From 96c276dda909ddf12714b9e64b7207156e8fd4bb Mon Sep 17 00:00:00 2001 From: =?utf8?q?Alejandro=20Pi=C3=B1eiro?= Date: Wed, 23 Mar 2016 12:20:05 +0100 Subject: [PATCH] i965/fs: half exec_size when dealing with 64 bits attributes The HW has a restriction that only vertical stride may cross register boundaries. Until now this was only handled on VGRFs at rw_reg_from_fs_reg, but it is also needed for attributes. v2: * Remove reference to commit id on commit message (Juan Suarez) * Simplify code that compute final exec_size (Ian Romanick) * Use REG_SIZE on that same code (Kenneth Graunke) Reviewed-by: Kenneth Graunke --- src/mesa/drivers/dri/i965/brw_fs.cpp | 21 +++++++++++++++++++-- 1 file changed, 19 insertions(+), 2 deletions(-) diff --git a/src/mesa/drivers/dri/i965/brw_fs.cpp b/src/mesa/drivers/dri/i965/brw_fs.cpp index 06a5de1785f..e8baf6c0706 100644 --- a/src/mesa/drivers/dri/i965/brw_fs.cpp +++ b/src/mesa/drivers/dri/i965/brw_fs.cpp @@ -1729,11 +1729,28 @@ fs_visitor::convert_attr_sources_to_hw_regs(fs_inst *inst) inst->src[i].nr + inst->src[i].reg_offset; - unsigned width = inst->src[i].stride == 0 ? 1 : inst->exec_size; + /* As explained at brw_reg_from_fs_reg, From the Haswell PRM: + * + * VertStride must be used to cross GRF register boundaries. This + * rule implies that elements within a 'Width' cannot cross GRF + * boundaries. + * + * So, for registers that are large enough, we have to split the exec + * size in two and trust the compression state to sort it out. + */ + unsigned total_size = inst->exec_size * + inst->src[i].stride * + type_sz(inst->src[i].type); + + assert(total_size <= 2 * REG_SIZE); + const unsigned exec_size = + (total_size <= REG_SIZE) ? inst->exec_size : inst->exec_size / 2; + + unsigned width = inst->src[i].stride == 0 ? 1 : exec_size; struct brw_reg reg = stride(byte_offset(retype(brw_vec8_grf(grf, 0), inst->src[i].type), inst->src[i].subreg_offset), - inst->exec_size * inst->src[i].stride, + exec_size * inst->src[i].stride, width, inst->src[i].stride); reg.abs = inst->src[i].abs; reg.negate = inst->src[i].negate; -- 2.30.2