intel/fs: Add support for CS to group invocations in quads

When using quads, instead of mapping the elements to the next 4 local
invocation indices, we map the two next in the "current" row and two
next in the "next row".  A side effect is that a thread will execute
the indices in a different order.

We now perform the lowering of both local invocation ID and index
together -- and don't rely anymore on lowering done by
nir_lower_system_values.  That is convenient when doing the math for
quads, because we need X and Y to get the right invocation index.

When the pass progresses, fold the constants and clean up to reduce
the noise from the indexing math.

This implements the derivative_group_quadsNV semantics from
NV_compute_shader_derivatives.

v2: Take subgroup_id into account, otherwise only values in the first
    subgroup would be used. (Jason)

v3: Calculate invocation index and ID together, to avoid duplicating
    some math in the quads case when both index and ID are used. (Jason)

v4: Don't call cleanup passes as part of the lowering, let that to the
    call site. (Jason)
    Change calculation to use less instructions. (Jason)

Reviewed-by: Ian Romanick <[email protected]> (v3)
Reviewed-by: Jason Ekstrand <[email protected]>

3 files changed, 103 insertions, 16 deletions

diff --git a/src/intel/compiler/brw_compiler.c b/src/intel/compiler/brw_compiler.c
index a3a0a393fad..d3f8c7ef1e0 100644
--- a/src/intel/compiler/brw_compiler.c
+++ b/src/intel/compiler/brw_compiler.c
@@ -45,7 +45,6 @@
    .lower_flrp64 = true,                                                      \
    .lower_isign = true,                                                       \
    .lower_ldexp = true,                                                       \
-   .lower_cs_local_id_from_index = true,                                      \
    .lower_device_index_to_zero = true,                                        \
    .native_integers = true,                                                   \
    .use_interpolated_input_intrinsics = true,                                 \
diff --git a/src/intel/compiler/brw_fs.cpp b/src/intel/compiler/brw_fs.cpp
index 0c2439d9daf..a637ee3422f 100644
--- a/src/intel/compiler/brw_fs.cpp
+++ b/src/intel/compiler/brw_fs.cpp
@@ -8017,6 +8017,11 @@ compile_cs_to_nir(const struct brw_compiler *compiler,
    nir_shader *shader = nir_shader_clone(mem_ctx, src_shader);
    shader = brw_nir_apply_sampler_key(shader, compiler, &key->tex, true);
    brw_nir_lower_cs_intrinsics(shader, dispatch_width);
+
+   /* Clean up after the local index and ID calculations. */
+   nir_opt_constant_folding(shader);
+   nir_opt_dce(shader);
+
    return brw_postprocess_nir(shader, compiler, true);
 }
 
diff --git a/src/intel/compiler/brw_nir_lower_cs_intrinsics.c b/src/intel/compiler/brw_nir_lower_cs_intrinsics.c
index fab5edc893f..6180e6a706b 100644
--- a/src/intel/compiler/brw_nir_lower_cs_intrinsics.c
+++ b/src/intel/compiler/brw_nir_lower_cs_intrinsics.c
@@ -41,6 +41,10 @@ lower_cs_intrinsics_convert_block(struct lower_intrinsics_state *state,
    nir_builder *b = &state->builder;
    nir_shader *nir = state->nir;
 
+   /* Reuse calculated values inside the block. */
+   nir_ssa_def *local_index = NULL;
+   nir_ssa_def *local_id = NULL;
+
    nir_foreach_instr_safe(instr, block) {
       if (instr->type != nir_instr_type_intrinsic)
          continue;
@@ -51,22 +55,91 @@ lower_cs_intrinsics_convert_block(struct lower_intrinsics_state *state,
 
