diff options
author | Marek Olšák <[email protected]> | 2020-06-24 08:09:48 -0400 |
---|---|---|
committer | Marek Olšák <[email protected]> | 2020-07-07 22:02:06 -0400 |
commit | 6aea39641a6c32f780c40e3344096d3e0eccb590 (patch) | |
tree | 658f3ff6fab47884ad61637560d2488a6a0a5a6a | |
parent | 7f00d4dac813836d9767efa381c43514b2a999cc (diff) |
glsl: lower mediump temporaries to 16 bits except structures (v2)
Without this, NIR contains non-lowerable 32-bit phis for mediump variables.
Structures are not lowered yet.
v2: add the LowerPrecisionTemporaries option
Reviewed-by: Alyssa Rosenzweig <[email protected]> (v1)
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5746>
-rw-r--r-- | src/compiler/glsl/lower_precision.cpp | 376 | ||||
-rw-r--r-- | src/compiler/glsl/standalone.cpp | 1 | ||||
-rw-r--r-- | src/compiler/glsl/tests/lower_precision_test.py | 647 | ||||
-rw-r--r-- | src/mesa/main/mtypes.h | 6 |
4 files changed, 1028 insertions, 2 deletions
diff --git a/src/compiler/glsl/lower_precision.cpp b/src/compiler/glsl/lower_precision.cpp index b39cda8b800..d03c66a5dc0 100644 --- a/src/compiler/glsl/lower_precision.cpp +++ b/src/compiler/glsl/lower_precision.cpp @@ -949,6 +949,375 @@ find_precision_visitor::~find_precision_visitor() } } +/* Lowering opcodes to 16 bits is not enough for programs with control flow + * (and the ?: operator, which is represented by if-then-else in the IR), + * because temporary variables, which are used for passing values between + * code blocks, are not lowered, resulting in 32-bit phis in NIR. + * + * First change the variable types to 16 bits, then change all ir_dereference + * types to 16 bits. + */ +class lower_variables_visitor : public ir_rvalue_enter_visitor { +public: + lower_variables_visitor(const struct gl_shader_compiler_options *options) + : options(options) { + lower_vars = _mesa_pointer_set_create(NULL); + } + + virtual ~lower_variables_visitor() + { + _mesa_set_destroy(lower_vars, NULL); + } + + virtual ir_visitor_status visit(ir_variable *var); + virtual ir_visitor_status visit_enter(ir_assignment *ir); + virtual ir_visitor_status visit_enter(ir_return *ir); + virtual ir_visitor_status visit_enter(ir_call *ir); + virtual void handle_rvalue(ir_rvalue **rvalue); + + void fix_types_in_deref_chain(ir_dereference *ir); + void convert_split_assignment(ir_dereference *lhs, ir_rvalue *rhs, + bool insert_before); + + const struct gl_shader_compiler_options *options; + set *lower_vars; +}; + +static void +lower_constant(ir_constant *ir) +{ + if (ir->type->is_array()) { + for (int i = 0; i < ir->type->array_size(); i++) + lower_constant(ir->get_array_element(i)); + + ir->type = lower_glsl_type(ir->type); + return; + } + + ir->type = lower_glsl_type(ir->type); + ir_constant_data value; + + if (ir->type->base_type == GLSL_TYPE_FLOAT16) { + for (unsigned i = 0; i < ARRAY_SIZE(value.f16); i++) + value.f16[i] = _mesa_float_to_half(ir->value.f[i]); + } else if (ir->type->base_type == GLSL_TYPE_INT16) { + for (unsigned i = 0; i < ARRAY_SIZE(value.i16); i++) + value.i16[i] = ir->value.i[i]; + } else if (ir->type->base_type == GLSL_TYPE_UINT16) { + for (unsigned i = 0; i < ARRAY_SIZE(value.u16); i++) + value.u16[i] = ir->value.u[i]; + } else { + unreachable("invalid type"); + } + + ir->value = value; +} + +ir_visitor_status +lower_variables_visitor::visit(ir_variable *var) +{ + if ((var->data.mode != ir_var_temporary && + var->data.mode != ir_var_auto) || + !var->type->without_array()->is_32bit() || + (var->data.precision != GLSL_PRECISION_MEDIUM && + var->data.precision != GLSL_PRECISION_LOW) || + !can_lower_type(options, var->type)) + return