i965/ir: Skip eliminate_find_live_channel() for stages with sparse thread dispatch.

The eliminate_find_live_channel optimization eliminates
FIND_LIVE_CHANNEL instructions in cases where control flow is known to
be uniform, and replaces them with 'MOV 0', which in turn unblocks
subsequent elimination of the BROADCAST instruction frequently used on
the result of FIND_LIVE_CHANNEL.  This is however not correct in
per-sample fragment shader dispatch because the PSD can dispatch a
fully unlit sample under certain conditions.  Disable the optimization
in that case.

Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
v2: Add devinfo argument to brw_stage_has_packed_dispatch() to
    implement hardware generation check.

diff --git a/src/mesa/drivers/dri/i965/brw_compiler.h b/src/mesa/drivers/dri/i965/brw_compiler.h
index 84d3dde8c995cbc1a1b1b67adbdc346718f817ff..445c166082ea861d1fe72c04196dc6fd976e885f 100644 (file)
--- a/src/mesa/drivers/dri/i965/brw_compiler.h
+++ b/src/mesa/drivers/dri/i965/brw_compiler.h
@@ -868,6 +868,55 @@ encode_slm_size(unsigned gen, uint32_t bytes)
    return slm_size;
 }
 
+/**
+ * Return true if the given shader stage is dispatched contiguously by the
+ * relevant fixed function starting from channel 0 of the SIMD thread, which
+ * implies that the dispatch mask of a thread can be assumed to have the form
+ * '2^n - 1' for some n.
+ */
+static inline bool
+brw_stage_has_packed_dispatch(const struct gen_device_info *devinfo,
+                              gl_shader_stage stage,
+                              const struct brw_stage_prog_data *prog_data)
+{
+   /* The code below makes assumptions about the hardware's thread dispatch
+    * behavior that could be proven wrong in future generations -- Make sure
+    * to do a full test run with brw_fs_test_dispatch_packing() hooked up to
+    * the NIR front-end before changing this assertion.
+    */
+   assert(devinfo->gen <= 9);
+
+   switch (stage) {
+   case MESA_SHADER_FRAGMENT: {
+      /* The PSD discards subspans coming in with no lit samples, which in the
+       * per-pixel shading case implies that each subspan will either be fully
+       * lit (due to the VMask being used to allow derivative computations),
+       * or not dispatched at all.  In per-sample dispatch mode individual
+       * samples from the same subspan have a fixed relative location within
+       * the SIMD thread, so dispatch of unlit samples cannot be avoided in
+       * general and we should return false.
+       */
+      const struct brw_wm_prog_data *wm_prog_data =
+         (const struct brw_wm_prog_data *)prog_data;
+      return !wm_prog_data->persample_dispatch;
+   }
+   case MESA_SHADER_COMPUTE:
+      /* Compute shaders will be spawned with either a fully enabled dispatch
+       * mask or with whatever bottom/right execution mask was given to the
+       * GPGPU walker command to be used along the workgroup edges -- In both
+       * cases the dispatch mask is required to be tightly packed for our
+       * invocation index calculations to work.
+       */
+      return true;
+   default:
+      /* Most remaining fixed functions are limited to use a packed dispatch
+       * mask due to the hardware representation of the dispatch mask as a
+       * single counter representing the number of enabled channels.
+       */
+      return true;
+   }
+}
+
 #ifdef __cplusplus
 } /* extern "C" */
 #endif

diff --git a/src/mesa/drivers/dri/i965/brw_fs.cpp b/src/mesa/drivers/dri/i965/brw_fs.cpp
index 5c44007ba884c848af3915c0941d3b63fd99ccf1..b60ec71888a0ef784ef4ea1e6225c2d122fa166a 100644 (file)
--- a/src/mesa/drivers/dri/i965/brw_fs.cpp
+++ b/src/mesa/drivers/dri/i965/brw_fs.cpp
@@ -2835,6 +2835,14 @@ fs_visitor::eliminate_find_live_channel()
    bool progress = false;
    unsigned depth = 0;
 
+   if (!brw_stage_has_packed_dispatch(devinfo, stage, stage_prog_data)) {
+      /* The optimization below assumes that channel zero is live on thread
+       * dispatch, which may not be the case if the fixed function dispatches
+       * threads sparsely.
+       */
+      return false;
+   }
+
    foreach_block_and_inst_safe(block, fs_inst, inst, cfg) {
       switch (inst->opcode) {
       case BRW_OPCODE_IF:

diff --git a/src/mesa/drivers/dri/i965/brw_vec4.cpp b/src/mesa/drivers/dri/i965/brw_vec4.cpp
index 58c8a8a5bdbe7bce2a5b3195ac120b391304438e..6aa91028fb7f2bb2945016e6cfc77acd59c5e851 100644 (file)
--- a/src/mesa/drivers/dri/i965/brw_vec4.cpp
+++ b/src/mesa/drivers/dri/i965/brw_vec4.cpp
@@ -1291,6 +1291,14 @@ vec4_visitor::eliminate_find_live_channel()
    bool progress = false;
    unsigned depth = 0;
 
+   if (!brw_stage_has_packed_dispatch(devinfo, stage, stage_prog_data)) {
+      /* The optimization below assumes that channel zero is live on thread
+       * dispatch, which may not be the case if the fixed function dispatches
+       * threads sparsely.
+       */
+      return false;
+   }
+
    foreach_block_and_inst_safe(block, vec4_instruction, inst, cfg) {
       switch (inst->opcode) {
       case BRW_OPCODE_IF: