#include <sys/types.h>
+#include "main/hash_table.h"
#include "main/macros.h"
#include "main/shaderobj.h"
#include "main/uniforms.h"
}
#include "brw_fs.h"
#include "glsl/glsl_types.h"
-#include "glsl/ir_print_visitor.h"
void
fs_inst::init()
this->src[0] = reg_undef;
this->src[1] = reg_undef;
this->src[2] = reg_undef;
+
+ /* This will be the case for almost all instructions. */
+ this->regs_written = 1;
}
fs_inst::fs_inst()
return new(mem_ctx) fs_inst(BRW_OPCODE_##op, dst, src0, src1); \
}
+#define ALU3(op) \
+ fs_inst * \
+ fs_visitor::op(fs_reg dst, fs_reg src0, fs_reg src1, fs_reg src2) \
+ { \
+ return new(mem_ctx) fs_inst(BRW_OPCODE_##op, dst, src0, src1, src2);\
+ }
+
ALU1(NOT)
ALU1(MOV)
ALU1(FRC)
ALU2(SHL)
ALU2(SHR)
ALU2(ASR)
+ALU3(LRP)
+ALU1(BFREV)
+ALU3(BFE)
+ALU2(BFI1)
+ALU3(BFI2)
+ALU1(FBH)
+ALU1(FBL)
+ALU1(CBIT)
/** Gen4 predicated IF. */
fs_inst *
*/
if (intel->gen == 4) {
dst.type = src0.type;
- if (dst.file == FIXED_HW_REG)
+ if (dst.file == HW_REG)
dst.fixed_hw_reg.type = dst.type;
}
return inst;
}
+exec_list
+fs_visitor::VARYING_PULL_CONSTANT_LOAD(fs_reg dst, fs_reg surf_index,
+ fs_reg varying_offset,
+ uint32_t const_offset)
+{
+ exec_list instructions;
+ fs_inst *inst;
+
+ /* We have our constant surface use a pitch of 4 bytes, so our index can
+ * be any component of a vector, and then we load 4 contiguous
+ * components starting from that.
+ *
+ * We break down the const_offset to a portion added to the variable
+ * offset and a portion done using reg_offset, which means that if you
+ * have GLSL using something like "uniform vec4 a[20]; gl_FragColor =
+ * a[i]", we'll temporarily generate 4 vec4 loads from offset i * 4, and
+ * CSE can later notice that those loads are all the same and eliminate
+ * the redundant ones.
+ */
+ fs_reg vec4_offset = fs_reg(this, glsl_type::int_type);
+ instructions.push_tail(ADD(vec4_offset,
+ varying_offset, const_offset & ~3));
+
+ int scale = 1;
+ if (intel->gen == 4 && dispatch_width == 8) {
+ /* Pre-gen5, we can either use a SIMD8 message that requires (header,
+ * u, v, r) as parameters, or we can just use the SIMD16 message
+ * consisting of (header, u). We choose the second, at the cost of a
+ * longer return length.
+ */
+ scale = 2;
+ }
+
+ enum opcode op;
+ if (intel->gen >= 7)
+ op = FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GEN7;
+ else
+ op = FS_OPCODE_VARYING_PULL_CONSTANT_LOAD;
+ fs_reg vec4_result = fs_reg(GRF, virtual_grf_alloc(4 * scale), dst.type);
+ inst = new(mem_ctx) fs_inst(op, vec4_result, surf_index, vec4_offset);
+ inst->regs_written = 4 * scale;
+ instructions.push_tail(inst);
+
+ if (intel->gen < 7) {
+ inst->base_mrf = 13;
+ inst->header_present = true;
+ if (intel->gen == 4)
+ inst->mlen = 3;
+ else
+ inst->mlen = 1 + dispatch_width / 8;
+ }
+
+ vec4_result.reg_offset += (const_offset & 3) * scale;
+ instructions.push_tail(MOV(dst, vec4_result));
+
+ return instructions;
+}
+
+/**
+ * A helper for MOV generation for fixing up broken hardware SEND dependency
+ * handling.
+ */
+fs_inst *
+fs_visitor::DEP_RESOLVE_MOV(int grf)
+{
+ fs_inst *inst = MOV(brw_null_reg(), fs_reg(GRF, grf, BRW_REGISTER_TYPE_F));
+
+ inst->ir = NULL;
+ inst->annotation = "send dependency resolve";
+
+ /* The caller always wants uncompressed to emit the minimal extra
+ * dependencies, and to avoid having to deal with aligning its regs to 2.
+ */
+ inst->force_uncompressed = true;
+
+ return inst;
+}
+
bool
fs_inst::equals(fs_inst *inst)
{
offset == inst->offset);
}
-int
-fs_inst::regs_written()
-{
- if (is_tex())
- return 4;
-
- /* The SINCOS and INT_DIV_QUOTIENT_AND_REMAINDER math functions return 2,
- * but we don't currently use them...nor do we have an opcode for them.
- */
-
- return 1;
-}
-
bool
fs_inst::overwrites_reg(const fs_reg ®)
{
return (reg.file == dst.file &&
reg.reg == dst.reg &&
reg.reg_offset >= dst.reg_offset &&
- reg.reg_offset < dst.reg_offset + regs_written());
+ reg.reg_offset < dst.reg_offset + regs_written);
}
bool
-fs_inst::is_tex()
+fs_inst::is_send_from_grf()
{
- return (opcode == SHADER_OPCODE_TEX ||
- opcode == FS_OPCODE_TXB ||
- opcode == SHADER_OPCODE_TXD ||
- opcode == SHADER_OPCODE_TXF ||
- opcode == SHADER_OPCODE_TXL ||
- opcode == SHADER_OPCODE_TXS);
+ return (opcode == FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GEN7 ||
+ opcode == SHADER_OPCODE_SHADER_TIME_ADD ||
+ (opcode == FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD &&
+ src[1].file == GRF));
}
bool
-fs_inst::is_math()
+fs_visitor::can_do_source_mods(fs_inst *inst)
{
- return (opcode == SHADER_OPCODE_RCP ||
- opcode == SHADER_OPCODE_RSQ ||
- opcode == SHADER_OPCODE_SQRT ||
- opcode == SHADER_OPCODE_EXP2 ||
- opcode == SHADER_OPCODE_LOG2 ||
- opcode == SHADER_OPCODE_SIN ||
- opcode == SHADER_OPCODE_COS ||
- opcode == SHADER_OPCODE_INT_QUOTIENT ||
- opcode == SHADER_OPCODE_INT_REMAINDER ||
- opcode == SHADER_OPCODE_POW);
+ if (intel->gen == 6 && inst->is_math())
+ return false;
+
+ if (inst->is_send_from_grf())
+ return false;
+
+ return true;
}
void
fs_reg::fs_reg(struct brw_reg fixed_hw_reg)
{
init();
- this->file = FIXED_HW_REG;
+ this->file = HW_REG;
this->fixed_hw_reg = fixed_hw_reg;
this->type = fixed_hw_reg.type;
}
type == r.type &&
negate == r.negate &&
abs == r.abs &&
+ !reladdr && !r.reladdr &&
memcmp(&fixed_hw_reg, &r.fixed_hw_reg,
sizeof(fixed_hw_reg)) == 0 &&
smear == r.smear &&
return type == BRW_REGISTER_TYPE_F ? imm.f == 1.0 : imm.i == 1;
}
+bool
+fs_reg::is_valid_3src() const
+{
+ return file == GRF || file == UNIFORM;
+}
+
int
fs_visitor::type_size(const struct glsl_type *type)
{
* link time.
