i965: Disable ARB_fragment_shader_interlock for platforms prior to GEN9

ARB_fragment_shader_interlock depends on memory fences to
ensure fragment ordering and this ordering guarantee is
only supported from GEN9 onwards.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=109980
Fixes: 939312702e35 "i965: Add ARB_fragment_shader_interlock support."
Signed-off-by: Plamena Manolova <plamena.n.manolova@gmail.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>

diff --git a/src/intel/compiler/brw_fs_generator.cpp b/src/intel/compiler/brw_fs_generator.cpp
index 649dd99..c24d4eb 100644 (file)
--- a/src/intel/compiler/brw_fs_generator.cpp
+++ b/src/intel/compiler/brw_fs_generator.cpp
@@ -2100,6 +2100,7 @@ fs_generator::generate_code(const cfg_t *cfg, int dispatch_width)
          break;
 
       case SHADER_OPCODE_INTERLOCK:
+         assert(devinfo->gen >= 9);
          /* The interlock is basically a memory fence issued via sendc */
          brw_memory_fence(p, dst, BRW_OPCODE_SENDC);
          break;

diff --git a/src/mesa/drivers/dri/i965/intel_extensions.c b/src/mesa/drivers/dri/i965/intel_extensions.c
index d7080fb..afca881 100644 (file)
--- a/src/mesa/drivers/dri/i965/intel_extensions.c
+++ b/src/mesa/drivers/dri/i965/intel_extensions.c
@@ -254,7 +254,6 @@ intelInitExtensions(struct gl_context *ctx)
       ctx->Extensions.EXT_shader_samples_identical = true;
       ctx->Extensions.OES_primitive_bounding_box = true;
       ctx->Extensions.OES_texture_buffer = true;
-      ctx->Extensions.ARB_fragment_shader_interlock = true;
 
       if (can_do_pipelined_register_writes(brw->screen)) {
          ctx->Extensions.ARB_draw_indirect = true;
@@ -327,6 +326,30 @@ intelInitExtensions(struct gl_context *ctx)
       ctx->Extensions.KHR_blend_equation_advanced_coherent = true;
       ctx->Extensions.KHR_texture_compression_astc_ldr = true;
       ctx->Extensions.KHR_texture_compression_astc_sliced_3d = true;
+
+      /*
+       * From the Skylake PRM Vol. 7 (Memory Fence Message, page 221):
+       *  "A memory fence message issued by a thread causes further messages
+       *   issued by the thread to be blocked until all previous data port
+       *   messages have completed, or the results can be globally observed from
+       *   the point of view of other threads in the system."
+       *
+       * From the Haswell PRM Vol. 7 (Memory Fence, page 256):
+       *  "A memory fence message issued by a thread causes further messages
+       *   issued by the thread to be blocked until all previous messages issued
+       *   by the thread to that data port (data cache or render cache) have
+       *   been globally observed from the point of view of other threads in the
+       *   system."
+       *
+       * Summarized: For ARB_fragment_shader_interlock to work, we need to
+       * ensure memory access ordering for all messages to the dataport from
+       * all threads. Memory fence messages prior to SKL only provide memory
+       * access ordering for messages from the same thread, so we can only
+       * support the feature from Gen9 onwards.
+       *
+       */
+
+      ctx->Extensions.ARB_fragment_shader_interlock = true;
    }
 
    if (gen_device_info_is_9lp(devinfo))