[Mesa-dev] [PATCH 44/51] glsl: WIP: Add lowering pass for treating mediump as float16
Topi Pohjolainen
topi.pohjolainen at gmail.com
Fri Nov 24 12:27:11 UTC 2017
At least the following need more thought:
1) Converting right-hand-side of assignments from 16-bits to 32-bits
- More correct thing to do is to treat rhs as 32-bits latest in the
expression producing the value
2) Texture arguments except coordinates are not handled at all
- Moreover, coordinates are always converted into 32-bits due to
logic missing in the Intel compiler backend.
Signed-off-by: Topi Pohjolainen <topi.pohjolainen at intel.com>
---
src/compiler/Makefile.sources | 1 +
src/compiler/glsl/ir_optimization.h | 1 +
src/compiler/glsl/lower_mediump.cpp | 273 ++++++++++++++++++++++++++++++++++++
3 files changed, 275 insertions(+)
create mode 100644 src/compiler/glsl/lower_mediump.cpp
diff --git a/src/compiler/Makefile.sources b/src/compiler/Makefile.sources
index 2ab8e163a2..47bde4fb78 100644
--- a/src/compiler/Makefile.sources
+++ b/src/compiler/Makefile.sources
@@ -94,6 +94,7 @@ LIBGLSL_FILES = \
glsl/lower_int64.cpp \
glsl/lower_jumps.cpp \
glsl/lower_mat_op_to_vec.cpp \
+ glsl/lower_mediump.cpp \
glsl/lower_noise.cpp \
glsl/lower_offset_array.cpp \
glsl/lower_packed_varyings.cpp \
diff --git a/src/compiler/glsl/ir_optimization.h b/src/compiler/glsl/ir_optimization.h
index 2b8c195151..09c4d664e0 100644
--- a/src/compiler/glsl/ir_optimization.h
+++ b/src/compiler/glsl/ir_optimization.h
@@ -132,6 +132,7 @@ bool do_vec_index_to_swizzle(exec_list *instructions);
bool lower_discard(exec_list *instructions);
void lower_discard_flow(exec_list *instructions);
bool lower_instructions(exec_list *instructions, unsigned what_to_lower);
+bool lower_mediump(struct gl_linked_shader *shader);
bool lower_noise(exec_list *instructions);
bool lower_variable_index_to_cond_assign(gl_shader_stage stage,
exec_list *instructions, bool lower_input, bool lower_output,
diff --git a/src/compiler/glsl/lower_mediump.cpp b/src/compiler/glsl/lower_mediump.cpp
new file mode 100644
index 0000000000..89eed8b294
--- /dev/null
+++ b/src/compiler/glsl/lower_mediump.cpp
@@ -0,0 +1,273 @@
+/*
+ * Copyright 2017 Intel Corporation
+ *
+ * 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.
+ */
+
+/**
+ * \file lower_mediump.cpp
+ *
+ */
+
+#include "compiler/glsl_types.h"
+#include "ir.h"
+#include "ir_rvalue_visitor.h"
+#include "ast.h"
+
+static const glsl_type *
+get_mediump(const glsl_type *highp)
+{
+ if (highp->is_float())
+ return glsl_type::get_instance(GLSL_TYPE_FLOAT16,
+ highp->vector_elements,
+ highp->matrix_columns);
+
+ if (highp->is_array() && highp->fields.array->is_float())
+ return glsl_type::get_array_instance(
+ glsl_type::get_instance(GLSL_TYPE_FLOAT16,
+ highp->fields.array->vector_elements,
+ highp->fields.array->matrix_columns),
+ highp->length);
+
+ return highp;
+}
+
+static bool
+is_16_bit(const ir_rvalue *ir)
+{
+ return ir->type->get_scalar_type()->base_type == GLSL_TYPE_FLOAT16;
+}
+
+static bool
+refers_16_bit_float(const ir_rvalue *ir)
+{
+ ir_variable *var = ir->variable_referenced();
+
+ /* Only variables have the mediump property, constants need conversion. */
+ if (!var)
+ return false;
+
+ return var->type->get_scalar_type()->base_type == GLSL_TYPE_FLOAT16;
+}
+
+static ir_rvalue *
+convert(ir_rvalue *ir, enum ir_expression_operation op)
+{
+ if (ir->ir_type == ir_type_constant) {
+ assert(op == ir_unop_f2h);
+ ir->type = get_mediump(ir->type);
+ return ir;
+ }
+
+ void *ctx = ralloc_parent(ir);
+ return new(ctx) ir_expression(op, ir);
+}
+
+class lower_mediump_visitor : public ir_rvalue_visitor {
+public:
+ lower_mediump_visitor() : progress(false) {}
+
+ virtual ir_visitor_status visit(ir_variable *ir);
+ virtual ir_visitor_status visit(ir_dereference_variable *ir);
+
+ virtual ir_visitor_status visit_leave(ir_assignment *ir);
+ virtual ir_visitor_status visit_leave(ir_expression *ir);
+ virtual ir_visitor_status visit_leave(ir_swizzle *ir);
+
+ virtual void handle_rvalue(ir_rvalue **rvalue);
+
+ bool progress;
+
+private:
+ bool can_be_lowered(const ir_variable *var) const;
+
+ void retype_to_float16(const glsl_type **t);
+};
+
+bool
+lower_mediump_visitor::can_be_lowered(const ir_variable *var) const
+{
+ if (!var->type->get_scalar_type()->is_float())
+ return false;
+
+ return var->data.precision == ast_precision_low ||
+ var->data.precision == ast_precision_medium;
+}
+
+void
+lower_mediump_visitor::retype_to_float16(const glsl_type **t)
+{
+ const glsl_type * const mediump = get_mediump(*t);
+ progress = mediump != *t;
+ *t = mediump;
+}
+
+ir_visitor_status
+lower_mediump_visitor::visit(ir_variable *ir)
+{
+ if (!can_be_lowered(ir))
+ return visit_continue;
+
+ const glsl_type * const mediump = get_mediump(ir->type);
+ progress = mediump != ir->type;
+ ir->type = mediump;
+
+ return visit_continue;
+}
+
+ir_visitor_status
+lower_mediump_visitor::visit(ir_dereference_variable *ir)
+{
+ if (!refers_16_bit_float(ir))
+ return visit_continue;
+
+ const glsl_type * const mediump = get_mediump(ir->type);
+ progress = mediump != ir->type;
+ ir->type = mediump;
+
+ return visit_continue;
+}
+
+ir_visitor_status
+lower_mediump_visitor::visit_leave(ir_assignment *ir)
+{
+ ir_rvalue_visitor::visit_leave(ir);
+
+ if (!ir->lhs || !ir->rhs)
+ return visit_continue;
+
+ ir_variable *lhs = ir->lhs->variable_referenced();
+ if (lhs == NULL)
+ return visit_continue;
+
+ /* Consider if the dereference targets 16-bit variable. In such case lower
+ * the dereference type and emit conversion for the source if needed.