*/
return 0;
- default:
+ case GLSL_TYPE_VOID:
+ case GLSL_TYPE_ERROR:
+ case GLSL_TYPE_INTERFACE:
assert(!"not reached");
- return 0;
+ break;
+ }
+
+ return 0;
+}
+
+fs_reg
+fs_visitor::get_timestamp()
+{
+ assert(intel->gen >= 7);
+
+ fs_reg ts = fs_reg(retype(brw_vec1_reg(BRW_ARCHITECTURE_REGISTER_FILE,
+ BRW_ARF_TIMESTAMP,
+ 0),
+ BRW_REGISTER_TYPE_UD));
+
+ fs_reg dst = fs_reg(this, glsl_type::uint_type);
+
+ fs_inst *mov = emit(MOV(dst, ts));
+ /* We want to read the 3 fields we care about (mostly field 0, but also 2)
+ * even if it's not enabled in the dispatch.
+ */
+ mov->force_writemask_all = true;
+ mov->force_uncompressed = true;
+
+ /* The caller wants the low 32 bits of the timestamp. Since it's running
+ * at the GPU clock rate of ~1.2ghz, it will roll over every ~3 seconds,
+ * which is plenty of time for our purposes. It is identical across the
+ * EUs, but since it's tracking GPU core speed it will increment at a
+ * varying rate as render P-states change.
+ *
+ * The caller could also check if render P-states have changed (or anything
+ * else that might disrupt timing) by setting smear to 2 and checking if
+ * that field is != 0.
+ */
+ dst.smear = 0;
+
+ return dst;
+}
+
+void
+fs_visitor::emit_shader_time_begin()
+{
+ current_annotation = "shader time start";
+ shader_start_time = get_timestamp();
+}
+
+void
+fs_visitor::emit_shader_time_end()
+{
+ current_annotation = "shader time end";
+
+ enum shader_time_shader_type type, written_type, reset_type;
+ if (dispatch_width == 8) {
+ type = ST_FS8;
+ written_type = ST_FS8_WRITTEN;
+ reset_type = ST_FS8_RESET;
+ } else {
+ assert(dispatch_width == 16);
+ type = ST_FS16;
+ written_type = ST_FS16_WRITTEN;
+ reset_type = ST_FS16_RESET;
}
+
+ fs_reg shader_end_time = get_timestamp();
+
+ /* Check that there weren't any timestamp reset events (assuming these
+ * were the only two timestamp reads that happened).
+ */
+ fs_reg reset = shader_end_time;
+ reset.smear = 2;
+ fs_inst *test = emit(AND(reg_null_d, reset, fs_reg(1u)));
+ test->conditional_mod = BRW_CONDITIONAL_Z;
+ emit(IF(BRW_PREDICATE_NORMAL));
+
+ push_force_uncompressed();
+ fs_reg start = shader_start_time;
+ start.negate = true;
+ fs_reg diff = fs_reg(this, glsl_type::uint_type);
+ emit(ADD(diff, start, shader_end_time));
+
+ /* If there were no instructions between the two timestamp gets, the diff
+ * is 2 cycles. Remove that overhead, so I can forget about that when
+ * trying to determine the time taken for single instructions.
+ */
+ emit(ADD(diff, diff, fs_reg(-2u)));
+
+ emit_shader_time_write(type, diff);
+ emit_shader_time_write(written_type, fs_reg(1u));
+ emit(BRW_OPCODE_ELSE);
+ emit_shader_time_write(reset_type, fs_reg(1u));
+ emit(BRW_OPCODE_ENDIF);
+
+ pop_force_uncompressed();
+}
+
+void
+fs_visitor::emit_shader_time_write(enum shader_time_shader_type type,
+ fs_reg value)
+{
+ int shader_time_index =
+ brw_get_shader_time_index(brw, shader_prog, &fp->Base, type);
+ fs_reg offset = fs_reg(shader_time_index * SHADER_TIME_STRIDE);
+
+ fs_reg payload;
+ if (dispatch_width == 8)
+ payload = fs_reg(this, glsl_type::uvec2_type);
+ else
+ payload = fs_reg(this, glsl_type::uint_type);
+
+ emit(fs_inst(SHADER_OPCODE_SHADER_TIME_ADD,
+ fs_reg(), payload, offset, value));
}
void
assert(force_sechalf_stack >= 0);
}
+/**
+ * Returns true if the instruction has a flag that means it won't
+ * update an entire destination register.
+ *
+ * For example, dead code elimination and live variable analysis want to know
+ * when a write to a variable screens off any preceding values that were in
+ * it.
+ */
+bool
+fs_inst::is_partial_write()
+{
+ return (this->predicate ||
+ this->force_uncompressed ||
+ this->force_sechalf);
+}
+
/**
* Returns how many MRFs an FS opcode will write over.
*
case FS_OPCODE_TXB:
case SHADER_OPCODE_TXD:
case SHADER_OPCODE_TXF:
+ case SHADER_OPCODE_TXF_MS:
case SHADER_OPCODE_TXL:
case SHADER_OPCODE_TXS:
+ case SHADER_OPCODE_LOD:
return 1;
case FS_OPCODE_FB_WRITE:
return 2;
case FS_OPCODE_UNSPILL:
return 1;
case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD:
- return inst->header_present;
+ return inst->mlen;
case FS_OPCODE_SPILL:
return 2;
default:
import_uniforms_callback,
variable_ht);
this->params_remap = v->params_remap;
+ this->nr_params_remap = v->nr_params_remap;
}
/* Our support for uniforms is piggy-backed on the struct
* get stored, rather than in some global gl_shader_program uniform
* store.
*/
-int
-fs_visitor::setup_uniform_values(int loc, const glsl_type *type)
+void
+fs_visitor::setup_uniform_values(ir_variable *ir)
{
- unsigned int offset = 0;
-
- if (type->is_matrix()) {
- const glsl_type *column = glsl_type::get_instance(GLSL_TYPE_FLOAT,
- type->vector_elements,
- 1);
-
- for (unsigned int i = 0; i < type->matrix_columns; i++) {
- offset += setup_uniform_values(loc + offset, column);
- }
-
- return offset;
- }
-
- 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->vector_elements; i++) {
- unsigned int param = c->prog_data.nr_params++;
+ int namelen = strlen(ir->name);
- this->param_index[param] = loc;
- this->param_offset[param] = i;
+ /* The data for our (non-builtin) uniforms is stored in a series of
+ * gl_uniform_driver_storage structs for each subcomponent that
+ * glGetUniformLocation() could name. We know it's been set up in the same
+ * order we'd walk the type, so walk the list of storage and find anything
+ * with our name, or the prefix of a component that starts with our name.
+ */
+ unsigned params_before = c->prog_data.nr_params;
+ for (unsigned u = 0; u < shader_prog->NumUserUniformStorage; u++) {
+ struct gl_uniform_storage *storage = &shader_prog->UniformStorage[u];
+
+ if (strncmp(ir->name, storage->name, namelen) != 0 ||
+ (storage->name[namelen] != 0 &&
+ storage->name[namelen] != '.' &&
+ storage->name[namelen] != '[')) {
+ continue;
}
- return 1;
- case GLSL_TYPE_STRUCT:
- for (unsigned int i = 0; i < type->length; i++) {
- offset += setup_uniform_values(loc + offset,
- type->fields.structure[i].type);
- }
- return offset;
+ unsigned slots = storage->type->component_slots();
+ if (storage->array_elements)
+ slots *= storage->array_elements;
- case GLSL_TYPE_ARRAY:
- for (unsigned int i = 0; i < type->length; i++) {
- offset += setup_uniform_values(loc + offset, type->fields.array);
+ for (unsigned i = 0; i < slots; i++) {
+ c->prog_data.param[c->prog_data.nr_params++] =
+ &storage->storage[i].f;
}
- return offset;
-
- case GLSL_TYPE_SAMPLER:
- /* The sampler takes up a slot, but we don't use any values from it. */
- return 1;
-
- default:
- assert(!"not reached");
- return 0;
}
+
+ /* Make sure we actually initialized the right amount of stuff here. */
+ assert(params_before + ir->type->component_slots() ==
+ c->prog_data.nr_params);
+ (void)params_before;
}
break;
last_swiz = swiz;
- this->param_index[c->prog_data.nr_params] = index;
- this->param_offset[c->prog_data.nr_params] = swiz;
- c->prog_data.nr_params++;
+ c->prog_data.param[c->prog_data.nr_params++] =
+ &fp->Base.Parameters->ParameterValues[index][swiz].f;
}
}
}
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));
+ interp_reg(VARYING_SLOT_POS, 2));
}
wpos.reg_offset++;
bool is_centroid)
{
brw_wm_barycentric_interp_mode barycoord_mode;
- if (is_centroid) {
- if (interpolation_mode == INTERP_QUALIFIER_SMOOTH)
- barycoord_mode = BRW_WM_PERSPECTIVE_CENTROID_BARYCENTRIC;
- else
- barycoord_mode = BRW_WM_NONPERSPECTIVE_CENTROID_BARYCENTRIC;
+ if (intel->gen >= 6) {
+ if (is_centroid) {
+ if (interpolation_mode == INTERP_QUALIFIER_SMOOTH)
+ barycoord_mode = BRW_WM_PERSPECTIVE_CENTROID_BARYCENTRIC;
+ else
+ barycoord_mode = BRW_WM_NONPERSPECTIVE_CENTROID_BARYCENTRIC;
+ } else {
+ if (interpolation_mode == INTERP_QUALIFIER_SMOOTH)
+ barycoord_mode = BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC;
+ else
+ barycoord_mode = BRW_WM_NONPERSPECTIVE_PIXEL_BARYCENTRIC;
+ }
} else {
- if (interpolation_mode == INTERP_QUALIFIER_SMOOTH)
- barycoord_mode = BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC;
- else
- barycoord_mode = BRW_WM_NONPERSPECTIVE_PIXEL_BARYCENTRIC;
+ /* On Ironlake and below, there is only one interpolation mode.
