freedreno/ir3/nir: lower if/else
authorRob Clark <robclark@freedesktop.org>
Wed, 15 Apr 2015 14:16:31 +0000 (10:16 -0400)
committerRob Clark <robclark@freedesktop.org>
Fri, 17 Apr 2015 15:40:14 +0000 (11:40 -0400)
For now, completely flatten if/else blocks.  That will almost certainly
change once we have flow control.

Signed-off-by: Rob Clark <robclark@freedesktop.org>
src/gallium/drivers/freedreno/Makefile.sources
src/gallium/drivers/freedreno/freedreno_screen.c
src/gallium/drivers/freedreno/ir3/ir3_compiler.h
src/gallium/drivers/freedreno/ir3/ir3_compiler_nir.c
src/gallium/drivers/freedreno/ir3/ir3_nir.h [new file with mode: 0644]
src/gallium/drivers/freedreno/ir3/ir3_nir_lower_if_else.c [new file with mode: 0644]

index 3224efc9994d51ad6f91d383c5da6c4873e32780..a565a9c4e4da8cb27ddc04deac3a7097ec735853 100644 (file)
@@ -130,6 +130,8 @@ ir3_SOURCES := \
        ir3/ir3_group.c \
        ir3/ir3.h \
        ir3/ir3_legalize.c \
+       ir3/ir3_nir.h \
+       ir3/ir3_nir_lower_if_else.c \
        ir3/ir3_ra.c \
        ir3/ir3_sched.c \
        ir3/ir3_shader.c \
index fda60eda6fd420eaa58cac5667c0ae969989194d..15ae287d7b0d2c8847f5e5c1b39a2d583e0056f9 100644 (file)
@@ -346,10 +346,6 @@ fd_screen_get_shader_param(struct pipe_screen *pscreen, unsigned shader,
        case PIPE_SHADER_CAP_MAX_TEX_INDIRECTIONS:
                return 16384;
        case PIPE_SHADER_CAP_MAX_CONTROL_FLOW_DEPTH:
-               /* for now, let someone else flatten if/else when using NIR: */
-               if ((fd_mesa_debug & FD_DBG_NIR) &&
-                               (is_a3xx(screen) || is_a4xx(screen)))
-                       return 0;
                return 8; /* XXX */
        case PIPE_SHADER_CAP_MAX_INPUTS:
        case PIPE_SHADER_CAP_MAX_OUTPUTS:
index ed9637bd2d0c5cf83e7024d0cc1730e46733201e..9213386e00c341cc04fb089591cf1f4d28e7294d 100644 (file)
@@ -26,8 +26,8 @@
  *    Rob Clark <robclark@freedesktop.org>
  */
 
-#ifndef FD3_COMPILER_H_
-#define FD3_COMPILER_H_
+#ifndef IR3_COMPILER_H_
+#define IR3_COMPILER_H_
 
 #include "ir3_shader.h"
 
@@ -39,4 +39,4 @@ int ir3_compile_shader(struct ir3_shader_variant *so,
                const struct tgsi_token *tokens,
                struct ir3_shader_key key, bool cp);
 
-#endif /* FD3_COMPILER_H_ */
+#endif /* IR3_COMPILER_H_ */
index 13ae7c2cac46eeb59851fa8c457bec8251b96933..0b42cd70381c7cfcd6adc5607d29a27dd1e7971d 100644 (file)
@@ -42,6 +42,7 @@
 
 #include "ir3_compiler.h"
 #include "ir3_shader.h"
+#include "ir3_nir.h"
 
 #include "instr-a3xx.h"
 #include "ir3.h"
@@ -149,7 +150,7 @@ static struct nir_shader *to_nir(const struct tgsi_token *tokens)
                progress |= nir_copy_prop(s);
                progress |= nir_opt_dce(s);
                progress |= nir_opt_cse(s);
-               progress |= nir_opt_peephole_select(s);
+               progress |= ir3_nir_lower_if_else(s);
                progress |= nir_opt_algebraic(s);
                progress |= nir_opt_constant_folding(s);
 
