Skip to content

Commit 14946bd

Browse files
aras-paras-p
committed
Fixed translation of matrix+scalar, matrix-scalar, matrix/scalar operations on Metal.
1 parent 0f4cba3 commit 14946bd

File tree

3 files changed

+93
-12
lines changed

3 files changed

+93
-12
lines changed

‎Changelog.md‎

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -8,6 +8,7 @@ Fixes:
88

99
* Fixed some cases of different precision matrix assignments being miscompiled on Metal.
1010
* Fixed yet more issues with translation of weird loops.
11+
* Fixed translation of matrix+scalar, matrix-scalar, matrix/scalar operations on Metal.
1112

1213

1314
2015 05

‎src/glsl/ir_print_metal_visitor.cpp‎

Lines changed: 78 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -101,6 +101,7 @@ struct metal_print_context
101101
, paramsStr(ralloc_strdup(buffer, ""))
102102
, writingParams(false)
103103
, matrixCastsDone(false)
104+
, matrixConstructorsDone(false)
104105
, shadowSamplerDone(false)
105106
, textureCounter(0)
106107
, attributeCounter(0)
@@ -118,6 +119,7 @@ struct metal_print_context
118119
string_buffer paramsStr;
119120
bool writingParams;
120121
bool matrixCastsDone;
122+
bool matrixConstructorsDone;
121123
bool shadowSamplerDone;
122124
int textureCounter;
123125
int attributeCounter;
@@ -940,21 +942,57 @@ void ir_print_metal_visitor::visit(ir_expression *ir)
940942
bool op0cast = ir->operands[0] && is_different_precision(arg_prec, ir->operands[0]->get_precision());
941943
bool op1cast = ir->operands[1] && is_different_precision(arg_prec, ir->operands[1]->get_precision());
942944
bool op2cast = ir->operands[2] && is_different_precision(arg_prec, ir->operands[2]->get_precision());
945+
const bool op0matrix = ir->operands[0] && ir->operands[0]->type->is_matrix();
946+
const bool op1matrix = ir->operands[1] && ir->operands[1]->type->is_matrix();
947+
bool op0castTo1 = false;
948+
bool op1castTo0 = false;
943949

944950
// Metal does not support matrix precision casts, so when any of the arguments is a matrix,
945951
// take precision from it. This isn't fully robust now, but oh well.
946-
if (op0cast && ir->operands[0]->type->is_matrix() && !op1cast)
952+
if (op0cast && op0matrix && !op1cast)
947953
{
948954
op0cast = false;
949955
arg_prec = ir->operands[0]->get_precision();
950956
op1cast = ir->operands[1] && is_different_precision(arg_prec, ir->operands[1]->get_precision());
951957
}
952-
if (op1cast && ir->operands[1]->type->is_matrix() && !op0cast)
958+
if (op1cast && op1matrix && !op0cast)
953959
{
954960
op1cast = false;
955961
arg_prec = ir->operands[1]->get_precision();
956962
op0cast = ir->operands[0] && is_different_precision(arg_prec, ir->operands[0]->get_precision());
957963
}
964+
965+
// Metal does not have matrix+scalar and matrix-scalar operations; we need to create matrices
966+
// out of the non-matrix argument.
967+
if (ir->operation == ir_binop_add || ir->operation == ir_binop_sub)
968+
{
969+
if (op0matrix && !op1matrix)
970+
{
971+
op1cast = true;
972+
op1castTo0 = true;
973+
}
974+
if (op1matrix && !op0matrix)
975+
{
976+
op0cast = true;
977+
op0castTo1 = true;
978+
}
979+
if (op1castTo0 || op0castTo1)
980+
{
981+
if (!ctx.matrixConstructorsDone)
982+
{
983+
ctx.prefixStr.asprintf_append(
984+
"inline float4x4 _xlinit_float4x4(float v) { return float4x4(float4(v), float4(v), float4(v), float4(v)); }\n"
985+
"inline float3x3 _xlinit_float3x3(float v) { return float3x3(float3(v), float3(v), float3(v)); }\n"
986+
"inline float2x2 _xlinit_float2x2(float v) { return float2x2(float2(v), float2(v)); }\n"
987+
"inline half4x4 _xlinit_half4x4(half v) { return half4x4(half4(v), half4(v), half4(v), half4(v)); }\n"
988+
"inline half3x3 _xlinit_half3x3(half v) { return half3x3(half3(v), half3(v), half3(v)); }\n"
989+
"inline half2x2 _xlinit_half2x2(half v) { return half2x2(half2(v), half2(v)); }\n"
990+
);
991+
ctx.matrixConstructorsDone = true;
992+
}
993+
}
994+
}
995+
958996

