* 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 brw_fs.cpp
*
- * Authors:
- * Eric Anholt <eric@anholt.net>
- *
+ * This file drives the GLSL IR -> LIR translation, contains the
+ * optimizations on the LIR, and drives the generation of native code
+ * from the LIR.
*/
extern "C" {
}
#include "brw_shader.h"
#include "brw_fs.h"
-#include "../glsl/glsl_types.h"
-#include "../glsl/ir_print_visitor.h"
+#include "glsl/glsl_types.h"
+#include "glsl/ir_print_visitor.h"
#define MAX_INSTRUCTION (1 << 30)
-static int
-type_size(const struct glsl_type *type)
+int
+fs_visitor::type_size(const struct glsl_type *type)
{
unsigned int size, i;
void
fs_visitor::fail(const char *format, ...)
{
- if (!failed) {
- failed = true;
+ va_list va;
+ char *msg;
- if (INTEL_DEBUG & DEBUG_WM) {
- fprintf(stderr, "FS compile failed: ");
+ if (failed)
+ return;
- va_list va;
- va_start(va, format);
- vfprintf(stderr, format, va);
- va_end(va);
- }
+ failed = true;
+
+ va_start(va, format);
+ msg = ralloc_vasprintf(mem_ctx, format, va);
+ va_end(va);
+ msg = ralloc_asprintf(mem_ctx, "FS compile failed: %s\n", msg);
+
+ this->fail_msg = msg;
+
+ if (INTEL_DEBUG & DEBUG_WM) {
+ fprintf(stderr, "%s", msg);
}
}
return 0;
switch (inst->opcode) {
- case FS_OPCODE_RCP:
- case FS_OPCODE_RSQ:
- case FS_OPCODE_SQRT:
- case FS_OPCODE_EXP2:
- case FS_OPCODE_LOG2:
- case FS_OPCODE_SIN:
- case FS_OPCODE_COS:
+ case SHADER_OPCODE_RCP:
+ case SHADER_OPCODE_RSQ:
+ case SHADER_OPCODE_SQRT:
+ case SHADER_OPCODE_EXP2:
+ case SHADER_OPCODE_LOG2:
+ case SHADER_OPCODE_SIN:
+ case SHADER_OPCODE_COS:
return 1 * c->dispatch_width / 8;
- case FS_OPCODE_POW:
+ case SHADER_OPCODE_POW:
+ case SHADER_OPCODE_INT_QUOTIENT:
+ case SHADER_OPCODE_INT_REMAINDER:
return 2 * c->dispatch_width / 8;
- case FS_OPCODE_TEX:
+ case SHADER_OPCODE_TEX:
case FS_OPCODE_TXB:
- case FS_OPCODE_TXD:
- case FS_OPCODE_TXL:
+ case SHADER_OPCODE_TXD:
+ case SHADER_OPCODE_TXF:
+ case SHADER_OPCODE_TXL:
+ case SHADER_OPCODE_TXS:
return 1;
case FS_OPCODE_FB_WRITE:
return 2;
virtual_grf_array_size *= 2;
virtual_grf_sizes = reralloc(mem_ctx, virtual_grf_sizes, int,
virtual_grf_array_size);
-
- /* This slot is always unused. */
- virtual_grf_sizes[0] = 0;
}
virtual_grf_sizes[virtual_grf_next] = size;
return virtual_grf_next++;
}
/** Fixed HW reg constructor. */
-fs_reg::fs_reg(enum register_file file, int hw_reg)
+fs_reg::fs_reg(enum register_file file, int reg)
{
init();
this->file = file;
- this->hw_reg = hw_reg;
+ this->reg = reg;
this->type = BRW_REGISTER_TYPE_F;
}
/** Fixed HW reg constructor. */
-fs_reg::fs_reg(enum register_file file, int hw_reg, uint32_t type)
+fs_reg::fs_reg(enum register_file file, int reg, uint32_t type)
{
init();
this->file = file;
- this->hw_reg = hw_reg;
+ this->reg = reg;
this->type = type;
}
init();
this->file = GRF;
- this->reg = v->virtual_grf_alloc(type_size(type));
+ this->reg = v->virtual_grf_alloc(v->type_size(type));
this->reg_offset = 0;
this->type = brw_type_for_base_type(type);
}
* This brings in those uniform definitions
*/
void
-fs_visitor::import_uniforms(struct hash_table *src_variable_ht)
+fs_visitor::import_uniforms(fs_visitor *v)
{
- hash_table_call_foreach(src_variable_ht,
+ hash_table_call_foreach(v->variable_ht,
import_uniforms_callback,
variable_ht);
+ this->params_remap = v->params_remap;
}
/* Our support for uniforms is piggy-backed on the struct
assert(param < ARRAY_SIZE(c->prog_data.param));
- switch (type->base_type) {
- case GLSL_TYPE_FLOAT:
- c->prog_data.param_convert[param] = PARAM_NO_CONVERT;
- break;
- case GLSL_TYPE_UINT:
- c->prog_data.param_convert[param] = PARAM_CONVERT_F2U;
- break;
- case GLSL_TYPE_INT:
- c->prog_data.param_convert[param] = PARAM_CONVERT_F2I;
- break;
- case GLSL_TYPE_BOOL:
- c->prog_data.param_convert[param] = PARAM_CONVERT_F2B;
- break;
- default:
- assert(!"not reached");
+ if (ctx->Const.NativeIntegers) {
c->prog_data.param_convert[param] = PARAM_NO_CONVERT;
- break;
+ } else {
+ switch (type->base_type) {
+ case GLSL_TYPE_FLOAT:
+ c->prog_data.param_convert[param] = PARAM_NO_CONVERT;
+ break;
+ case GLSL_TYPE_UINT:
+ c->prog_data.param_convert[param] = PARAM_CONVERT_F2U;
+ break;
+ case GLSL_TYPE_INT:
+ c->prog_data.param_convert[param] = PARAM_CONVERT_F2I;
+ break;
+ case GLSL_TYPE_BOOL:
+ c->prog_data.param_convert[param] = PARAM_CONVERT_F2B;
+ break;
+ default:
+ assert(!"not reached");
+ c->prog_data.param_convert[param] = PARAM_NO_CONVERT;
+ break;
+ }
}
this->param_index[param] = loc;
this->param_offset[param] = i;
emit(BRW_OPCODE_MOV, wpos,
fs_reg(brw_vec8_grf(c->source_depth_reg, 0)));
} else {
- emit(FS_OPCODE_LINTERP, wpos, this->delta_x, this->delta_y,
- interp_reg(FRAG_ATTRIB_WPOS, 2));
+ emit(FS_OPCODE_LINTERP, wpos,
+ this->delta_x[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC],
+ this->delta_y[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC],
+ interp_reg(FRAG_ATTRIB_WPOS, 2));
}
wpos.reg_offset++;
fs_visitor::emit_general_interpolation(ir_variable *ir)
{
fs_reg *reg = new(this->mem_ctx) fs_reg(this, ir->type);
- /* Interpolation is always in floating point regs. */
- reg->type = BRW_REGISTER_TYPE_F;
+ reg->type = brw_type_for_base_type(ir->type->get_scalar_type());
fs_reg attr = *reg;
unsigned int array_elements;
type = ir->type;
}
+ glsl_interp_qualifier interpolation_mode =
+ ir->determine_interpolation_mode(c->key.flat_shade);
+
int location = ir->location;
for (unsigned int i = 0; i < array_elements; i++) {
for (unsigned int j = 0; j < type->matrix_columns; j++) {
attr.reg_offset += type->vector_elements;
location++;
continue;
- }
-
- bool is_gl_Color =
- location == FRAG_ATTRIB_COL0 || location == FRAG_ATTRIB_COL1;
-
- if (c->key.flat_shade && is_gl_Color) {
- /* Constant interpolation (flat shading) case. The SF has
- * handed us defined values in only the constant offset
- * field of the setup reg.
- */
- for (unsigned int k = 0; k < type->vector_elements; k++) {
- struct brw_reg interp = interp_reg(location, k);
- interp = suboffset(interp, 3);
- emit(FS_OPCODE_CINTERP, attr, fs_reg(interp));
- attr.reg_offset++;
- }
- } else {
- /* Perspective interpolation case. */
- for (unsigned int k = 0; k < type->vector_elements; k++) {
- struct brw_reg interp = interp_reg(location, k);
- emit(FS_OPCODE_LINTERP, attr,
- this->delta_x, this->delta_y, fs_reg(interp));
- attr.reg_offset++;
- }
-
- if (intel->gen < 6) {
- attr.reg_offset -= type->vector_elements;
- for (unsigned int k = 0; k < type->vector_elements; k++) {
- emit(BRW_OPCODE_MUL, attr, attr, this->pixel_w);
- attr.reg_offset++;
- }
- }
- }
- location++;
- }
- }
-
- return reg;
-}
-
-fs_reg *
-fs_visitor::emit_frontfacing_interpolation(ir_variable *ir)
-{
- fs_reg *reg = new(this->mem_ctx) fs_reg(this, ir->type);
-
- /* The frontfacing comes in as a bit in the thread payload. */
- if (intel->gen >= 6) {
- emit(BRW_OPCODE_ASR, *reg,
- fs_reg(retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_D)),
- fs_reg(15));
- emit(BRW_OPCODE_NOT, *reg, *reg);
- emit(BRW_OPCODE_AND, *reg, *reg, fs_reg(1));
- } else {
- struct brw_reg r1_6ud = retype(brw_vec1_grf(1, 6), BRW_REGISTER_TYPE_UD);
- /* bit 31 is "primitive is back face", so checking < (1 << 31) gives
- * us front face
- */
- fs_inst *inst = emit(BRW_OPCODE_CMP, *reg,
- fs_reg(r1_6ud),
- fs_reg(1u << 31));
- inst->conditional_mod = BRW_CONDITIONAL_L;
- emit(BRW_OPCODE_AND, *reg, *reg, fs_reg(1u));
- }
-
- return reg;
-}
-
-fs_inst *
-fs_visitor::emit_math(fs_opcodes opcode, fs_reg dst, fs_reg src)
-{
- switch (opcode) {
- case FS_OPCODE_RCP:
- case FS_OPCODE_RSQ:
- case FS_OPCODE_SQRT:
- case FS_OPCODE_EXP2:
- case FS_OPCODE_LOG2:
- case FS_OPCODE_SIN:
- case FS_OPCODE_COS:
- break;
- default:
- assert(!"not reached: bad math opcode");
- return NULL;
- }
-
- /* Can't do hstride == 0 args to gen6 math, so expand it out. We
- * might be able to do better by doing execsize = 1 math and then
- * expanding that result out, but we would need to be careful with
- * masking.
- *
- * The hardware ignores source modifiers (negate and abs) on math
- * instructions, so we also move to a temp to set those up.
- */
- if (intel->gen >= 6 && (src.file == UNIFORM ||
- src.abs ||
- src.negate)) {
- fs_reg expanded = fs_reg(this, glsl_type::float_type);
- emit(BRW_OPCODE_MOV, expanded, src);
- src = expanded;
- }
-
- fs_inst *inst = emit(opcode, dst, src);
-
- if (intel->gen < 6) {
- inst->base_mrf = 2;
- inst->mlen = c->dispatch_width / 8;
- }
-
- return inst;
-}
-
-fs_inst *
-fs_visitor::emit_math(fs_opcodes opcode, fs_reg dst, fs_reg src0, fs_reg src1)
-{
- int base_mrf = 2;
- fs_inst *inst;
-
- assert(opcode == FS_OPCODE_POW);
-
- if (intel->gen >= 6) {
- /* Can't do hstride == 0 args to gen6 math, so expand it out.
- *
- * The hardware ignores source modifiers (negate and abs) on math
- * instructions, so we also move to a temp to set those up.