+ * Centroid interpolation doesn't mean anything on this hardware --
+ * there is no multisampling.
+ */
+ barycoord_mode = BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC;
}
return emit(FS_OPCODE_LINTERP, attr,
this->delta_x[barycoord_mode],
* 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);
- emit_linterp(attr, fs_reg(interp), interpolation_mode,
- ir->centroid);
- if (brw->needs_unlit_centroid_workaround && ir->centroid) {
- /* Get the pixel/sample mask into f0 so that we know
- * which pixels are lit. Then, for each channel that is
- * unlit, replace the centroid data with non-centroid
- * data.
- */
- emit(FS_OPCODE_MOV_DISPATCH_TO_FLAGS, attr);
- fs_inst *inst = emit_linterp(attr, fs_reg(interp),
- interpolation_mode, false);
- inst->predicate = BRW_PREDICATE_NORMAL;
- inst->predicate_inverse = true;
- }
- if (intel->gen < 6) {
- emit(BRW_OPCODE_MUL, attr, attr, this->pixel_w);
- }
- }
+ struct brw_reg interp = interp_reg(location, k);
+ emit_linterp(attr, fs_reg(interp), interpolation_mode,
+ ir->centroid);
+ if (brw->needs_unlit_centroid_workaround && ir->centroid) {
+ /* Get the pixel/sample mask into f0 so that we know
+ * which pixels are lit. Then, for each channel that is
+ * unlit, replace the centroid data with non-centroid
+ * data.
+ */
+ emit(FS_OPCODE_MOV_DISPATCH_TO_FLAGS);
+ fs_inst *inst = emit_linterp(attr, fs_reg(interp),
+ interpolation_mode, false);
+ inst->predicate = BRW_PREDICATE_NORMAL;
+ inst->predicate_inverse = true;
+ }
+ if (intel->gen < 6) {
+ emit(BRW_OPCODE_MUL, attr, attr, this->pixel_w);
+ }
attr.reg_offset++;
}
return reg;
}
+fs_reg
+fs_visitor::fix_math_operand(fs_reg src)
+{
+ /* Can't do hstride == 0 args on 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.file != IMM &&
+ !src.abs && !src.negate)
+ return src;
+
+ /* Gen7 relaxes most of the above restrictions, but still can't use IMM
+ * operands to math
+ */
+ if (intel->gen >= 7 && src.file != IMM)
+ return src;
+
+ fs_reg expanded = fs_reg(this, glsl_type::float_type);
+ expanded.type = src.type;
+ emit(BRW_OPCODE_MOV, expanded, src);
+ return expanded;
+}
+
fs_inst *
fs_visitor::emit_math(enum opcode opcode, fs_reg dst, fs_reg src)
{
* 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;
- }
+ if (intel->gen >= 6)
+ src = fix_math_operand(src);
fs_inst *inst = emit(opcode, dst, src);
fs_inst *inst;
switch (opcode) {
- case SHADER_OPCODE_POW:
case SHADER_OPCODE_INT_QUOTIENT:
case SHADER_OPCODE_INT_REMAINDER:
+ if (intel->gen >= 7 && dispatch_width == 16)
+ fail("16-wide INTDIV unsupported\n");
+ break;
+ case SHADER_OPCODE_POW:
break;
default:
assert(!"not reached: unsupported binary math opcode.");
return NULL;
}
- 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 (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;
- }
+ if (intel->gen >= 6) {
+ src0 = fix_math_operand(src0);
+ src1 = fix_math_operand(src1);
inst = emit(opcode, dst, src0, src1);
} else {
return inst;
}
-/**
- * To be called after the last _mesa_add_state_reference() call, to
- * set up prog_data.param[] for assign_curb_setup() and
- * setup_pull_constants().
- */
-void
-fs_visitor::setup_paramvalues_refs()
-{
- if (dispatch_width != 8)
- return;
-
- /* 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] =
- (const float *)fp->Base.Parameters->ParameterValues[this->param_index[i]] +
- this->param_offset[i];
- }
-}
-
void
fs_visitor::assign_curb_setup()
{
constant_nr / 8,
constant_nr % 8);
- inst->src[i].file = FIXED_HW_REG;
+ inst->src[i].file = HW_REG;
inst->src[i].fixed_hw_reg = retype(brw_reg, inst->src[i].type);
}
}
void
fs_visitor::calculate_urb_setup()
{
- for (unsigned int i = 0; i < FRAG_ATTRIB_MAX; i++) {
+ for (unsigned int i = 0; i < VARYING_SLOT_MAX; i++) {
urb_setup[i] = -1;
}
int urb_next = 0;
/* Figure out where each of the incoming setup attributes lands. */
if (intel->gen >= 6) {
- for (unsigned int i = 0; i < FRAG_ATTRIB_MAX; i++) {
+ for (unsigned int i = 0; i < VARYING_SLOT_MAX; i++) {
if (fp->Base.InputsRead & BITFIELD64_BIT(i)) {
urb_setup[i] = urb_next++;
}
}
} else {
/* FINISHME: The sf doesn't map VS->FS inputs for us very well. */
- for (unsigned int i = 0; i < VERT_RESULT_MAX; i++) {
+ for (unsigned int i = 0; i < VARYING_SLOT_MAX; i++) {
/* Point size is packed into the header, not as a general attribute */
- if (i == VERT_RESULT_PSIZ)
+ if (i == VARYING_SLOT_PSIZ)
continue;
- if (c->key.vp_outputs_written & BITFIELD64_BIT(i)) {
- int fp_index = _mesa_vert_result_to_frag_attrib((gl_vert_result) i);
-
+ if (c->key.input_slots_valid & BITFIELD64_BIT(i)) {
/* The back color slot is skipped when the front color is
* also written to. In addition, some slots can be
* written in the vertex shader and not read in the
* fragment shader. So the register number must always be
* incremented, mapped or not.
*/
- if (fp_index >= 0)
- urb_setup[fp_index] = urb_next;
+ if (_mesa_varying_slot_in_fs((gl_varying_slot) i))
+ urb_setup[i] = urb_next;
urb_next++;
}
}
*
* See compile_sf_prog() for more info.
