glsl: Add support for float16 types in the IR tree

Reviewed-by: Marek Olšák <[email protected]>
Reviewed-by: Alyssa Rosenzweig <[email protected]>
Reviewed-by: Ian Romanick <[email protected]>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/3929>

5 files changed, 86 insertions, 12 deletions

diff --git a/src/compiler/glsl/glsl_to_nir.cpp b/src/compiler/glsl/glsl_to_nir.cpp
index fed456a36ea..8f15cd55fcb 100644
--- a/src/compiler/glsl/glsl_to_nir.cpp
+++ b/src/compiler/glsl/glsl_to_nir.cpp
@@ -316,6 +316,7 @@ nir_visitor::constant_copy(ir_constant *ir, void *mem_ctx)
       break;
 
    case GLSL_TYPE_FLOAT:
+   case GLSL_TYPE_FLOAT16:
    case GLSL_TYPE_DOUBLE:
       if (cols > 1) {
          ret->elements = ralloc_array(mem_ctx, nir_constant *, cols);
@@ -329,6 +330,11 @@ nir_visitor::constant_copy(ir_constant *ir, void *mem_ctx)
                   col_const->values[r].f32 = ir->value.f[c * rows + r];
                break;
 
+            case GLSL_TYPE_FLOAT16:
+               for (unsigned r = 0; r < rows; r++)
+                  col_const->values[r].u16 = ir->value.f16[c * rows + r];
+               break;
+
             case GLSL_TYPE_DOUBLE:
                for (unsigned r = 0; r < rows; r++)
                   col_const->values[r].f64 = ir->value.d[c * rows + r];
@@ -346,6 +352,11 @@ nir_visitor::constant_copy(ir_constant *ir, void *mem_ctx)
                ret->values[r].f32 = ir->value.f[r];
             break;
 
+         case GLSL_TYPE_FLOAT16:
+            for (unsigned r = 0; r < rows; r++)
+               ret->values[r].u16 = ir->value.f16[r];
+            break;
+
          case GLSL_TYPE_DOUBLE:
             for (unsigned r = 0; r < rows; r++)
                ret->values[r].f64 = ir->value.d[r];
diff --git a/src/compiler/glsl/ir.cpp b/src/compiler/glsl/ir.cpp
index 57551850757..d9bc83501cd 100644
--- a/src/compiler/glsl/ir.cpp
+++ b/src/compiler/glsl/ir.cpp
@@ -22,6 +22,7 @@
  */
 #include <string.h>
 #include "ir.h"
+#include "util/half_float.h"
 #include "compiler/glsl_types.h"
 #include "glsl_parser_extras.h"
 