- */
- if (src0.file == UNIFORM || src0.abs || src0.negate) {
- fs_reg expanded = fs_reg(this, glsl_type::float_type);
- emit(BRW_OPCODE_MOV, expanded, src0);
- src0 = expanded;
- }
-
- if (src1.file == UNIFORM || src1.abs || src1.negate) {
- fs_reg expanded = fs_reg(this, glsl_type::float_type);
- emit(BRW_OPCODE_MOV, expanded, src1);
- src1 = expanded;
- }
-
- inst = emit(opcode, dst, src0, src1);
- } else {
- emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + 1), src1);
- inst = emit(opcode, dst, src0, reg_null_f);
-
- inst->base_mrf = base_mrf;
- inst->mlen = 2 * c->dispatch_width / 8;
- }
- return inst;
-}
-
-void
-fs_visitor::visit(ir_variable *ir)
-{
- fs_reg *reg = NULL;
-
- if (variable_storage(ir))
- return;
-
- if (strcmp(ir->name, "gl_FragColor") == 0) {
- this->frag_color = ir;
- } else if (strcmp(ir->name, "gl_FragData") == 0) {
- this->frag_data = ir;
- } else if (strcmp(ir->name, "gl_FragDepth") == 0) {
- this->frag_depth = ir;
- }
-
- if (ir->mode == ir_var_in) {
- if (!strcmp(ir->name, "gl_FragCoord")) {
- reg = emit_fragcoord_interpolation(ir);
- } else if (!strcmp(ir->name, "gl_FrontFacing")) {
- reg = emit_frontfacing_interpolation(ir);
- } else {
- reg = emit_general_interpolation(ir);
- }
- assert(reg);
- hash_table_insert(this->variable_ht, reg, ir);
- return;
- }
-
- if (ir->mode == ir_var_uniform) {
- int param_index = c->prog_data.nr_params;
-
- if (c->dispatch_width == 16) {
- if (!variable_storage(ir)) {
- fail("Failed to find uniform '%s' in 16-wide\n", ir->name);
- }
- return;
- }
-
- if (!strncmp(ir->name, "gl_", 3)) {
- setup_builtin_uniform_values(ir);
- } else {
- setup_uniform_values(ir->location, ir->type);
- }
-
- reg = new(this->mem_ctx) fs_reg(UNIFORM, param_index);
- reg->type = brw_type_for_base_type(ir->type);
- }
-
- if (!reg)
- reg = new(this->mem_ctx) fs_reg(this, ir->type);
-
- hash_table_insert(this->variable_ht, reg, ir);
-}
-
-void
-fs_visitor::visit(ir_dereference_variable *ir)
-{
- fs_reg *reg = variable_storage(ir->var);
- this->result = *reg;
-}
-
-void
-fs_visitor::visit(ir_dereference_record *ir)
-{
- const glsl_type *struct_type = ir->record->type;
-
- ir->record->accept(this);
-
- unsigned int offset = 0;
- for (unsigned int i = 0; i < struct_type->length; i++) {
- if (strcmp(struct_type->fields.structure[i].name, ir->field) == 0)
- break;
- offset += type_size(struct_type->fields.structure[i].type);
- }
- this->result.reg_offset += offset;
- this->result.type = brw_type_for_base_type(ir->type);
-}
-
-void
-fs_visitor::visit(ir_dereference_array *ir)
-{
- ir_constant *index;
- int element_size;
-
- ir->array->accept(this);
- index = ir->array_index->as_constant();
-
- element_size = type_size(ir->type);
- this->result.type = brw_type_for_base_type(ir->type);
-
- if (index) {
- assert(this->result.file == UNIFORM ||
- (this->result.file == GRF &&
- this->result.reg != 0));
- this->result.reg_offset += index->value.i[0] * element_size;
- } else {
- assert(!"FINISHME: non-constant array element");
- }
-}
-
-/* Instruction selection: Produce a MOV.sat instead of
- * MIN(MAX(val, 0), 1) when possible.
- */
-bool
-fs_visitor::try_emit_saturate(ir_expression *ir)
-{
- ir_rvalue *sat_val = ir->as_rvalue_to_saturate();
-
- if (!sat_val)
- return false;
-
- this->result = reg_undef;
- sat_val->accept(this);
- fs_reg src = this->result;
-
- this->result = fs_reg(this, ir->type);
- fs_inst *inst = emit(BRW_OPCODE_MOV, this->result, src);
- inst->saturate = true;
-
- return true;
-}
-
-void
-fs_visitor::visit(ir_expression *ir)
-{
- unsigned int operand;
- fs_reg op[2], temp;
- fs_inst *inst;
-
- assert(ir->get_num_operands() <= 2);
-
- if (try_emit_saturate(ir))
- return;
-
- /* This is where our caller would like us to put the result, if possible. */
- fs_reg saved_result_storage = this->result;
-
- for (operand = 0; operand < ir->get_num_operands(); operand++) {
- this->result = reg_undef;
- ir->operands[operand]->accept(this);
- if (this->result.file == BAD_FILE) {
- ir_print_visitor v;
- fail("Failed to get tree for expression operand:\n");
- ir->operands[operand]->accept(&v);
- }
- op[operand] = this->result;
-
- /* Matrix expression operands should have been broken down to vector
- * operations already.
- */
- assert(!ir->operands[operand]->type->is_matrix());
- /* And then those vector operands should have been broken down to scalar.
- */
- assert(!ir->operands[operand]->type->is_vector());
- }
-
- /* Inherit storage from our parent if possible, and otherwise we
- * alloc a temporary.
- */
- if (saved_result_storage.file == BAD_FILE) {
- this->result = fs_reg(this, ir->type);
- } else {
- this->result = saved_result_storage;
- }
-
- switch (ir->operation) {
- case ir_unop_logic_not:
- /* Note that BRW_OPCODE_NOT is not appropriate here, since it is
- * ones complement of the whole register, not just bit 0.
- */
- emit(BRW_OPCODE_XOR, this->result, op[0], fs_reg(1));
- break;
- case ir_unop_neg:
- op[0].negate = !op[0].negate;
- this->result = op[0];
- break;
- case ir_unop_abs:
- op[0].abs = true;
- op[0].negate = false;
- this->result = op[0];
- break;
- case ir_unop_sign:
- temp = fs_reg(this, ir->type);
-
- /* Unalias the destination. (imagine a = sign(a)) */
- this->result = fs_reg(this, ir->type);
-
- emit(BRW_OPCODE_MOV, this->result, fs_reg(0.0f));
-
- inst = emit(BRW_OPCODE_CMP, reg_null_f, op[0], fs_reg(0.0f));
- inst->conditional_mod = BRW_CONDITIONAL_G;
- inst = emit(BRW_OPCODE_MOV, this->result, fs_reg(1.0f));
- inst->predicated = true;
-
- inst = emit(BRW_OPCODE_CMP, reg_null_f, op[0], fs_reg(0.0f));
- inst->conditional_mod = BRW_CONDITIONAL_L;
- inst = emit(BRW_OPCODE_MOV, this->result, fs_reg(-1.0f));
- inst->predicated = true;
-
- break;
- case ir_unop_rcp:
- emit_math(FS_OPCODE_RCP, this->result, op[0]);
- break;
-
- case ir_unop_exp2:
- emit_math(FS_OPCODE_EXP2, this->result, op[0]);
- break;
- case ir_unop_log2:
- emit_math(FS_OPCODE_LOG2, this->result, op[0]);
- break;
- case ir_unop_exp:
- case ir_unop_log:
- assert(!"not reached: should be handled by ir_explog_to_explog2");
- break;
- case ir_unop_sin:
- case ir_unop_sin_reduced:
- emit_math(FS_OPCODE_SIN, this->result, op[0]);
- break;
- case ir_unop_cos:
- case ir_unop_cos_reduced:
- emit_math(FS_OPCODE_COS, this->result, op[0]);
- break;
-
- case ir_unop_dFdx:
- emit(FS_OPCODE_DDX, this->result, op[0]);
- break;
- case ir_unop_dFdy:
- emit(FS_OPCODE_DDY, this->result, op[0]);
- break;
-
- case ir_binop_add:
- emit(BRW_OPCODE_ADD, this->result, op[0], op[1]);
- break;
- case ir_binop_sub:
- assert(!"not reached: should be handled by ir_sub_to_add_neg");
- break;
-
- case ir_binop_mul:
- emit(BRW_OPCODE_MUL, this->result, op[0], op[1]);
- break;
- case ir_binop_div:
- assert(!"not reached: should be handled by ir_div_to_mul_rcp");
- break;
- case ir_binop_mod:
- assert(!"ir_binop_mod should have been converted to b * fract(a/b)");
- break;
-
- case ir_binop_less:
- case ir_binop_greater:
- case ir_binop_lequal:
- case ir_binop_gequal:
- case ir_binop_equal:
- case ir_binop_all_equal:
- case ir_binop_nequal:
- case ir_binop_any_nequal:
- temp = this->result;
- /* original gen4 does implicit conversion before comparison. */
- if (intel->gen < 5)
- temp.type = op[0].type;
-
- inst = emit(BRW_OPCODE_CMP, temp, op[0], op[1]);
- inst->conditional_mod = brw_conditional_for_comparison(ir->operation);
- emit(BRW_OPCODE_AND, this->result, this->result, fs_reg(0x1));
- break;
-
- case ir_binop_logic_xor:
- emit(BRW_OPCODE_XOR, this->result, op[0], op[1]);
- break;
-
- case ir_binop_logic_or:
- emit(BRW_OPCODE_OR, this->result, op[0], op[1]);
- break;
-
- case ir_binop_logic_and:
- emit(BRW_OPCODE_AND, this->result, op[0], op[1]);
- break;
-
- case ir_binop_dot:
- case ir_unop_any:
- assert(!"not reached: should be handled by brw_fs_channel_expressions");
- break;
-
- case ir_unop_noise:
- assert(!"not reached: should be handled by lower_noise");
- break;
-
- case ir_quadop_vector:
- assert(!"not reached: should be handled by lower_quadop_vector");
- break;
-
- case ir_unop_sqrt:
- emit_math(FS_OPCODE_SQRT, this->result, op[0]);
- break;
-
- case ir_unop_rsq:
- emit_math(FS_OPCODE_RSQ, this->result, op[0]);
- break;
-
- case ir_unop_i2f:
- case ir_unop_b2f:
- case ir_unop_b2i:
- case ir_unop_f2i:
- emit(BRW_OPCODE_MOV, this->result, op[0]);
- break;
- case ir_unop_f2b:
- case ir_unop_i2b:
- temp = this->result;
- /* original gen4 does implicit conversion before comparison. */
- if (intel->gen < 5)
- temp.type = op[0].type;
-
- inst = emit(BRW_OPCODE_CMP, temp, op[0], fs_reg(0.0f));
- inst->conditional_mod = BRW_CONDITIONAL_NZ;
- inst = emit(BRW_OPCODE_AND, this->result, this->result, fs_reg(1));
- break;
-
- case ir_unop_trunc:
- emit(BRW_OPCODE_RNDZ, this->result, op[0]);
- break;
- case ir_unop_ceil:
- op[0].negate = !op[0].negate;
- inst = emit(BRW_OPCODE_RNDD, this->result, op[0]);
- this->result.negate = true;
- break;
- case ir_unop_floor:
- inst = emit(BRW_OPCODE_RNDD, this->result, op[0]);
- break;
- case ir_unop_fract:
- inst = emit(BRW_OPCODE_FRC, this->result, op[0]);
- break;
- case ir_unop_round_even:
- emit(BRW_OPCODE_RNDE, this->result, op[0]);
- break;
-
- case ir_binop_min:
- /* Unalias the destination */
- this->result = fs_reg(this, ir->type);
-
- inst = emit(BRW_OPCODE_CMP, this->result, op[0], op[1]);
- inst->conditional_mod = BRW_CONDITIONAL_L;
-
- inst = emit(BRW_OPCODE_SEL, this->result, op[0], op[1]);
- inst->predicated = true;
- break;
- case ir_binop_max:
- /* Unalias the destination */
- this->result = fs_reg(this, ir->type);
-
- inst = emit(BRW_OPCODE_CMP, this->result, op[0], op[1]);
- inst->conditional_mod = BRW_CONDITIONAL_G;
-
- inst = emit(BRW_OPCODE_SEL, this->result, op[0], op[1]);
- inst->predicated = true;
- break;
-
- case ir_binop_pow:
- emit_math(FS_OPCODE_POW, this->result, op[0], op[1]);
- break;
-
- case ir_unop_bit_not:
- inst = emit(BRW_OPCODE_NOT, this->result, op[0]);
- break;
- case ir_binop_bit_and:
- inst = emit(BRW_OPCODE_AND, this->result, op[0], op[1]);
- break;
- case ir_binop_bit_xor:
- inst = emit(BRW_OPCODE_XOR, this->result, op[0], op[1]);
- break;
- case ir_binop_bit_or:
- inst = emit(BRW_OPCODE_OR, this->result, op[0], op[1]);
- break;
-
- case ir_unop_u2f:
- case ir_binop_lshift:
- case ir_binop_rshift:
- assert(!"GLSL 1.30 features unsupported");
- break;
- }
-}
-
-void
-fs_visitor::emit_assignment_writes(fs_reg &l, fs_reg &r,
- const glsl_type *type, bool predicated)
-{
- switch (type->base_type) {
- case GLSL_TYPE_FLOAT:
- case GLSL_TYPE_UINT:
- case GLSL_TYPE_INT:
- case GLSL_TYPE_BOOL:
- for (unsigned int i = 0; i < type->components(); i++) {
- l.type = brw_type_for_base_type(type);
- r.type = brw_type_for_base_type(type);
-
- if (predicated || !l.equals(&r)) {
- fs_inst *inst = emit(BRW_OPCODE_MOV, l, r);
- inst->predicated = predicated;
- }
-
- l.reg_offset++;
- r.reg_offset++;
- }
- break;
- case GLSL_TYPE_ARRAY:
- for (unsigned int i = 0; i < type->length; i++) {
- emit_assignment_writes(l, r, type->fields.array, predicated);
- }
- break;
-
- case GLSL_TYPE_STRUCT:
- for (unsigned int i = 0; i < type->length; i++) {
- emit_assignment_writes(l, r, type->fields.structure[i].type,
- predicated);
- }
- break;
-
- case GLSL_TYPE_SAMPLER:
- break;
-
- default:
- assert(!"not reached");
- break;
- }
-}
-
-void
-fs_visitor::visit(ir_assignment *ir)
-{
- struct fs_reg l, r;
- fs_inst *inst;
-
- /* FINISHME: arrays on the lhs */
- this->result = reg_undef;
- ir->lhs->accept(this);
- l = this->result;
-
- /* If we're doing a direct assignment, an RHS expression could
- * drop its result right into our destination. Otherwise, tell it
- * not to.
