intel/fs: Use split sends for surface writes on gen9+

Surface reads don't need them because they just have the one address
payload.  With surface writes, on the other hand, we can put the address
and the data in the different halves and avoid building the payload all
together.

The decrease in register pressure and added freedom in register
allocation resulting from this change reduces spilling enough to improve
the performance of one customer benchmark by about 2x.

Reviewed-by: Iago Toral Quiroga <[email protected]>

2 files changed, 47 insertions, 18 deletions

diff --git a/src/intel/compiler/brw_fs.cpp b/src/intel/compiler/brw_fs.cpp
index d82572a5b70..8dd3b94fbd5 100644
--- a/src/intel/compiler/brw_fs.cpp
+++ b/src/intel/compiler/brw_fs.cpp
@@ -5016,28 +5016,42 @@ lower_surface_logical_send(const fs_builder &bld, fs_inst *inst)
     * Gen11+ the header has been removed so we can only use predication.
     */
    const unsigned header_sz = devinfo->gen < 9 && is_typed_access ? 1 : 0;
-   const unsigned sz = header_sz + addr_sz + src_sz;
-
-   /* Allocate space for the payload. */
-   fs_reg *const components = new fs_reg[sz];
-   const fs_reg payload = bld.vgrf(BRW_REGISTER_TYPE_UD, sz);
-   unsigned n = 0;
 
    const bool has_side_effects = inst->has_side_effects();
    fs_reg sample_mask = has_side_effects ? bld.sample_mask_reg() :
                                            fs_reg(brw_imm_d(0xffff));
 
-   /* Construct the payload. */
-   if (header_sz)
-      components[n++] = emit_surface_header(bld, sample_mask);
+   fs_reg payload, payload2;
+   unsigned mlen, ex_mlen = 0;
+   if (devinfo->gen >= 9) {
+      /* We have split sends on gen9 and above */
+      assert(header_sz == 0);
+      payload = bld.move_to_vgrf(addr, addr_sz);
+      payload2 = bld.move_to_vgrf(src, src_sz);
+      mlen = addr_sz * (inst->exec_size / 8);
+      ex_mlen = src_sz * (inst->exec_size / 8);
+   } else {
+      /* Allocate space for the payload. */
+      const unsigned sz = header_sz + addr_sz + src_sz;
+      payload = bld.vgrf(BRW_REGISTER_TYPE_UD, sz);
+      fs_reg *const components = new fs_reg[sz];
+      unsigned n = 0;
+
+      /* Construct the payload. */
+      if (header_sz)
+         components[n++] = emit_surface_header(bld, sample_mask);
+
+      for (unsigned i = 0; i < addr_sz; i++)
+         components[n++] = offset(addr, bld, i);
 
-   for (unsigned i = 0; i < addr_sz; i++)
-      components[n++] = offset(addr, bld, i);
+      for (unsigned i = 0; i < src_sz; i++)
+         components[n++] = offset(src, bld, i);
 
-   for (unsigned i = 0; i < src_sz; i++)
-      components[n++] = offset(src, bld, i);
+      bld.LOAD_PAYLOAD(payload, components, sz, header_sz);
+      mlen = header_sz + (addr_sz + src_sz) * inst->exec_size / 8;
 
-   bld.LOAD_PAYLOAD(payload, components, sz, header_sz);
+      delete[] components;
+   }
 
    /* Predicate the instruction on the sample mask if no header is
     * provided.
@@ -5162,7 +5176,8 @@ lower_surface_logical_send(const fs_builder &bld, fs_inst *inst)
 
    /* Update the original instruction. */
    inst->opcode = SHADER_OPCODE_SEND;
-   inst->mlen = header_sz + (addr_sz + src_sz) * inst->exec_size / 8;
+   inst->mlen = mlen;
+   inst->ex_mlen = ex_mlen;
    inst->header_size = header_sz;
    inst->send_has_side_effects = has_side_effects;
    inst->send_is_volatile = !has_side_effects;
@@ -5183,10 +5198,9 @@ lower_surface_logical_send(const fs_builder &bld, fs_inst *inst)
 
    /* Finally, the payload */
    inst->src[2] = payload;
+   inst->src[3] = payload2;
 
-   inst->resize_sources(3);
-
-   delete[] components;
+   inst->resize_sources(4);
 }
 
 static void
diff --git a/src/intel/compiler/brw_fs_builder.h b/src/intel/compiler/brw_fs_builder.h
index c50af4c1f55..3fc6e5f2c25 100644
--- a/src/intel/compiler/brw_fs_builder.h
+++ b/src/intel/compiler/brw_fs_builder.h
@@ -426,6 +426,21 @@ namespace brw {
          return src_reg(component(dst, 0));
       }
 
+      src_reg
+      move_to_vgrf(const src_reg &src, unsigned num_components) const
+      {
+         src_reg *const src_comps = new src_reg[num_components];
+         for (unsigned i = 0; i < num_components; i++)
+            src_comps[i] = offset(src, dispatch_width(), i);
+
+         const dst_reg dst = vgrf(src.type, num_components);
+         LOAD_PAYLOAD(dst, src_comps, num_components, 0);
+
+         delete[] src_comps;
+
+         return src_reg(dst);
+      }
+
       void
       emit_scan(enum opcode opcode, const dst_reg &tmp,
                 unsigned cluster_size, brw_conditional_mod mod) const