i965: Add memory fence opcode.

Acked-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>

6 files changed, 87 insertions, 0 deletions

diff --git a/src/mesa/drivers/dri/i965/brw_defines.h b/src/mesa/drivers/dri/i965/brw_defines.h
index edb113b..634885b 100644
--- a/src/mesa/drivers/dri/i965/brw_defines.h
+++ b/src/mesa/drivers/dri/i965/brw_defines.h
@@ -910,6 +910,8 @@ enum opcode {
    SHADER_OPCODE_TYPED_SURFACE_READ,
    SHADER_OPCODE_TYPED_SURFACE_WRITE,
 
+   SHADER_OPCODE_MEMORY_FENCE,
+
    SHADER_OPCODE_GEN4_SCRATCH_READ,
    SHADER_OPCODE_GEN4_SCRATCH_WRITE,
    SHADER_OPCODE_GEN7_SCRATCH_READ,
diff --git a/src/mesa/drivers/dri/i965/brw_eu.h b/src/mesa/drivers/dri/i965/brw_eu.h
index 3ee5de7..84c1e57 100644
--- a/src/mesa/drivers/dri/i965/brw_eu.h
+++ b/src/mesa/drivers/dri/i965/brw_eu.h
@@ -448,6 +448,10 @@ brw_typed_surface_write(struct brw_codegen *p,
                         unsigned num_channels);
 
 void
+brw_memory_fence(struct brw_codegen *p,
+                 struct brw_reg dst);
+
+void
 brw_pixel_interpolator_query(struct brw_codegen *p,
                              struct brw_reg dest,
                              struct brw_reg mrf,
diff --git a/src/mesa/drivers/dri/i965/brw_eu_emit.c b/src/mesa/drivers/dri/i965/brw_eu_emit.c
index 92d97de..3039662 100644
--- a/src/mesa/drivers/dri/i965/brw_eu_emit.c
+++ b/src/mesa/drivers/dri/i965/brw_eu_emit.c
@@ -3114,6 +3114,76 @@ brw_typed_surface_write(struct brw_codegen *p,
       p, insn, num_channels);
 }
 
+static void
+brw_set_memory_fence_message(struct brw_codegen *p,
+                             struct brw_inst *insn,
+                             enum brw_message_target sfid,
+                             bool commit_enable)
+{
+   const struct brw_device_info *devinfo = p->devinfo;
+
+   brw_set_message_descriptor(p, insn, sfid,
+                              1 /* message length */,
+                              (commit_enable ? 1 : 0) /* response length */,
+                              true /* header present */,
+                              false);
+
+   switch (sfid) {
+   case GEN6_SFID_DATAPORT_RENDER_CACHE:
+      brw_inst_set_dp_msg_type(devinfo, insn, GEN7_DATAPORT_RC_MEMORY_FENCE);
+      break;
+   case GEN7_SFID_DATAPORT_DATA_CACHE:
+      brw_inst_set_dp_msg_type(devinfo, insn, GEN7_DATAPORT_DC_MEMORY_FENCE);
+      break;
+   default:
+      unreachable("Not reached");
+   }
+
+   if (commit_enable)
+      brw_inst_set_dp_msg_control(devinfo, insn, 1 << 5);
+}
+
+void
+brw_memory_fence(struct brw_codegen *p,
+                 struct brw_reg dst)
+{
+   const struct brw_device_info *devinfo = p->devinfo;
+   const bool commit_enable = devinfo->gen == 7 && !devinfo->is_haswell;
+   struct brw_inst *insn;
+
+   /* Set dst as destination for dependency tracking, the MEMORY_FENCE
+    * message doesn't write anything back.
+    */
+   insn = next_insn(p, BRW_OPCODE_SEND);
+   brw_set_dest(p, insn, dst);
+   brw_set_src0(p, insn, dst);
+   brw_set_memory_fence_message(p, insn, GEN7_SFID_DATAPORT_DATA_CACHE,
+                                commit_enable);
+
+   if (devinfo->gen == 7 && !devinfo->is_haswell) {
+      /* IVB does typed surface access through the render cache, so we need to
+       * flush it too.  Use a different register so both flushes can be
+       * pipelined by the hardware.
+       */
+      insn = next_insn(p, BRW_OPCODE_SEND);
+      brw_set_dest(p, insn, offset(dst, 1));
+      brw_set_src0(p, insn, offset(dst, 1));
+      brw_set_memory_fence_message(p, insn, GEN6_SFID_DATAPORT_RENDER_CACHE,
+                                   commit_enable);
+
+      /* Now write the response of the second message into the response of the
+       * first to trigger a pipeline stall -- This way future render and data
+       * cache messages will be properly ordered with respect to past data and
+       * render cache messages.
+       */
+      brw_push_insn_state(p);
+      brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
+      brw_set_default_mask_control(p, BRW_MASK_DISABLE);
+      brw_MOV(p, dst, offset(dst, 1));
+      brw_pop_insn_state(p);
+   }
+}
+
 void
 brw_pixel_interpolator_query(struct brw_codegen *p,
                              struct brw_reg dest,
diff --git a/src/mesa/drivers/dri/i965/brw_fs_generator.cpp b/src/mesa/drivers/dri/i965/brw_fs_generator.cpp
index 5b9f95d..b81978c 100644
--- a/src/mesa/drivers/dri/i965/brw_fs_generator.cpp
+++ b/src/mesa/drivers/dri/i965/brw_fs_generator.cpp
@@ -2053,6 +2053,10 @@ fs_generator::generate_code(const cfg_t *cfg, int dispatch_width)
          brw_typed_surface_write(p, src[0], src[1], inst->mlen, src[2].dw1.ud);
          break;
 