- */
- if (ir->condition ||
- !(ir->lhs->type->is_scalar() ||
- (ir->lhs->type->is_vector() &&
- ir->write_mask == (1 << ir->lhs->type->vector_elements) - 1))) {
- this->result = reg_undef;
- }
-
- ir->rhs->accept(this);
- r = this->result;
-
- assert(l.file != BAD_FILE);
- assert(r.file != BAD_FILE);
-
- if (ir->condition) {
- emit_bool_to_cond_code(ir->condition);
- }
-
- if (ir->lhs->type->is_scalar() ||
- ir->lhs->type->is_vector()) {
- for (int i = 0; i < ir->lhs->type->vector_elements; i++) {
- if (ir->write_mask & (1 << i)) {
- if (ir->condition) {
- inst = emit(BRW_OPCODE_MOV, l, r);
- inst->predicated = true;
- } else if (!l.equals(&r)) {
- inst = emit(BRW_OPCODE_MOV, l, r);
- }
-
- r.reg_offset++;
- }
- l.reg_offset++;
- }
- } else {
- emit_assignment_writes(l, r, ir->lhs->type, ir->condition != NULL);
- }
-}
-
-fs_inst *
-fs_visitor::emit_texture_gen4(ir_texture *ir, fs_reg dst, fs_reg coordinate,
- int sampler)
-{
- int mlen;
- int base_mrf = 1;
- bool simd16 = false;
- fs_reg orig_dst;
-
- /* g0 header. */
- mlen = 1;
-
- if (ir->shadow_comparitor) {
- for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
- fs_inst *inst = emit(BRW_OPCODE_MOV,
- fs_reg(MRF, base_mrf + mlen + i), coordinate);
- if (i < 3 && c->key.gl_clamp_mask[i] & (1 << sampler))
- inst->saturate = true;
-
- coordinate.reg_offset++;
- }
- /* gen4's SIMD8 sampler always has the slots for u,v,r present. */
- mlen += 3;
-
- if (ir->op == ir_tex) {
- /* There's no plain shadow compare message, so we use shadow
- * compare with a bias of 0.0.
- */
- emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), fs_reg(0.0f));
- mlen++;
- } else if (ir->op == ir_txb) {
- this->result = reg_undef;
- ir->lod_info.bias->accept(this);
- emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result);
- mlen++;
- } else {
- assert(ir->op == ir_txl);
- this->result = reg_undef;
- ir->lod_info.lod->accept(this);
- emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result);
- mlen++;
- }
-
- this->result = reg_undef;
- ir->shadow_comparitor->accept(this);
- emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result);
- mlen++;
- } else if (ir->op == ir_tex) {
- for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
- fs_inst *inst = emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen + i),
- coordinate);
- if (i < 3 && c->key.gl_clamp_mask[i] & (1 << sampler))
- inst->saturate = true;
- coordinate.reg_offset++;
- }
- /* gen4's SIMD8 sampler always has the slots for u,v,r present. */
- mlen += 3;
- } else if (ir->op == ir_txd) {
- assert(!"TXD isn't supported on gen4 yet.");
- } else {
- /* Oh joy. gen4 doesn't have SIMD8 non-shadow-compare bias/lod
- * instructions. We'll need to do SIMD16 here.
- */
- assert(ir->op == ir_txb || ir->op == ir_txl);
-
- for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
- fs_inst *inst = emit(BRW_OPCODE_MOV, fs_reg(MRF,
- base_mrf + mlen + i * 2),
- coordinate);
- if (i < 3 && c->key.gl_clamp_mask[i] & (1 << sampler))
- inst->saturate = true;
- coordinate.reg_offset++;
- }
-
- /* lod/bias appears after u/v/r. */
- mlen += 6;
-
- if (ir->op == ir_txb) {
- this->result = reg_undef;
- ir->lod_info.bias->accept(this);
- emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result);
- mlen++;
- } else {
- this->result = reg_undef;
- ir->lod_info.lod->accept(this);
- emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result);
- mlen++;
- }
-
- /* The unused upper half. */
- mlen++;
-
- /* Now, since we're doing simd16, the return is 2 interleaved
- * vec4s where the odd-indexed ones are junk. We'll need to move
- * this weirdness around to the expected layout.
- */
- simd16 = true;
- orig_dst = dst;
- dst = fs_reg(this, glsl_type::get_array_instance(glsl_type::vec4_type,
- 2));
- dst.type = BRW_REGISTER_TYPE_F;
- }
-
- fs_inst *inst = NULL;
- switch (ir->op) {
- case ir_tex:
- inst = emit(FS_OPCODE_TEX, dst);
- break;
- case ir_txb:
- inst = emit(FS_OPCODE_TXB, dst);
- break;
- case ir_txl:
- inst = emit(FS_OPCODE_TXL, dst);
- break;
- case ir_txd:
- inst = emit(FS_OPCODE_TXD, dst);
- break;
- case ir_txf:
- assert(!"GLSL 1.30 features unsupported");
- break;
- }
- inst->base_mrf = base_mrf;
- inst->mlen = mlen;
- inst->header_present = true;
-
- if (simd16) {
- for (int i = 0; i < 4; i++) {
- emit(BRW_OPCODE_MOV, orig_dst, dst);
- orig_dst.reg_offset++;
- dst.reg_offset += 2;
- }
- }
-
- return inst;
-}
-
-/* gen5's sampler has slots for u, v, r, array index, then optional
- * parameters like shadow comparitor or LOD bias. If optional
- * parameters aren't present, those base slots are optional and don't
- * need to be included in the message.
- *
- * We don't fill in the unnecessary slots regardless, which may look
- * surprising in the disassembly.
- */
-fs_inst *
-fs_visitor::emit_texture_gen5(ir_texture *ir, fs_reg dst, fs_reg coordinate,
- int sampler)
-{
- int mlen = 0;
- int base_mrf = 2;
- int reg_width = c->dispatch_width / 8;
- bool header_present = false;
-
- if (ir->offset) {
- /* The offsets set up by the ir_texture visitor are in the
- * m1 header, so we can't go headerless.
- */
- header_present = true;
- mlen++;
- base_mrf--;
- }
-
- for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
- fs_inst *inst = emit(BRW_OPCODE_MOV,
- fs_reg(MRF, base_mrf + mlen + i * reg_width),
- coordinate);
- if (i < 3 && c->key.gl_clamp_mask[i] & (1 << sampler))
- inst->saturate = true;
- coordinate.reg_offset++;
- }
- mlen += ir->coordinate->type->vector_elements * reg_width;
-
- if (ir->shadow_comparitor) {
- mlen = MAX2(mlen, header_present + 4 * reg_width);
-
- this->result = reg_undef;
- ir->shadow_comparitor->accept(this);
- emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result);
- mlen += reg_width;
- }
-
- fs_inst *inst = NULL;
- switch (ir->op) {
- case ir_tex:
- inst = emit(FS_OPCODE_TEX, dst);
- break;
- case ir_txb:
- this->result = reg_undef;
- ir->lod_info.bias->accept(this);
- mlen = MAX2(mlen, header_present + 4 * reg_width);
- emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result);
- mlen += reg_width;
-
- inst = emit(FS_OPCODE_TXB, dst);
-
- break;
- case ir_txl:
- this->result = reg_undef;
- ir->lod_info.lod->accept(this);
- mlen = MAX2(mlen, header_present + 4 * reg_width);
- emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result);
- mlen += reg_width;
-
- inst = emit(FS_OPCODE_TXL, dst);
- break;
- case ir_txd:
- case ir_txf:
- assert(!"GLSL 1.30 features unsupported");
- break;
- }
- inst->base_mrf = base_mrf;
- inst->mlen = mlen;
- inst->header_present = header_present;
-
- if (mlen > 11) {
- fail("Message length >11 disallowed by hardware\n");
- }
-
- return inst;
-}
-
-fs_inst *
-fs_visitor::emit_texture_gen7(ir_texture *ir, fs_reg dst, fs_reg coordinate,
- int sampler)
-{
- int mlen = 0;
- int base_mrf = 2;
- int reg_width = c->dispatch_width / 8;
- bool header_present = false;
-
- if (ir->offset) {
- /* The offsets set up by the ir_texture visitor are in the
- * m1 header, so we can't go headerless.
- */
- header_present = true;
- mlen++;
- base_mrf--;
- }
-
- if (ir->shadow_comparitor) {
- ir->shadow_comparitor->accept(this);
- emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result);
- mlen += reg_width;
- }
-
- /* Set up the LOD info */
- switch (ir->op) {
- case ir_tex:
- break;
- case ir_txb:
- ir->lod_info.bias->accept(this);
- emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result);
- mlen += reg_width;
- break;
- case ir_txl:
- ir->lod_info.lod->accept(this);
- emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result);
- mlen += reg_width;
- break;
- case ir_txd:
- case ir_txf:
- assert(!"GLSL 1.30 features unsupported");
- break;
- }
-
- /* Set up the coordinate */
- for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
- fs_inst *inst = emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen),
- coordinate);
- if (i < 3 && c->key.gl_clamp_mask[i] & (1 << sampler))
- inst->saturate = true;
- coordinate.reg_offset++;
- mlen += reg_width;
- }
-
- /* Generate the SEND */
- fs_inst *inst = NULL;
- switch (ir->op) {
- case ir_tex: inst = emit(FS_OPCODE_TEX, dst); break;
- case ir_txb: inst = emit(FS_OPCODE_TXB, dst); break;
- case ir_txl: inst = emit(FS_OPCODE_TXL, dst); break;
- case ir_txd: inst = emit(FS_OPCODE_TXD, dst); break;
- case ir_txf: assert(!"TXF unsupported.");
- }
- inst->base_mrf = base_mrf;
- inst->mlen = mlen;
- inst->header_present = header_present;
-
- if (mlen > 11) {
- fail("Message length >11 disallowed by hardware\n");
- }
-
- return inst;
-}
-
-void
-fs_visitor::visit(ir_texture *ir)
-{
- int sampler;
- fs_inst *inst = NULL;
-
- this->result = reg_undef;
- ir->coordinate->accept(this);
- fs_reg coordinate = this->result;
-
- if (ir->offset != NULL) {
- ir_constant *offset = ir->offset->as_constant();
- assert(offset != NULL);
-
- signed char offsets[3];
- for (unsigned i = 0; i < ir->offset->type->vector_elements; i++)
- offsets[i] = (signed char) offset->value.i[i];
-
- /* Combine all three offsets into a single unsigned dword:
- *
- * bits 11:8 - U Offset (X component)
- * bits 7:4 - V Offset (Y component)
- * bits 3:0 - R Offset (Z component)
- */
- unsigned offset_bits = 0;
- for (unsigned i = 0; i < ir->offset->type->vector_elements; i++) {
- const unsigned shift = 4 * (2 - i);
- offset_bits |= (offsets[i] << shift) & (0xF << shift);
- }
-
- /* Explicitly set up the message header by copying g0 to msg reg m1. */
- emit(BRW_OPCODE_MOV, fs_reg(MRF, 1, BRW_REGISTER_TYPE_UD),
- fs_reg(GRF, 0, BRW_REGISTER_TYPE_UD));
-
- /* Then set the offset bits in DWord 2 of the message header. */
- emit(BRW_OPCODE_MOV,
- fs_reg(retype(brw_vec1_reg(BRW_MESSAGE_REGISTER_FILE, 1, 2),
- BRW_REGISTER_TYPE_UD)),
- fs_reg(brw_imm_uw(offset_bits)));
- }
-
- /* Should be lowered by do_lower_texture_projection */
- assert(!ir->projector);
-
- sampler = _mesa_get_sampler_uniform_value(ir->sampler,
- ctx->Shader.CurrentFragmentProgram,
- &brw->fragment_program->Base);
- sampler = c->fp->program.Base.SamplerUnits[sampler];
-
- /* The 965 requires the EU to do the normalization of GL rectangle
- * texture coordinates. We use the program parameter state
- * tracking to get the scaling factor.