diff --git a/src/gallium/drivers/freedreno/ir3/ir3_nir.h b/src/gallium/drivers/freedreno/ir3/ir3_nir.h
new file mode 100644 (file)
index 0000000..f3d3075
--- /dev/null
@@ -0,0 +1,36 @@
+/* -*- mode: C; c-file-style: "k&r"; tab-width 4; indent-tabs-mode: t; -*- */
+
+/*
+ * Copyright (C) 2015 Rob Clark <robclark@freedesktop.org>
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ * Authors:
+ *    Rob Clark <robclark@freedesktop.org>
+ */
+
+#ifndef IR3_NIR_H_
+#define IR3_NIR_H_
+
+#include "glsl/nir/nir.h"
+
+bool ir3_nir_lower_if_else(nir_shader *shader);
+
+#endif /* IR3_NIR_H_ */
diff --git a/src/gallium/drivers/freedreno/ir3/ir3_nir_lower_if_else.c b/src/gallium/drivers/freedreno/ir3/ir3_nir_lower_if_else.c
new file mode 100644 (file)
index 0000000..ae36019
--- /dev/null
@@ -0,0 +1,338 @@
+/*
+ * Copyright © 2014 Intel Corporation
+ * Copyright © 2015 Red Hat
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ *
+ * Authors:
+ *    Jason Ekstrand (jason@jlekstrand.net)
+ *    Rob Clark (robclark@freedesktop.org)
+ *
+ */
+
+#include "ir3_nir.h"
+#include "glsl/nir/nir_builder.h"
+
+/* Based on nir_opt_peephole_select, and hacked up to more aggressively
+ * flatten anything that can be flattened
+ *
+ * This *might* be something that other drivers could use.  On the other
+ * hand, I think most other hw has predicated instructions or similar
+ * to select which side of if/else writes back result (and therefore
+ * not having to assign unique registers to both sides of the if/else.
+ * (And hopefully those drivers don't also have crazy scheduling reqs
+ * and can more easily do this in their backend.)
+ *
+ * TODO eventually when we have proper flow control in the backend:
+ *
+ *  + Probably weight differently normal ALUs vs SFUs (cos/rcp/exp)
+ *    since executing extra SFUs for the branch-not-taken path will
+ *    generally be much more expensive.
+ *
+ *    Possibly what constitutes an ALU vs SFU differs between hw
+ *    backends.. but that seems doubtful.
+ *
+ *  + Account for texture fetch and memory accesses (incl UBOs)
+ *    since these will be more expensive..
+ *
+ *  + When if-condition is const (or uniform) or we have some way
+ *    to know that all threads in the warp take the same branch
+ *    then we should prefer to not flatten the if/else..
+ */
+
+struct lower_state {
+       nir_builder b;
+       void *mem_ctx;
+       bool progress;
+};
+
+static bool
+valid_dest(nir_block *block, nir_dest *dest)
+{
+       /* It must be SSA */
+       if (!dest->is_ssa)
+               return false;
+
+       /* We only lower blocks that do not contain other blocks
+        * (so this is run iteratively in a loop).  Therefore if
+        * we get this far, it should not have any if_uses:
+        */
+       assert(dest->ssa.if_uses->entries == 0);
+
+       /* The only uses of this definition must be phi's in the
+        * successor or in the current block
+        */
+       struct set_entry *entry;
+       set_foreach(dest->ssa.uses, entry) {
+               const nir_instr *dest_instr = entry->key;
+               if (dest_instr->block == block)
+                       continue;
+               if ((dest_instr->type == nir_instr_type_phi) &&
+                               (dest_instr->block == block->successors[0]))
+                       continue;
+               return false;
+       }
+
+       return true;
+}
+
+static bool
+block_check_for_allowed_instrs(nir_block *block)
+{
+       nir_foreach_instr(block, instr) {
+               switch (instr->type) {
+               case nir_instr_type_intrinsic: {
+                       nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
+                       const nir_intrinsic_info *info =
+                                       &nir_intrinsic_infos[intr->intrinsic];
+
+                       switch (intr->intrinsic) {
+                       case nir_intrinsic_discard_if:
+                               /* to simplify things, we want discard_if src in ssa: */
+                               if (!intr->src[0].is_ssa)
+                                       return false;
+                               /* fallthrough */
+                       case nir_intrinsic_discard:
+                               /* discard/discard_if can be reordered, but only
+                                * with some special care
+                                */
+                               break;
+                       case nir_intrinsic_store_output:
+                               /* TODO technically, if both if and else store