*/
- if (fp->Base.InputsRead & BITFIELD64_BIT(FRAG_ATTRIB_PNTC))
- urb_setup[FRAG_ATTRIB_PNTC] = urb_next++;
+ if (fp->Base.InputsRead & BITFIELD64_BIT(VARYING_SLOT_PNTC))
+ urb_setup[VARYING_SLOT_PNTC] = urb_next++;
}
/* Each attribute is 4 setup channels, each of which is half a reg. */
fs_inst *inst = (fs_inst *)node;
if (inst->opcode == FS_OPCODE_LINTERP) {
- assert(inst->src[2].file == FIXED_HW_REG);
+ assert(inst->src[2].file == HW_REG);
inst->src[2].fixed_hw_reg.nr += urb_start;
}
if (inst->opcode == FS_OPCODE_CINTERP) {
- assert(inst->src[0].file == FIXED_HW_REG);
+ assert(inst->src[0].file == HW_REG);
inst->src[0].fixed_hw_reg.nr += urb_start;
}
}
/* If there's a SEND message that requires contiguous destination
* registers, no splitting is allowed.
*/
- if (inst->regs_written() > 1) {
+ if (inst->regs_written > 1) {
split_grf[inst->dst.reg] = false;
}
+
+ /* If we're sending from a GRF, don't split it, on the assumption that
+ * the send is reading the whole thing.
+ */
+ if (inst->is_send_from_grf()) {
+ split_grf[inst->src[0].reg] = false;
+ }
}
/* Allocate new space for split regs. Note that the virtual
remap_table[i] = new_index;
virtual_grf_sizes[new_index] = virtual_grf_sizes[i];
if (live_intervals_valid) {
- virtual_grf_use[new_index] = virtual_grf_use[i];
- virtual_grf_def[new_index] = virtual_grf_def[i];
+ virtual_grf_start[new_index] = virtual_grf_start[i];
+ virtual_grf_end[new_index] = virtual_grf_end[i];
}
++new_index;
}
{
if (dispatch_width == 8) {
this->params_remap = ralloc_array(mem_ctx, int, c->prog_data.nr_params);
+ this->nr_params_remap = c->prog_data.nr_params;
for (unsigned int i = 0; i < c->prog_data.nr_params; i++)
this->params_remap[i] = -1;
if (inst->src[i].file != UNIFORM)
continue;
- assert(constant_nr < (int)c->prog_data.nr_params);
+ /* Section 5.11 of the OpenGL 4.3 spec says:
+ *
+ * "Out-of-bounds reads return undefined values, which include
+ * values from other variables of the active program or zero."
+ */
+ if (constant_nr < 0 || constant_nr >= (int)c->prog_data.nr_params) {
+ constant_nr = 0;
+ }
/* For now, set this to non-negative. We'll give it the
* actual new number in a moment, in order to keep the
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];
}
if (inst->src[i].file != UNIFORM)
continue;
+ /* as above alias to 0 */
+ if (constant_nr < 0 || constant_nr >= (int)this->nr_params_remap) {
+ constant_nr = 0;
+ }
assert(this->params_remap[constant_nr] != -1);
inst->src[i].reg = this->params_remap[constant_nr];
inst->src[i].reg_offset = 0;
return true;
}
+/*
+ * Implements array access of uniforms by inserting a
+ * PULL_CONSTANT_LOAD instruction.
+ *
+ * Unlike temporary GRF array access (where we don't support it due to
+ * the difficulty of doing relative addressing on instruction
+ * destinations), we could potentially do array access of uniforms
+ * that were loaded in GRF space as push constants. In real-world
+ * usage we've seen, though, the arrays being used are always larger
+ * than we could load as push constants, so just always move all
+ * uniform array access out to a pull constant buffer.
+ */
+void
+fs_visitor::move_uniform_array_access_to_pull_constants()
+{
+ int pull_constant_loc[c->prog_data.nr_params];
+
+ for (unsigned int i = 0; i < c->prog_data.nr_params; i++) {
+ pull_constant_loc[i] = -1;
+ }
+
+ /* Walk through and find array access of uniforms. Put a copy of that
+ * uniform in the pull constant buffer.
+ *
+ * Note that we don't move constant-indexed accesses to arrays. No
+ * testing has been done of the performance impact of this choice.
+ */
+ foreach_list_safe(node, &this->instructions) {
+ fs_inst *inst = (fs_inst *)node;
+
+ for (int i = 0 ; i < 3; i++) {
+ if (inst->src[i].file != UNIFORM || !inst->src[i].reladdr)
+ continue;
+
+ int uniform = inst->src[i].reg;
+
+ /* If this array isn't already present in the pull constant buffer,
+ * add it.
+ */
+ if (pull_constant_loc[uniform] == -1) {
+ const float **values = &c->prog_data.param[uniform];
+
+ pull_constant_loc[uniform] = c->prog_data.nr_pull_params;
+
+ assert(param_size[uniform]);
+
+ for (int j = 0; j < param_size[uniform]; j++) {
+ c->prog_data.pull_param[c->prog_data.nr_pull_params++] =
+ values[j];
+ }
+ }
+
+ /* Set up the annotation tracking for new generated instructions. */
+ base_ir = inst->ir;
+ current_annotation = inst->annotation;
+
+ fs_reg surf_index = fs_reg((unsigned)SURF_INDEX_FRAG_CONST_BUFFER);
+ fs_reg temp = fs_reg(this, glsl_type::float_type);
+ exec_list list = VARYING_PULL_CONSTANT_LOAD(temp,
+ surf_index,
+ *inst->src[i].reladdr,
+ pull_constant_loc[uniform] +
+ inst->src[i].reg_offset);
+ inst->insert_before(&list);
+
+ inst->src[i].file = temp.file;
+ inst->src[i].reg = temp.reg;
+ inst->src[i].reg_offset = temp.reg_offset;
+ inst->src[i].reladdr = NULL;
+ }
+ }
+}
+
/**
* Choose accesses from the UNIFORM file to demote to using the pull
* constant buffer.
/* Just demote the end of the list. We could probably do better
* here, demoting things that are rarely used in the program first.
*/
- int pull_uniform_base = max_uniform_components;
- int pull_uniform_count = c->prog_data.nr_params - pull_uniform_base;
+ unsigned int pull_uniform_base = max_uniform_components;
+
+ int pull_constant_loc[c->prog_data.nr_params];
+ for (unsigned int i = 0; i < c->prog_data.nr_params; i++) {
+ if (i < pull_uniform_base) {
+ pull_constant_loc[i] = -1;
+ } else {
+ pull_constant_loc[i] = -1;
+ /* If our constant is already being uploaded for reladdr purposes,
+ * reuse it.
+ */
+ for (unsigned int j = 0; j < c->prog_data.nr_pull_params; j++) {
+ if (c->prog_data.pull_param[j] == c->prog_data.param[i]) {
+ pull_constant_loc[i] = j;
+ break;
+ }
+ }
+ if (pull_constant_loc[i] == -1) {
+ int pull_index = c->prog_data.nr_pull_params++;
+ c->prog_data.pull_param[pull_index] = c->prog_data.param[i];
+ pull_constant_loc[i] = pull_index;;
+ }
+ }
+ }
+ c->prog_data.nr_params = pull_uniform_base;
foreach_list(node, &this->instructions) {
fs_inst *inst = (fs_inst *)node;
if (inst->src[i].file != UNIFORM)
continue;
- int uniform_nr = inst->src[i].reg + inst->src[i].reg_offset;
- if (uniform_nr < pull_uniform_base)
+ int pull_index = pull_constant_loc[inst->src[i].reg +
+ inst->src[i].reg_offset];
+ if (pull_index == -1)
continue;
+ assert(!inst->src[i].reladdr);
+
fs_reg dst = fs_reg(this, glsl_type::float_type);
fs_reg index = fs_reg((unsigned)SURF_INDEX_FRAG_CONST_BUFFER);
- fs_reg offset = fs_reg((unsigned)(((uniform_nr -
- pull_uniform_base) * 4) & ~15));
+ fs_reg offset = fs_reg((unsigned)(pull_index * 4) & ~15);
fs_inst *pull =
new(mem_ctx) fs_inst(FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD,
dst, index, offset);
pull->ir = inst->ir;
pull->annotation = inst->annotation;
- pull->base_mrf = 14;
- pull->mlen = 1;
inst->insert_before(pull);
inst->src[i].file = GRF;
inst->src[i].reg = dst.reg;
inst->src[i].reg_offset = 0;
- inst->src[i].smear = (uniform_nr - pull_uniform_base) & 3;
+ inst->src[i].smear = pull_index & 3;
}
}
-
- for (int i = 0; i < pull_uniform_count; i++) {
- c->prog_data.pull_param[i] = c->prog_data.param[pull_uniform_base + i];
- }
- c->prog_data.nr_params -= pull_uniform_count;
- c->prog_data.nr_pull_params = pull_uniform_count;
}
bool
}
/**
- * Must be called after calculate_live_intervales() to remove unused
- * writes to registers -- register allocation will fail otherwise
- * because something deffed but not used won't be considered to
- * interfere with other regs.