       nir_ssa_def *sysval;
       switch (intrinsic->intrinsic) {
-      case nir_intrinsic_load_local_invocation_index: {
-         /* We construct the local invocation index from:
-          *
-          *    gl_LocalInvocationIndex =
-          *       cs_thread_local_id + subgroup_invocation;
-          */
-         nir_ssa_def *subgroup_id;
-         if (state->local_workgroup_size <= state->dispatch_width)
-            subgroup_id = nir_imm_int(b, 0);
-         else
-            subgroup_id = nir_load_subgroup_id(b);
+      case nir_intrinsic_load_local_invocation_index:
+      case nir_intrinsic_load_local_invocation_id: {
+         /* First time we are using those, so let's calculate them. */
+         if (!local_index) {
+            assert(!local_id);
+
+            nir_ssa_def *subgroup_id;
+            if (state->local_workgroup_size <= state->dispatch_width)
+               subgroup_id = nir_imm_int(b, 0);
+            else
+               subgroup_id = nir_load_subgroup_id(b);
+
+            nir_ssa_def *thread_local_id =
+               nir_imul_imm(b, subgroup_id, state->dispatch_width);
+            nir_ssa_def *channel = nir_load_subgroup_invocation(b);
+            nir_ssa_def *linear = nir_iadd(b, channel, thread_local_id);
+
+            nir_ssa_def *size_x = nir_imm_int(b, nir->info.cs.local_size[0]);
+            nir_ssa_def *size_y = nir_imm_int(b, nir->info.cs.local_size[1]);
+
+            /* The local invocation index and ID must respect the following
+             *
+             *    gl_LocalInvocationID.x =
+             *       gl_LocalInvocationIndex % gl_WorkGroupSize.x;
+             *    gl_LocalInvocationID.y =
+             *       (gl_LocalInvocationIndex / gl_WorkGroupSize.x) %
+             *       gl_WorkGroupSize.y;
+             *    gl_LocalInvocationID.z =
+             *       (gl_LocalInvocationIndex /
+             *        (gl_WorkGroupSize.x * gl_WorkGroupSize.y)) %
+             *       gl_WorkGroupSize.z;
+             *
+             * However, the final % gl_WorkGroupSize.z does nothing unless we
+             * accidentally end up with a gl_LocalInvocationIndex that is too
+             * large so it can safely be omitted.
+             */
+
+            if (state->nir->info.cs.derivative_group != DERIVATIVE_GROUP_QUADS) {
+               /* If we are not grouping in quads, just set the local invocatio
+                * index linearly, and calculate local invocation ID from that.
+                */
+               local_index = linear;
+
+               nir_ssa_def *id_x, *id_y, *id_z;
+               id_x = nir_umod(b, local_index, size_x);
+               id_y = nir_umod(b, nir_udiv(b, local_index, size_x), size_y);
+               id_z = nir_udiv(b, local_index, nir_imul(b, size_x, size_y));
+               local_id = nir_vec3(b, id_x, id_y, id_z);
+            } else {
+               /* For quads, first we figure out the 2x2 grid the invocation
+                * belongs to -- treating extra Z layers as just more rows.
+                * Then map that into local invocation ID (trivial) and local
+                * invocation index.  Skipping Z simplify index calculation.
+                */
+
+               nir_ssa_def *one = nir_imm_int(b, 1);
+               nir_ssa_def *double_size_x = nir_ishl(b, size_x, one);
+
+               /* ID within a pair of rows, where each group of 4 is 2x2 quad. */
+               nir_ssa_def *row_pair_id = nir_umod(b, linear, double_size_x);
+               nir_ssa_def *y_row_pairs = nir_udiv(b, linear, double_size_x);
+
+               nir_ssa_def *x =
+                  nir_ior(b,
+                          nir_iand(b, row_pair_id, one),
+                          nir_iand(b, nir_ishr(b, row_pair_id, one),
+                                   nir_imm_int(b, 0xfffffffe)));
+               nir_ssa_def *y =
+                  nir_ior(b,
+                          nir_ishl(b, y_row_pairs, one),
+                          nir_iand(b, nir_ishr(b, row_pair_id, one), one));
+
+               local_id = nir_vec3(b, x,
+                                   nir_umod(b, y, size_y),
+                                   nir_udiv(b, y, size_y));
+               local_index = nir_iadd(b, x, nir_imul(b, y, size_x));
+            }
+         }
 
-         nir_ssa_def *thread_local_id =
-            nir_imul(b, subgroup_id, nir_imm_int(b, state->dispatch_width));
-         nir_ssa_def *channel = nir_load_subgroup_invocation(b);
-         sysval = nir_iadd(b, channel, thread_local_id);
+         assert(local_id);
+         assert(local_index);
+         if (intrinsic->intrinsic == nir_intrinsic_load_local_invocation_id)
+            sysval = local_id;
+         else
+            sysval = local_index;
          break;
       }
 
@@ -125,10 +198,20 @@ brw_nir_lower_cs_intrinsics(nir_shader *nir,
    memset(&state, 0, sizeof(state));
    state.nir = nir;
    state.dispatch_width = dispatch_width;
+
+   assert(!nir->info.cs.local_size_variable);
    state.local_workgroup_size = nir->info.cs.local_size[0] *
                                 nir->info.cs.local_size[1] *
                                 nir->info.cs.local_size[2];
 
+   /* Constraints from NV_compute_shader_derivatives. */
+   if (nir->info.cs.derivative_group == DERIVATIVE_GROUP_QUADS) {
+      assert(nir->info.cs.local_size[0] % 2 == 0);
+      assert(nir->info.cs.local_size[1] % 2 == 0);
+   } else if (nir->info.cs.derivative_group == DERIVATIVE_GROUP_LINEAR) {
+      assert(state.local_workgroup_size % 4 == 0);
+   }
+
    do {
       state.progress = false;
       nir_foreach_function(function, nir) {