+      case SHADER_OPCODE_MEMORY_FENCE:
+         brw_memory_fence(p, dst);
+         break;
+
       case FS_OPCODE_SET_SIMD4X2_OFFSET:
          generate_set_simd4x2_offset(inst, dst, src[0]);
          break;
diff --git a/src/mesa/drivers/dri/i965/brw_shader.cpp b/src/mesa/drivers/dri/i965/brw_shader.cpp
index a2cb39d..20588e9 100644
--- a/src/mesa/drivers/dri/i965/brw_shader.cpp
+++ b/src/mesa/drivers/dri/i965/brw_shader.cpp
@@ -502,6 +502,8 @@ brw_instruction_name(enum opcode op)
       return "typed_surface_read";
    case SHADER_OPCODE_TYPED_SURFACE_WRITE:
       return "typed_surface_write";
+   case SHADER_OPCODE_MEMORY_FENCE:
+      return "memory_fence";
 
    case SHADER_OPCODE_LOAD_PAYLOAD:
       return "load_payload";
@@ -1049,6 +1051,7 @@ backend_instruction::has_side_effects() const
    case SHADER_OPCODE_UNTYPED_SURFACE_WRITE:
    case SHADER_OPCODE_TYPED_ATOMIC:
    case SHADER_OPCODE_TYPED_SURFACE_WRITE:
+   case SHADER_OPCODE_MEMORY_FENCE:
    case SHADER_OPCODE_URB_WRITE_SIMD8:
    case FS_OPCODE_FB_WRITE:
       return true;
diff --git a/src/mesa/drivers/dri/i965/brw_vec4_generator.cpp b/src/mesa/drivers/dri/i965/brw_vec4_generator.cpp
index 324f795..c15fa16 100644
--- a/src/mesa/drivers/dri/i965/brw_vec4_generator.cpp
+++ b/src/mesa/drivers/dri/i965/brw_vec4_generator.cpp
@@ -1508,6 +1508,10 @@ vec4_generator::generate_code(const cfg_t *cfg)
                                  src[2].dw1.ud);
          break;
 
+      case SHADER_OPCODE_MEMORY_FENCE:
+         brw_memory_fence(p, dst);
+         break;
+
       case VS_OPCODE_UNPACK_FLAGS_SIMD4X2:
          generate_unpack_flags(dst);
          break;