959997
const bool rescast = is_different_precision(arg_prec, res_prec) && !ir->type->is_boolean();
960998
if (rescast)
@@ -1000,6 +1038,7 @@ void ir_print_metal_visitor::visit(ir_expression *ir)
10001038
}
10011039
else if (is_binop_func_like(ir->operation, ir->type))
10021040
{
1041+
// binary operation that must be printed like a function, "foo(a,b)"
10031042
if (ir->operation == ir_binop_mod)
10041043
{
10051044
buffer.asprintf_append ("(");
@@ -1025,23 +1064,58 @@ void ir_print_metal_visitor::visit(ir_expression *ir)
10251064
if (ir->operation == ir_binop_mod)
10261065
buffer.asprintf_append ("))");
10271066
}
1067+
else if (ir->get_num_operands() == 2 && ir->operation == ir_binop_div && op0matrix && !op1matrix)
1068+
{
1069+
// "matrix/scalar" - Metal does not have it, so print multiply by inverse instead
1070+
buffer.asprintf_append ("(");
1071+
ir->operands[0]->accept(this);
1072+
const bool halfCast = (arg_prec == glsl_precision_medium || arg_prec == glsl_precision_low);
1073+
buffer.asprintf_append (halfCast ? " * (1.0h/half(" : " * (1.0/(");
1074+
ir->operands[1]->accept(this);
1075+
buffer.asprintf_append (")))");
1076+
}
10281077
else if (ir->get_num_operands() == 2)
10291078
{
1079+
// regular binary operator
10301080
buffer.asprintf_append ("(");
10311081
if (ir->operands[0])
10321082
{
1033-
if (op0cast)
1083+
if (op0castTo1)
1084+
{
1085+
buffer.asprintf_append ("_xlinit_");
1086+
print_type_precision(buffer, ir->operands[1]->type, arg_prec, false);
1087+
buffer.asprintf_append ("(");
1088+
}
1089+
else if (op0cast)
1090+
{
10341091
print_cast (buffer, arg_prec, ir->operands[0]);
1092+
}
10351093
ir->operands[0]->accept(this);
1094+
if (op0castTo1)
1095+
{
1096+
buffer.asprintf_append (")");
1097+
}
10361098
}
10371099

10381100
buffer.asprintf_append (" %s ", operator_glsl_strs[ir->operation]);
10391101

10401102
if (ir->operands[1])
10411103
{
1042-
if (op1cast)
1104+
if (op1castTo0)
1105+
{
1106+
buffer.asprintf_append ("_xlinit_");
1107+
print_type_precision(buffer, ir->operands[0]->type, arg_prec, false);
1108+
buffer.asprintf_append ("(");
1109+
}
1110+
else if (op1cast)
1111+
{
10431112
print_cast (buffer, arg_prec, ir->operands[1]);
1113+
}
10441114
ir->operands[1]->accept(this);
1115+
if (op1castTo0)
1116+
{
1117+
buffer.asprintf_append (")");
1118+
}
10451119
}
10461120
buffer.asprintf_append (")");
10471121
}

‎tests/fragment/matrix-ops-outES3Metal.txt‎

Lines changed: 14 additions & 8 deletions
Original file line numberDiff line numberDiff line change
@@ -1,5 +1,11 @@
11
#include <metal_stdlib>
22
using namespace metal;
3+
inline float4x4 _xlinit_float4x4(float v) { return float4x4(float4(v), float4(v), float4(v), float4(v)); }
4+
inline float3x3 _xlinit_float3x3(float v) { return float3x3(float3(v), float3(v), float3(v)); }
5+
inline float2x2 _xlinit_float2x2(float v) { return float2x2(float2(v), float2(v)); }
6+
inline half4x4 _xlinit_half4x4(half v) { return half4x4(half4(v), half4(v), half4(v), half4(v)); }
7+
inline half3x3 _xlinit_half3x3(half v) { return half3x3(half3(v), half3(v), half3(v)); }
8+
inline half2x2 _xlinit_half2x2(half v) { return half2x2(half2(v), half2(v)); }
39
inline float4x4 _xlcast_float4x4(half4x4 v) { return float4x4(float4(v[0]), float4(v[1]), float4(v[2]), float4(v[3])); }
410
inline float3x3 _xlcast_float3x3(half3x3 v) { return float3x3(float3(v[0]), float3(v[1]), float3(v[2])); }
511
inline float2x2 _xlcast_float2x2(half2x2 v) { return float2x2(float2(v[0]), float2(v[1])); }
@@ -33,21 +39,21 @@ fragment xlatMtlShaderOutput xlatMtlMain (xlatMtlShaderInput _mtl_i [[stage_in]]
3339
tmpvar_6[1].x = -(tmpvar_3);
3440
tmpvar_6[1].y = tmpvar_5;
3541
rotationMatrix_2 = (tmpvar_6 * 2.0);
36-
rotationMatrix_2 = (rotationMatrix_2 - 1.0);
37-
rotationMatrix_2 = (tmpvar_3 - rotationMatrix_2);
38-
rotationMatrix_2 = (tmpvar_5 + rotationMatrix_2);
39-
rotationMatrix_2 = (rotationMatrix_2 / tmpvar_3);
42+
rotationMatrix_2 = (rotationMatrix_2 - _xlinit_float2x2(1.0));
43+
rotationMatrix_2 = (_xlinit_float2x2(tmpvar_3) - rotationMatrix_2);
44+
rotationMatrix_2 = (_xlinit_float2x2(tmpvar_5) + rotationMatrix_2);
45+
rotationMatrix_2 = (rotationMatrix_2 * (1.0/(tmpvar_3)));
4046
float2x2 tmpvar_7;
4147
tmpvar_7[0].x = tmpvar_5;
4248
tmpvar_7[0].y = tmpvar_3;
4349
tmpvar_7[1].x = -(tmpvar_3);
4450
tmpvar_7[1].y = tmpvar_5;
4551
halfMatrix_1 = _xlcast_half2x2(tmpvar_7);
4652
halfMatrix_1 = (halfMatrix_1 * (half)2.0);
47-
halfMatrix_1 = (halfMatrix_1 - (half)1.0);
48-
halfMatrix_1 = _xlcast_half2x2(((float2x2)((half)tmpvar_3 - halfMatrix_1)));
49-
halfMatrix_1 = _xlcast_half2x2(((float2x2)((half)tmpvar_5 + halfMatrix_1)));
50-
halfMatrix_1 = _xlcast_half2x2(((float2x2)(halfMatrix_1 / (half)tmpvar_3)));
53+
halfMatrix_1 = (halfMatrix_1 - _xlinit_half2x2(1.0));
54+
halfMatrix_1 = _xlcast_half2x2(((float2x2)(_xlinit_half2x2(tmpvar_3) - halfMatrix_1)));
55+
halfMatrix_1 = _xlcast_half2x2(((float2x2)(_xlinit_half2x2(tmpvar_5) + halfMatrix_1)));
56+
halfMatrix_1 = _xlcast_half2x2(((float2x2)(halfMatrix_1 * (1.0h/half(tmpvar_3)))));
5157
float4 tmpvar_8;
5258
tmpvar_8.xy = (rotationMatrix_2 * _mtl_i.uv);
5359
tmpvar_8.zw = ((float2)(halfMatrix_1 * (half2)_mtl_i.uv));

0 commit comments

Comments
 (0)