- */
- if (ir->sampler->type->sampler_dimensionality == GLSL_SAMPLER_DIM_RECT) {
- struct gl_program_parameter_list *params = c->fp->program.Base.Parameters;
- int tokens[STATE_LENGTH] = {
- STATE_INTERNAL,
- STATE_TEXRECT_SCALE,
- sampler,
- 0,
- 0
- };
-
- if (c->dispatch_width == 16) {
- fail("rectangle scale uniform setup not supported on 16-wide\n");
- this->result = fs_reg(this, ir->type);
- return;
- }
-
- c->prog_data.param_convert[c->prog_data.nr_params] =
- PARAM_NO_CONVERT;
- c->prog_data.param_convert[c->prog_data.nr_params + 1] =
- PARAM_NO_CONVERT;
-
- fs_reg scale_x = fs_reg(UNIFORM, c->prog_data.nr_params);
- fs_reg scale_y = fs_reg(UNIFORM, c->prog_data.nr_params + 1);
- GLuint index = _mesa_add_state_reference(params,
- (gl_state_index *)tokens);
-
- this->param_index[c->prog_data.nr_params] = index;
- this->param_offset[c->prog_data.nr_params] = 0;
- c->prog_data.nr_params++;
- this->param_index[c->prog_data.nr_params] = index;
- this->param_offset[c->prog_data.nr_params] = 1;
- c->prog_data.nr_params++;
-
- fs_reg dst = fs_reg(this, ir->coordinate->type);
- fs_reg src = coordinate;
- coordinate = dst;
-
- emit(BRW_OPCODE_MUL, dst, src, scale_x);
- dst.reg_offset++;
- src.reg_offset++;
- emit(BRW_OPCODE_MUL, dst, src, scale_y);
- }
-
- /* Writemasking doesn't eliminate channels on SIMD8 texture
- * samples, so don't worry about them.
- */
- fs_reg dst = fs_reg(this, glsl_type::vec4_type);
-
- if (intel->gen >= 7) {
- inst = emit_texture_gen7(ir, dst, coordinate, sampler);
- } else if (intel->gen >= 5) {
- inst = emit_texture_gen5(ir, dst, coordinate, sampler);
- } else {
- inst = emit_texture_gen4(ir, dst, coordinate, sampler);
- }
-
- /* If there's an offset, we already set up m1. To avoid the implied move,
- * use the null register. Otherwise, we want an implied move from g0.
- */
- if (ir->offset != NULL || !inst->header_present)
- inst->src[0] = reg_undef;
- else
- inst->src[0] = fs_reg(retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UW));
-
- inst->sampler = sampler;
-
- this->result = dst;
-
- if (ir->shadow_comparitor)
- inst->shadow_compare = true;
-
- if (ir->type == glsl_type::float_type) {
- /* Ignore DEPTH_TEXTURE_MODE swizzling. */
- assert(ir->sampler->type->sampler_shadow);
- } else if (c->key.tex_swizzles[inst->sampler] != SWIZZLE_NOOP) {
- fs_reg swizzle_dst = fs_reg(this, glsl_type::vec4_type);
-
- for (int i = 0; i < 4; i++) {
- int swiz = GET_SWZ(c->key.tex_swizzles[inst->sampler], i);
- fs_reg l = swizzle_dst;
- l.reg_offset += i;
-
- if (swiz == SWIZZLE_ZERO) {
- emit(BRW_OPCODE_MOV, l, fs_reg(0.0f));
- } else if (swiz == SWIZZLE_ONE) {
- emit(BRW_OPCODE_MOV, l, fs_reg(1.0f));
- } else {
- fs_reg r = dst;
- r.reg_offset += GET_SWZ(c->key.tex_swizzles[inst->sampler], i);
- emit(BRW_OPCODE_MOV, l, r);
- }
- }
- this->result = swizzle_dst;
- }
-}
-
-void
-fs_visitor::visit(ir_swizzle *ir)
-{
- this->result = reg_undef;
- ir->val->accept(this);
- fs_reg val = this->result;
-
- if (ir->type->vector_elements == 1) {
- this->result.reg_offset += ir->mask.x;
- return;
- }
-
- fs_reg result = fs_reg(this, ir->type);
- this->result = result;
-
- for (unsigned int i = 0; i < ir->type->vector_elements; i++) {
- fs_reg channel = val;
- int swiz = 0;
-
- switch (i) {
- case 0:
- swiz = ir->mask.x;
- break;
- case 1:
- swiz = ir->mask.y;
- break;
- case 2:
- swiz = ir->mask.z;
- break;
- case 3:
- swiz = ir->mask.w;
- break;
- }
-
- channel.reg_offset += swiz;
- emit(BRW_OPCODE_MOV, result, channel);
- result.reg_offset++;
- }
-}
-
-void
-fs_visitor::visit(ir_discard *ir)
-{
- assert(ir->condition == NULL); /* FINISHME */
-
- emit(FS_OPCODE_DISCARD);
- kill_emitted = true;
-}
-
-void
-fs_visitor::visit(ir_constant *ir)
-{
- /* Set this->result to reg at the bottom of the function because some code
- * paths will cause this visitor to be applied to other fields. This will
- * cause the value stored in this->result to be modified.
- *
- * Make reg constant so that it doesn't get accidentally modified along the
- * way. Yes, I actually had this problem. :(
- */
- const fs_reg reg(this, ir->type);
- fs_reg dst_reg = reg;
-
- if (ir->type->is_array()) {
- const unsigned size = type_size(ir->type->fields.array);
-
- for (unsigned i = 0; i < ir->type->length; i++) {
- this->result = reg_undef;
- ir->array_elements[i]->accept(this);
- fs_reg src_reg = this->result;
-
- dst_reg.type = src_reg.type;
- for (unsigned j = 0; j < size; j++) {
- emit(BRW_OPCODE_MOV, dst_reg, src_reg);
- src_reg.reg_offset++;
- dst_reg.reg_offset++;
- }
- }
- } else if (ir->type->is_record()) {
- foreach_list(node, &ir->components) {
- ir_instruction *const field = (ir_instruction *) node;
- const unsigned size = type_size(field->type);
-
- this->result = reg_undef;
- field->accept(this);
- fs_reg src_reg = this->result;
-
- dst_reg.type = src_reg.type;
- for (unsigned j = 0; j < size; j++) {
- emit(BRW_OPCODE_MOV, dst_reg, src_reg);
- src_reg.reg_offset++;
- dst_reg.reg_offset++;
- }
- }
- } else {
- const unsigned size = type_size(ir->type);
-
- for (unsigned i = 0; i < size; i++) {
- switch (ir->type->base_type) {
- case GLSL_TYPE_FLOAT:
- emit(BRW_OPCODE_MOV, dst_reg, fs_reg(ir->value.f[i]));
- break;
- case GLSL_TYPE_UINT:
- emit(BRW_OPCODE_MOV, dst_reg, fs_reg(ir->value.u[i]));
- break;
- case GLSL_TYPE_INT:
- emit(BRW_OPCODE_MOV, dst_reg, fs_reg(ir->value.i[i]));
- break;
- case GLSL_TYPE_BOOL:
- emit(BRW_OPCODE_MOV, dst_reg, fs_reg((int)ir->value.b[i]));
- break;
- default:
- assert(!"Non-float/uint/int/bool constant");
- }
- dst_reg.reg_offset++;
- }
- }
-
- this->result = reg;
-}
-
-void
-fs_visitor::emit_bool_to_cond_code(ir_rvalue *ir)
-{
- ir_expression *expr = ir->as_expression();
-
- if (expr) {
- fs_reg op[2];
- fs_inst *inst;
-
- assert(expr->get_num_operands() <= 2);
- for (unsigned int i = 0; i < expr->get_num_operands(); i++) {
- assert(expr->operands[i]->type->is_scalar());
-
- this->result = reg_undef;
- expr->operands[i]->accept(this);
- op[i] = this->result;
- }
-
- switch (expr->operation) {
- case ir_unop_logic_not:
- inst = emit(BRW_OPCODE_AND, reg_null_d, op[0], fs_reg(1));
- inst->conditional_mod = BRW_CONDITIONAL_Z;
- break;
-
- case ir_binop_logic_xor:
- inst = emit(BRW_OPCODE_XOR, reg_null_d, op[0], op[1]);
- inst->conditional_mod = BRW_CONDITIONAL_NZ;
- break;
-
- case ir_binop_logic_or:
- inst = emit(BRW_OPCODE_OR, reg_null_d, op[0], op[1]);
- inst->conditional_mod = BRW_CONDITIONAL_NZ;
- break;
-
- case ir_binop_logic_and:
- inst = emit(BRW_OPCODE_AND, reg_null_d, op[0], op[1]);
- inst->conditional_mod = BRW_CONDITIONAL_NZ;
- break;
-
- case ir_unop_f2b:
- if (intel->gen >= 6) {
- inst = emit(BRW_OPCODE_CMP, reg_null_d, op[0], fs_reg(0.0f));
- } else {
- inst = emit(BRW_OPCODE_MOV, reg_null_f, op[0]);
- }
- inst->conditional_mod = BRW_CONDITIONAL_NZ;
- break;
-
- case ir_unop_i2b:
- if (intel->gen >= 6) {
- inst = emit(BRW_OPCODE_CMP, reg_null_d, op[0], fs_reg(0));
- } else {
- inst = emit(BRW_OPCODE_MOV, reg_null_d, op[0]);
- }
- inst->conditional_mod = BRW_CONDITIONAL_NZ;
- break;
-
- case ir_binop_greater:
- case ir_binop_gequal:
- case ir_binop_less:
- case ir_binop_lequal:
- case ir_binop_equal:
- case ir_binop_all_equal:
- case ir_binop_nequal:
- case ir_binop_any_nequal:
- inst = emit(BRW_OPCODE_CMP, reg_null_cmp, op[0], op[1]);
- inst->conditional_mod =
- brw_conditional_for_comparison(expr->operation);
- break;
-
- default:
- assert(!"not reached");
- fail("bad cond code\n");
- break;
- }
- return;
- }
-
- this->result = reg_undef;
- ir->accept(this);
-
- if (intel->gen >= 6) {
- fs_inst *inst = emit(BRW_OPCODE_AND, reg_null_d, this->result, fs_reg(1));
- inst->conditional_mod = BRW_CONDITIONAL_NZ;
- } else {
- fs_inst *inst = emit(BRW_OPCODE_MOV, reg_null_d, this->result);
- inst->conditional_mod = BRW_CONDITIONAL_NZ;
- }
-}
-
-/**
- * Emit a gen6 IF statement with the comparison folded into the IF
- * instruction.