@@ -796,6 +797,7 @@ ir_constant::ir_constant(const ir_constant *c, unsigned i)
    case GLSL_TYPE_UINT:  this->value.u[0] = c->value.u[i]; break;
    case GLSL_TYPE_INT:   this->value.i[0] = c->value.i[i]; break;
    case GLSL_TYPE_FLOAT: this->value.f[0] = c->value.f[i]; break;
+   case GLSL_TYPE_FLOAT16: this->value.f16[0] = c->value.f16[i]; break;
    case GLSL_TYPE_BOOL:  this->value.b[0] = c->value.b[i]; break;
    case GLSL_TYPE_DOUBLE: this->value.d[0] = c->value.d[i]; break;
    default:              assert(!"Should not get here."); break;
@@ -841,14 +843,23 @@ ir_constant::ir_constant(const struct glsl_type *type, exec_list *value_list)
    if (value->type->is_scalar() && value->next->is_tail_sentinel()) {
       if (type->is_matrix()) {
 	 /* Matrix - fill diagonal (rest is already set to 0) */
-         assert(type->is_float() || type->is_double());
          for (unsigned i = 0; i < type->matrix_columns; i++) {
-            if (type->is_float())
+            switch (type->base_type) {
+            case GLSL_TYPE_FLOAT:
                this->value.f[i * type->vector_elements + i] =
                   value->value.f[0];
-            else
+               break;
+            case GLSL_TYPE_DOUBLE:
                this->value.d[i * type->vector_elements + i] =
                   value->value.d[0];
+               break;
+            case GLSL_TYPE_FLOAT16:
+               this->value.f16[i * type->vector_elements + i] =
+                  value->value.f16[0];
+               break;
+            default:
+               assert(!"unexpected matrix base type");
+            }
          }
       } else {
 	 /* Vector or scalar - fill all components */
@@ -862,6 +873,10 @@ ir_constant::ir_constant(const struct glsl_type *type, exec_list *value_list)
 	    for (unsigned i = 0; i < type->components(); i++)
 	       this->value.f[i] = value->value.f[0];
 	    break;
+	 case GLSL_TYPE_FLOAT16:
+	    for (unsigned i = 0; i < type->components(); i++)
+	       this->value.f16[i] = value->value.f16[0];
+	    break;
 	 case GLSL_TYPE_DOUBLE:
 	    for (unsigned i = 0; i < type->components(); i++)
 	       this->value.d[i] = value->value.d[0];
@@ -931,6 +946,9 @@ ir_constant::ir_constant(const struct glsl_type *type, exec_list *value_list)
 	 case GLSL_TYPE_FLOAT:
 	    this->value.f[i] = value->get_float_component(j);
 	    break;
+	 case GLSL_TYPE_FLOAT16:
+	    this->value.f16[i] = value->get_float16_component(j);
+	    break;
 	 case GLSL_TYPE_BOOL:
 	    this->value.b[i] = value->get_bool_component(j);
 	    break;
@@ -996,6 +1014,7 @@ ir_constant::get_bool_component(unsigned i) const
    case GLSL_TYPE_UINT:  return this->value.u[i] != 0;
    case GLSL_TYPE_INT:   return this->value.i[i] != 0;
    case GLSL_TYPE_FLOAT: return ((int)this->value.f[i]) != 0;
+   case GLSL_TYPE_FLOAT16: return ((int)_mesa_half_to_float(this->value.f16[i])) != 0;
    case GLSL_TYPE_BOOL:  return this->value.b[i];
    case GLSL_TYPE_DOUBLE: return this->value.d[i] != 0.0;
    case GLSL_TYPE_SAMPLER:
@@ -1018,6 +1037,7 @@ ir_constant::get_float_component(unsigned i) const
    case GLSL_TYPE_UINT:  return (float) this->value.u[i];
    case GLSL_TYPE_INT:   return (float) this->value.i[i];
    case GLSL_TYPE_FLOAT: return this->value.f[i];
+   case GLSL_TYPE_FLOAT16: return _mesa_half_to_float(this->value.f16[i]);
    case GLSL_TYPE_BOOL:  return this->value.b[i] ? 1.0f : 0.0f;
    case GLSL_TYPE_DOUBLE: return (float) this->value.d[i];
    case GLSL_TYPE_SAMPLER:
@@ -1033,6 +1053,15 @@ ir_constant::get_float_component(unsigned i) const
    return 0.0;
 }
 
