ntt: Use scoped barriers

In addition to bringing us one backend closer to the scoped-only future, this
improves the generated code in cases like:

   memoryBarrierBuffer();
   memoryBarrierShared();
   controlBarrier();

With scoped_barriers + nir_opt_combine_barriers, we now emit only one MEMBAR
instruction (and a BARRIER) rather than two MEMBARs.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/23191>

6 files changed, 56 insertions, 38 deletions

diff --git a/src/gallium/auxiliary/nir/nir_to_tgsi.c b/src/gallium/auxiliary/nir/nir_to_tgsi.c
index b25eb3b484c..35465d2dfad 100644
--- a/src/gallium/auxiliary/nir/nir_to_tgsi.c
+++ b/src/gallium/auxiliary/nir/nir_to_tgsi.c
@@ -2341,6 +2341,50 @@ ntt_emit_load_sysval(struct ntt_compile *c, nir_intrinsic_instr *instr)
 }
 
 static void
+ntt_emit_barrier(struct ntt_compile *c, nir_intrinsic_instr *intr)
+{
+   bool compute = gl_shader_stage_is_compute(c->s->info.stage);
+
+   if (nir_intrinsic_memory_scope(intr) != NIR_SCOPE_NONE) {
+      nir_variable_mode modes = nir_intrinsic_memory_modes(intr);
+      unsigned membar = 0;
+
+      if (modes & nir_var_image)
+         membar |= TGSI_MEMBAR_SHADER_IMAGE;
+
+      if (modes & nir_var_mem_shared)
+         membar |= TGSI_MEMBAR_SHARED;
+
+      /* Atomic counters are lowered to SSBOs, there's no NIR mode corresponding
+       * exactly to atomics. Take the closest match.
+       */
+      if (modes & nir_var_mem_ssbo)
+         membar |= TGSI_MEMBAR_SHADER_BUFFER | TGSI_MEMBAR_ATOMIC_BUFFER;
+
+      if (modes & nir_var_mem_global)
+         membar |= TGSI_MEMBAR_SHADER_BUFFER;
+
+      /* If we only need workgroup scope (not device-scope), we might be able to
+       * optimize a bit.
+       */
+      if (membar && compute &&
+          nir_intrinsic_memory_scope(intr) == NIR_SCOPE_WORKGROUP) {
+
+         membar |= TGSI_MEMBAR_THREAD_GROUP;
+      }
+
+      /* Only emit a memory barrier if there are any relevant modes */
+      if (membar)
+         ntt_MEMBAR(c, ureg_imm1u(c->ureg, membar));
+   }
+
+   if (nir_intrinsic_execution_scope(intr) != NIR_SCOPE_NONE) {
+      assert(compute || c->s->info.stage == MESA_SHADER_TESS_CTRL);
+      ntt_BARRIER(c);
+   }
+}
+
+static void
 ntt_emit_intrinsic(struct ntt_compile *c, nir_intrinsic_instr *instr)
 {
    switch (instr->intrinsic) {
@@ -2491,42 +2535,8 @@ ntt_emit_intrinsic(struct ntt_compile *c, nir_intrinsic_instr *instr)
       ntt_emit_image_load_store(c, instr);
       break;
 
-   case nir_intrinsic_control_barrier:
-   case nir_intrinsic_memory_barrier_tcs_patch:
-      ntt_BARRIER(c);
-      break;
-
-   case nir_intrinsic_memory_barrier:
-      ntt_MEMBAR(c, ureg_imm1u(c->ureg,
-                               TGSI_MEMBAR_SHADER_BUFFER |
-                               TGSI_MEMBAR_ATOMIC_BUFFER |
-                               TGSI_MEMBAR_SHADER_IMAGE |
-                               TGSI_MEMBAR_SHARED));
-      break;
-
-   case nir_intrinsic_memory_barrier_atomic_counter:
-      ntt_MEMBAR(c, ureg_imm1u(c->ureg, TGSI_MEMBAR_ATOMIC_BUFFER));
-      break;
-
-   case nir_intrinsic_memory_barrier_buffer:
-      ntt_MEMBAR(c, ureg_imm1u(c->ureg, TGSI_MEMBAR_SHADER_BUFFER));
-      break;
-
-   case nir_intrinsic_memory_barrier_image:
-      ntt_MEMBAR(c, ureg_imm1u(c->ureg, TGSI_MEMBAR_SHADER_IMAGE));
-      break;
-
-   case nir_intrinsic_memory_barrier_shared:
-      ntt_MEMBAR(c, ureg_imm1u(c->ureg, TGSI_MEMBAR_SHARED));
-      break;
-
-   case nir_intrinsic_group_memory_barrier:
-      ntt_MEMBAR(c, ureg_imm1u(c->ureg,
-                               TGSI_MEMBAR_SHADER_BUFFER |
-                               TGSI_MEMBAR_ATOMIC_BUFFER |
-                               TGSI_MEMBAR_SHADER_IMAGE |
-                               TGSI_MEMBAR_SHARED |
-                               TGSI_MEMBAR_THREAD_GROUP));
+   case nir_intrinsic_scoped_barrier:
+      ntt_emit_barrier(c, instr);
       break;
 
    case nir_intrinsic_end_primitive:
@@ -3824,6 +3834,8 @@ const void *nir_to_tgsi_options(struct nir_shader *s,
       }
    } while (progress);
 