- */
-void
-fs_visitor::emit_if_gen6(ir_if *ir)
-{
- ir_expression *expr = ir->condition->as_expression();
-
- if (expr) {
- fs_reg op[2];
- fs_inst *inst;
- fs_reg temp;
-
- assert(expr->get_num_operands() <= 2);
- for (unsigned int i = 0; i < expr->get_num_operands(); i++) {
- assert(expr->operands[i]->type->is_scalar());
-
- this->result = reg_undef;
- expr->operands[i]->accept(this);
- op[i] = this->result;
- }
-
- switch (expr->operation) {
- case ir_unop_logic_not:
- inst = emit(BRW_OPCODE_IF, temp, op[0], fs_reg(0));
- inst->conditional_mod = BRW_CONDITIONAL_Z;
- return;
-
- case ir_binop_logic_xor:
- inst = emit(BRW_OPCODE_IF, reg_null_d, op[0], op[1]);
- inst->conditional_mod = BRW_CONDITIONAL_NZ;
- return;
-
- case ir_binop_logic_or:
- temp = fs_reg(this, glsl_type::bool_type);
- emit(BRW_OPCODE_OR, temp, op[0], op[1]);
- inst = emit(BRW_OPCODE_IF, reg_null_d, temp, fs_reg(0));
- inst->conditional_mod = BRW_CONDITIONAL_NZ;
- return;
-
- case ir_binop_logic_and:
- temp = fs_reg(this, glsl_type::bool_type);
- emit(BRW_OPCODE_AND, temp, op[0], op[1]);
- inst = emit(BRW_OPCODE_IF, reg_null_d, temp, fs_reg(0));
- inst->conditional_mod = BRW_CONDITIONAL_NZ;
- return;
-
- case ir_unop_f2b:
- inst = emit(BRW_OPCODE_IF, reg_null_f, op[0], fs_reg(0));
- inst->conditional_mod = BRW_CONDITIONAL_NZ;
- return;
-
- case ir_unop_i2b:
- inst = emit(BRW_OPCODE_IF, reg_null_d, op[0], fs_reg(0));
- inst->conditional_mod = BRW_CONDITIONAL_NZ;
- return;
-
- case ir_binop_greater:
- case ir_binop_gequal:
- case ir_binop_less:
- case ir_binop_lequal:
- case ir_binop_equal:
- case ir_binop_all_equal:
- case ir_binop_nequal:
- case ir_binop_any_nequal:
- inst = emit(BRW_OPCODE_IF, reg_null_d, op[0], op[1]);
- inst->conditional_mod =
- brw_conditional_for_comparison(expr->operation);
- return;
- default:
- assert(!"not reached");
- inst = emit(BRW_OPCODE_IF, reg_null_d, op[0], fs_reg(0));
- inst->conditional_mod = BRW_CONDITIONAL_NZ;
- fail("bad condition\n");
- return;
- }
- return;
- }
-
- this->result = reg_undef;
- ir->condition->accept(this);
-
- fs_inst *inst = emit(BRW_OPCODE_IF, reg_null_d, this->result, fs_reg(0));
- inst->conditional_mod = BRW_CONDITIONAL_NZ;
-}
-
-void
-fs_visitor::visit(ir_if *ir)
-{
- fs_inst *inst;
-
- if (intel->gen != 6 && c->dispatch_width == 16) {
- fail("Can't support (non-uniform) control flow on 16-wide\n");
- }
-
- /* Don't point the annotation at the if statement, because then it plus
- * the then and else blocks get printed.
- */
- this->base_ir = ir->condition;
-
- if (intel->gen == 6) {
- emit_if_gen6(ir);
- } else {
- emit_bool_to_cond_code(ir->condition);
-
- inst = emit(BRW_OPCODE_IF);
- inst->predicated = true;
- }
-
- foreach_iter(exec_list_iterator, iter, ir->then_instructions) {
- ir_instruction *ir = (ir_instruction *)iter.get();
- this->base_ir = ir;
- this->result = reg_undef;
- ir->accept(this);
- }
-
- if (!ir->else_instructions.is_empty()) {
- emit(BRW_OPCODE_ELSE);
-
- foreach_iter(exec_list_iterator, iter, ir->else_instructions) {
- ir_instruction *ir = (ir_instruction *)iter.get();
- this->base_ir = ir;
- this->result = reg_undef;
- ir->accept(this);
- }
- }
-
- emit(BRW_OPCODE_ENDIF);
-}
-
-void
-fs_visitor::visit(ir_loop *ir)
-{
- fs_reg counter = reg_undef;
-
- if (c->dispatch_width == 16) {
- fail("Can't support (non-uniform) control flow on 16-wide\n");
- }
-
- if (ir->counter) {
- this->base_ir = ir->counter;
- ir->counter->accept(this);
- counter = *(variable_storage(ir->counter));
-
- if (ir->from) {
- this->result = counter;
-
- this->base_ir = ir->from;
- this->result = counter;
- ir->from->accept(this);
-
- if (!this->result.equals(&counter))
- emit(BRW_OPCODE_MOV, counter, this->result);
- }
- }
-
- emit(BRW_OPCODE_DO);
-
- if (ir->to) {
- this->base_ir = ir->to;
- this->result = reg_undef;
- ir->to->accept(this);
-
- fs_inst *inst = emit(BRW_OPCODE_CMP, reg_null_cmp, counter, this->result);
- inst->conditional_mod = brw_conditional_for_comparison(ir->cmp);
-
- inst = emit(BRW_OPCODE_BREAK);
- inst->predicated = true;
- }
-
- foreach_iter(exec_list_iterator, iter, ir->body_instructions) {
- ir_instruction *ir = (ir_instruction *)iter.get();
-
- this->base_ir = ir;
- this->result = reg_undef;
- ir->accept(this);
- }
-
- if (ir->increment) {
- this->base_ir = ir->increment;
- this->result = reg_undef;
- ir->increment->accept(this);
- emit(BRW_OPCODE_ADD, counter, counter, this->result);
- }
-
- emit(BRW_OPCODE_WHILE);
-}
-
-void
-fs_visitor::visit(ir_loop_jump *ir)
-{
- switch (ir->mode) {
- case ir_loop_jump::jump_break:
- emit(BRW_OPCODE_BREAK);
- break;
- case ir_loop_jump::jump_continue:
- emit(BRW_OPCODE_CONTINUE);
- break;
- }
-}
-
-void
-fs_visitor::visit(ir_call *ir)
-{
- assert(!"FINISHME");
-}
-
-void
-fs_visitor::visit(ir_return *ir)
-{
- assert(!"FINISHME");
-}
-
-void
-fs_visitor::visit(ir_function *ir)
-{
- /* Ignore function bodies other than main() -- we shouldn't see calls to
- * them since they should all be inlined before we get to ir_to_mesa.
- */
- if (strcmp(ir->name, "main") == 0) {
- const ir_function_signature *sig;
- exec_list empty;
-
- sig = ir->matching_signature(&empty);
-
- assert(sig);
-
- foreach_iter(exec_list_iterator, iter, sig->body) {
- ir_instruction *ir = (ir_instruction *)iter.get();
- this->base_ir = ir;
- this->result = reg_undef;
- ir->accept(this);
- }
- }
-}
-
-void
-fs_visitor::visit(ir_function_signature *ir)
-{
- assert(!"not reached");
- (void)ir;
-}
-
-fs_inst *
-fs_visitor::emit(fs_inst inst)
-{
- fs_inst *list_inst = new(mem_ctx) fs_inst;
- *list_inst = inst;
-
- if (force_uncompressed_stack > 0)
- list_inst->force_uncompressed = true;
- else if (force_sechalf_stack > 0)
- list_inst->force_sechalf = true;
-
- list_inst->annotation = this->current_annotation;
- list_inst->ir = this->base_ir;
-
- this->instructions.push_tail(list_inst);
-
- return list_inst;
-}
-
-/** Emits a dummy fragment shader consisting of magenta for bringup purposes. */
-void
-fs_visitor::emit_dummy_fs()
-{
- /* Everyone's favorite color. */
- emit(BRW_OPCODE_MOV, fs_reg(MRF, 2), fs_reg(1.0f));
- emit(BRW_OPCODE_MOV, fs_reg(MRF, 3), fs_reg(0.0f));
- emit(BRW_OPCODE_MOV, fs_reg(MRF, 4), fs_reg(1.0f));
- emit(BRW_OPCODE_MOV, fs_reg(MRF, 5), fs_reg(0.0f));
-
- fs_inst *write;
- write = emit(FS_OPCODE_FB_WRITE, fs_reg(0), fs_reg(0));
- write->base_mrf = 0;
-}
-
-/* The register location here is relative to the start of the URB
- * data. It will get adjusted to be a real location before
- * generate_code() time.
- */
-struct brw_reg
-fs_visitor::interp_reg(int location, int channel)
-{
- int regnr = urb_setup[location] * 2 + channel / 2;
- int stride = (channel & 1) * 4;
-
- assert(urb_setup[location] != -1);
-
- return brw_vec1_grf(regnr, stride);
-}
-
-/** Emits the interpolation for the varying inputs. */
-void
-fs_visitor::emit_interpolation_setup_gen4()
-{
- this->current_annotation = "compute pixel centers";
- this->pixel_x = fs_reg(this, glsl_type::uint_type);
- this->pixel_y = fs_reg(this, glsl_type::uint_type);
- this->pixel_x.type = BRW_REGISTER_TYPE_UW;
- this->pixel_y.type = BRW_REGISTER_TYPE_UW;
-
- emit(FS_OPCODE_PIXEL_X, this->pixel_x);
- emit(FS_OPCODE_PIXEL_Y, this->pixel_y);
-
- this->current_annotation = "compute pixel deltas from v0";
- if (brw->has_pln) {
- this->delta_x = fs_reg(this, glsl_type::vec2_type);
- this->delta_y = this->delta_x;
- this->delta_y.reg_offset++;
- } else {
- this->delta_x = fs_reg(this, glsl_type::float_type);
- this->delta_y = fs_reg(this, glsl_type::float_type);
- }
- emit(BRW_OPCODE_ADD, this->delta_x,
- this->pixel_x, fs_reg(negate(brw_vec1_grf(1, 0))));
- emit(BRW_OPCODE_ADD, this->delta_y,
- this->pixel_y, fs_reg(negate(brw_vec1_grf(1, 1))));
-
- this->current_annotation = "compute pos.w and 1/pos.w";
- /* Compute wpos.w. It's always in our setup, since it's needed to
- * interpolate the other attributes.
- */
- this->wpos_w = fs_reg(this, glsl_type::float_type);
- emit(FS_OPCODE_LINTERP, wpos_w, this->delta_x, this->delta_y,
- interp_reg(FRAG_ATTRIB_WPOS, 3));
- /* Compute the pixel 1/W value from wpos.w. */
- this->pixel_w = fs_reg(this, glsl_type::float_type);
- emit_math(FS_OPCODE_RCP, this->pixel_w, wpos_w);
- this->current_annotation = NULL;
-}
-
-/** Emits the interpolation for the varying inputs. */
-void
-fs_visitor::emit_interpolation_setup_gen6()
-{
- struct brw_reg g1_uw = retype(brw_vec1_grf(1, 0), BRW_REGISTER_TYPE_UW);
-
- /* If the pixel centers end up used, the setup is the same as for gen4. */
- this->current_annotation = "compute pixel centers";
- fs_reg int_pixel_x = fs_reg(this, glsl_type::uint_type);
- fs_reg int_pixel_y = fs_reg(this, glsl_type::uint_type);
- int_pixel_x.type = BRW_REGISTER_TYPE_UW;
- int_pixel_y.type = BRW_REGISTER_TYPE_UW;
- emit(BRW_OPCODE_ADD,
- int_pixel_x,
- fs_reg(stride(suboffset(g1_uw, 4), 2, 4, 0)),
- fs_reg(brw_imm_v(0x10101010)));
- emit(BRW_OPCODE_ADD,
- int_pixel_y,
- fs_reg(stride(suboffset(g1_uw, 5), 2, 4, 0)),
- fs_reg(brw_imm_v(0x11001100)));
-
- /* As of gen6, we can no longer mix float and int sources. We have
- * to turn the integer pixel centers into floats for their actual
- * use.
- */
- this->pixel_x = fs_reg(this, glsl_type::float_type);
- this->pixel_y = fs_reg(this, glsl_type::float_type);
- emit(BRW_OPCODE_MOV, this->pixel_x, int_pixel_x);
- emit(BRW_OPCODE_MOV, this->pixel_y, int_pixel_y);
+ }
- this->current_annotation = "compute pos.w";
- this->pixel_w = fs_reg(brw_vec8_grf(c->source_w_reg, 0));
- this->wpos_w = fs_reg(this, glsl_type::float_type);
- emit_math(FS_OPCODE_RCP, this->wpos_w, this->pixel_w);
+ if (interpolation_mode == INTERP_QUALIFIER_FLAT) {
+ /* Constant interpolation (flat shading) case. The SF has
+ * handed us defined values in only the constant offset
+ * field of the setup reg.