+                                * the same output, we can hoist that out to
+                                * the end of the block w/ a phi..
+                                * In practice, the tgsi shaders we already get
+                                * do this for us, so I think we don't need to
+                                */
+                       default:
+                               if (!(info->flags & NIR_INTRINSIC_CAN_REORDER))
+                                       return false;
+                       }
+
+                       break;
+               }
+
+               case nir_instr_type_tex: {
+                       nir_tex_instr *tex = nir_instr_as_tex(instr);
+                       if (!valid_dest(block, &tex->dest))
+                               return false;
+                       break;
+               }
+               case nir_instr_type_phi: {
+                       nir_phi_instr *phi = nir_instr_as_phi(instr);
+                       if (!valid_dest(block, &phi->dest))
+                               return false;
+                       break;
+               }
+               case nir_instr_type_alu: {
+                       nir_alu_instr *alu = nir_instr_as_alu(instr);
+                       if (!valid_dest(block, &alu->dest.dest))
+                               return false;
+                       break;
+               }
+
+               case nir_instr_type_load_const:
+               case nir_instr_type_ssa_undef:
+                       break; /* always ssa dest */
+
+               default:
+                       return false;
+               }
+       }
+
+       return true;
+}
+
+/* flatten an then or else block: */
+static void
+flatten_block(nir_builder *bld, nir_block *if_block, nir_block *prev_block,
+               nir_ssa_def *condition, bool invert)
+{
+       nir_foreach_instr_safe(if_block, instr) {
+               if (instr->type == nir_instr_type_intrinsic) {
+                       nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
+                       if ((intr->intrinsic == nir_intrinsic_discard) ||
+                                       (intr->intrinsic == nir_intrinsic_discard_if)) {
+                               nir_ssa_def *discard_cond;
+
+                               nir_builder_insert_after_instr(bld,
+                                               nir_block_last_instr(prev_block));
+
+                               if (invert) {
+                                       condition = nir_inot(bld, condition);
+                                       invert = false;
+                               }
+
+                               if (intr->intrinsic == nir_intrinsic_discard) {
+                                       discard_cond = condition;
+                               } else {
+                                       assert(intr->src[0].is_ssa);
+                                       /* discard_if gets re-written w/ src and'd: */
+                                       discard_cond = nir_iand(bld, condition, intr->src[0].ssa);
+                               }
+
+                               nir_intrinsic_instr *discard_if =
+                                               nir_intrinsic_instr_create(bld->shader,
+                                                               nir_intrinsic_discard_if);
+                               discard_if->src[0] = nir_src_for_ssa(discard_cond);
+
+                               nir_instr_insert_after(nir_block_last_instr(prev_block),
+                                               &discard_if->instr);
+                               nir_instr_remove(instr);
+                               instr = NULL;
+                       }
+               }
+               /* if not an handled specially, just move to prev block: */
+               if (instr) {
+                       /* NOTE: exec_node_remove() is safe here (vs nir_instr_remove()
+                        * since we are re-adding the instructin back in to the prev
+                        * block (so no dangling SSA uses)
+                        */
+                       exec_node_remove(&instr->node);
+                       instr->block = prev_block;
+                       exec_list_push_tail(&prev_block->instr_list, &instr->node);
+               }
+       }
+}
+
+static bool
+lower_if_else_block(nir_block *block, void *void_state)
+{
+       struct lower_state *state = void_state;
+
+       /* If the block is empty, then it certainly doesn't have any phi nodes,
+        * so we can skip it.  This also ensures that we do an early skip on the
+        * end block of the function which isn't actually attached to the CFG.
+        */
+       if (exec_list_is_empty(&block->instr_list))
+               return true;
+
+       if (nir_cf_node_is_first(&block->cf_node))
+               return true;
+
+       nir_cf_node *prev_node = nir_cf_node_prev(&block->cf_node);
+       if (prev_node->type != nir_cf_node_if)
+               return true;
+
+       nir_if *if_stmt = nir_cf_node_as_if(prev_node);
+       nir_cf_node *then_node = nir_if_first_then_node(if_stmt);
+       nir_cf_node *else_node = nir_if_first_else_node(if_stmt);
+
+       /* We can only have one block in each side ... */
+       if (nir_if_last_then_node(if_stmt) != then_node ||
+                       nir_if_last_else_node(if_stmt) != else_node)
+               return true;
+
+       nir_block *then_block = nir_cf_node_as_block(then_node);
+       nir_block *else_block = nir_cf_node_as_block(else_node);
+
+       /* ... and those blocks must only contain "allowed" instructions. */
+       if (!block_check_for_allowed_instrs(then_block) ||
+                       !block_check_for_allowed_instrs(else_block))
+               return true;
+
+       /* condition should be ssa too, which simplifies flatten_block: */
+       if (!if_stmt->condition.is_ssa)
+               return true;
+
+       /* At this point, we know that the previous CFG node is an if-then
+        * statement containing only moves to phi nodes in this block.  We can
+        * just remove that entire CF node and replace all of the phi nodes with
+        * selects.
+        */
+
+       nir_block *prev_block = nir_cf_node_as_block(nir_cf_node_prev(prev_node));
+       assert(prev_block->cf_node.type == nir_cf_node_block);
+
+       /* First, we move the remaining instructions from the blocks to the
+        * block before.  There are a few things that need handling specially
+        * like discard/discard_if.
+        */
+       flatten_block(&state->b, then_block, prev_block,
+                       if_stmt->condition.ssa, false);
+       flatten_block(&state->b, else_block, prev_block,
+                       if_stmt->condition.ssa, true);
+
+       nir_foreach_instr_safe(block, instr) {
+               if (instr->type != nir_instr_type_phi)
+                       break;
+
+               nir_phi_instr *phi = nir_instr_as_phi(instr);
+               nir_alu_instr *sel = nir_alu_instr_create(state->mem_ctx, nir_op_bcsel);
+               nir_src_copy(&sel->src[0].src, &if_stmt->condition, state->mem_ctx);
+               /* Splat the condition to all channels */
+               memset(sel->src[0].swizzle, 0, sizeof sel->src[0].swizzle);
+
+               assert(exec_list_length(&phi->srcs) == 2);
+               nir_foreach_phi_src(phi, src) {
+                       assert(src->pred == then_block || src->pred == else_block);
+                       assert(src->src.is_ssa);
+
+                       unsigned idx = src->pred == then_block ? 1 : 2;
+                       nir_src_copy(&sel->src[idx].src, &src->src, state->mem_ctx);
+               }
+
+               nir_ssa_dest_init(&sel->instr, &sel->dest.dest,
+                               phi->dest.ssa.num_components, phi->dest.ssa.name);
+               sel->dest.write_mask = (1 << phi->dest.ssa.num_components) - 1;
+
+               nir_ssa_def_rewrite_uses(&phi->dest.ssa,
+                               nir_src_for_ssa(&sel->dest.dest.ssa),
+                               state->mem_ctx);
+
+               nir_instr_insert_before(&phi->instr, &sel->instr);
+               nir_instr_remove(&phi->instr);
+       }
+
+       nir_cf_node_remove(&if_stmt->cf_node);
+       state->progress = true;
+
+       return true;
+}
+
+static bool
+lower_if_else_impl(nir_function_impl *impl)
+{
+       struct lower_state state;
+
+       state.mem_ctx = ralloc_parent(impl);
+       state.progress = false;
+       nir_builder_init(&state.b, impl);
+
+       nir_foreach_block(impl, lower_if_else_block, &state);
+
+       if (state.progress)
+               nir_metadata_preserve(impl, nir_metadata_none);
+
+       return state.progress;
+}
+
+bool
+ir3_nir_lower_if_else(nir_shader *shader)
+{
+       bool progress = false;
+
+       nir_foreach_overload(shader, overload) {
+               if (overload->impl)
+                       progress |= lower_if_else_impl(overload->impl);
+       }
+
+       return progress;
+}