad_swap_horizontal: { const fs_reg value = get_nir_src(instr->src[0]); const fs_reg tmp = bld.vgrf(value.type); - const fs_builder ubld = bld.exec_all().group(dispatch_width / 2, 0); + if (devinfo->gen <= 7) { + /* The hardware doesn't seem to support these crazy regions with + * compressed instructions on gen7 and earlier so we fall back to + * using quad swizzles. Fortunately, we don't support 64-bit + * anything in Vulkan on gen7. + */ + assert(nir_src_bit_size(instr->src[0]) == 32); + const fs_builder ubld = bld.exec_all(); + ubld.emit(SHADER_OPCODE_QUAD_SWIZZLE, tmp, value, + brw_imm_ud(BRW_SWIZZLE4(1,0,3,2))); + bld.MOV(retype(dest, value.type), tmp); + } else { + const fs_builder ubld = bld.exec_all().group(dispatch_width / 2, 0); - const fs_reg src_left = horiz_stride(value, 2); - const fs_reg src_right = horiz_stride(horiz_offset(value, 1), 2); - const fs_reg tmp_left = horiz_stride(tmp, 2); - const fs_reg tmp_right = horiz_stride(horiz_offset(tmp, 1), 2); + const fs_reg src_left = horiz_stride(value, 2); + const fs_reg src_right = horiz_stride(horiz_offset(value, 1), 2); + const fs_reg tmp_left = horiz_stride(tmp, 2); + const fs_reg tmp_right = horiz_stride(horiz_offset(tmp, 1), 2); - ubld.MOV(tmp_left, src_right); - ubld.MOV(tmp_right, src_left); + ubld.MOV(tmp_left, src_right); + ubld.MOV(tmp_right, src_left); + } bld.MOV(retype(dest, value.type), tmp); break; } -- 2.30.2