From d8dd49419aebf025fece1a2342dd880e7504c2e5 Mon Sep 17 00:00:00 2001 From: Ian Romanick Date: Tue, 12 Jul 2016 12:44:18 -0700 Subject: [PATCH] glsl: Generate code for constant ir_binop_mul expressions v2: 'for (a, b) in d' => 'for a, b in d'. Suggested by Dylan. Signed-off-by: Ian Romanick Reviewed-by: Matt Turner Acked-by: Dylan Baker --- src/compiler/glsl/ir_expression_operation.py | 56 +++++++++++++++++++- 1 file changed, 54 insertions(+), 2 deletions(-) diff --git a/src/compiler/glsl/ir_expression_operation.py b/src/compiler/glsl/ir_expression_operation.py index de3b866bb96..7d82ea80dd2 100644 --- a/src/compiler/glsl/ir_expression_operation.py +++ b/src/compiler/glsl/ir_expression_operation.py @@ -190,6 +190,56 @@ constant_template_vector_scalar = mako.template.Template("""\ } break;""") +# This template is for multiplication. It is unique because it has to support +# matrix * vector and matrix * matrix operations, and those are just different. +constant_template_mul = mako.template.Template("""\ + case ${op.get_enum_name()}: + /* Check for equal types, or unequal types involving scalars */ + if ((op[0]->type == op[1]->type && !op[0]->type->is_matrix()) + || op0_scalar || op1_scalar) { + for (unsigned c = 0, c0 = 0, c1 = 0; + c < components; + c0 += c0_inc, c1 += c1_inc, c++) { + + switch (op[0]->type->base_type) { + % for dst_type, src_types in op.signatures(): + case ${src_types[0].glsl_type}: + data.${dst_type.union_field}[c] = ${op.get_c_expression(src_types, ("c0", "c1", "c2"))}; + break; + % endfor + default: + assert(0); + } + } + } else { + assert(op[0]->type->is_matrix() || op[1]->type->is_matrix()); + + /* Multiply an N-by-M matrix with an M-by-P matrix. Since either + * matrix can be a GLSL vector, either N or P can be 1. + * + * For vec*mat, the vector is treated as a row vector. This + * means the vector is a 1-row x M-column matrix. + * + * For mat*vec, the vector is treated as a column vector. Since + * matrix_columns is 1 for vectors, this just works. + */ + const unsigned n = op[0]->type->is_vector() + ? 1 : op[0]->type->vector_elements; + const unsigned m = op[1]->type->vector_elements; + const unsigned p = op[1]->type->matrix_columns; + for (unsigned j = 0; j < p; j++) { + for (unsigned i = 0; i < n; i++) { + for (unsigned k = 0; k < m; k++) { + if (op[0]->type->base_type == GLSL_TYPE_DOUBLE) + data.d[i+n*j] += op[0]->value.d[i+n*k]*op[1]->value.d[k+m*j]; + else + data.f[i+n*j] += op[0]->value.f[i+n*k]*op[1]->value.f[k+m*j]; + } + } + } + } + break;""") + # This template is for operations that are horizontal and either have only a # single type or the implementation for all types is identical. That is, the # operation consumes a vector and produces a scalar. @@ -285,7 +335,9 @@ class operation(object): else: return constant_template3.render(op=self) elif self.num_operands == 2: - if vector_scalar_operation in self.flags: + if self.name == "mul": + return constant_template_mul.render(op=self) + elif vector_scalar_operation in self.flags: return constant_template_vector_scalar.render(op=self) elif horizontal_operation in self.flags and types_identical_operation in self.flags: return constant_template_horizontal_single_implementation.render(op=self) @@ -454,7 +506,7 @@ ir_expression_operation = [ operation("add", 2, printable_name="+", source_types=numeric_types, c_expression="{src0} + {src1}", flags=vector_scalar_operation), operation("sub", 2, printable_name="-", source_types=numeric_types, c_expression="{src0} - {src1}", flags=vector_scalar_operation), # "Floating-point or low 32-bit integer multiply." - operation("mul", 2, printable_name="*"), + operation("mul", 2, printable_name="*", source_types=numeric_types, c_expression="{src0} * {src1}"), operation("imul_high", 2), # Calculates the high 32-bits of a 64-bit multiply. operation("div", 2, printable_name="/", source_types=numeric_types, c_expression={'u': "{src1} == 0 ? 0 : {src0} / {src1}", 'i': "{src1} == 0 ? 0 : {src0} / {src1}", 'default': "{src0} / {src1}"}, flags=vector_scalar_operation), -- 2.30.2