+ */
+ for (unsigned int k = 0; k < type->vector_elements; k++) {
+ struct brw_reg interp = interp_reg(location, k);
+ interp = suboffset(interp, 3);
+ interp.type = reg->type;
+ emit(FS_OPCODE_CINTERP, attr, fs_reg(interp));
+ attr.reg_offset++;
+ }
+ } else {
+ /* Smooth/noperspective interpolation case. */
+ for (unsigned int k = 0; k < type->vector_elements; k++) {
+ /* FINISHME: At some point we probably want to push
+ * this farther by giving similar treatment to the
+ * other potentially constant components of the
+ * attribute, as well as making brw_vs_constval.c
+ * handle varyings other than gl_TexCoord.
+ */
+ if (location >= FRAG_ATTRIB_TEX0 &&
+ location <= FRAG_ATTRIB_TEX7 &&
+ k == 3 && !(c->key.proj_attrib_mask & (1 << location))) {
+ emit(BRW_OPCODE_MOV, attr, fs_reg(1.0f));
+ } else {
+ struct brw_reg interp = interp_reg(location, k);
+ brw_wm_barycentric_interp_mode barycoord_mode;
+ if (interpolation_mode == INTERP_QUALIFIER_SMOOTH)
+ barycoord_mode = BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC;
+ else
+ barycoord_mode = BRW_WM_NONPERSPECTIVE_PIXEL_BARYCENTRIC;
+ emit(FS_OPCODE_LINTERP, attr,
+ this->delta_x[barycoord_mode],
+ this->delta_y[barycoord_mode], fs_reg(interp));
+ }
+ attr.reg_offset++;
+ }
- this->delta_x = fs_reg(brw_vec8_grf(2, 0));
- this->delta_y = fs_reg(brw_vec8_grf(3, 0));
+ if (intel->gen < 6) {
+ attr.reg_offset -= type->vector_elements;
+ for (unsigned int k = 0; k < type->vector_elements; k++) {
+ emit(BRW_OPCODE_MUL, attr, attr, this->pixel_w);
+ attr.reg_offset++;
+ }
+ }
+ }
+ location++;
+ }
+ }
- this->current_annotation = NULL;
+ return reg;
}
-void
-fs_visitor::emit_color_write(int index, int first_color_mrf, fs_reg color)
+fs_reg *
+fs_visitor::emit_frontfacing_interpolation(ir_variable *ir)
{
- int reg_width = c->dispatch_width / 8;
-
- if (c->dispatch_width == 8 || intel->gen == 6) {
- /* SIMD8 write looks like:
- * m + 0: r0
- * m + 1: r1
- * m + 2: g0
- * m + 3: g1
- *
- * gen6 SIMD16 DP write looks like:
- * m + 0: r0
- * m + 1: r1
- * m + 2: g0
- * m + 3: g1
- * m + 4: b0
- * m + 5: b1
- * m + 6: a0
- * m + 7: a1
- */
- emit(BRW_OPCODE_MOV, fs_reg(MRF, first_color_mrf + index * reg_width),
- color);
+ fs_reg *reg = new(this->mem_ctx) fs_reg(this, ir->type);
+
+ /* The frontfacing comes in as a bit in the thread payload. */
+ if (intel->gen >= 6) {
+ emit(BRW_OPCODE_ASR, *reg,
+ fs_reg(retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_D)),
+ fs_reg(15));
+ emit(BRW_OPCODE_NOT, *reg, *reg);
+ emit(BRW_OPCODE_AND, *reg, *reg, fs_reg(1));
} else {
- /* pre-gen6 SIMD16 single source DP write looks like:
- * m + 0: r0
- * m + 1: g0
- * m + 2: b0
- * m + 3: a0
- * m + 4: r1
- * m + 5: g1
- * m + 6: b1
- * m + 7: a1
+ struct brw_reg r1_6ud = retype(brw_vec1_grf(1, 6), BRW_REGISTER_TYPE_UD);
+ /* bit 31 is "primitive is back face", so checking < (1 << 31) gives
+ * us front face
*/
- if (brw->has_compr4) {
- /* By setting the high bit of the MRF register number, we
- * indicate that we want COMPR4 mode - instead of doing the
- * usual destination + 1 for the second half we get
- * destination + 4.
- */
- emit(BRW_OPCODE_MOV,
- fs_reg(MRF, BRW_MRF_COMPR4 + first_color_mrf + index), color);
- } else {
- push_force_uncompressed();
- emit(BRW_OPCODE_MOV, fs_reg(MRF, first_color_mrf + index), color);
- pop_force_uncompressed();
-
- push_force_sechalf();
- color.sechalf = true;
- emit(BRW_OPCODE_MOV, fs_reg(MRF, first_color_mrf + index + 4), color);
- pop_force_sechalf();
- color.sechalf = false;
- }
+ fs_inst *inst = emit(BRW_OPCODE_CMP, *reg,
+ fs_reg(r1_6ud),
+ fs_reg(1u << 31));
+ inst->conditional_mod = BRW_CONDITIONAL_L;
+ emit(BRW_OPCODE_AND, *reg, *reg, fs_reg(1u));
}
+
+ return reg;
}
-void
-fs_visitor::emit_fb_writes()
+fs_inst *
+fs_visitor::emit_math(enum opcode opcode, fs_reg dst, fs_reg src)
{
- this->current_annotation = "FB write header";
- GLboolean header_present = GL_TRUE;
- int nr = 0;
- int reg_width = c->dispatch_width / 8;
-
- if (intel->gen >= 6 &&
- !this->kill_emitted &&
- c->key.nr_color_regions == 1) {
- header_present = false;
- }
-
- if (header_present) {
- /* m0, m1 header */
- nr += 2;
+ switch (opcode) {
+ case SHADER_OPCODE_RCP:
+ case SHADER_OPCODE_RSQ:
+ case SHADER_OPCODE_SQRT:
+ case SHADER_OPCODE_EXP2:
+ case SHADER_OPCODE_LOG2:
+ case SHADER_OPCODE_SIN:
+ case SHADER_OPCODE_COS:
+ break;
+ default:
+ assert(!"not reached: bad math opcode");
+ return NULL;
}
- if (c->aa_dest_stencil_reg) {
- push_force_uncompressed();
- emit(BRW_OPCODE_MOV, fs_reg(MRF, nr++),
- fs_reg(brw_vec8_grf(c->aa_dest_stencil_reg, 0)));
- pop_force_uncompressed();
+ /* Can't do hstride == 0 args to gen6 math, so expand it out. We
+ * might be able to do better by doing execsize = 1 math and then
+ * expanding that result out, but we would need to be careful with
+ * masking.
+ *
+ * Gen 6 hardware ignores source modifiers (negate and abs) on math
+ * instructions, so we also move to a temp to set those up.
+ */
+ if (intel->gen == 6 && (src.file == UNIFORM ||
+ src.abs ||
+ src.negate)) {
+ fs_reg expanded = fs_reg(this, glsl_type::float_type);
+ emit(BRW_OPCODE_MOV, expanded, src);
+ src = expanded;
}
- /* Reserve space for color. It'll be filled in per MRT below. */
- int color_mrf = nr;
- nr += 4 * reg_width;
-
- if (c->source_depth_to_render_target) {
- if (intel->gen == 6 && c->dispatch_width == 16) {
- /* For outputting oDepth on gen6, SIMD8 writes have to be
- * used. This would require 8-wide moves of each half to
- * message regs, kind of like pre-gen5 SIMD16 FB writes.
- * Just bail on doing so for now.
- */
- fail("Missing support for simd16 depth writes on gen6\n");
- }
-
- if (c->computes_depth) {
- /* Hand over gl_FragDepth. */
- assert(this->frag_depth);
- fs_reg depth = *(variable_storage(this->frag_depth));
+ fs_inst *inst = emit(opcode, dst, src);
- emit(BRW_OPCODE_MOV, fs_reg(MRF, nr), depth);
- } else {
- /* Pass through the payload depth. */
- emit(BRW_OPCODE_MOV, fs_reg(MRF, nr),
- fs_reg(brw_vec8_grf(c->source_depth_reg, 0)));
- }
- nr += reg_width;
+ if (intel->gen < 6) {
+ inst->base_mrf = 2;
+ inst->mlen = c->dispatch_width / 8;
}
- if (c->dest_depth_reg) {
- emit(BRW_OPCODE_MOV, fs_reg(MRF, nr),
- fs_reg(brw_vec8_grf(c->dest_depth_reg, 0)));
- nr += reg_width;
- }
+ return inst;
+}
+
+fs_inst *
+fs_visitor::emit_math(enum opcode opcode, fs_reg dst, fs_reg src0, fs_reg src1)
+{
+ int base_mrf = 2;
+ fs_inst *inst;
- fs_reg color = reg_undef;
- if (this->frag_color)
- color = *(variable_storage(this->frag_color));
- else if (this->frag_data) {
- color = *(variable_storage(this->frag_data));
- color.type = BRW_REGISTER_TYPE_F;
+ switch (opcode) {
+ case SHADER_OPCODE_POW:
+ case SHADER_OPCODE_INT_QUOTIENT:
+ case SHADER_OPCODE_INT_REMAINDER:
+ break;
+ default:
+ assert(!"not reached: unsupported binary math opcode.");
+ return NULL;
}
- for (int target = 0; target < c->key.nr_color_regions; target++) {
- this->current_annotation = ralloc_asprintf(this->mem_ctx,
- "FB write target %d",
- target);
- if (this->frag_color || this->frag_data) {
- for (int i = 0; i < 4; i++) {
- emit_color_write(i, color_mrf, color);
- color.reg_offset++;
- }
+ if (intel->gen >= 7) {
+ inst = emit(opcode, dst, src0, src1);
+ } else if (intel->gen == 6) {
+ /* Can't do hstride == 0 args to gen6 math, so expand it out.
+ *
+ * The hardware ignores source modifiers (negate and abs) on math
+ * instructions, so we also move to a temp to set those up.
+ */
+ if (src0.file == UNIFORM || src0.abs || src0.negate) {
+ fs_reg expanded = fs_reg(this, glsl_type::float_type);
+ expanded.type = src0.type;
+ emit(BRW_OPCODE_MOV, expanded, src0);
+ src0 = expanded;
}
- if (this->frag_color)
- color.reg_offset -= 4;
+ if (src1.file == UNIFORM || src1.abs || src1.negate) {
+ fs_reg expanded = fs_reg(this, glsl_type::float_type);
+ expanded.type = src1.type;
+ emit(BRW_OPCODE_MOV, expanded, src1);
+ src1 = expanded;
+ }
- fs_inst *inst = emit(FS_OPCODE_FB_WRITE);
- inst->target = target;
- inst->base_mrf = 0;
- inst->mlen = nr;
- if (target == c->key.nr_color_regions - 1)
- inst->eot = true;
- inst->header_present = header_present;
- }
+ inst = emit(opcode, dst, src0, src1);
+ } else {
+ /* From the Ironlake PRM, Volume 4, Part 1, Section 6.1.13
+ * "Message Payload":
+ *
+ * "Operand0[7]. For the INT DIV functions, this operand is the
+ * denominator."
+ * ...
+ * "Operand1[7]. For the INT DIV functions, this operand is the
+ * numerator."
+ */
+ bool is_int_div = opcode != SHADER_OPCODE_POW;
+ fs_reg &op0 = is_int_div ? src1 : src0;
+ fs_reg &op1 = is_int_div ? src0 : src1;
- if (c->key.nr_color_regions == 0) {
- if (c->key.alpha_test && (this->frag_color || this->frag_data)) {
- /* If the alpha test is enabled but there's no color buffer,
- * we still need to send alpha out the pipeline to our null
- * renderbuffer.
