Mesa (master): aco: add Info::{operand_size,definition_size}

[GitLab Mirror]

Module: Mesa
Branch: master
Commit: 207c35cbe85db944d4c58b54f7f9a289bd50e18f
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=207c35cbe85db944d4c58b54f7f9a289bd50e18f

Author: Rhys Perry <pendingchaos02 at gmail.com>
Date:   Mon May 18 15:37:33 2020 +0100

aco: add Info::{operand_size,definition_size}

No shader-db changes.

Signed-off-by: Rhys Perry <pendingchaos02 at gmail.com>
Reviewed-by: Daniel Schürmann <daniel at schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5040>

---

 src/amd/compiler/aco_ir.h           |  3 +++
 src/amd/compiler/aco_opcodes.py     | 31 +++++++++++++++++++++++++++++++
 src/amd/compiler/aco_opcodes_cpp.py | 10 ++++++++++
 3 files changed, 44 insertions(+)

diff --git a/src/amd/compiler/aco_ir.h b/src/amd/compiler/aco_ir.h
index 988ae6195f1..6ce1e8d0044 100644
--- a/src/amd/compiler/aco_ir.h
+++ b/src/amd/compiler/aco_ir.h
@@ -1566,6 +1566,9 @@ typedef struct {
    const std::bitset<static_cast<int>(aco_opcode::num_opcodes)> is_atomic;
    const char *name[static_cast<int>(aco_opcode::num_opcodes)];
    const aco::Format format[static_cast<int>(aco_opcode::num_opcodes)];
+   /* sizes used for input/output modifiers and constants */
+   const unsigned operand_size[static_cast<int>(aco_opcode::num_opcodes)];
+   const unsigned definition_size[static_cast<int>(aco_opcode::num_opcodes)];
 } Info;
 
 extern const Info instr_info;
diff --git a/src/amd/compiler/aco_opcodes.py b/src/amd/compiler/aco_opcodes.py
index 564a8309ccf..1396b8c3af4 100644
--- a/src/amd/compiler/aco_opcodes.py
+++ b/src/amd/compiler/aco_opcodes.py
@@ -191,6 +191,37 @@ class Opcode(object):
       self.is_atomic = "1" if is_atomic else "0"
       self.format = format
 
+      parts = name.replace('_e64', '').rsplit('_', 2)
+      op_dtype = parts[-1]
+      def_dtype = parts[-2] if len(parts) > 1 else parts[-1]
+      dtype_sizes = {'{}{}'.format(prefix, size) : size for prefix in 'biuf' for size in [64, 32, 24, 16]}
+      self.operand_size = dtype_sizes.get(op_dtype, 0)
+      self.definition_size = dtype_sizes.get(def_dtype, self.operand_size)
+
+      # exceptions
+      if self.operand_size == 24:
+        self.operand_size = 32
+      elif name in ['s_sext_i32_i8', 's_sext_i32_i16', 'v_msad_u8', 'v_cvt_pk_u16_u32', 'v_cvt_pk_i16_i32']:
+         self.operand_size = 32
+      elif name in ['v_qsad_pk_u16_u8', 'v_mqsad_pk_u16_u8', 'v_mqsad_u32_u8']:
+         self.definition_size = 0
+         self.operand_size = 0
+      elif name in ['v_mad_u64_u32', 'v_mad_i64_i32']:
+         self.operand_size = 0
+      elif name.replace('_e64', '') in ['v_lshrrev_b16', 'v_ashrrev_i16', 'v_lshlrev_b16']:
+         # v_lshlrev_b16 tested on GFX10 with 1/2 PI inline constant
+         self.operand_size = 32
+      elif '_pk_' in name or name in ['v_lerp_u8', 'v_sad_u8', 'v_sad_u16',
+                                      'v_cvt_f32_ubyte0', 'v_cvt_f32_ubyte1',
+                                      'v_cvt_f32_ubyte2', 'v_cvt_f32_ubyte3']:
+         self.operand_size = 32
+         self.definition_size = 32
+      elif '_pknorm_' in name:
+         self.definition_size = 32
+      elif format == Format.PSEUDO_REDUCTION:
+         # 64-bit reductions can have a larger definition size, but get_subdword_definition_info() handles that
+         self.definition_size = 32
+
 
 # global dictionary of opcodes
 opcodes = {}
diff --git a/src/amd/compiler/aco_opcodes_cpp.py b/src/amd/compiler/aco_opcodes_cpp.py
index 8309dee18c2..6678b102adc 100644
--- a/src/amd/compiler/aco_opcodes_cpp.py
+++ b/src/amd/compiler/aco_opcodes_cpp.py
@@ -65,6 +65,16 @@ extern const aco::Info instr_info = {
       aco::Format::${str(opcodes[name].format.name)},
       % endfor
    },
+   .operand_size = {
+      % for name in opcode_names:
+      ${opcodes[name].operand_size},
+      % endfor
+   },
+   .definition_size = {
+      % for name in opcode_names:
+      ${opcodes[name].definition_size},
+      % endfor
+   },
 };
 
 }