Mesa (master): microsoft/compiler: Add a lowering pass to emit double [un]pack instructions

[GitLab Mirror]

Module: Mesa
Branch: master
Commit: 5578fdd2c49c06e34bce789343f587c666321975
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=5578fdd2c49c06e34bce789343f587c666321975

Author: Jesse Natalie <jenatali at microsoft.com>
Date:   Tue Apr  6 10:44:55 2021 -0700

microsoft/compiler: Add a lowering pass to emit double [un]pack instructions

Whenever we have an ALU op that's operating on a double, we'll unpack
it as an integer, then repack it as a float. When we have an ALU op that
returns a double, we'll unpack it as a double, then repack it as an integer.

Then, simple algebraic opts will remove any redundant unpack/repack ops,
so we should be left with constructing and deconstructing doubles using
the right operations.

Reviewed-by: Enrico Galli <enrico.galli at intel.com>
Reviewed-by: Michael Tang <tangm at microsoft.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10063>

---

 src/microsoft/compiler/dxil_nir.c | 63 +++++++++++++++++++++++++++++++++++++++
 src/microsoft/compiler/dxil_nir.h |  1 +
 2 files changed, 64 insertions(+)

diff --git a/src/microsoft/compiler/dxil_nir.c b/src/microsoft/compiler/dxil_nir.c
index fe25f23552b..97ad6f1cf67 100644
--- a/src/microsoft/compiler/dxil_nir.c
+++ b/src/microsoft/compiler/dxil_nir.c
@@ -1230,3 +1230,66 @@ dxil_nir_split_clip_cull_distance(nir_shader *shader)
 
    return new_var != NULL;
 }
+
+bool
+dxil_nir_lower_double_math(nir_shader *shader)
+{
+   bool progress = false;
+   nir_foreach_function(func, shader) {
+      bool func_progress = false;
+      if (!func->impl)
+         continue;
+
+      nir_builder b;
+      nir_builder_init(&b, func->impl);
+      nir_foreach_block(block, func->impl) {
+         nir_foreach_instr_safe(instr, block) {
+            if (instr->type != nir_instr_type_alu)
+               continue;
+
+            nir_alu_instr *alu = nir_instr_as_alu(instr);
+
+            /* TODO: See if we can apply this explicitly to packs/unpacks that are then
+             * used as a double. As-is, if we had an app explicitly do a 64bit integer op,
+             * then try to bitcast to double (not expressible in HLSL, but it is in other
+             * source languages), this would unpack the integer and repack as a double, when
+             * we probably want to just send the bitcast through to the backend.
+             */
+
+            b.cursor = nir_before_instr(&alu->instr);
+
+            for (unsigned i = 0; i < nir_op_infos[alu->op].num_inputs; ++i) {
+               if (nir_alu_type_get_base_type(nir_op_infos[alu->op].input_types[i]) == nir_type_float &&
+                   alu->src[i].src.ssa->bit_size == 64) {
+                  nir_ssa_def *packed_double = nir_channel(&b, alu->src[i].src.ssa, alu->src[i].swizzle[0]);
+                  nir_ssa_def *unpacked_double = nir_unpack_64_2x32(&b, packed_double);
+                  nir_ssa_def *repacked_double = nir_pack_double_2x32_dxil(&b, unpacked_double);
+                  nir_instr_rewrite_src_ssa(instr, &alu->src[i].src, repacked_double);
+                  memset(alu->src[i].swizzle, 0, ARRAY_SIZE(alu->src[i].swizzle));
+                  func_progress = true;
+               }
+            }
+
+            if (nir_alu_type_get_base_type(nir_op_infos[alu->op].output_type) == nir_type_float &&
+                alu->dest.dest.ssa.bit_size == 64) {
+               b.cursor = nir_after_instr(&alu->instr);
+               nir_ssa_def *packed_double = &alu->dest.dest.ssa;
+               nir_ssa_def *unpacked_double = nir_unpack_double_2x32_dxil(&b, packed_double);
+               nir_ssa_def *repacked_double = nir_pack_64_2x32(&b, unpacked_double);
+               nir_ssa_def_rewrite_uses_after(packed_double, repacked_double, unpacked_double->parent_instr);
+               func_progress = true;
+            }
+         }
+      }
+      
+      if (func_progress)
+         nir_metadata_preserve(func->impl, nir_metadata_block_index |
+                                           nir_metadata_dominance |
+                                           nir_metadata_loop_analysis);
+      else
+         nir_metadata_preserve(func->impl, nir_metadata_all);
+      progress |= func_progress;
+   }
+
+   return progress;
+}
diff --git a/src/microsoft/compiler/dxil_nir.h b/src/microsoft/compiler/dxil_nir.h
index 6c11d117d78..fe657efdd0c 100644
--- a/src/microsoft/compiler/dxil_nir.h
+++ b/src/microsoft/compiler/dxil_nir.h
@@ -45,6 +45,7 @@ bool dxil_nir_lower_memcpy_deref(nir_shader *shader);
 bool dxil_nir_lower_upcast_phis(nir_shader *shader, unsigned min_bit_size);
 bool dxil_nir_lower_fp16_casts(nir_shader *shader);
 bool dxil_nir_split_clip_cull_distance(nir_shader *shader);
+bool dxil_nir_lower_double_math(nir_shader *shader);
 
 nir_ssa_def *
 build_load_ubo_dxil(nir_builder *b, nir_ssa_def *buffer,