+   NIR_PASS_V(s, nir_opt_combine_barriers, NULL, NULL);
+
    if (screen->get_shader_param(screen,
                                 pipe_shader_type_from_mesa(s->info.stage),
                                 PIPE_SHADER_CAP_INTEGERS)) {
@@ -3942,6 +3954,7 @@ static const nir_shader_compiler_options nir_to_tgsi_compiler_options = {
    .lower_vector_cmp = true,
    .lower_int64_options = nir_lower_imul_2x32_64,
    .use_interpolated_input_intrinsics = true,
+   .use_scoped_barrier = true,
 
    /* TGSI doesn't have a semantic for local or global index, just local and
     * workgroup id.
diff --git a/src/gallium/drivers/i915/i915_screen.c b/src/gallium/drivers/i915/i915_screen.c
index 01f2cb8f909..fd9bece25dd 100644
--- a/src/gallium/drivers/i915/i915_screen.c
+++ b/src/gallium/drivers/i915/i915_screen.c
@@ -120,6 +120,7 @@ static const nir_shader_compiler_options i915_compiler_options = {
    .lower_uniforms_to_ubo = true,
    .lower_vector_cmp = true,
    .use_interpolated_input_intrinsics = true,
+   .use_scoped_barrier = true,
    .force_indirect_unrolling = nir_var_all,
    .force_indirect_unrolling_sampler = true,
    .max_unroll_iterations = 32,
diff --git a/src/gallium/drivers/nouveau/nv30/nv30_screen.c b/src/gallium/drivers/nouveau/nv30/nv30_screen.c
index 78597f7936e..f1872ca09b9 100644
--- a/src/gallium/drivers/nouveau/nv30/nv30_screen.c
+++ b/src/gallium/drivers/nouveau/nv30/nv30_screen.c
@@ -490,6 +490,7 @@ static const nir_shader_compiler_options nv30_base_compiler_options = {
    .no_integers = true,
 
    .use_interpolated_input_intrinsics = true,
+   .use_scoped_barrier = true,
 };
 
 static const void *
diff --git a/src/gallium/drivers/r300/r300_screen.c b/src/gallium/drivers/r300/r300_screen.c
index c096dbc0b2c..2eeb61d3172 100644
--- a/src/gallium/drivers/r300/r300_screen.c
+++ b/src/gallium/drivers/r300/r300_screen.c
@@ -507,7 +507,8 @@ static int r300_get_video_param(struct pipe_screen *screen,
    .lower_uniforms_to_ubo = true,             \
    .lower_vector_cmp = true,                  \
    .no_integers = true,                       \
-   .use_interpolated_input_intrinsics = true
+   .use_interpolated_input_intrinsics = true, \
+   .use_scoped_barrier = true
 
 static const nir_shader_compiler_options r500_vs_compiler_options = {
    COMMON_NIR_OPTIONS,
diff --git a/src/gallium/drivers/softpipe/sp_screen.c b/src/gallium/drivers/softpipe/sp_screen.c
index 7e096a02162..10468333ff0 100644
--- a/src/gallium/drivers/softpipe/sp_screen.c
+++ b/src/gallium/drivers/softpipe/sp_screen.c
@@ -91,6 +91,7 @@ static const nir_shader_compiler_options sp_compiler_options = {
    .lower_int64_options = nir_lower_imul_2x32_64,
    .max_unroll_iterations = 32,
    .use_interpolated_input_intrinsics = true,
+   .use_scoped_barrier = true,
 
    /* TGSI doesn't have a semantic for local or global index, just local and
     * workgroup id.
diff --git a/src/gallium/drivers/svga/svga_screen.c b/src/gallium/drivers/svga/svga_screen.c
index e23eb07e7c8..5f6eb52c729 100644
--- a/src/gallium/drivers/svga/svga_screen.c
+++ b/src/gallium/drivers/svga/svga_screen.c
@@ -752,7 +752,8 @@ vgpu10_get_shader_param(struct pipe_screen *screen,
    .lower_vector_cmp = true,                                                  \
    .lower_cs_local_index_to_id = true,                                        \
    .max_unroll_iterations = 32,                                               \
-   .use_interpolated_input_intrinsics = true
+   .use_interpolated_input_intrinsics = true,                                 \
+   .use_scoped_barrier = true
 
 #define VGPU10_OPTIONS                                                        \
    .lower_doubles_options = nir_lower_dfloor,                                 \