pan/va: Optimize add with imm to ADD_IMM

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15223>

diff --git a/src/panfrost/bifrost/bifrost_compile.c b/src/panfrost/bifrost/bifrost_compile.c
index 663a655fae6b7d121ed8f0273d1cc4379864b3a0..88e7f1bb12947be0487e04037872aa6459dd387b 100644 (file)
--- a/src/panfrost/bifrost/bifrost_compile.c
+++ b/src/panfrost/bifrost/bifrost_compile.c
@@ -4123,6 +4123,10 @@ bi_compile_variant_nir(nir_shader *nir,
                 bi_lower_opt_instruction(I);
         }
 
+        if (ctx->arch >= 9) {
+                va_optimize(ctx);
+        }
+
         bi_foreach_block(ctx, block) {
                 bi_lower_branch(ctx, block);
         }

diff --git a/src/panfrost/bifrost/compiler.h b/src/panfrost/bifrost/compiler.h
index e074421c3dbd140598e9a877d27d3e503c2b7651..920e18900f05531cdfa004f39f434276b06407aa 100644 (file)
--- a/src/panfrost/bifrost/compiler.h
+++ b/src/panfrost/bifrost/compiler.h
@@ -1038,6 +1038,7 @@ void bi_lower_swizzle(bi_context *ctx);
 void bi_lower_fau(bi_context *ctx);
 void bi_assign_scoreboard(bi_context *ctx);
 void bi_register_allocate(bi_context *ctx);
+void va_optimize(bi_context *ctx);
 
 void bi_lower_opt_instruction(bi_instr *I);
 

diff --git a/src/panfrost/bifrost/meson.build b/src/panfrost/bifrost/meson.build
index ea9e7c751ea16b1f326a31a15a981aa43e67226b..3191d24ec65157ebcce22d2f9e45909f607d51dc 100644 (file)
--- a/src/panfrost/bifrost/meson.build
+++ b/src/panfrost/bifrost/meson.build
@@ -45,6 +45,7 @@ libpanfrost_bifrost_files = files(
   'bi_validate.c',
   'bir.c',
   'bifrost_compile.c',
+  'valhall/va_optimize.c',
   'valhall/va_pack.c',
 )
 

diff --git a/src/panfrost/bifrost/valhall/va_compiler.h b/src/panfrost/bifrost/valhall/va_compiler.h
index a6409ae6fee91bc078e949a616e7024c836bf979..d7207cd77b4ed524c153e1e1622394a279f53cd7 100644 (file)
--- a/src/panfrost/bifrost/valhall/va_compiler.h
+++ b/src/panfrost/bifrost/valhall/va_compiler.h
@@ -34,6 +34,7 @@
 extern "C" {
 #endif
 
+void va_fuse_add_imm(bi_instr *I);
 uint64_t va_pack_instr(const bi_instr *I, unsigned flow);
 
 static inline unsigned

diff --git a/src/panfrost/bifrost/valhall/va_optimize.c b/src/panfrost/bifrost/valhall/va_optimize.c
new file mode 100644 (file)
index 0000000..118321c
--- /dev/null
+++ b/src/panfrost/bifrost/valhall/va_optimize.c
@@ -0,0 +1,113 @@
+/*
+ * Copyright (C) 2021 Collabora Ltd.
+ *
+ * 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 "va_compiler.h"
+
+/* Valhall specific instruction selection optimizations */
+
+static enum bi_opcode
+va_op_add_imm(enum bi_opcode op)
+{
+   switch (op) {
+   case BI_OPCODE_FADD_F32:   return BI_OPCODE_FADD_IMM_F32;
+   case BI_OPCODE_FADD_V2F16: return BI_OPCODE_FADD_IMM_V2F16;
+   case BI_OPCODE_IADD_S32:
+   case BI_OPCODE_IADD_U32:   return BI_OPCODE_IADD_IMM_I32;
+   case BI_OPCODE_IADD_V2S16:
+   case BI_OPCODE_IADD_V2U16: return BI_OPCODE_IADD_IMM_V2I16;
+   case BI_OPCODE_IADD_V4S8:
+   case BI_OPCODE_IADD_V4U8:  return BI_OPCODE_IADD_IMM_V4I8;
+   default: return 0;
+   }
+}
+
+static bool
+va_is_add_imm(bi_instr *I, unsigned s)
+{
+   return I->src[s].swizzle == BI_SWIZZLE_H01 &&
+          !I->src[s].abs && !I->src[s].neg && !I->clamp && !I->round;
+}
+
+static unsigned
+va_choose_imm(bi_instr *I)
+{
+   for (unsigned i = 0; i < 2; ++i) {
+      if (I->src[i].type == BI_INDEX_CONSTANT)
+         return i;
+   }
+
+   return ~0;
+}
+
+/* Lower MOV.i32 #constant --> IADD_IMM.i32 0x0, #constant */
+static void
+va_lower_mov_imm(bi_instr *I)
+{
+   if (I->src[0].type == BI_INDEX_CONSTANT) {
+      I->op = BI_OPCODE_IADD_IMM_I32;
+      I->index = I->src[0].value;
+      I->src[0] = bi_zero();
+   }
+}
+
+void
+va_fuse_add_imm(bi_instr *I)
+{
+   if (I->op == BI_OPCODE_MOV_I32) {
+      va_lower_mov_imm(I);
+      return;
+   }
+
+   enum bi_opcode op = va_op_add_imm(I->op);
+   if (!op) return;
+
+   unsigned s = va_choose_imm(I);
+   if (s > 1) return;
+   if (!va_is_add_imm(I, 1 - s)) return;
+
+   I->op = op;
+   I->index = bi_apply_swizzle(I->src[s].value, I->src[s].swizzle);
+
+   assert(!I->src[s].abs && "redundant .abs set");
+
+   /* If the constant is negated, flip the sign bit */
+   if (I->src[s].neg) {
+      if (I->op == BI_OPCODE_FADD_IMM_F32)
+         I->index ^= (1 << 31);
+      else if (I->op == BI_OPCODE_FADD_IMM_V2F16)
+         I->index ^= (1 << 31) | (1 << 15);
+      else
+         unreachable("unexpected .neg");
+   }
+
+   I->src[0] = I->src[1 - s];
+   I->src[1] = bi_null();
+}
+
+void
+va_optimize(bi_context *ctx)
+{
+   bi_foreach_instr_global(ctx, I) {
+      va_fuse_add_imm(I);
+   }
+}