- */
- color.reg_offset += 3;
- emit_color_write(3, color_mrf, color);
- }
+ emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + 1, op1.type), op1);
+ inst = emit(opcode, dst, op0, reg_null_f);
- fs_inst *inst = emit(FS_OPCODE_FB_WRITE);
- inst->base_mrf = 0;
- inst->mlen = nr;
- inst->eot = true;
- inst->header_present = header_present;
+ inst->base_mrf = base_mrf;
+ inst->mlen = 2 * c->dispatch_width / 8;
}
-
- this->current_annotation = NULL;
+ return inst;
}
/**
/* Set up the pointers to ParamValues now that that array is finalized. */
for (unsigned int i = 0; i < c->prog_data.nr_params; i++) {
c->prog_data.param[i] =
- fp->Base.Parameters->ParameterValues[this->param_index[i]] +
+ (const float *)fp->Base.Parameters->ParameterValues[this->param_index[i]] +
this->param_offset[i];
}
}
}
/* Map the offsets in the UNIFORM file to fixed HW regs. */
- foreach_iter(exec_list_iterator, iter, this->instructions) {
- fs_inst *inst = (fs_inst *)iter.get();
+ foreach_list(node, &this->instructions) {
+ fs_inst *inst = (fs_inst *)node;
for (unsigned int i = 0; i < 3; i++) {
if (inst->src[i].file == UNIFORM) {
- int constant_nr = inst->src[i].hw_reg + inst->src[i].reg_offset;
+ int constant_nr = inst->src[i].reg + inst->src[i].reg_offset;
struct brw_reg brw_reg = brw_vec1_grf(c->nr_payload_regs +
constant_nr / 8,
constant_nr % 8);
/* Figure out where each of the incoming setup attributes lands. */
if (intel->gen >= 6) {
for (unsigned int i = 0; i < FRAG_ATTRIB_MAX; i++) {
- if (brw->fragment_program->Base.InputsRead & BITFIELD64_BIT(i)) {
+ if (fp->Base.InputsRead & BITFIELD64_BIT(i)) {
urb_setup[i] = urb_next++;
}
}
/* FINISHME: The sf doesn't map VS->FS inputs for us very well. */
for (unsigned int i = 0; i < VERT_RESULT_MAX; i++) {
if (c->key.vp_outputs_written & BITFIELD64_BIT(i)) {
- int fp_index;
-
- if (i >= VERT_RESULT_VAR0)
- fp_index = i - (VERT_RESULT_VAR0 - FRAG_ATTRIB_VAR0);
- else if (i <= VERT_RESULT_TEX7)
- fp_index = i;
- else
- fp_index = -1;
+ int fp_index = _mesa_vert_result_to_frag_attrib((gl_vert_result) i);
if (fp_index >= 0)
urb_setup[fp_index] = urb_next++;
/* Offset all the urb_setup[] index by the actual position of the
* setup regs, now that the location of the constants has been chosen.
*/
- foreach_iter(exec_list_iterator, iter, this->instructions) {
- fs_inst *inst = (fs_inst *)iter.get();
+ foreach_list(node, &this->instructions) {
+ fs_inst *inst = (fs_inst *)node;
if (inst->opcode == FS_OPCODE_LINTERP) {
assert(inst->src[2].file == FIXED_HW_REG);
split_grf[i] = false;
}
- if (brw->has_pln) {
- /* PLN opcodes rely on the delta_xy being contiguous. */
- split_grf[this->delta_x.reg] = false;
+ if (brw->has_pln &&
+ this->delta_x[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC].file == GRF) {
+ /* PLN opcodes rely on the delta_xy being contiguous. We only have to
+ * check this for BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC, because prior to
+ * Gen6, that was the only supported interpolation mode, and since Gen6,
+ * delta_x and delta_y are in fixed hardware registers.
+ */
+ split_grf[this->delta_x[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC].reg] =
+ false;
}
- foreach_iter(exec_list_iterator, iter, this->instructions) {
- fs_inst *inst = (fs_inst *)iter.get();
+ foreach_list(node, &this->instructions) {
+ fs_inst *inst = (fs_inst *)node;
/* Texturing produces 4 contiguous registers, so no splitting. */
if (inst->is_tex()) {
}
}
- foreach_iter(exec_list_iterator, iter, this->instructions) {
- fs_inst *inst = (fs_inst *)iter.get();
+ foreach_list(node, &this->instructions) {
+ fs_inst *inst = (fs_inst *)node;
if (inst->dst.file == GRF &&
split_grf[inst->dst.reg] &&
this->live_intervals_valid = false;
}
+bool
+fs_visitor::remove_dead_constants()
+{
+ if (c->dispatch_width == 8) {
+ this->params_remap = ralloc_array(mem_ctx, int, c->prog_data.nr_params);
+
+ for (unsigned int i = 0; i < c->prog_data.nr_params; i++)
+ this->params_remap[i] = -1;
+
+ /* Find which params are still in use. */
+ foreach_list(node, &this->instructions) {
+ fs_inst *inst = (fs_inst *)node;
+
+ for (int i = 0; i < 3; i++) {
+ int constant_nr = inst->src[i].reg + inst->src[i].reg_offset;
+
+ if (inst->src[i].file != UNIFORM)
+ continue;
+
+ assert(constant_nr < (int)c->prog_data.nr_params);
+
+ /* For now, set this to non-negative. We'll give it the
+ * actual new number in a moment, in order to keep the
+ * register numbers nicely ordered.
+ */
+ this->params_remap[constant_nr] = 0;
+ }
+ }
+
+ /* Figure out what the new numbers for the params will be. At some
+ * point when we're doing uniform array access, we're going to want
+ * to keep the distinction between .reg and .reg_offset, but for
+ * now we don't care.
+ */
+ unsigned int new_nr_params = 0;
+ for (unsigned int i = 0; i < c->prog_data.nr_params; i++) {
+ if (this->params_remap[i] != -1) {
+ this->params_remap[i] = new_nr_params++;
+ }
+ }
+
+ /* Update the list of params to be uploaded to match our new numbering. */
+ for (unsigned int i = 0; i < c->prog_data.nr_params; i++) {
+ int remapped = this->params_remap[i];
+
+ if (remapped == -1)
+ continue;
+
+ /* We've already done setup_paramvalues_refs() so no need to worry
+ * about param_index and param_offset.
+ */
+ c->prog_data.param[remapped] = c->prog_data.param[i];
+ c->prog_data.param_convert[remapped] = c->prog_data.param_convert[i];
+ }
+
+ c->prog_data.nr_params = new_nr_params;
+ } else {
+ /* This should have been generated in the 8-wide pass already. */
+ assert(this->params_remap);
+ }
+
+ /* Now do the renumbering of the shader to remove unused params. */
+ foreach_list(node, &this->instructions) {
+ fs_inst *inst = (fs_inst *)node;
+
+ for (int i = 0; i < 3; i++) {
+ int constant_nr = inst->src[i].reg + inst->src[i].reg_offset;
+
+ if (inst->src[i].file != UNIFORM)
+ continue;
+
+ assert(this->params_remap[constant_nr] != -1);
+ inst->src[i].reg = this->params_remap[constant_nr];
+ inst->src[i].reg_offset = 0;
+ }
+ }
+
+ return true;
+}
+
/**
* Choose accesses from the UNIFORM file to demote to using the pull
* constant buffer.
int pull_uniform_base = max_uniform_components;
int pull_uniform_count = c->prog_data.nr_params - pull_uniform_base;
- foreach_iter(exec_list_iterator, iter, this->instructions) {
- fs_inst *inst = (fs_inst *)iter.get();
+ foreach_list(node, &this->instructions) {
+ fs_inst *inst = (fs_inst *)node;
for (int i = 0; i < 3; i++) {
if (inst->src[i].file != UNIFORM)
continue;
- int uniform_nr = inst->src[i].hw_reg + inst->src[i].reg_offset;
+ int uniform_nr = inst->src[i].reg + inst->src[i].reg_offset;
if (uniform_nr < pull_uniform_base)
continue;
}
int ip = 0;
- foreach_iter(exec_list_iterator, iter, this->instructions) {
- fs_inst *inst = (fs_inst *)iter.get();
+ foreach_list(node, &this->instructions) {
+ fs_inst *inst = (fs_inst *)node;
if (inst->opcode == BRW_OPCODE_DO) {
if (loop_depth++ == 0)
}
} else {
for (unsigned int i = 0; i < 3; i++) {
- if (inst->src[i].file == GRF && inst->src[i].reg != 0) {
+ if (inst->src[i].file == GRF) {
int reg = inst->src[i].reg;
if (!loop_depth) {
}
}
}
- if (inst->dst.file == GRF && inst->dst.reg != 0) {
+ if (inst->dst.file == GRF) {
int reg = inst->dst.reg;
if (!loop_depth) {
calculate_live_intervals();
- foreach_iter(exec_list_iterator, iter, this->instructions) {
- fs_inst *inst = (fs_inst *)iter.get();
+ foreach_list(node, &this->instructions) {
+ fs_inst *inst = (fs_inst *)node;
if (inst->opcode != BRW_OPCODE_MOV ||
inst->predicated ||
/* Found a move of a constant to a GRF. Find anything else using the GRF
* before it's written, and replace it with the constant if we can.
*/
- exec_list_iterator scan_iter = iter;
- scan_iter.next();
- for (; scan_iter.has_next(); scan_iter.next()) {
- fs_inst *scan_inst = (fs_inst *)scan_iter.get();
-
+ for (fs_inst *scan_inst = (fs_inst *)inst->next;
+ !scan_inst->is_tail_sentinel();
+ scan_inst = (fs_inst *)scan_inst->next) {
if (scan_inst->opcode == BRW_OPCODE_DO ||
scan_inst->opcode == BRW_OPCODE_WHILE ||
scan_inst->opcode == BRW_OPCODE_ELSE ||
scan_inst->src[i] = inst->src[0];
progress = true;
} else if (i == 0 && scan_inst->src[1].file != IMM) {
- /* Fit this constant in by commuting the operands */
+ /* Fit this constant in by commuting the operands.
+ * Exception: we can't do this for 32-bit integer MUL
+ * because it's asymmetric.
+ */
+ if (scan_inst->opcode == BRW_OPCODE_MUL &&
+ (scan_inst->src[1].type == BRW_REGISTER_TYPE_D ||
+ scan_inst->src[1].type == BRW_REGISTER_TYPE_UD))
+ break;
scan_inst->src[0] = scan_inst->src[1];
scan_inst->src[1] = inst->src[0];
progress = true;
break;
case BRW_OPCODE_CMP:
+ case BRW_OPCODE_IF:
if (i == 1) {
scan_inst->src[i] = inst->src[0];
progress = true;
scan_inst->src[i] = inst->src[0];
progress = true;
} else if (i == 0 && scan_inst->src[1].file != IMM) {
- /* Fit this constant in by swapping the operands and
- * flipping the predicate
- */
scan_inst->src[0] = scan_inst->src[1];
scan_inst->src[1] = inst->src[0];
- scan_inst->predicate_inverse = !scan_inst->predicate_inverse;
+
+ /* If this was predicated, flipping operands means
+ * we also need to flip the predicate.
+ */
+ if (scan_inst->conditional_mod == BRW_CONDITIONAL_NONE) {
+ scan_inst->predicate_inverse =
+ !scan_inst->predicate_inverse;
+ }
+ progress = true;
+ }
+ break;
+
+ case SHADER_OPCODE_RCP:
+ /* The hardware doesn't do math on immediate values
+ * (because why are you doing that, seriously?), but
+ * the correct answer is to just constant fold it
+ * anyway.
+ */
+ assert(i == 0);
+ if (inst->src[0].imm.f != 0.0f) {
+ scan_inst->opcode = BRW_OPCODE_MOV;
+ scan_inst->src[0] = inst->src[0];
+ scan_inst->src[0].imm.f = 1.0f / scan_inst->src[0].imm.f;
progress = true;
}
break;
+
+ default:
+ break;
}
}
return progress;
}
+
+
+/**
+ * Attempts to move immediate constants into the immediate
+ * constant slot of following instructions.
+ *
+ * Immediate constants are a bit tricky -- they have to be in the last
+ * operand slot, you can't do abs/negate on them,
+ */
+
+bool
+fs_visitor::opt_algebraic()
+{
+ bool progress = false;
+
+ calculate_live_intervals();
+
+ foreach_list(node, &this->instructions) {
+ fs_inst *inst = (fs_inst *)node;
+
+ switch (inst->opcode) {
+ case BRW_OPCODE_MUL:
+ if (inst->src[1].file != IMM)
+ continue;
+
+ /* a * 1.0 = a */
+ if (inst->src[1].type == BRW_REGISTER_TYPE_F &&
+ inst->src[1].imm.f == 1.0) {
+ inst->opcode = BRW_OPCODE_MOV;
+ inst->src[1] = reg_undef;
+ progress = true;
+ break;
+ }
+
+ break;
+ default:
+ break;
+ }
+ }
+
+ return progress;
+}
+
/**
* Must be called after calculate_live_intervales() to remove unused
* writes to registers -- register allocation will fail otherwise
calculate_live_intervals();
- foreach_iter(exec_list_iterator, iter, this->instructions) {
- fs_inst *inst = (fs_inst *)iter.get();
+ foreach_list_safe(node, &this->instructions) {
+ fs_inst *inst = (fs_inst *)node;
if (inst->dst.file == GRF && this->virtual_grf_use[inst->dst.reg] <= pc) {
inst->remove();
int if_depth = 0;
int loop_depth = 0;
- foreach_iter(exec_list_iterator, iter, this->instructions) {
- fs_inst *inst = (fs_inst *)iter.get();
+ foreach_list_safe(node, &this->instructions) {
+ fs_inst *inst = (fs_inst *)node;
/* Make sure that we dominate the instructions we're going to
* scan for interfering with our coalescing, or we won't have
case BRW_OPCODE_ENDIF:
if_depth--;
break;
+ default:
+ break;
}
if (loop_depth || if_depth)
continue;
if (inst->opcode != BRW_OPCODE_MOV ||
inst->predicated ||
inst->saturate ||
- inst->dst.file != GRF || inst->src[0].file != GRF ||
+ inst->dst.file != GRF || (inst->src[0].file != GRF &&
+ inst->src[0].file != UNIFORM)||
inst->dst.type != inst->src[0].type)
continue;
* program.