visit_continue; + + /* Lower constant initializers. */ + if (var->constant_value && + var->type == var->constant_value->type) { + var->constant_value = + var->constant_value->clone(ralloc_parent(var), NULL); + lower_constant(var->constant_value); + } + + if (var->constant_initializer && + var->type == var->constant_initializer->type) { + var->constant_initializer = + var->constant_initializer->clone(ralloc_parent(var), NULL); + lower_constant(var->constant_initializer); + } + + var->type = lower_glsl_type(var->type); + _mesa_set_add(lower_vars, var); + + return visit_continue; +} + +void +lower_variables_visitor::fix_types_in_deref_chain(ir_dereference *ir) +{ + assert(ir->type->without_array()->is_32bit()); + assert(_mesa_set_search(lower_vars, ir->variable_referenced())); + + /* Fix the type in the dereference node. */ + ir->type = lower_glsl_type(ir->type); + + /* If it's an array, fix the types in the whole dereference chain. */ + for (ir_dereference_array *deref_array = ir->as_dereference_array(); + deref_array; + deref_array = deref_array->array->as_dereference_array()) { + assert(deref_array->array->type->without_array()->is_32bit()); + deref_array->array->type = lower_glsl_type(deref_array->array->type); + } +} + +void +lower_variables_visitor::convert_split_assignment(ir_dereference *lhs, + ir_rvalue *rhs, + bool insert_before) +{ + void *mem_ctx = ralloc_parent(lhs); + + if (lhs->type->is_array()) { + for (unsigned i = 0; i < lhs->type->length; i++) { + ir_dereference *l, *r; + + l = new(mem_ctx) ir_dereference_array(lhs->clone(mem_ctx, NULL), + new(mem_ctx) ir_constant(i)); + r = new(mem_ctx) ir_dereference_array(rhs->clone(mem_ctx, NULL), + new(mem_ctx) ir_constant(i)); + convert_split_assignment(l, r, insert_before); + } + return; + } + + assert(lhs->type->is_16bit() || lhs->type->is_32bit()); + assert(rhs->type->is_16bit() || rhs->type->is_32bit()); + assert(lhs->type->is_16bit() != rhs->type->is_16bit()); + + ir_assignment *assign = + new(mem_ctx) ir_assignment(lhs, convert_precision(lhs->type->is_32bit(), rhs)); + + if (insert_before) + base_ir->insert_before(assign); + else + base_ir->insert_after(assign); +} + +ir_visitor_status +lower_variables_visitor::visit_enter(ir_assignment *ir) +{ + ir_dereference *lhs = ir->lhs; + ir_variable *var = lhs->variable_referenced(); + ir_dereference *rhs_deref = ir->rhs->as_dereference(); + ir_variable *rhs_var = rhs_deref ? rhs_deref->variable_referenced() : NULL; + ir_constant *rhs_const = ir->rhs->as_constant(); + + /* Legalize array assignments between lowered and non-lowered variables. */ + if (lhs->type->is_array() && + (rhs_var || rhs_const) && + (!rhs_var || + var->type->without_array()->is_16bit() != + rhs_var->type->without_array()->is_16bit()) && + (!rhs_const || + (var->type->without_array()->is_16bit() && + rhs_const->type->without_array()->is_32bit()))) { + assert(ir->rhs->type->is_array()); + + /* Fix array assignments from lowered to non-lowered. */ + if (rhs_var && _mesa_set_search(lower_vars, rhs_var)) { + fix_types_in_deref_chain(rhs_deref); + /* Convert to 32 bits for LHS. */ + convert_split_assignment(lhs, rhs_deref, true); + ir->remove(); + return visit_continue; + } + + /* Fix array assignments from non-lowered to lowered. */ + if (_mesa_set_search(lower_vars, var) && + ir->rhs->type->without_array()->is_32bit()) { + fix_types_in_deref_chain(lhs); + /* Convert to 16 bits for LHS. */ + convert_split_assignment(lhs, ir->rhs, true); + ir->remove(); + return visit_continue; + } + } + + /* Fix assignment types. */ + if (_mesa_set_search(lower_vars, var)) { + /* Fix the LHS type. */ + if (lhs->type->without_array()->is_32bit()) + fix_types_in_deref_chain(lhs); + + /* Fix the RHS type if it's a lowered variable. */ + if (rhs_var && + _mesa_set_search(lower_vars, rhs_var) && + rhs_deref->type->without_array()->is_32bit()) + fix_types_in_deref_chain(rhs_deref); + + /* Fix the RHS type if it's a non-array expression. */ + if (ir->rhs->type->is_32bit()) { + ir_expression *expr = ir->rhs->as_expression(); + + /* Convert the RHS to the LHS type. */ + if (expr && + (expr->operation == ir_unop_f162f || + expr->operation == ir_unop_i2i || + expr->operation == ir_unop_u2u) && + expr->operands[0]->type->is_16bit()) { + /* If there is an "up" conversion, just remove it. + * This is optional. We could as well execute the else statement and + * let NIR eliminate the up+down conversions. + */ + ir->rhs = expr->operands[0]; + } else { + /* Add a "down" conversion operation to fix the type of RHS. */ + ir->rhs = convert_precision(false, ir->rhs); + } + } + } + + return ir_rvalue_enter_visitor::visit_enter(ir); +} + +ir_visitor_status +lower_variables_visitor::visit_enter(ir_return *ir) +{ + void *mem_ctx = ralloc_parent(ir); + + ir_dereference *deref = ir->value ? ir->value->as_dereference() : NULL; + if (deref) { + ir_variable *var = deref->variable_referenced(); + + /* Fix the type of the return value. */ + if (_mesa_set_search(lower_vars, var) && + deref->type->without_array()->is_32bit()) { + /* Create a 32-bit temporary variable. */ + ir_variable *new_var = + new(mem_ctx) ir_variable(deref->type, "lowerp", ir_var_temporary); + base_ir->insert_before(new_var); + + /* Fix types in dereferences. */ + fix_types_in_deref_chain(deref); + + /* Convert to 32 bits for the return value. */ + convert_split_assignment(new(mem_ctx) ir_dereference_variable(new_var), + deref, true); + ir->value = new(mem_ctx) ir_dereference_variable(new_var); + } + } + + return ir_rvalue_enter_visitor::visit_enter(ir); +} + +void lower_variables_visitor::handle_rvalue(ir_rvalue **rvalue) +{ + ir_rvalue *ir = *rvalue; + + if (in_assignee || ir == NULL) + return; + + ir_expression *expr = ir->as_expression(); + ir_dereference *expr_op0_deref = expr ? expr->operands[0]->as_dereference() : NULL; + + /* Remove f2fmp(float16). Same for int16 and uint16. */ + if (expr && + expr_op0_deref && + (expr->operation == ir_unop_f2fmp || + expr->operation == ir_unop_i2imp || + expr->operation == ir_unop_u2ump || + expr->operation == ir_unop_f2f16 || + expr->operation == ir_unop_i2i || + expr->operation == ir_unop_u2u) && + expr->type->without_array()->is_16bit() && + expr_op0_deref->type->without_array()->is_32bit() && + _mesa_set_search(lower_vars, expr_op0_deref->variable_referenced())) { + fix_types_in_deref_chain(expr_op0_deref); + + /* Remove f2fmp/i2imp/u2ump. */ + *rvalue = expr_op0_deref; + return; + } + + ir_dereference *deref = ir->as_dereference(); + + if (deref) { + ir_variable *var = deref->variable_referenced(); + assert(var); + + if (_mesa_set_search(lower_vars, var) && + deref->type->without_array()->is_32bit()) { + fix_types_in_deref_chain(deref); + + /* Then convert the type up. Optimizations should eliminate this. */ + *rvalue = convert_precision(true, deref); + } + } +} + +ir_visitor_status +lower_variables_visitor::visit_enter(ir_call *ir) +{ + void *mem_ctx = ralloc_parent(ir); + + /* We can't pass 16-bit variables as 32-bit inout/out parameters. */ + foreach_two_lists(formal_node, &ir->callee->parameters, + actual_node, &ir->actual_parameters) { + ir_dereference *param_deref = + ((ir_rvalue *)actual_node)->as_dereference(); + ir_variable *param = (ir_variable *)formal_node; + + if (!param_deref) + continue; + + ir_variable *var = param_deref->variable_referenced(); + + if (_mesa_set_search(lower_vars, var) && + param->type->without_array()->is_32bit()) { + fix_types_in_deref_chain(param_deref); + + /* Create a 32-bit temporary variable for the parameter. */ + ir_variable *new_var = + new(mem_ctx) ir_variable(param->type, "lowerp", ir_var_temporary); + base_ir->insert_before(new_var); + + /* Replace the parameter. */ + actual_node->replace_with(new(mem_ctx) ir_dereference_variable(new_var)); + + if (param->data.mode == ir_var_function_in || + param->data.mode == ir_var_function_inout) { + /* Convert to 32 bits for passing in. */ + convert_split_assignment(new(mem_ctx) ir_dereference_variable(new_var), + param_deref->clone(mem_ctx, NULL), true); + } + if (param->data.mode == ir_var_function_out || + param->data.mode == ir_var_function_inout) { + /* Convert to 16 bits after returning. */ + convert_split_assignment(param_deref, + new(mem_ctx) ir_dereference_variable(new_var), + false); + } + } + } + + /* Fix the type of return value dereferencies. */ + ir_dereference_variable *ret_deref = ir->return_deref; + ir_variable *ret_var = ret_deref ? ret_deref->variable_referenced() : NULL; + + if (ret_var && + _mesa_set_search(lower_vars, ret_var) && + ret_deref->type->without_array()->is_32bit()) { + /* Create a 32-bit temporary variable. */ + ir_variable *new_var = + new(mem_ctx) ir_variable(ir->callee->return_type, "lowerp", + ir_var_temporary); + base_ir->insert_before(new_var); + + /* Replace the return variable. */ + ret_deref->var = new_var; + + /* Convert to 16 bits after returning. */ + convert_split_assignment(new(mem_ctx) ir_dereference_variable(ret_var), + new(mem_ctx) ir_dereference_variable(new_var), + false); + } + + return ir_rvalue_enter_visitor::visit_enter(ir); +} + } void @@ -956,8 +1325,11 @@ lower_precision(const struct gl_shader_compiler_options *options, exec_list *instructions) { find_precision_visitor v(options); - find_lowerable_rvalues(options, instructions, v.lowerable_rvalues); - visit_list_elements(&v, instructions); + + if (options->LowerPrecisionTemporaries) { + lower_variables_visitor vars(options); + visit_list_elements(&vars, instructions); + } } diff --git a/src/compiler/glsl/standalone.cpp b/src/compiler/glsl/standalone.cpp index 5b34297df59..f0305f735bb 100644 --- a/src/compiler/glsl/standalone.cpp +++ b/src/compiler/glsl/standalone.cpp @@ -441,6 +441,7 @@ standalone_compile_shader(const struct standalone_options *_options, options->LowerPrecisionFloat16 = true; options->LowerPrecisionInt16 = true; options->LowerPrecisionDerivatives = true; + options->LowerPrecisionTemporaries = true; } } diff --git a/src/compiler/glsl/tests/lower_precision_test.py b/src/compiler/glsl/tests/lower_precision_test.py index a6ad9503135..2878f8c6c90 100644 --- a/src/compiler/glsl/tests/lower_precision_test.py +++ b/src/compiler/glsl/tests/lower_precision_test.py @@ -1628,6 +1628,653 @@ TESTS = [ } """, r'expression f16vec4 \+ \(expression f16vec4 f2fmp \(expression vec4 unpackSnorm4x8 \(var_ref val'), + Test("f32 csel", + """ + #version 300 es + precision mediump float; + + in vec4 var; + out vec4 color; + + void main() + { + color = (var.x > var.y) ? var : vec4(10.0); + } + """, + r'\(constant +f16vec4 \(10'), + Test("i32 csel", + """ + #version 310 es + precision mediump int; + + in flat ivec4 var; + out ivec4 color; + + void main() + { + color = (var.x > var.y) ? var : ivec4(10); + } + """, + r'\(constant +i16vec4 \(10'), + Test("u32 csel", + """ + #version 310 es + precision mediump int; + + in flat uvec4 var; + out