+ * Removes any instructions writing a VGRF where that VGRF is not used by any
+ * later instruction.
*/
bool
fs_visitor::dead_code_eliminate()
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();
- progress = true;
+ if (inst->dst.file == GRF) {
+ assert(this->virtual_grf_end[inst->dst.reg] >= pc);
+ if (this->virtual_grf_end[inst->dst.reg] == pc) {
+ inst->remove();
+ progress = true;
+ }
}
pc++;
return progress;
}
+struct dead_code_hash_key
+{
+ int vgrf;
+ int reg_offset;
+};
+
+static bool
+dead_code_hash_compare(const void *a, const void *b)
+{
+ return memcmp(a, b, sizeof(struct dead_code_hash_key)) == 0;
+}
+
+static void
+clear_dead_code_hash(struct hash_table *ht)
+{
+ struct hash_entry *entry;
+
+ hash_table_foreach(ht, entry) {
+ _mesa_hash_table_remove(ht, entry);
+ }
+}
+
+static void
+insert_dead_code_hash(struct hash_table *ht,
+ int vgrf, int reg_offset, fs_inst *inst)
+{
+ /* We don't bother freeing keys, because they'll be GCed with the ht. */
+ struct dead_code_hash_key *key = ralloc(ht, struct dead_code_hash_key);
+
+ key->vgrf = vgrf;
+ key->reg_offset = reg_offset;
+
+ _mesa_hash_table_insert(ht, _mesa_hash_data(key, sizeof(*key)), key, inst);
+}
+
+static struct hash_entry *
+get_dead_code_hash_entry(struct hash_table *ht, int vgrf, int reg_offset)
+{
+ struct dead_code_hash_key key;
+
+ key.vgrf = vgrf;
+ key.reg_offset = reg_offset;
+
+ return _mesa_hash_table_search(ht, _mesa_hash_data(&key, sizeof(key)), &key);
+}
+
+static void
+remove_dead_code_hash(struct hash_table *ht,
+ int vgrf, int reg_offset)
+{
+ struct hash_entry *entry = get_dead_code_hash_entry(ht, vgrf, reg_offset);
+ if (!entry)
+ return;
+
+ _mesa_hash_table_remove(ht, entry);
+}
+
+/**
+ * Walks basic blocks, removing any regs that are written but not read before
+ * being redefined.
+ *
+ * The dead_code_eliminate() function implements a global dead code
+ * elimination, but it only handles the removing the last write to a register
+ * if it's never read. This one can handle intermediate writes, but only
+ * within a basic block.
+ */
+bool
+fs_visitor::dead_code_eliminate_local()
+{
+ struct hash_table *ht;
+ bool progress = false;
+
+ ht = _mesa_hash_table_create(mem_ctx, dead_code_hash_compare);
+
+ foreach_list_safe(node, &this->instructions) {
+ fs_inst *inst = (fs_inst *)node;
+
+ /* At a basic block, empty the HT since we don't understand dataflow
+ * here.
+ */
+ if (inst->is_control_flow()) {
+ clear_dead_code_hash(ht);
+ continue;
+ }
+
+ /* Clear the HT of any instructions that got read. */
+ for (int i = 0; i < 3; i++) {
+ fs_reg src = inst->src[i];
+ if (src.file != GRF)
+ continue;
+
+ int read = 1;
+ if (inst->is_send_from_grf())
+ read = virtual_grf_sizes[src.reg] - src.reg_offset;
+
+ for (int reg_offset = src.reg_offset;
+ reg_offset < src.reg_offset + read;
+ reg_offset++) {
+ remove_dead_code_hash(ht, src.reg, reg_offset);
+ }
+ }
+
+ /* Add any update of a GRF to the HT, removing a previous write if it
+ * wasn't read.
+ */
+ if (inst->dst.file == GRF) {
+ if (inst->regs_written > 1) {
+ /* We don't know how to trim channels from an instruction's
+ * writes, so we can't incrementally remove unread channels from
+ * it. Just remove whatever it overwrites from the table
+ */
+ for (int i = 0; i < inst->regs_written; i++) {
+ remove_dead_code_hash(ht,
+ inst->dst.reg,
+ inst->dst.reg_offset + i);
+ }
+ } else {
+ struct hash_entry *entry =
+ get_dead_code_hash_entry(ht, inst->dst.reg,
+ inst->dst.reg_offset);
+
+ if (inst->is_partial_write()) {
+ /* For a partial write, we can't remove any previous dead code
+ * candidate, since we're just modifying their result, but we can
+ * be dead code eliminiated ourselves.
+ */
+ if (entry) {
+ entry->data = inst;
+ } else {
+ insert_dead_code_hash(ht, inst->dst.reg, inst->dst.reg_offset,
+ inst);
+ }
+ } else {
+ if (entry) {
+ /* We're completely updating a channel, and there was a
+ * previous write to the channel that wasn't read. Kill it!
+ */
+ fs_inst *inst = (fs_inst *)entry->data;
+ inst->remove();
+ progress = true;
+ _mesa_hash_table_remove(ht, entry);
+ }
+
+ insert_dead_code_hash(ht, inst->dst.reg, inst->dst.reg_offset,
+ inst);
+ }
+ }
+ }
+ }
+
+ _mesa_hash_table_destroy(ht, NULL);
+
+ if (progress)
+ live_intervals_valid = false;
+
+ return progress;
+}
+
/**
* Implements a second type of register coalescing: This one checks if
* the two regs involved in a raw move don't interfere, in which case
fs_inst *inst = (fs_inst *)node;
if (inst->opcode != BRW_OPCODE_MOV ||
- inst->predicate ||
+ inst->is_partial_write() ||
inst->saturate ||
inst->src[0].file != GRF ||
inst->src[0].negate ||
int reg_to = inst->dst.reg;
int reg_to_offset = inst->dst.reg_offset;
- foreach_list_safe(node, &this->instructions) {
+ foreach_list(node, &this->instructions) {
fs_inst *scan_inst = (fs_inst *)node;
if (scan_inst->dst.file == GRF &&
}
inst->remove();
+
+ /* We don't need to recalculate live intervals inside the loop despite
+ * flagging live_intervals_valid because we only use live intervals for
+ * the interferes test, and we must have had a situation where the
+ * intervals were:
+ *
+ * from to
+ * ^
+ * |
+ * v
+ * ^
+ * |
+ * v
+ *
+ * Some register R that might get coalesced with one of these two could
+ * only be referencing "to", otherwise "from"'s range would have been
+ * longer. R's range could also only start at the end of "to" or later,
+ * otherwise it will conflict with "to" when we try to coalesce "to"
+ * into Rw anyway.