*/
bool interfered = false;
- exec_list_iterator scan_iter = iter;
- scan_iter.next();
- for (; scan_iter.has_next(); scan_iter.next()) {
- fs_inst *scan_inst = (fs_inst *)scan_iter.get();
+ for (fs_inst *scan_inst = (fs_inst *)inst->next;
+ !scan_inst->is_tail_sentinel();
+ scan_inst = (fs_inst *)scan_inst->next) {
if (scan_inst->dst.file == GRF) {
if (scan_inst->dst.reg == inst->dst.reg &&
(scan_inst->dst.reg_offset == inst->dst.reg_offset ||
interfered = true;
break;
}
- if (scan_inst->dst.reg == inst->src[0].reg &&
+ if (inst->src[0].file == GRF &&
+ scan_inst->dst.reg == inst->src[0].reg &&
(scan_inst->dst.reg_offset == inst->src[0].reg_offset ||
scan_inst->is_tex())) {
interfered = true;
}
}
- /* The gen6 MATH instruction can't handle source modifiers, so avoid
- * coalescing those for now. We should do something more specific.
+ /* The gen6 MATH instruction can't handle source modifiers or
+ * unusual register regions, so avoid coalescing those for
+ * now. We should do something more specific.
+ */
+ if (intel->gen >= 6 &&
+ scan_inst->is_math() &&
+ (has_source_modifiers || inst->src[0].file == UNIFORM)) {
+ interfered = true;
+ break;
+ }
+
+ /* The accumulator result appears to get used for the
+ * conditional modifier generation. When negating a UD
+ * value, there is a 33rd bit generated for the sign in the
+ * accumulator value, so now you can't check, for example,
+ * equality with a 32-bit value. See piglit fs-op-neg-uint.
*/
- if (intel->gen >= 6 && scan_inst->is_math() && has_source_modifiers) {
+ if (scan_inst->conditional_mod &&
+ inst->src[0].negate &&
+ inst->src[0].type == BRW_REGISTER_TYPE_UD) {
interfered = true;
break;
}
/* Rewrite the later usage to point at the source of the move to
* be removed.
*/
- for (exec_list_iterator scan_iter = iter; scan_iter.has_next();
- scan_iter.next()) {
- fs_inst *scan_inst = (fs_inst *)scan_iter.get();
-
+ for (fs_inst *scan_inst = inst;
+ !scan_inst->is_tail_sentinel();
+ scan_inst = (fs_inst *)scan_inst->next) {
for (int i = 0; i < 3; i++) {
if (scan_inst->src[i].file == GRF &&
scan_inst->src[i].reg == inst->dst.reg &&
scan_inst->src[i].reg_offset == inst->dst.reg_offset) {
- scan_inst->src[i].reg = inst->src[0].reg;
- scan_inst->src[i].reg_offset = inst->src[0].reg_offset;
- scan_inst->src[i].abs |= inst->src[0].abs;
- scan_inst->src[i].negate ^= inst->src[0].negate;
- scan_inst->src[i].smear = inst->src[0].smear;
+ fs_reg new_src = inst->src[0];
+ if (scan_inst->src[i].abs) {
+ new_src.negate = 0;
+ new_src.abs = 1;
+ }
+ new_src.negate ^= scan_inst->src[i].negate;
+ scan_inst->src[i] = new_src;
}
}
}
calculate_live_intervals();
- foreach_iter(exec_list_iterator, iter, this->instructions) {
- fs_inst *inst = (fs_inst *)iter.get();
+ foreach_list_safe(node, &this->instructions) {
+ fs_inst *inst = (fs_inst *)node;
int ip = next_ip;
next_ip++;
/* Work out which hardware MRF registers are written by this
* instruction.
*/
- int mrf_low = inst->dst.hw_reg & ~BRW_MRF_COMPR4;
+ int mrf_low = inst->dst.reg & ~BRW_MRF_COMPR4;
int mrf_high;
- if (inst->dst.hw_reg & BRW_MRF_COMPR4) {
+ if (inst->dst.reg & BRW_MRF_COMPR4) {
mrf_high = mrf_low + 4;
} else if (c->dispatch_width == 16 &&
(!inst->force_uncompressed && !inst->force_sechalf)) {
if (scan_inst->dst.reg_offset == inst->src[0].reg_offset) {
/* Found the creator of our MRF's source value. */
scan_inst->dst.file = MRF;
- scan_inst->dst.hw_reg = inst->dst.hw_reg;
+ scan_inst->dst.reg = inst->dst.reg;
scan_inst->saturate |= inst->saturate;
inst->remove();
progress = true;
/* If somebody else writes our MRF here, we can't
* compute-to-MRF before that.
*/
- int scan_mrf_low = scan_inst->dst.hw_reg & ~BRW_MRF_COMPR4;
+ int scan_mrf_low = scan_inst->dst.reg & ~BRW_MRF_COMPR4;
int scan_mrf_high;
- if (scan_inst->dst.hw_reg & BRW_MRF_COMPR4) {
+ if (scan_inst->dst.reg & BRW_MRF_COMPR4) {
scan_mrf_high = scan_mrf_low + 4;
} else if (c->dispatch_width == 16 &&
(!scan_inst->force_uncompressed &&
memset(last_mrf_move, 0, sizeof(last_mrf_move));
- foreach_iter(exec_list_iterator, iter, this->instructions) {
- fs_inst *inst = (fs_inst *)iter.get();
+ foreach_list_safe(node, &this->instructions) {
+ fs_inst *inst = (fs_inst *)node;
switch (inst->opcode) {
case BRW_OPCODE_DO:
if (inst->opcode == BRW_OPCODE_MOV &&
inst->dst.file == MRF) {
- fs_inst *prev_inst = last_mrf_move[inst->dst.hw_reg];
+ fs_inst *prev_inst = last_mrf_move[inst->dst.reg];
if (prev_inst && inst->equals(prev_inst)) {
inst->remove();
progress = true;
/* Clear out the last-write records for MRFs that were overwritten. */
if (inst->dst.file == MRF) {
- last_mrf_move[inst->dst.hw_reg] = NULL;
+ last_mrf_move[inst->dst.reg] = NULL;
}
if (inst->mlen > 0) {
inst->dst.file == MRF &&
inst->src[0].file == GRF &&
!inst->predicated) {
- last_mrf_move[inst->dst.hw_reg] = inst;
+ last_mrf_move[inst->dst.reg] = inst;
}
}
/* Generate FS IR for main(). (the visitor only descends into
* functions called "main").
*/
- foreach_iter(exec_list_iterator, iter, *shader->ir) {
- ir_instruction *ir = (ir_instruction *)iter.get();
+ foreach_list(node, &*shader->ir) {
+ ir_instruction *ir = (ir_instruction *)node;
base_ir = ir;
this->result = reg_undef;
ir->accept(this);
}
+ if (failed)
+ return false;
emit_fb_writes();
progress = remove_duplicate_mrf_writes() || progress;
progress = propagate_constants() || progress;
+ progress = opt_algebraic() || progress;
progress = register_coalesce() || progress;
progress = compute_to_mrf() || progress;
progress = dead_code_eliminate() || progress;
} while (progress);
+ remove_dead_constants();
+
schedule_instructions();
assign_curb_setup();
if (0) {
/* Debug of register spilling: Go spill everything. */
int virtual_grf_count = virtual_grf_next;
- for (int i = 1; i < virtual_grf_count; i++) {
+ for (int i = 0; i < virtual_grf_count; i++) {
spill_reg(i);
}
}
/* Make sure we didn't try to sneak in an extra uniform */
assert(orig_nr_params == c->prog_data.nr_params);
+ (void) orig_nr_params;
}
return !failed;
}
bool
-brw_wm_fs_emit(struct brw_context *brw, struct brw_wm_compile *c)
+brw_wm_fs_emit(struct brw_context *brw, struct brw_wm_compile *c,
+ struct gl_shader_program *prog)
{
struct intel_context *intel = &brw->intel;
- struct gl_context *ctx = &intel->ctx;
- struct gl_shader_program *prog = ctx->Shader.CurrentFragmentProgram;
if (!prog)
return false;
*/
c->dispatch_width = 8;
- fs_visitor v(c, shader);
+ fs_visitor v(c, prog, shader);
if (!v.run()) {
- /* FINISHME: Cleanly fail, test at link time, etc. */
- assert(!"not reached");
+ prog->LinkStatus = false;
+ ralloc_strcat(&prog->InfoLog, v.fail_msg);
+
return false;
}
if (intel->gen >= 5 && c->prog_data.nr_pull_params == 0) {
c->dispatch_width = 16;
- fs_visitor v2(c, shader);
- v2.import_uniforms(v.variable_ht);
+ fs_visitor v2(c, prog, shader);
+ v2.import_uniforms(&v);
v2.run();
}
return true;
}
+
+bool
+brw_fs_precompile(struct gl_context *ctx, struct gl_shader_program *prog)
+{
+ struct brw_context *brw = brw_context(ctx);
+ struct brw_wm_prog_key key;
+
+ if (!prog->_LinkedShaders[MESA_SHADER_FRAGMENT])
+ return true;
+
+ struct gl_fragment_program *fp = (struct gl_fragment_program *)
+ prog->_LinkedShaders[MESA_SHADER_FRAGMENT]->Program;
+ struct brw_fragment_program *bfp = brw_fragment_program(fp);
+
+ memset(&key, 0, sizeof(key));
+
+ if (fp->UsesKill)
+ key.iz_lookup |= IZ_PS_KILL_ALPHATEST_BIT;
+
+ if (fp->Base.OutputsWritten & BITFIELD64_BIT(FRAG_RESULT_DEPTH))
+ key.iz_lookup |= IZ_PS_COMPUTES_DEPTH_BIT;
+
+ /* Just assume depth testing. */
+ key.iz_lookup |= IZ_DEPTH_TEST_ENABLE_BIT;
+ key.iz_lookup |= IZ_DEPTH_WRITE_ENABLE_BIT;
+
+ key.vp_outputs_written |= BITFIELD64_BIT(FRAG_ATTRIB_WPOS);
+ for (int i = 0; i < FRAG_ATTRIB_MAX; i++) {
+ if (!(fp->Base.InputsRead & BITFIELD64_BIT(i)))
+ continue;
+
+ key.proj_attrib_mask |= 1 << i;
+
+ int vp_index = _mesa_vert_result_to_frag_attrib((gl_vert_result) i);
+
+ if (vp_index >= 0)
+ key.vp_outputs_written |= BITFIELD64_BIT(vp_index);
+ }
+
+ key.clamp_fragment_color = true;
+
+ for (int i = 0; i < BRW_MAX_TEX_UNIT; i++) {
+ if (fp->Base.ShadowSamplers & (1 << i))
+ key.tex.compare_funcs[i] = GL_LESS;
+
+ /* FINISHME: depth compares might use (0,0,0,W) for example */
+ key.tex.swizzles[i] = SWIZZLE_XYZW;
+ }
+
+ if (fp->Base.InputsRead & FRAG_BIT_WPOS) {
+ key.drawable_height = ctx->DrawBuffer->Height;
+ key.render_to_fbo = ctx->DrawBuffer->Name != 0;
+ }
+
+ key.nr_color_regions = 1;
+
+ key.program_string_id = bfp->id;
+
+ uint32_t old_prog_offset = brw->wm.prog_offset;
+ struct brw_wm_prog_data *old_prog_data = brw->wm.prog_data;
+
+ bool success = do_wm_prog(brw, prog, bfp, &key);
+
+ brw->wm.prog_offset = old_prog_offset;
+ brw->wm.prog_data = old_prog_data;
+
+ return success;
+}