uvec4 color; + + void main() + { + color = (var.x > var.y) ? var : uvec4(10); + } + """, + r'\(constant +u16vec4 \(10'), + Test("f32 loop counter", + """ + #version 300 es + precision mediump float; + + uniform float n, incr; + out float color; + + void main() + { + color = 0.0; + for (float x = 0.0; x < n; x += incr) + color += x; + } + """, + r'\(assign \(x\) \(var_ref x\) \(expression float16_t \+ \(var_ref x\) \(expression float16_t f2fmp \(var_ref incr'), + Test("i32 loop counter", + """ + #version 310 es + precision mediump float; + precision mediump int; + + uniform int n, incr; + out int color; + + void main() + { + color = 0; + for (int x = 0; x < n; x += incr) + color += x; + } + """, + r'\(assign \(x\) \(var_ref x\) \(expression int16_t \+ \(var_ref x\) \(expression int16_t i2imp \(var_ref incr'), + Test("u32 loop counter", + """ + #version 310 es + precision mediump float; + precision mediump int; + + uniform uint n, incr; + out uint color; + + void main() + { + color = 0u; + for (uint x = 0u; x < n; x += incr) + color += x; + } + """, + r'\(assign \(x\) \(var_ref x\) \(expression uint16_t \+ \(var_ref x\) \(expression uint16_t u2ump \(var_ref incr'), + Test("f32 temp array", + """ + #version 300 es + precision mediump float; + + uniform float x,y; + out float color; + + void main() + { + float a[2] = float[2](x, y); + if (x > 0.0) + a[1] = 3.0; + color = a[0] + a[1]; + } + """, + r'\(constant float16_t \(3'), + Test("i32 temp array", + """ + #version 310 es + precision mediump int; + + uniform int x,y; + out int color; + + void main() + { + int a[2] = int[2](x, y); + if (x > 0) + a[1] = 3; + color = a[0] + a[1]; + } + """, + r'\(constant int16_t \(3'), + Test("u32 temp array", + """ + #version 310 es + precision mediump int; + + uniform uint x,y; + out uint color; + + void main() + { + uint a[2] = uint[2](x, y); + if (x > 0u) + a[1] = 3u; + color = a[0] + a[1]; + } + """, + r'\(constant uint16_t \(3'), + Test("f32 temp array of array", + """ + #version 310 es + precision mediump float; + + uniform float x,y; + out float color; + + void main() + { + float a[2][2] = float[2][2](float[2](x, y), float[2](x, y)); + if (x > 0.0) + a[1][1] = 3.0; + color = a[0][0] + a[1][1]; + } + """, + r'\(constant float16_t \(3'), + Test("i32 temp array of array", + """ + #version 310 es + precision mediump int; + + uniform int x,y; + out int color; + + void main() + { + int a[2][2] = int[2][2](int[2](x, y), int[2](x, y)); + if (x > 0) + a[1][1] = 3; + color = a[0][0] + a[1][1]; + } + """, + r'\(constant int16_t \(3'), + Test("u32 temp array of array", + """ + #version 310 es + precision mediump int; + + uniform uint x,y; + out uint color; + + void main() + { + uint a[2][2] = uint[2][2](uint[2](x, y), uint[2](x, y)); + if (x > 0u) + a[1][1] = 3u; + color = a[0][0] + a[1][1]; + } + """, + r'\(constant uint16_t \(3'), + Test("f32 temp array of array assigned from highp", + """ + #version 310 es + precision mediump float; + + uniform float x,y; + out float color; + + void main() + { + highp float b[2][2] = float[2][2](float[2](x, y), float[2](x, y)); + float a[2][2]; + a = b; + if (x > 0.0) + a[1][1] = 3.0; + color = a[0][0] + a[1][1]; + } + """, + r'\(constant float16_t \(3'), + Test("i32 temp array of array assigned from highp", + """ + #version 310 es + precision mediump int; + + uniform int x,y; + out int color; + + void main() + { + highp int b[2][2] = int[2][2](int[2](x, y), int[2](x, y)); + int a[2][2]; + a = b; + if (x > 0) + a[1][1] = 3; + color = a[0][0] + a[1][1]; + } + """, + r'\(constant int16_t \(3'), + Test("u32 temp array of array assigned from highp", + """ + #version 310 es + precision mediump int; + + uniform uint x,y; + out uint color; + + void main() + { + highp uint b[2][2] = uint[2][2](uint[2](x, y), uint[2](x, y)); + uint a[2][2]; + a = b; + if (x > 0u) + a[1][1] = 3u; + color = a[0][0] + a[1][1]; + } + """, + r'\(constant uint16_t \(3'), + Test("f32 temp array of array assigned to highp", + """ + #version 310 es + precision mediump float; + + uniform float x,y; + out float color; + + void main() + { + float a[2][2] = float[2][2](float[2](x, y), float[2](x, y)); + highp float b[2][2]; + b = a; + a = b; + if (x > 0.0) + a[1][1] = 3.0; + color = a[0][0] + a[1][1]; + } + """, + r'\(constant float16_t \(3'), + Test("i32 temp array of array assigned to highp", + """ + #version 310 es + precision mediump int; + + uniform int x,y; + out int color; + + void main() + { + int a[2][2] = int[2][2](int[2](x, y), int[2](x, y)); + highp int b[2][2]; + b = a; + a = b; + if (x > 0) + a[1][1] = 3; + color = a[0][0] + a[1][1]; + } + """, + r'\(constant int16_t \(3'), + Test("u32 temp array of array assigned to highp", + """ + #version 310 es + precision mediump int; + + uniform uint x,y; + out uint color; + + void main() + { + uint a[2][2] = uint[2][2](uint[2](x, y), uint[2](x, y)); + highp uint b[2][2]; + b = a; + a = b; + if (x > 0u) + a[1][1] = 3u; + color = a[0][0] + a[1][1]; + } + """, + r'\(constant uint16_t \(3'), + Test("f32 temp array of array returned by function", + """ + #version 310 es + precision mediump float; + + uniform float x,y; + out float color; + + float[2][2] f(void) + { + return float[2][2](float[2](x, y), float[2](x, y)); + } + + void main() + { + float a[2][2] = f(); + if (x > 0.0) + a[1][1] = 3.0; + color = a[0][0] + a[1][1]; + } + """, + r'\(constant float16_t \(3'), + Test("i32 temp array of array returned by function", + """ + #version 310 es + precision mediump int; + + uniform int x,y; + out int color; + + int[2][2] f(void) + { + return int[2][2](int[2](x, y), int[2](x, y)); + } + + void main() + { + int a[2][2] = f(); + if (x > 0) + a[1][1] = 3; + color = a[0][0] + a[1][1]; + } + """, + r'\(constant int16_t \(3'), + Test("u32 temp array of array returned by function", + """ + #version 310 es + precision mediump int; + + uniform uint x,y; + out uint color; + + uint[2][2] f(void) + { + return uint[2][2](uint[2](x, y), uint[2](x, y)); + } + + void main() + { + uint a[2][2] = f(); + if (x > 0u) + a[1][1] = 3u; + color = a[0][0] + a[1][1]; + } + """, + r'\(constant uint16_t \(3'), + Test("f32 temp array of array as function out", + """ + #version 310 es + precision mediump float; + + uniform float x,y; + out float color; + + void f(out float[2][2] v) + { + v = float[2][2](float[2](x, y), float[2](x, y)); + } + + void main() + { + float a[2][2]; + f(a); + if (x > 0.0) + a[1][1] = 3.0; + color = a[0][0] + a[1][1]; + } + """, + r'\(constant float16_t \(3'), + Test("i32 temp array of array as function out", + """ + #version 310 es + precision mediump int; + + uniform int x,y; + out int color; + + void f(out int[2][2] v) + { + v = int[2][2](int[2](x, y), int[2](x, y)); + } + + void main() + { + int a[2][2]; + f(a); + if (x > 0) + a[1][1] = 3; + color = a[0][0] + a[1][1]; + } + """, + r'\(constant int16_t \(3'), + Test("u32 temp array of array as function out", + """ + #version 310 es + precision mediump int; + + uniform uint x,y; + out uint color; + + void f(out uint[2][2] v) + { + v = uint[2][2](uint[2](x, y), uint[2](x, y)); + } + + void main() + { + uint a[2][2]; + f(a); + if (x > 0u) + a[1][1] = 3u; + color = a[0][0] + a[1][1]; + } + """, + r'\(constant uint16_t \(3'), + Test("f32 temp array