+ */
live_intervals_valid = false;
+
progress = true;
continue;
}
continue;
if (inst->opcode != BRW_OPCODE_MOV ||
- inst->predicate ||
+ inst->is_partial_write() ||
inst->saturate ||
inst->dst.file != GRF || (inst->src[0].file != GRF &&
inst->src[0].file != UNIFORM)||
inst->dst.type != inst->src[0].type)
continue;
- bool has_source_modifiers = inst->src[0].abs || inst->src[0].negate;
+ bool has_source_modifiers = (inst->src[0].abs ||
+ inst->src[0].negate ||
+ inst->src[0].smear != -1 ||
+ inst->src[0].file == UNIFORM);
/* Found a move of a GRF to a GRF. Let's see if we can coalesce
* them: check for no writes to either one until the exit of the
* 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;
+ if (has_source_modifiers && !can_do_source_mods(scan_inst)) {
+ interfered = true;
break;
}
next_ip++;
if (inst->opcode != BRW_OPCODE_MOV ||
- inst->predicate ||
+ inst->is_partial_write() ||
inst->dst.file != MRF || inst->src[0].file != GRF ||
inst->dst.type != inst->src[0].type ||
inst->src[0].abs || inst->src[0].negate || inst->src[0].smear != -1)
/* Can't compute-to-MRF this GRF if someone else was going to
* read it later.
*/
- if (this->virtual_grf_use[inst->src[0].reg] > ip)
+ if (this->virtual_grf_end[inst->src[0].reg] > ip)
continue;
/* Found a move of a GRF to a MRF. Let's see if we can go
* into a compute-to-MRF.
*/
- /* SENDs can only write to GRFs, so no compute-to-MRF. */
- if (scan_inst->mlen) {
- break;
- }
-
- /* If it's predicated, it (probably) didn't populate all
- * the channels. We might be able to rewrite everything
+ /* If this one instruction didn't populate all the
+ * channels, bail. We might be able to rewrite everything
* that writes that reg, but it would require smarter
* tracking to delay the rewriting until complete success.
*/
- if (scan_inst->predicate)
+ if (scan_inst->is_partial_write())
break;
- /* If it's half of register setup and not the same half as
- * our MOV we're trying to remove, bail for now.
- */
- if (scan_inst->force_uncompressed != inst->force_uncompressed ||
- scan_inst->force_sechalf != inst->force_sechalf) {
- break;
- }
+ /* Things returning more than one register would need us to
+ * understand coalescing out more than one MOV at a time.
+ */
+ if (scan_inst->regs_written > 1)
+ break;
/* SEND instructions can't have MRF as a destination. */
if (scan_inst->mlen)
break;
- if (intel->gen >= 6) {
+ if (intel->gen == 6) {
/* gen6 math instructions must have the destination be
* GRF, so no compute-to-MRF for them.
*/
break;
}
- /* We don't handle flow control here. Most computation of
+ /* We don't handle control flow here. Most computation of
* values that end up in MRFs are shortly before the MRF
* write anyway.
*/
- if (scan_inst->opcode == BRW_OPCODE_DO ||
- scan_inst->opcode == BRW_OPCODE_WHILE ||
- scan_inst->opcode == BRW_OPCODE_ELSE ||
- scan_inst->opcode == BRW_OPCODE_ENDIF) {
+ if (scan_inst->is_control_flow() && scan_inst->opcode != BRW_OPCODE_IF)
break;
- }
/* You can't read from an MRF, so if someone else reads our
* MRF's source GRF that we wanted to rewrite, that stops us.
foreach_list_safe(node, &this->instructions) {
fs_inst *inst = (fs_inst *)node;
- switch (inst->opcode) {
- case BRW_OPCODE_DO:
- case BRW_OPCODE_WHILE:
- case BRW_OPCODE_IF:
- case BRW_OPCODE_ELSE:
- case BRW_OPCODE_ENDIF:
+ if (inst->is_control_flow()) {
memset(last_mrf_move, 0, sizeof(last_mrf_move));
- continue;
- default:
- break;
}
if (inst->opcode == BRW_OPCODE_MOV &&
if (inst->opcode == BRW_OPCODE_MOV &&
inst->dst.file == MRF &&
inst->src[0].file == GRF &&
- !inst->predicate) {
+ !inst->is_partial_write()) {
last_mrf_move[inst->dst.reg] = inst;
}
}
return progress;
}
+static void
+clear_deps_for_inst_src(fs_inst *inst, int dispatch_width, bool *deps,
+ int first_grf, int grf_len)
+{
+ bool inst_16wide = (dispatch_width > 8 &&
+ !inst->force_uncompressed &&
+ !inst->force_sechalf);
+
+ /* Clear the flag for registers that actually got read (as expected). */
+ for (int i = 0; i < 3; i++) {
+ int grf;
+ if (inst->src[i].file == GRF) {
+ grf = inst->src[i].reg;
+ } else if (inst->src[i].file == HW_REG &&
+ inst->src[i].fixed_hw_reg.file == BRW_GENERAL_REGISTER_FILE) {
+ grf = inst->src[i].fixed_hw_reg.nr;
+ } else {
+ continue;
+ }
+
+ if (grf >= first_grf &&
+ grf < first_grf + grf_len) {
+ deps[grf - first_grf] = false;
+ if (inst_16wide)
+ deps[grf - first_grf + 1] = false;
+ }
+ }
+}
+
+/**
+ * Implements this workaround for the original 965:
+ *
+ * "[DevBW, DevCL] Implementation Restrictions: As the hardware does not
+ * check for post destination dependencies on this instruction, software
+ * must ensure that there is no destination hazard for the case of ‘write
+ * followed by a posted write’ shown in the following example.
+ *
+ * 1. mov r3 0
+ * 2. send r3.xy <rest of send instruction>
+ * 3. mov r2 r3
+ *
+ * Due to no post-destination dependency check on the ‘send’, the above
+ * code sequence could have two instructions (1 and 2) in flight at the
+ * same time that both consider ‘r3’ as the target of their final writes.
+ */
void
-fs_visitor::dump_instruction(fs_inst *inst)
+fs_visitor::insert_gen4_pre_send_dependency_workarounds(fs_inst *inst)
{
- if (inst->opcode < ARRAY_SIZE(opcode_descs) &&
- opcode_descs[inst->opcode].name) {
- printf("%s", opcode_descs[inst->opcode].name);
- } else {
- printf("op%d", inst->opcode);
+ int reg_size = dispatch_width / 8;
+ int write_len = inst->regs_written * reg_size;
+ int first_write_grf = inst->dst.reg;
+ bool needs_dep[BRW_MAX_MRF];
+ assert(write_len < (int)sizeof(needs_dep) - 1);
+
+ memset(needs_dep, false, sizeof(needs_dep));
+ memset(needs_dep, true, write_len);
+
+ clear_deps_for_inst_src(inst, dispatch_width,
+ needs_dep, first_write_grf, write_len);
+
+ /* Walk backwards looking for writes to registers we're writing which
+ * aren't read since being written. If we hit the start of the program,
+ * we assume that there are no outstanding dependencies on entry to the
+ * program.
+ */
+ for (fs_inst *scan_inst = (fs_inst *)inst->prev;
+ scan_inst != NULL;
+ scan_inst = (fs_inst *)scan_inst->prev) {
+
+ /* If we hit control flow, assume that there *are* outstanding
+ * dependencies, and force their cleanup before our instruction.
+ */
+ if (scan_inst->is_control_flow()) {
+ for (int i = 0; i < write_len; i++) {
+ if (needs_dep[i]) {
+ inst->insert_before(DEP_RESOLVE_MOV(first_write_grf + i));
+ }
+ }
+ return;
+ }
+
+ bool scan_inst_16wide = (dispatch_width > 8 &&
+ !scan_inst->force_uncompressed &&
+ !scan_inst->force_sechalf);
+
+ /* We insert our reads as late as possible on the assumption that any
+ * instruction but a MOV that might have left us an outstanding
+ * dependency has more latency than a MOV.
