i965/fs: Add remove_extra_rounding_modes optimization

Although from SPIR-V point of view, rounding modes are attached to the
operation/destination, on i965 it is a status, so we don't need to
explicitly set the rounding mode if the one we want is already set.

Taking into account that the default mode is RTE, one possible
optimization would be optimize out the first RTE set for each
block. For in order to work, we would need to take into account block
interrelationships. At this point, it is not worth to complicate the
optimization for such small gain.

v2: Use a single SHADER_OPCODE_RND_MODE opcode taking an immediate
    with the rounding mode (Curro)
v3: Reset optimization for every block. (Jason Ekstrand)

Signed-off-by: Jose Maria Casanova Crespo <jmcasanova@igalia.com>
Signed-off-by: Alejandro Piñeiro <apinheiro@igalia.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>

src/intel/compiler/brw_eu_defines.h

@@ -1252,6 +1252,7 @@ enum PACKED brw_rnd_mode {
    BRW_RND_MODE_RU = 1,    /* Round Up, toward +inf */
    BRW_RND_MODE_RD = 2,    /* Round Down, toward -inf */
    BRW_RND_MODE_RTZ = 3,   /* Round Toward Zero */
+   BRW_RND_MODE_UNSPECIFIED,  /* Unspecified rounding mode */
 };
 
 #endif /* BRW_EU_DEFINES_H */

src/intel/compiler/brw_fs.cpp

@@ -3092,6 +3092,42 @@ fs_visitor::remove_duplicate_mrf_writes()
    return progress;
 }
 
+/**
+ * Rounding modes for conversion instructions are included for each
+ * conversion, but right now it is a state. So once it is set,
+ * we don't need to call it again for subsequent calls.
+ *
+ * This is useful for vector/matrices conversions, as setting the
+ * mode once is enough for the full vector/matrix
+ */
+bool
+fs_visitor::remove_extra_rounding_modes()
+{
+   bool progress = false;
+
+   foreach_block (block, cfg) {
+      brw_rnd_mode prev_mode = BRW_RND_MODE_UNSPECIFIED;
+
+      foreach_inst_in_block_safe (fs_inst, inst, block) {
+         if (inst->opcode == SHADER_OPCODE_RND_MODE) {
+            assert(inst->src[0].file == BRW_IMMEDIATE_VALUE);
+            const brw_rnd_mode mode = (brw_rnd_mode) inst->src[0].d;
+            if (mode == prev_mode) {
+               inst->remove(block);
+               progress = true;
+            } else {
+               prev_mode = mode;
+            }
+         }
+      }
+   }
+
+   if (progress)
+      invalidate_live_intervals();
+
+   return progress;
+}
+
 static void
 clear_deps_for_inst_src(fs_inst *inst, bool *deps, int first_grf, int grf_len)
 {
@@ -5808,6 +5844,7 @@ fs_visitor::optimize()
    int pass_num = 0;
 
    OPT(opt_drop_redundant_mov_to_flags);
+   OPT(remove_extra_rounding_modes);
 
    do {
       progress = false;

src/intel/compiler/brw_fs.h

@@ -150,6 +150,7 @@ public:
    bool eliminate_find_live_channel();
    bool dead_code_eliminate();
    bool remove_duplicate_mrf_writes();
+   bool remove_extra_rounding_modes();
 
    bool opt_sampler_eot();
    bool virtual_grf_interferes(int a, int b);