of array as function in", + """ + #version 310 es + precision mediump float; + + uniform float x,y; + out float color; + + float[2][2] f(in float[2][2] v) + { + float t[2][2] = v; + return t; + } + + void main() + { + float a[2][2]; + a = f(a); + if (x > 0.0) + a[1][1] = 3.0; + color = a[0][0] + a[1][1]; + } + """, + r'\(constant float16_t \(3'), + Test("i32 temp array of array as function in", + """ + #version 310 es + precision mediump int; + + uniform int x,y; + out int color; + + int[2][2] f(in int[2][2] v) + { + int t[2][2] = v; + return t; + } + + void main() + { + int a[2][2]; + a = f(a); + if (x > 0) + a[1][1] = 3; + color = a[0][0] + a[1][1]; + } + """, + r'\(constant int16_t \(3'), + Test("u32 temp array of array as function in", + """ + #version 310 es + precision mediump int; + + uniform uint x,y; + out uint color; + + uint[2][2] f(in uint[2][2] v) + { + uint t[2][2] = v; + return t; + } + + void main() + { + uint a[2][2]; + a = f(a); + if (x > 0u) + a[1][1] = 3u; + color = a[0][0] + a[1][1]; + } + """, + r'\(constant uint16_t \(3'), + Test("f32 temp array of array as function inout", + """ + #version 310 es + precision mediump float; + + uniform float x,y; + out float color; + + void f(inout float[2][2] v) + { + float t[2][2] = v; + v = t; + } + + void main() + { + float a[2][2]; + f(a); + if (x > 0.0) + a[1][1] = 3.0; + color = a[0][0] + a[1][1]; + } + """, + r'\(constant float16_t \(3'), + Test("i32 temp array of array as function inout", + """ + #version 310 es + precision mediump int; + + uniform int x,y; + out int color; + + void f(inout int[2][2] v) + { + int t[2][2] = v; + v = t; + } + + void main() + { + int a[2][2]; + f(a); + if (x > 0) + a[1][1] = 3; + color = a[0][0] + a[1][1]; + } + """, + r'\(constant int16_t \(3'), + Test("u32 temp array of array as function inout", + """ + #version 310 es + precision mediump int; + + uniform uint x,y; + out uint color; + + void f(inout uint[2][2] v) + { + uint t[2][2] = v; + v = t; + } + + void main() + { + uint a[2][2]; + f(a); + if (x > 0u) + a[1][1] = 3u; + color = a[0][0] + a[1][1]; + } + """, + r'\(constant uint16_t \(3'), + Test("f32 temp struct (not lowered in the presence of control flow - TODO)", + """ + #version 300 es + precision mediump float; + + uniform float x,y; + out float color; + + void main() + { + struct { float x,y; } s; + s.x = x; + s.y = y; + if (x > 0.0) + s.y = 3.0; + color = s.x + s.y; + } + """, + r'\(constant float \(3'), # should be float16_t + Test("i32 temp struct (not lowered in the presence of control flow - TODO)", + """ + #version 300 es + precision mediump int; + + uniform int x,y; + out int color; + + void main() + { + struct { int x,y; } s; + s.x = x; + s.y = y; + if (x > 0) + s.y = 3; + color = s.x + s.y; + } + """, + r'\(constant int \(3'), # should be int16_t + Test("u32 temp struct (not lowered in the presence of control flow - TODO)", + """ + #version 300 es + precision mediump int; + + uniform uint x,y; + out uint color; + + void main() + { + struct { uint x,y; } s; + s.x = x; + s.y = y; + if (x > 0u) + s.y = 3u; + color = s.x + s.y; + } + """, + r'\(constant uint \(3'), # should be uint16_t ] diff --git a/src/mesa/main/mtypes.h b/src/mesa/main/mtypes.h index 836961b077f..a6af3fbca17 100644 --- a/src/mesa/main/mtypes.h +++ b/src/mesa/main/mtypes.h @@ -3211,6 +3211,12 @@ struct gl_shader_compiler_options GLboolean LowerPrecisionDerivatives; /** + * This enables 16-bit phis in NIR, 16-bit loop counters, 16-bit indirect + * arrays, etc. + */ + GLboolean LowerPrecisionTemporaries; + + /** * \name Forms of indirect addressing the driver cannot do. */ /*@{*/ |