I had some very minor comments but, with those addressed 1-4 are Reviewed-by: Jason Ekstrand <<a href="mailto:jason.ekstrand@intel.com">jason.ekstrand@intel.com</a>> <div class="gmail_quote">On Dec 11, 2015 1:24 PM, "Kenneth Graunke" <<a href="mailto:kenneth@whitecape.org">kenneth@whitecape.org</a>> wrote: <blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">Instead of performing the read-modify-write cycle in glsl->nir, we can simply emit a partial writemask.Â For locals, nir_lower_vars_to_ssa will do the equivalent read-modify-write cycle for us, so we continue to get the same SSA values we had before. Because glsl_to_nir calls nir_lower_outputs_to_temporaries, all outputs are shadowed with temporary values, and written out as whole vectors at the end of the shader.Â So, most consumers will still not see partial writemasks. However, nir_lower_outputs_to_temporaries bails for tessellation control shader outputs.Â So those remain actual variables, and stores to those variables now get a writemask.Â nir_lower_io passes that through.Â This means that TCS outputs should actually work now. This is a functional change for tessellation control shaders. v2: Relax the nir_validate assert to allow partial writemasks. Signed-off-by: Kenneth Graunke <<a href="mailto:kenneth@whitecape.org">kenneth@whitecape.org</a>> --- Â src/glsl/nir/glsl_to_nir.cpp | 39 +++++++++------------------------------ Â src/glsl/nir/nir_validate.cÂ |Â 3 +-- Â 2 files changed, 10 insertions(+), 32 deletions(-) diff --git a/src/glsl/nir/glsl_to_nir.cpp b/src/glsl/nir/glsl_to_nir.cpp index 14f3153..a45e2a0 100644 --- a/src/glsl/nir/glsl_to_nir.cpp +++ b/src/glsl/nir/glsl_to_nir.cpp @@ -1130,44 +1130,23 @@ nir_visitor::visit(ir_assignment *ir) Â Â nir_ssa_def *src = evaluate_rvalue(ir->rhs); Â Â if (ir->write_mask != (1 << num_components) - 1 && ir->write_mask != 0) { -Â Â Â /* -Â Â Â Â * We have no good way to update only part of a variable, so just load -Â Â Â Â * the LHS and do a vec operation to combine the old with the new, and -Â Â Â Â * then store it -Â Â Â Â * back into the LHS. Copy propagation should get rid of the mess. +Â Â Â /* GLSL IR will give us the input to the write-masked assignment in a +Â Â Â Â * single packed vector.Â So, for example, if the writemask is xzw, then +Â Â Â Â * we have to swizzle x -> x, y -> z, and z -> w and get the y component +Â Â Â Â * from the load. Â Â Â Â */ - -Â Â Â nir_intrinsic_instr *load = -Â Â Â Â Â nir_intrinsic_instr_create(this->shader, nir_intrinsic_load_var); -Â Â Â load->num_components = ir->lhs->type->vector_elements; -Â Â Â nir_ssa_dest_init(&load->instr, &load->dest, num_components, NULL); -Â Â Â load->variables[0] = lhs_deref; -Â Â Â ralloc_steal(load, load->variables[0]); -Â Â Â nir_builder_instr_insert(&b, &load->instr); - -Â Â Â nir_ssa_def *srcs[4]; - +Â Â Â unsigned swiz[4]; Â Â Â Â unsigned component = 0; -Â Â Â for (unsigned i = 0; i < ir->lhs->type->vector_elements; i++) { -Â Â Â Â Â if (ir->write_mask & (1 << i)) { -Â Â Â Â Â Â /* GLSL IR will give us the input to the write-masked assignment -Â Â Â Â Â Â Â * in a single packed vector.Â So, for example, if the -Â Â Â Â Â Â Â * writemask is xzw, then we have to swizzle x -> x, y -> z, -Â Â Â Â Â Â Â * and z -> w and get the y component from the load. -Â Â Â Â Â Â Â */ -Â Â Â Â Â Â srcs[i] = nir_channel(&b, src, component++); -Â Â Â Â Â } else { -Â Â Â Â Â Â srcs[i] = nir_channel(&b, &load->dest.ssa, i); -Â Â Â Â Â } +Â Â Â for (unsigned i = 0; i < 4; i++) { +Â Â Â Â Â swiz[i] = ir->write_mask & (1 << i) ? component++ : 0; Â Â Â Â } - -Â Â Â src = nir_vec(&b, srcs, ir->lhs->type->vector_elements); +Â Â Â src = nir_swizzle(&b, src, swiz, num_components, !supports_ints); Â Â } Â Â nir_intrinsic_instr *store = Â Â Â Â nir_intrinsic_instr_create(this->shader, nir_intrinsic_store_var); Â Â store->num_components = ir->lhs->type->vector_elements; -Â Â store->const_index[0] = (1 << store->num_components) - 1; +Â Â store->const_index[0] = ir->write_mask; Â Â nir_deref *store_deref = nir_copy_deref(store, &lhs_deref->deref); Â Â store->variables[0] = nir_deref_as_var(store_deref); Â Â store->src[0] = nir_src_for_ssa(src); diff --git a/src/glsl/nir/nir_validate.c b/src/glsl/nir/nir_validate.c index 89cf0b8..da92055 100644 --- a/src/glsl/nir/nir_validate.c +++ b/src/glsl/nir/nir_validate.c @@ -417,8 +417,7 @@ validate_intrinsic_instr(nir_intrinsic_instr *instr, validate_state *state) Â Â Â Â assert(instr->variables[0]->var->data.mode != nir_var_shader_in && Â Â Â Â Â Â Â instr->variables[0]->var->data.mode != nir_var_uniform && Â Â Â Â Â Â Â instr->variables[0]->var->data.mode != nir_var_shader_storage); -Â Â Â /* Currently, writemasks must cover the entire value */ -Â Â Â assert(instr->const_index[0] == (1 << instr->num_components) - 1); +Â Â Â assert((instr->const_index[0] & ~((1 << instr->num_components) - 1)) == 0); Â Â Â Â break; Â Â } Â Â case nir_intrinsic_copy_var: -- 2.6.3 _______________________________________________ mesa-dev mailing list <a href="mailto:mesa-dev@lists.freedesktop.org">mesa-dev@lists.freedesktop.org</a> <a href="http://lists.freedesktop.org/mailman/listinfo/mesa-dev" rel="noreferrer" target="_blank">http://lists.freedesktop.org/mailman/listinfo/mesa-dev</a> </blockquote></div>