+ */
+ if (refers_16_bit_float(ir->lhs)) {
+ retype_to_float16(&ir->lhs->type);
+
+ /* Consider if source needs conversion to float16. */
+ if (!is_16_bit(ir->rhs))
+ ir->rhs = convert(ir->rhs, ir_unop_f2h);
+
+ return visit_continue;
+ }
+
+ /* If in turn target is 32-bit but source is 16-bit, emit conversion. */
+ if (is_16_bit(ir->rhs))
+ ir->rhs = convert(ir->rhs, ir_unop_h2f);
+
+ return visit_continue;
+}
+
+ir_visitor_status
+lower_mediump_visitor::visit_leave(ir_swizzle *ir)
+{
+ ir_rvalue_visitor::visit_leave(ir);
+
+ if (is_16_bit(ir->val))
+ retype_to_float16(&ir->type);
+
+ return visit_continue;
+}
+
+/* Consider expressions in depth-first order according to the rules found in
+ * GLES 3.2 Specification, 4.7.3 Precision Qualifiers:
+ *
+ * In cases where operands do not have a precision qualifier, the precision
+ * qualification will come from the other operands. If no operands have a
+ * precision qualifier, then the precision qualifications of the operands of
+ * the next consuming operation in the expression will be used. This rule can
+ * be applied recursively until a precision qualified operand is found. If
+ * necessary, it will also include the precision qualification of l-values for
+ * assignments, of the declared variable for initializers, of formal
+ * parameters for function call arguments, or of function return types for
+ * function return values. If the precision cannot be determined by this
+ * method e.g. if an entire expression is composed only of operands with no
+ * precision qualifier, and the result is not assigned or passed as an
+ * argument, then it is evaluated at the default precision of the type or
+ * greater. When this occurs in the fragment shader, the default precision
+ * must be defined.
+ *
+ * For example, consider the statements:
+ *
+ * uniform highp float h1;
+ * highp float h2 = 2.3 * 4.7; // operation and result are highp precision
+ * mediump float m;
+ * m = 3.7 * h1 * h2; // all operations are highp precision
+ * h2 = m * h1; // operation is highp precision
+ * m = h2 - h1; // operation is highp precision
+ * h2 = m + m; // addition and result at mediump precision
+ * void f(highp float p);
+ * f(3.3); // 3.3 will be passed in at highp precisi
+ */
+ir_visitor_status
+lower_mediump_visitor::visit_leave(ir_expression *ir)
+{
+ ir_rvalue_visitor::visit_leave(ir);
+
+ bool has_32_bit_src = false;
+ bool has_16_bit_src = false;
+ for (unsigned i = 0; i < ir->num_operands; i++) {
+ if (is_16_bit(ir->operands[i]))
+ has_16_bit_src = true;
+ else
+ has_32_bit_src = true;
+ }
+
+ if (!has_16_bit_src)
+ return visit_continue;
+
+ /* If there are no 32-bit source operands the expression can be evaluated
+ * in lower precision.
+ */
+ if (!has_32_bit_src &&
+ ir->operation != ir_triop_lrp) {
+ retype_to_float16(&ir->type);
+ return visit_continue;
+ }
+
+ /* All other 16-bit operands need to be converted to 32-bits in order to
+ * perform the operation in 32-bits.
+ */
+ for (unsigned i = 0; i < ir->num_operands; i++) {
+ if (is_16_bit(ir->operands[i]))
+ ir->operands[i] = convert(ir->operands[i], ir_unop_h2f);
+ }
+
+ return visit_continue;
+}
+
+void
+lower_mediump_visitor::handle_rvalue(ir_rvalue **rvalue)
+{
+ if (*rvalue == NULL || !refers_16_bit_float(*rvalue))
+ return;
+
+ retype_to_float16(&(*rvalue)->type);
+}
+
+bool
+lower_mediump(struct gl_linked_shader *shader)
+{
+ lower_mediump_visitor v;
+ visit_list_elements(&v, shader->ir);
+ return v.progress;
+}
--
2.11.0
More information about the mesa-dev
mailing list