+ */
+ if (scan_inst->dst.file == GRF) {
+ for (int i = 0; i < scan_inst->regs_written; i++) {
+ int reg = scan_inst->dst.reg + i * reg_size;
+
+ if (reg >= first_write_grf &&
+ reg < first_write_grf + write_len &&
+ needs_dep[reg - first_write_grf]) {
+ inst->insert_before(DEP_RESOLVE_MOV(reg));
+ needs_dep[reg - first_write_grf] = false;
+ if (scan_inst_16wide)
+ needs_dep[reg - first_write_grf + 1] = false;
+ }
+ }
+ }
+
+ /* Clear the flag for registers that actually got read (as expected). */
+ clear_deps_for_inst_src(scan_inst, dispatch_width,
+ needs_dep, first_write_grf, write_len);
+
+ /* Continue the loop only if we haven't resolved all the dependencies */
+ int i;
+ for (i = 0; i < write_len; i++) {
+ if (needs_dep[i])
+ break;
+ }
+ if (i == write_len)
+ return;
+ }
+}
+
+/**
+ * Implements this workaround for the original 965:
+ *
+ * "[DevBW, DevCL] Errata: A destination register from a send can not be
+ * used as a destination register until after it has been sourced by an
+ * instruction with a different destination register.
+ */
+void
+fs_visitor::insert_gen4_post_send_dependency_workarounds(fs_inst *inst)
+{
+ int write_len = inst->regs_written * dispatch_width / 8;
+ int first_write_grf = inst->dst.reg;
+ bool needs_dep[BRW_MAX_MRF];
+ assert(write_len < (int)sizeof(needs_dep) - 1);
+
+ memset(needs_dep, false, sizeof(needs_dep));
+ memset(needs_dep, true, write_len);
+ /* Walk forwards looking for writes to registers we're writing which aren't
+ * read before being written.
+ */
+ for (fs_inst *scan_inst = (fs_inst *)inst->next;
+ !scan_inst->is_tail_sentinel();
+ scan_inst = (fs_inst *)scan_inst->next) {
+ /* If we hit control flow, force resolve all remaining dependencies. */
+ if (scan_inst->is_control_flow()) {
+ for (int i = 0; i < write_len; i++) {
+ if (needs_dep[i])
+ scan_inst->insert_before(DEP_RESOLVE_MOV(first_write_grf + i));
+ }
+ return;
+ }
+
+ /* Clear the flag for registers that actually got read (as expected). */
+ clear_deps_for_inst_src(scan_inst, dispatch_width,
+ needs_dep, first_write_grf, write_len);
+
+ /* We insert our reads as late as possible since they're reading the
+ * result of a SEND, which has massive latency.
+ */
+ if (scan_inst->dst.file == GRF &&
+ scan_inst->dst.reg >= first_write_grf &&
+ scan_inst->dst.reg < first_write_grf + write_len &&
+ needs_dep[scan_inst->dst.reg - first_write_grf]) {
+ scan_inst->insert_before(DEP_RESOLVE_MOV(scan_inst->dst.reg));
+ needs_dep[scan_inst->dst.reg - first_write_grf] = false;
+ }
+
+ /* Continue the loop only if we haven't resolved all the dependencies */
+ int i;
+ for (i = 0; i < write_len; i++) {
+ if (needs_dep[i])
+ break;
+ }
+ if (i == write_len)
+ return;
+ }
+
+ /* If we hit the end of the program, resolve all remaining dependencies out
+ * of paranoia.
+ */
+ fs_inst *last_inst = (fs_inst *)this->instructions.get_tail();
+ assert(last_inst->eot);
+ for (int i = 0; i < write_len; i++) {
+ if (needs_dep[i])
+ last_inst->insert_before(DEP_RESOLVE_MOV(first_write_grf + i));
+ }
+}
+
+void
+fs_visitor::insert_gen4_send_dependency_workarounds()
+{
+ if (intel->gen != 4 || intel->is_g4x)
+ return;
+
+ /* Note that we're done with register allocation, so GRF fs_regs always
+ * have a .reg_offset of 0.
+ */
+
+ foreach_list_safe(node, &this->instructions) {
+ fs_inst *inst = (fs_inst *)node;
+
+ if (inst->mlen != 0 && inst->dst.file == GRF) {
+ insert_gen4_pre_send_dependency_workarounds(inst);
+ insert_gen4_post_send_dependency_workarounds(inst);
+ }
+ }
+}
+
+/**
+ * Turns the generic expression-style uniform pull constant load instruction
+ * into a hardware-specific series of instructions for loading a pull
+ * constant.
+ *
+ * The expression style allows the CSE pass before this to optimize out
+ * repeated loads from the same offset, and gives the pre-register-allocation
+ * scheduling full flexibility, while the conversion to native instructions
+ * allows the post-register-allocation scheduler the best information
+ * possible.
+ *
+ * Note that execution masking for setting up pull constant loads is special:
+ * the channels that need to be written are unrelated to the current execution
+ * mask, since a later instruction will use one of the result channels as a
+ * source operand for all 8 or 16 of its channels.
+ */
+void
+fs_visitor::lower_uniform_pull_constant_loads()
+{
+ foreach_list(node, &this->instructions) {
+ fs_inst *inst = (fs_inst *)node;
+
+ if (inst->opcode != FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD)
+ continue;
+
+ if (intel->gen >= 7) {
+ /* The offset arg before was a vec4-aligned byte offset. We need to
+ * turn it into a dword offset.
+ */
+ fs_reg const_offset_reg = inst->src[1];
+ assert(const_offset_reg.file == IMM &&
+ const_offset_reg.type == BRW_REGISTER_TYPE_UD);
+ const_offset_reg.imm.u /= 4;
+ fs_reg payload = fs_reg(this, glsl_type::uint_type);
+
+ /* This is actually going to be a MOV, but since only the first dword
+ * is accessed, we have a special opcode to do just that one. Note
+ * that this needs to be an operation that will be considered a def
+ * by live variable analysis, or register allocation will explode.
+ */
+ fs_inst *setup = new(mem_ctx) fs_inst(FS_OPCODE_SET_SIMD4X2_OFFSET,
+ payload, const_offset_reg);
+ setup->force_writemask_all = true;
+
+ setup->ir = inst->ir;
+ setup->annotation = inst->annotation;
+ inst->insert_before(setup);
+
+ /* Similarly, this will only populate the first 4 channels of the
+ * result register (since we only use smear values from 0-3), but we
+ * don't tell the optimizer.
+ */
+ inst->opcode = FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD_GEN7;
+ inst->src[1] = payload;
+
+ this->live_intervals_valid = false;
+ } else {
+ /* Before register allocation, we didn't tell the scheduler about the
+ * MRF we use. We know it's safe to use this MRF because nothing
+ * else does except for register spill/unspill, which generates and
+ * uses its MRF within a single IR instruction.
+ */
+ inst->base_mrf = 14;
+ inst->mlen = 1;
+ }
+ }
+}
+
+void
+fs_visitor::dump_instruction(backend_instruction *be_inst)
+{
+ fs_inst *inst = (fs_inst *)be_inst;
+
+ if (inst->predicate) {
+ printf("(%cf0.%d) ",
+ inst->predicate_inverse ? '-' : '+',
+ inst->flag_subreg);
}
+
+ printf("%s", brw_instruction_name(inst->opcode));
if (inst->saturate)
printf(".sat");
+ if (inst->conditional_mod) {
+ printf(".cmod");
+ if (!inst->predicate &&
+ (intel->gen < 5 || (inst->opcode != BRW_OPCODE_SEL &&
+ inst->opcode != BRW_OPCODE_IF &&
+ inst->opcode != BRW_OPCODE_WHILE))) {
+ printf(".f0.%d\n", inst->flag_subreg);
+ }
+ }
printf(" ");
+
switch (inst->dst.file) {
case GRF:
printf("vgrf%d", inst->dst.reg);
case BAD_FILE:
printf("(null)");
break;
+ case IMM:
+ switch (inst->src[i].type) {
+ case BRW_REGISTER_TYPE_F:
+ printf("%ff", inst->src[i].imm.f);
+ break;
+ case BRW_REGISTER_TYPE_D:
+ printf("%dd", inst->src[i].imm.i);
+ break;
+ case BRW_REGISTER_TYPE_UD:
+ printf("%uu", inst->src[i].imm.u);
+ break;
+ default:
+ printf("???");
+ break;
+ }
+ break;
default:
printf("???");
break;
printf("\n");
}
-void
-fs_visitor::dump_instructions()
-{
- int ip = 0;
- foreach_list(node, &this->instructions) {
- fs_inst *inst = (fs_inst *)node;
- printf("%d: ", ip++);
- dump_instruction(inst);
- }
-}
-
/**
* Possibly returns an instruction that set up @param reg.