+uint16_t
+ir_constant::get_float16_component(unsigned i) const
+{
+   if (this->type->base_type == GLSL_TYPE_FLOAT16)
+      return this->value.f16[i];
+   else
+      return _mesa_float_to_half(get_float_component(i));
+}
+
 double
 ir_constant::get_double_component(unsigned i) const
 {
@@ -1040,6 +1069,7 @@ ir_constant::get_double_component(unsigned i) const
    case GLSL_TYPE_UINT:  return (double) this->value.u[i];
    case GLSL_TYPE_INT:   return (double) this->value.i[i];
    case GLSL_TYPE_FLOAT: return (double) this->value.f[i];
+   case GLSL_TYPE_FLOAT16: return (double) _mesa_half_to_float(this->value.f16[i]);
    case GLSL_TYPE_BOOL:  return this->value.b[i] ? 1.0 : 0.0;
    case GLSL_TYPE_DOUBLE: return this->value.d[i];
    case GLSL_TYPE_SAMPLER:
@@ -1062,6 +1092,7 @@ ir_constant::get_int_component(unsigned i) const
    case GLSL_TYPE_UINT:  return this->value.u[i];
    case GLSL_TYPE_INT:   return this->value.i[i];
    case GLSL_TYPE_FLOAT: return (int) this->value.f[i];
+   case GLSL_TYPE_FLOAT16: return (int) _mesa_half_to_float(this->value.f16[i]);
    case GLSL_TYPE_BOOL:  return this->value.b[i] ? 1 : 0;
    case GLSL_TYPE_DOUBLE: return (int) this->value.d[i];
    case GLSL_TYPE_SAMPLER:
@@ -1084,6 +1115,7 @@ ir_constant::get_uint_component(unsigned i) const
    case GLSL_TYPE_UINT:  return this->value.u[i];
    case GLSL_TYPE_INT:   return this->value.i[i];
    case GLSL_TYPE_FLOAT: return (unsigned) this->value.f[i];
+   case GLSL_TYPE_FLOAT16: return (unsigned) _mesa_half_to_float(this->value.f16[i]);
    case GLSL_TYPE_BOOL:  return this->value.b[i] ? 1 : 0;
    case GLSL_TYPE_DOUBLE: return (unsigned) this->value.d[i];
    case GLSL_TYPE_SAMPLER:
@@ -1106,6 +1138,7 @@ ir_constant::get_int64_component(unsigned i) const
    case GLSL_TYPE_UINT:  return this->value.u[i];
    case GLSL_TYPE_INT:   return this->value.i[i];
    case GLSL_TYPE_FLOAT: return (int64_t) this->value.f[i];
+   case GLSL_TYPE_FLOAT16: return (int64_t) _mesa_half_to_float(this->value.f16[i]);
    case GLSL_TYPE_BOOL:  return this->value.b[i] ? 1 : 0;
    case GLSL_TYPE_DOUBLE: return (int64_t) this->value.d[i];
    case GLSL_TYPE_SAMPLER:
@@ -1128,6 +1161,7 @@ ir_constant::get_uint64_component(unsigned i) const
    case GLSL_TYPE_UINT:  return this->value.u[i];
    case GLSL_TYPE_INT:   return this->value.i[i];
    case GLSL_TYPE_FLOAT: return (uint64_t) this->value.f[i];
+   case GLSL_TYPE_FLOAT16: return (uint64_t) _mesa_half_to_float(this->value.f16[i]);
    case GLSL_TYPE_BOOL:  return this->value.b[i] ? 1 : 0;
    case GLSL_TYPE_DOUBLE: return (uint64_t) this->value.d[i];
    case GLSL_TYPE_SAMPLER:
@@ -1182,6 +1216,7 @@ ir_constant::copy_offset(ir_constant *src, int offset)
    case GLSL_TYPE_UINT:
    case GLSL_TYPE_INT:
    case GLSL_TYPE_FLOAT:
+   case GLSL_TYPE_FLOAT16:
    case GLSL_TYPE_DOUBLE:
    case GLSL_TYPE_SAMPLER:
    case GLSL_TYPE_IMAGE:
@@ -1201,6 +1236,9 @@ ir_constant::copy_offset(ir_constant *src, int offset)
 	 case GLSL_TYPE_FLOAT:
 	    value.f[i+offset] = src->get_float_component(i);
 	    break;
+	 case GLSL_TYPE_FLOAT16:
+	    value.f16[i+offset] = src->get_float16_component(i);
+	    break;
 	 case GLSL_TYPE_BOOL:
 	    value.b[i+offset] = src->get_bool_component(i);
 	    break;
@@ -1260,6 +1298,9 @@ ir_constant::copy_masked_offset(ir_constant *src, int offset, unsigned int mask)
 	 case GLSL_TYPE_FLOAT:
 	    value.f[i+offset] = src->get_float_component(id++);
 	    break;
+	 case GLSL_TYPE_FLOAT16:
+	    value.f16[i+offset] = src->get_float16_component(id++);
+	    break;
 	 case GLSL_TYPE_BOOL:
 	    value.b[i+offset] = src->get_bool_component(id++);
 	    break;
@@ -1310,6 +1351,12 @@ ir_constant::has_value(const ir_constant *c) const
 	 if (this->value.f[i] != c->value.f[i])
 	    return false;
 	 break;
+      case GLSL_TYPE_FLOAT16:
+	/* Convert to float to make sure NaN and ±0.0 compares correctly */
+	 if (_mesa_half_to_float(this->value.f16[i]) !=
+             _mesa_half_to_float(c->value.f16[i]))
+	    return false;
+	 break;
       case GLSL_TYPE_BOOL:
 	 if (this->value.b[i] != c->value.b[i])
 	    return false;
@@ -1353,6 +1400,10 @@ ir_constant::is_value(float f, int i) const
 	 if (this->value.f[c] != f)
 	    return false;
 	 break;
+      case GLSL_TYPE_FLOAT16:
+         if (_mesa_half_to_float(this->value.f16[c]) != f)
+            return false;
+         break;
       case GLSL_TYPE_INT:
 	 if (this->value.i[c] != i)
 	    return false;
diff --git a/src/compiler/glsl/ir.h b/src/compiler/glsl/ir.h
index c43c03c89a7..2e20f4bee7a 100644
--- a/src/compiler/glsl/ir.h
+++ b/src/compiler/glsl/ir.h
@@ -2180,6 +2180,7 @@ union ir_constant_data {
       float f[16];
       bool b[16];
       double d[16];
+      uint16_t f16[16];
       uint64_t u64[16];
       int64_t i64[16];
 };
@@ -2244,6 +2245,7 @@ public:
    /*@{*/
    bool get_bool_component(unsigned i) const;
    float get_float_component(unsigned i) const;
+   uint16_t get_float16_component(unsigned i) const;
    double get_double_component(unsigned i) const;
    int get_int_component(unsigned i) const;
    unsigned get_uint_component(unsigned i) const;
diff --git a/src/compiler/glsl/ir_constant_expression.cpp b/src/compiler/glsl/ir_constant_expression.cpp
index d7f0abe8304..4b80282c266 100644
--- a/src/compiler/glsl/ir_constant_expression.cpp
+++ b/src/compiler/glsl/ir_constant_expression.cpp
@@ -773,6 +773,7 @@ ir_swizzle::constant_expression_value(void *mem_ctx,
          case GLSL_TYPE_UINT:
          case GLSL_TYPE_INT:   data.u[i] = v->value.u[swiz_idx[i]]; break;
          case GLSL_TYPE_FLOAT: data.f[i] = v->value.f[swiz_idx[i]]; break;
+         case GLSL_TYPE_FLOAT16: data.f16[i] = v->value.f16[swiz_idx[i]]; break;
          case GLSL_TYPE_BOOL:  data.b[i] = v->value.b[swiz_idx[i]]; break;
          case GLSL_TYPE_DOUBLE:data.d[i] = v->value.d[swiz_idx[i]]; break;
          case GLSL_TYPE_UINT64:data.u64[i] = v->value.u64[swiz_idx[i]]; break;
diff --git a/src/compiler/glsl/ir_print_visitor.cpp b/src/compiler/glsl/ir_print_visitor.cpp
index 86120e19b19..7533a52ee81 100644
--- a/src/compiler/glsl/ir_print_visitor.cpp
+++ b/src/compiler/glsl/ir_print_visitor.cpp
@@ -28,6 +28,7 @@
 #include "main/macros.h"
 #include "util/hash_table.h"
 #include "util/u_string.h"
+#include "util/half_float.h"
 
