ac/nir: Add vote_ieq/vote_feq lowering pass.
authorBas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Mon, 5 Feb 2018 21:54:18 +0000 (22:54 +0100)
committerBas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Wed, 7 Mar 2018 20:18:32 +0000 (21:18 +0100)
The old vote_eq implementation supported only booleans, but now
we have to support arbitrary values, so use the read_first_invocation
intrinsic + ballot.

I took this as an opportunity to figure out how easy it was to do this
in nir instead of in the nir_to_llvm pass, and it actually turned out
pretty okay IMO. Only creating the pass is some extra code.

Reviewed-by: Dave Airlie <airlied@redhat.com>
src/amd/Makefile.sources
src/amd/common/ac_lower_subgroups.c [new file with mode: 0644]
src/amd/common/ac_nir_to_llvm.c
src/amd/common/ac_nir_to_llvm.h
src/amd/common/meson.build

index 10c4827e19c9f2ffac7c2e22353dccbd78c86dbc..f57dd403c5f4eda2b77915bb2873ec2a60c0dbf5 100644 (file)
@@ -44,6 +44,7 @@ AMD_COMPILER_FILES = \
        common/ac_llvm_helper.cpp \
        common/ac_llvm_util.c \
        common/ac_llvm_util.h \
+       common/ac_lower_subgroups.c \
        common/ac_shader_abi.h \
        common/ac_shader_info.c \
        common/ac_shader_info.h \
diff --git a/src/amd/common/ac_lower_subgroups.c b/src/amd/common/ac_lower_subgroups.c
new file mode 100644 (file)
index 0000000..d0782b4
--- /dev/null
@@ -0,0 +1,92 @@
+/*
+ * Copyright © 2018 Google Inc.
+ *
+ * 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.
+ */
+
+#include "nir/nir.h"
+#include "nir/nir_builder.h"
+
+#include "ac_nir_to_llvm.h"
+
+static nir_ssa_def *ac_lower_subgroups_intrin(nir_builder *b, nir_intrinsic_instr *intrin)
+{
+       switch(intrin->intrinsic) {
+       case nir_intrinsic_vote_ieq:
+       case nir_intrinsic_vote_feq: {
+               nir_intrinsic_instr *rfi =
+                       nir_intrinsic_instr_create(b->shader, nir_intrinsic_read_first_invocation);
+               nir_ssa_dest_init(&rfi->instr, &rfi->dest,
+                                 1, intrin->src[0].ssa->bit_size, NULL);
+               nir_src_copy(&rfi->src[0], &intrin->src[0], rfi);
+               rfi->num_components = 1;
+
+               nir_ssa_def *is_ne;
+               if (intrin->intrinsic == nir_intrinsic_vote_feq)
+                       is_ne = nir_fne(b, &rfi->dest.ssa, intrin->src[0].ssa);
+               else
+                       is_ne = nir_ine(b, &rfi->dest.ssa, intrin->src[0].ssa);
+
+               nir_intrinsic_instr *ballot =
+                       nir_intrinsic_instr_create(b->shader, nir_intrinsic_ballot);
+               nir_ssa_dest_init(&ballot->instr, &ballot->dest,
+                                 1, 64, NULL);
+               ballot->src[0] = nir_src_for_ssa(is_ne);
+               ballot->num_components = 1;
+
+               return nir_ieq(b, &ballot->dest.ssa, nir_imm_int64(b, 0));
+       }
+       default:
+               return NULL;
+       }
+}
+
+bool ac_lower_subgroups(struct nir_shader *shader)
+{
+       bool progress = false;
+
+       nir_foreach_function(function, shader) {
+               if (!function->impl)
+                       continue;
+
+               nir_builder b;
+               nir_builder_init(&b, function->impl);
+
+               nir_foreach_block(block, function->impl) {
+                       nir_foreach_instr_safe(instr, block) {
+                               if (instr->type != nir_instr_type_intrinsic)
+                                       continue;
+
+                               nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
+                               b.cursor = nir_before_instr(instr);
+
+                               nir_ssa_def *lower = ac_lower_subgroups_intrin(&b, intrin);
+                               if (!lower)
+                                       continue;
+
+                               nir_ssa_def_rewrite_uses(&intrin->dest.ssa, nir_src_for_ssa(lower));
+                               nir_instr_remove(instr);
+                               progress = true;
+                       }
+               }
+       }
+
+       return progress;
+}
index 4b3b253c439bb00d74dc57099d3780a353239222..f6ad51a412a4815984f59bc5590dc53d25022d1d 100644 (file)
@@ -4563,11 +4563,6 @@ static void visit_intrinsic(struct ac_nir_context *ctx,
                result = LLVMBuildSExt(ctx->ac.builder, tmp, ctx->ac.i32, "");
                break;
        }
-       case nir_intrinsic_vote_ieq: {
-               LLVMValueRef tmp = ac_build_vote_eq(&ctx->ac, get_src(ctx, instr->src[0]));
-               result = LLVMBuildSExt(ctx->ac.builder, tmp, ctx->ac.i32, "");
-               break;
-       }
        default:
                fprintf(stderr, "Unknown intrinsic: ");
                nir_print_instr(&instr->instr, stderr);
@@ -6757,6 +6752,9 @@ void ac_nir_translate(struct ac_llvm_context *ac, struct ac_shader_abi *abi,
        struct ac_nir_context ctx = {};
        struct nir_function *func;
 
+       /* Last minute passes for both radv & radeonsi */
+       ac_lower_subgroups(nir);
+
        ctx.ac = *ac;
        ctx.abi = abi;
 
index 870e7f8988cab8086f6e8984fd04192528a7272b..983189eb93ade92430288edad18a0cf51a535a1a 100644 (file)
@@ -235,4 +235,6 @@ void ac_lower_indirect_derefs(struct nir_shader *nir, enum chip_class);
 void ac_nir_translate(struct ac_llvm_context *ac, struct ac_shader_abi *abi,
                      struct nir_shader *nir);
 
+bool ac_lower_subgroups(struct nir_shader *shader);
+
 #endif /* AC_NIR_TO_LLVM_H */
index 22c13b955fa626563cdae59c37c891df5d01ed90..6c5c2f45ac5e7d58089a9045a5cb2af41917ee20 100644 (file)
@@ -35,6 +35,7 @@ amd_common_files = files(
   'ac_llvm_helper.cpp',
   'ac_llvm_util.c',
   'ac_llvm_util.h',
+  'ac_lower_subgroups.c',
   'ac_shader_abi.h',
   'ac_shader_info.c',
   'ac_shader_info.h',