*
fs_reg reg)
{
if (end == start ||
- end->predicate ||
- end->force_uncompressed ||
- end->force_sechalf ||
+ end->is_partial_write() ||
+ reg.reladdr ||
!reg.equals(end->dst)) {
return NULL;
} else {
void
fs_visitor::setup_payload_gen6()
{
- struct intel_context *intel = &brw->intel;
bool uses_depth =
- (fp->Base.InputsRead & (1 << FRAG_ATTRIB_WPOS)) != 0;
+ (fp->Base.InputsRead & (1 << VARYING_SLOT_POS)) != 0;
unsigned barycentric_interp_modes = c->prog_data.barycentric_interp_modes;
assert(intel->gen >= 6);
bool
fs_visitor::run()
{
+ sanity_param_count = fp->Base.Parameters->NumParameters;
uint32_t orig_nr_params = c->prog_data.nr_params;
if (intel->gen >= 6)
if (0) {
emit_dummy_fs();
} else {
+ if (INTEL_DEBUG & DEBUG_SHADER_TIME)
+ emit_shader_time_begin();
+
calculate_urb_setup();
if (intel->gen < 6)
emit_interpolation_setup_gen4();
else
emit_interpolation_setup_gen6();
+ /* We handle discards by keeping track of the still-live pixels in f0.1.
+ * Initialize it with the dispatched pixels.
+ */
+ if (fp->UsesKill) {
+ fs_inst *discard_init = emit(FS_OPCODE_MOV_DISPATCH_TO_FLAGS);
+ discard_init->flag_subreg = 1;
+ }
+
/* Generate FS IR for main(). (the visitor only descends into
* functions called "main").
*/
if (failed)
return false;
+ emit(FS_OPCODE_PLACEHOLDER_HALT);
+
emit_fb_writes();
split_virtual_grfs();
- setup_paramvalues_refs();
+ move_uniform_array_access_to_pull_constants();
setup_pull_constants();
bool progress;
progress = opt_cse() || progress;
progress = opt_copy_propagate() || progress;
progress = dead_code_eliminate() || progress;
+ progress = dead_code_eliminate_local() || progress;
progress = register_coalesce() || progress;
progress = register_coalesce_2() || progress;
progress = compute_to_mrf() || progress;
remove_dead_constants();
- schedule_instructions();
+ schedule_instructions(false);
+
+ lower_uniform_pull_constant_loads();
assign_curb_setup();
assign_urb_setup();
assert(force_uncompressed_stack == 0);
assert(force_sechalf_stack == 0);
+ /* This must come after all optimization and register allocation, since
+ * it inserts dead code that happens to have side effects, and it does
+ * so based on the actual physical registers in use.
+ */
+ insert_gen4_send_dependency_workarounds();
+
if (failed)
return false;
+ schedule_instructions(true);
+
if (dispatch_width == 8) {
c->prog_data.reg_blocks = brw_register_blocks(grf_used);
} else {
(void) orig_nr_params;
}
+ /* If any state parameters were appended, then ParameterValues could have
+ * been realloced, in which case the driver uniform storage set up by
+ * _mesa_associate_uniform_storage() would point to freed memory. Make
+ * sure that didn't happen.
+ */
+ assert(sanity_param_count == fp->Base.Parameters->NumParameters);
+
return !failed;
}
bool start_busy = false;
float start_time = 0;
- if (unlikely(INTEL_DEBUG & DEBUG_PERF)) {
- start_busy = (intel->batch.last_bo &&
- drm_intel_bo_busy(intel->batch.last_bo));
+ if (unlikely(brw->perf_debug)) {
+ start_busy = (brw->batch.last_bo &&
+ drm_intel_bo_busy(brw->batch.last_bo));
start_time = get_time();
}
shader = (brw_shader *) prog->_LinkedShaders[MESA_SHADER_FRAGMENT];
if (unlikely(INTEL_DEBUG & DEBUG_WM)) {
- if (shader) {
+ if (prog) {
printf("GLSL IR for native fragment shader %d:\n", prog->Name);
_mesa_print_ir(shader->ir, NULL);
printf("\n\n");
*/
fs_visitor v(brw, c, prog, fp, 8);
if (!v.run()) {
- prog->LinkStatus = false;
- ralloc_strcat(&prog->InfoLog, v.fail_msg);
+ if (prog) {
+ prog->LinkStatus = false;
+ ralloc_strcat(&prog->InfoLog, v.fail_msg);
+ }
_mesa_problem(NULL, "Failed to compile fragment shader: %s\n",
- v.fail_msg);
+ v.fail_msg);
return NULL;
}
exec_list *simd16_instructions = NULL;
fs_visitor v2(brw, c, prog, fp, 16);
- if (intel->gen >= 5 && c->prog_data.nr_pull_params == 0) {
+ bool no16 = INTEL_DEBUG & DEBUG_NO16;
+ if (intel->gen >= 5 && c->prog_data.nr_pull_params == 0 && likely(!no16)) {
v2.import_uniforms(&v);
if (!v2.run()) {
perf_debug("16-wide shader failed to compile, falling back to "
simd16_instructions,
final_assembly_size);
- if (unlikely(INTEL_DEBUG & DEBUG_PERF) && shader) {
+ if (unlikely(brw->perf_debug) && shader) {
if (shader->compiled_once)
brw_wm_debug_recompile(brw, prog, &c->key);
shader->compiled_once = true;
- if (start_busy && !drm_intel_bo_busy(intel->batch.last_bo)) {
+ if (start_busy && !drm_intel_bo_busy(brw->batch.last_bo)) {
perf_debug("FS compile took %.03f ms and stalled the GPU\n",
(get_time() - start_time) * 1000);
}
key.iz_lookup |= IZ_DEPTH_WRITE_ENABLE_BIT;
}
- if (prog->Name != 0)
- key.proj_attrib_mask = 0xffffffff;
-
if (intel->gen < 6)
- key.vp_outputs_written |= BITFIELD64_BIT(FRAG_ATTRIB_WPOS);
+ key.input_slots_valid |= BITFIELD64_BIT(VARYING_SLOT_POS);
- for (int i = 0; i < FRAG_ATTRIB_MAX; i++) {
+ for (int i = 0; i < VARYING_SLOT_MAX; i++) {
if (!(fp->Base.InputsRead & BITFIELD64_BIT(i)))
continue;
- if (prog->Name == 0)
- key.proj_attrib_mask |= 1 << i;
-
if (intel->gen < 6) {
- 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);
+ if (_mesa_varying_slot_in_fs((gl_varying_slot) i))
+ key.input_slots_valid |= BITFIELD64_BIT(i);
}
}
- key.clamp_fragment_color = true;
+ key.clamp_fragment_color = ctx->API == API_OPENGL_COMPAT;
for (int i = 0; i < MAX_SAMPLERS; i++) {
if (fp->Base.ShadowSamplers & (1 << i)) {
}
}
- if (fp->Base.InputsRead & FRAG_BIT_WPOS) {
+ if (fp->Base.InputsRead & VARYING_BIT_POS) {
key.drawable_height = ctx->DrawBuffer->Height;
}
- if ((fp->Base.InputsRead & FRAG_BIT_WPOS) || program_uses_dfdy) {
+ if ((fp->Base.InputsRead & VARYING_BIT_POS) || program_uses_dfdy) {
key.render_to_fbo = _mesa_is_user_fbo(ctx->DrawBuffer);
}
projects / mesa.git / blobdiff