 static void print_type(FILE *f, const glsl_type *t);
 
@@ -460,6 +461,19 @@ void ir_print_visitor::visit(ir_assignment *ir)
    fprintf(f, ") ");
 }
 
+static void
+print_float_constant(FILE *f, float val)
+{
+   if (val == 0.0f)
+      /* 0.0 == -0.0, so print with %f to get the proper sign. */
+      fprintf(f, "%f", val);
+   else if (fabs(val) < 0.000001f)
+      fprintf(f, "%a", val);
+   else if (fabs(val) > 1000000.0f)
+      fprintf(f, "%e", val);
+   else
+      fprintf(f, "%f", val);
+}
 
 void ir_print_visitor::visit(ir_constant *ir)
 {
@@ -484,15 +498,10 @@ void ir_print_visitor::visit(ir_constant *ir)
 	 case GLSL_TYPE_UINT:  fprintf(f, "%u", ir->value.u[i]); break;
 	 case GLSL_TYPE_INT:   fprintf(f, "%d", ir->value.i[i]); break;
 	 case GLSL_TYPE_FLOAT:
-            if (ir->value.f[i] == 0.0f)
-               /* 0.0 == -0.0, so print with %f to get the proper sign. */
-               fprintf(f, "%f", ir->value.f[i]);
-            else if (fabs(ir->value.f[i]) < 0.000001f)
-               fprintf(f, "%a", ir->value.f[i]);
-            else if (fabs(ir->value.f[i]) > 1000000.0f)
-               fprintf(f, "%e", ir->value.f[i]);
-            else
-               fprintf(f, "%f", ir->value.f[i]);
+            print_float_constant(f, ir->value.f[i]);
+            break;
+	 case GLSL_TYPE_FLOAT16:
+            print_float_constant(f, _mesa_half_to_float(ir->value.f16[i]));
             break;
 	 case GLSL_TYPE_SAMPLER:
 	 case GLSL_TYPE_IMAGE: