#include "brw_cfg.h"
#include "brw_vs.h"
#include "brw_nir.h"
+#include "brw_vec4_builder.h"
#include "brw_vec4_live_variables.h"
#include "brw_dead_control_flow.h"
-
-extern "C" {
-#include "main/macros.h"
-#include "main/shaderobj.h"
-#include "program/prog_print.h"
#include "program/prog_parameter.h"
-}
-#include "main/context.h"
#define MAX_INSTRUCTION (1 << 30)
init();
}
-src_reg::src_reg(struct brw_reg reg) :
+src_reg::src_reg(struct ::brw_reg reg) :
backend_reg(reg)
{
this->reg_offset = 0;
}
src_reg::src_reg(const dst_reg ®) :
- backend_reg(static_cast<struct brw_reg>(reg))
+ backend_reg(reg)
{
- this->reg_offset = reg.reg_offset;
this->reladdr = reg.reladdr;
this->swizzle = brw_swizzle_for_mask(reg.writemask);
}
this->writemask = writemask;
}
-dst_reg::dst_reg(struct brw_reg reg) :
+dst_reg::dst_reg(struct ::brw_reg reg) :
backend_reg(reg)
{
this->reg_offset = 0;
}
dst_reg::dst_reg(const src_reg ®) :
- backend_reg(static_cast<struct brw_reg>(reg))
+ backend_reg(reg)
{
- this->reg_offset = reg.reg_offset;
this->writemask = brw_mask_for_swizzle(reg.swizzle);
this->reladdr = reg.reladdr;
}
bool
dst_reg::equals(const dst_reg &r) const
{
- return (memcmp((brw_reg *)this, (brw_reg *)&r, sizeof(brw_reg)) == 0 &&
- reg_offset == r.reg_offset &&
+ return (this->backend_reg::equals(r) &&
(reladdr == r.reladdr ||
(reladdr && r.reladdr && reladdr->equals(*r.reladdr))));
}
case SHADER_OPCODE_TYPED_ATOMIC:
case SHADER_OPCODE_TYPED_SURFACE_READ:
case SHADER_OPCODE_TYPED_SURFACE_WRITE:
+ case VEC4_OPCODE_URB_READ:
+ case TCS_OPCODE_URB_WRITE:
+ case TCS_OPCODE_RELEASE_INPUT:
+ case SHADER_OPCODE_BARRIER:
+ return true;
+ default:
+ return false;
+ }
+}
+
+/**
+ * Returns true if this instruction's sources and destinations cannot
+ * safely be the same register.
+ *
+ * In most cases, a register can be written over safely by the same
+ * instruction that is its last use. For a single instruction, the
+ * sources are dereferenced before writing of the destination starts
+ * (naturally).
+ *
+ * However, there are a few cases where this can be problematic:
+ *
+ * - Virtual opcodes that translate to multiple instructions in the
+ * code generator: if src == dst and one instruction writes the
+ * destination before a later instruction reads the source, then
+ * src will have been clobbered.
+ *
+ * The register allocator uses this information to set up conflicts between
+ * GRF sources and the destination.
+ */
+bool
+vec4_instruction::has_source_and_destination_hazard() const
+{
+ switch (opcode) {
+ case TCS_OPCODE_SET_INPUT_URB_OFFSETS:
+ case TCS_OPCODE_SET_OUTPUT_URB_OFFSETS:
+ case TES_OPCODE_ADD_INDIRECT_URB_OFFSET:
return true;
default:
return false;
case SHADER_OPCODE_TYPED_ATOMIC:
case SHADER_OPCODE_TYPED_SURFACE_READ:
case SHADER_OPCODE_TYPED_SURFACE_WRITE:
+ case TCS_OPCODE_URB_WRITE:
return arg == 0 ? mlen : 1;
case VS_OPCODE_PULL_CONSTANT_LOAD_GEN7:
return true;
}
+bool
+vec4_instruction::can_do_writemask(const struct brw_device_info *devinfo)
+{
+ switch (opcode) {
+ case SHADER_OPCODE_GEN4_SCRATCH_READ:
+ case VS_OPCODE_PULL_CONSTANT_LOAD:
+ case VS_OPCODE_PULL_CONSTANT_LOAD_GEN7:
+ case VS_OPCODE_SET_SIMD4X2_HEADER_GEN9:
+ case TCS_OPCODE_SET_INPUT_URB_OFFSETS:
+ case TCS_OPCODE_SET_OUTPUT_URB_OFFSETS:
+ case TES_OPCODE_CREATE_INPUT_READ_HEADER:
+ case TES_OPCODE_ADD_INDIRECT_URB_OFFSET:
+ case VEC4_OPCODE_URB_READ:
+ case SHADER_OPCODE_MOV_INDIRECT:
+ return false;
+ default:
+ /* The MATH instruction on Gen6 only executes in align1 mode, which does
+ * not support writemasking.
+ */
+ if (devinfo->gen == 6 && is_math())
+ return false;
+
+ if (is_tex())
+ return false;
+
+ return true;
+ }
+}
+
bool
vec4_instruction::can_change_types() const
{
case SHADER_OPCODE_INT_QUOTIENT:
case SHADER_OPCODE_INT_REMAINDER:
case SHADER_OPCODE_POW:
+ case TCS_OPCODE_THREAD_END:
return 2;
case VS_OPCODE_URB_WRITE:
return 1;
return 0;
case GS_OPCODE_FF_SYNC:
return 1;
+ case TCS_OPCODE_URB_WRITE:
+ return 0;
case SHADER_OPCODE_SHADER_TIME_ADD:
return 0;
case SHADER_OPCODE_TEX:
bool
src_reg::equals(const src_reg &r) const
{
- return (memcmp((brw_reg *)this, (brw_reg *)&r, sizeof(brw_reg)) == 0 &&
- reg_offset == r.reg_offset &&
+ return (this->backend_reg::equals(r) &&
!reladdr && !r.reladdr);
}
{
bool progress = false;
- int last_reg = -1, last_reg_offset = -1;
- enum brw_reg_file last_reg_file = BAD_FILE;
+ foreach_block(block, cfg) {
+ int last_reg = -1, last_reg_offset = -1;
+ enum brw_reg_file last_reg_file = BAD_FILE;
- int remaining_channels = 0;
- uint8_t imm[4];
- int inst_count = 0;
- vec4_instruction *imm_inst[4];
+ uint8_t imm[4] = { 0 };
+ int inst_count = 0;
+ vec4_instruction *imm_inst[4];
+ unsigned writemask = 0;
+ enum brw_reg_type dest_type = BRW_REGISTER_TYPE_F;
- foreach_block_and_inst_safe(block, vec4_instruction, inst, cfg) {
- if (last_reg != inst->dst.nr ||
- last_reg_offset != inst->dst.reg_offset ||
- last_reg_file != inst->dst.file) {
- last_reg = inst->dst.nr;
- last_reg_offset = inst->dst.reg_offset;
- last_reg_file = inst->dst.file;
- remaining_channels = WRITEMASK_XYZW;
-
- inst_count = 0;
- }
+ foreach_inst_in_block_safe(vec4_instruction, inst, block) {
+ int vf = -1;
+ enum brw_reg_type need_type;
- if (inst->opcode != BRW_OPCODE_MOV ||
- inst->dst.writemask == WRITEMASK_XYZW ||
- inst->src[0].file != IMM)
- continue;
+ /* Look for unconditional MOVs from an immediate with a partial
+ * writemask. Skip type-conversion MOVs other than integer 0,
+ * where the type doesn't matter. See if the immediate can be
+ * represented as a VF.
+ */
+ if (inst->opcode == BRW_OPCODE_MOV &&
+ inst->src[0].file == IMM &&
+ inst->predicate == BRW_PREDICATE_NONE &&
+ inst->dst.writemask != WRITEMASK_XYZW &&
+ (inst->src[0].type == inst->dst.type || inst->src[0].d == 0)) {
+
+ vf = brw_float_to_vf(inst->src[0].d);
+ need_type = BRW_REGISTER_TYPE_D;
+
+ if (vf == -1) {
+ vf = brw_float_to_vf(inst->src[0].f);
+ need_type = BRW_REGISTER_TYPE_F;
+ }
+ } else {
+ last_reg = -1;
+ }
- int vf = brw_float_to_vf(inst->src[0].f);
- if (vf == -1)
- continue;
+ /* If this wasn't a MOV, or the destination register doesn't match,
+ * or we have to switch destination types, then this breaks our
+ * sequence. Combine anything we've accumulated so far.
+ */
+ if (last_reg != inst->dst.nr ||
+ last_reg_offset != inst->dst.reg_offset ||
+ last_reg_file != inst->dst.file ||
+ (vf > 0 && dest_type != need_type)) {
+
+ if (inst_count > 1) {
+ unsigned vf;
+ memcpy(&vf, imm, sizeof(vf));
+ vec4_instruction *mov = MOV(imm_inst[0]->dst, brw_imm_vf(vf));
+ mov->dst.type = dest_type;
+ mov->dst.writemask = writemask;
+ inst->insert_before(block, mov);
+
+ for (int i = 0; i < inst_count; i++) {
+ imm_inst[i]->remove(block);
+ }
+
+ progress = true;
+ }
- if ((inst->dst.writemask & WRITEMASK_X) != 0)
- imm[0] = vf;
- if ((inst->dst.writemask & WRITEMASK_Y) != 0)
- imm[1] = vf;
- if ((inst->dst.writemask & WRITEMASK_Z) != 0)
- imm[2] = vf;
- if ((inst->dst.writemask & WRITEMASK_W) != 0)
- imm[3] = vf;
-
- imm_inst[inst_count++] = inst;
-
- remaining_channels &= ~inst->dst.writemask;
- if (remaining_channels == 0) {
- unsigned vf;
- memcpy(&vf, imm, sizeof(vf));
- vec4_instruction *mov = MOV(inst->dst, brw_imm_vf(vf));
- mov->dst.type = BRW_REGISTER_TYPE_F;
- mov->dst.writemask = WRITEMASK_XYZW;
- inst->insert_after(block, mov);
- last_reg = -1;
-
- for (int i = 0; i < inst_count; i++) {
- imm_inst[i]->remove(block);
+ inst_count = 0;
+ last_reg = -1;
+ writemask = 0;
+ dest_type = BRW_REGISTER_TYPE_F;
+
+ for (int i = 0; i < 4; i++) {
+ imm[i] = 0;
+ }
+ }
+
+ /* Record this instruction's value (if it was representable). */
+ if (vf != -1) {
+ if ((inst->dst.writemask & WRITEMASK_X) != 0)
+ imm[0] = vf;
+ if ((inst->dst.writemask & WRITEMASK_Y) != 0)
+ imm[1] = vf;
+ if ((inst->dst.writemask & WRITEMASK_Z) != 0)
+ imm[2] = vf;
+ if ((inst->dst.writemask & WRITEMASK_W) != 0)
+ imm[3] = vf;
+
+ writemask |= inst->dst.writemask;
+ imm_inst[inst_count++] = inst;
+
+ last_reg = inst->dst.nr;
+ last_reg_offset = inst->dst.reg_offset;
+ last_reg_file = inst->dst.file;
+ if (vf > 0)
+ dest_type = need_type;
}
- progress = true;
}
}
inst->src[i].reg_offset = 0;
}
}
-
- /* Update that everything is now vector-sized. */
- for (int i = 0; i < this->uniforms; i++) {
- this->uniform_size[i] = 1;
- }
}
void
BRW_GET_SWZ(inst->src[i].swizzle, c) + 1);
}
}
+
+ if (inst->opcode == SHADER_OPCODE_MOV_INDIRECT &&
+ inst->src[0].file == UNIFORM) {
+ assert(inst->src[2].file == BRW_IMMEDIATE_VALUE);
+ assert(inst->src[0].subnr == 0);
+
+ unsigned bytes_read = inst->src[2].ud;
+ assert(bytes_read % 4 == 0);
+ unsigned vec4s_read = DIV_ROUND_UP(bytes_read, 16);
+
+ /* We just mark every register touched by a MOV_INDIRECT as being
+ * fully used. This ensures that it doesn't broken up piecewise by
+ * the next part of our packing algorithm.
+ */
+ int reg = inst->src[0].nr;
+ for (unsigned i = 0; i < vec4s_read; i++)
+ chans_used[reg + i] = 4;
+ }
}
int new_uniform_count = 0;
* push constants.
*/
for (int src = 0; src < uniforms; src++) {
- assert(src < uniform_array_size);
int size = chans_used[src];
if (size == 0)
if (inst->dst.type != inst->src[0].type)
assert(!"unimplemented: saturate mixed types");
- if (brw_saturate_immediate(inst->dst.type, &inst->src[0])) {
+ if (brw_saturate_immediate(inst->dst.type,
+ &inst->src[0].as_brw_reg())) {
inst->saturate = false;
progress = true;
}
break;
}
break;
- case SHADER_OPCODE_RCP: {
- vec4_instruction *prev = (vec4_instruction *)inst->prev;
- if (prev->opcode == SHADER_OPCODE_SQRT) {
- if (inst->src[0].equals(src_reg(prev->dst))) {
- inst->opcode = SHADER_OPCODE_RSQ;
- inst->src[0] = prev->src[0];
- progress = true;
- }
- }
- break;
- }
case SHADER_OPCODE_BROADCAST:
if (is_uniform(inst->src[0]) ||
inst->src[1].is_zero()) {
dst_reg temp = dst_reg(this, glsl_type::vec4_type);
emit_pull_constant_load(block, inst, temp, inst->src[i],
- pull_constant_loc[uniform]);
+ pull_constant_loc[uniform], src_reg());
inst->src[i].file = temp.file;
inst->src[i].nr = temp.nr;
int swizzle_mask)
{
/* Gen6 MATH instructions can not execute in align16 mode, so swizzles
- * or writemasking are not allowed.
+ * are not allowed.
*/
- if (devinfo->gen == 6 && is_math() &&
- (swizzle != BRW_SWIZZLE_XYZW || dst_writemask != WRITEMASK_XYZW))
+ if (devinfo->gen == 6 && is_math() && swizzle != BRW_SWIZZLE_XYZW)
+ return false;
+
+ if (!can_do_writemask(devinfo) && dst_writemask != WRITEMASK_XYZW)
return false;
/* If this instruction sets anything not referenced by swizzle, then we'd
if (is_nop_mov) {
inst->remove(block);
+ progress = true;
continue;
}
}
pred_ctrl_align16[inst->predicate]);
}
- fprintf(file, "%s", brw_instruction_name(inst->opcode));
+ fprintf(file, "%s", brw_instruction_name(devinfo, inst->opcode));
if (inst->saturate)
fprintf(file, ".sat");
if (inst->conditional_mod) {
bool interleaved)
{
foreach_block_and_inst(block, vec4_instruction, inst, cfg) {
- /* We have to support ATTR as a destination for GL_FIXED fixup. */
- if (inst->dst.file == ATTR) {
- int grf = attribute_map[inst->dst.nr + inst->dst.reg_offset];
-
- /* All attributes used in the shader need to have been assigned a
- * hardware register by the caller
- */
- assert(grf != 0);
-
- struct brw_reg reg = attribute_to_hw_reg(grf, interleaved);
- reg.type = inst->dst.type;
- reg.writemask = inst->dst.writemask;
-
- inst->dst = reg;
- }
-
for (int i = 0; i < 3; i++) {
if (inst->src[i].file != ATTR)
continue;
vec4_vs_visitor::setup_attributes(int payload_reg)
{
int nr_attributes;
- int attribute_map[VERT_ATTRIB_MAX + 1];
+ int attribute_map[VERT_ATTRIB_MAX + 2];
memset(attribute_map, 0, sizeof(attribute_map));
nr_attributes = 0;
* don't represent it with a flag in inputs_read, so we call it
* VERT_ATTRIB_MAX.
*/
- if (vs_prog_data->uses_vertexid || vs_prog_data->uses_instanceid) {
+ if (vs_prog_data->uses_vertexid || vs_prog_data->uses_instanceid ||
+ vs_prog_data->uses_basevertex || vs_prog_data->uses_baseinstance) {
attribute_map[VERT_ATTRIB_MAX] = payload_reg + nr_attributes;
+ nr_attributes++;
+ }
+
+ if (vs_prog_data->uses_drawid) {
+ attribute_map[VERT_ATTRIB_MAX + 1] = payload_reg + nr_attributes;
+ nr_attributes++;
}
lower_attributes_to_hw_regs(attribute_map, false /* interleaved */);
* matter what, or the GPU would hang.
*/
if (devinfo->gen < 6 && this->uniforms == 0) {
- assert(this->uniforms < this->uniform_array_size);
-
stage_prog_data->param =
reralloc(NULL, stage_prog_data->param, const gl_constant_value *, 4);
for (unsigned int i = 0; i < 4; i++) {
this->first_non_payload_grf = reg;
}
+bool
+vec4_visitor::lower_minmax()
+{
+ assert(devinfo->gen < 6);
+
+ bool progress = false;
+
+ foreach_block_and_inst_safe(block, vec4_instruction, inst, cfg) {
+ const vec4_builder ibld(this, block, inst);
+
+ if (inst->opcode == BRW_OPCODE_SEL &&
+ inst->predicate == BRW_PREDICATE_NONE) {
+ /* FIXME: Using CMP doesn't preserve the NaN propagation semantics of
+ * the original SEL.L/GE instruction
+ */
+ ibld.CMP(ibld.null_reg_d(), inst->src[0], inst->src[1],
+ inst->conditional_mod);
+ inst->predicate = BRW_PREDICATE_NORMAL;
+ inst->conditional_mod = BRW_CONDITIONAL_NONE;
+
+ progress = true;
+ }
+ }
+
+ if (progress)
+ invalidate_live_intervals();
+
+ return progress;
+}
+
src_reg
vec4_visitor::get_timestamp()
{
case ATTR:
unreachable("not reached");
}
+
src = reg;
}
+ if (inst->is_3src(devinfo)) {
+ /* 3-src instructions with scalar sources support arbitrary subnr,
+ * but don't actually use swizzles. Convert swizzle into subnr.
+ */
+ for (int i = 0; i < 3; i++) {
+ if (inst->src[i].vstride == BRW_VERTICAL_STRIDE_0) {
+ assert(brw_is_single_value_swizzle(inst->src[i].swizzle));
+ inst->src[i].subnr += 4 * BRW_GET_SWZ(inst->src[i].swizzle, 0);
+ }
+ }
+ }
+
dst_reg &dst = inst->dst;
struct brw_reg reg;
case ARF:
case FIXED_GRF:
- reg = dst;
+ reg = dst.as_brw_reg();
break;
case BAD_FILE:
if (unlikely(INTEL_DEBUG & DEBUG_OPTIMIZER)) {
char filename[64];
- snprintf(filename, 64, "%s-%s-00-start",
+ snprintf(filename, 64, "%s-%s-00-00-start",
stage_abbrev, nir->info.name);
backend_shader::dump_instructions(filename);
OPT(dead_code_eliminate);
}
+ if (devinfo->gen <= 5 && OPT(lower_minmax)) {
+ OPT(opt_cmod_propagation);
+ OPT(opt_cse);
+ OPT(opt_copy_propagation);
+ OPT(dead_code_eliminate);
+ }
+
if (failed)
return false;
void *mem_ctx,
const struct brw_vs_prog_key *key,
struct brw_vs_prog_data *prog_data,
- const nir_shader *shader,
+ const nir_shader *src_shader,
gl_clip_plane *clip_planes,
bool use_legacy_snorm_formula,
int shader_time_index,
unsigned *final_assembly_size,
char **error_str)
{
+ const bool is_scalar = compiler->scalar_stage[MESA_SHADER_VERTEX];
+ nir_shader *shader = nir_shader_clone(mem_ctx, src_shader);
+ shader = brw_nir_apply_sampler_key(shader, compiler->devinfo, &key->tex,
+ is_scalar);
+ brw_nir_lower_vs_inputs(shader, compiler->devinfo, is_scalar,
+ use_legacy_snorm_formula, key->gl_attrib_wa_flags);
+ brw_nir_lower_vue_outputs(shader, is_scalar);
+ shader = brw_postprocess_nir(shader, compiler->devinfo, is_scalar);
+
const unsigned *assembly = NULL;
unsigned nr_attributes = _mesa_bitcount_64(prog_data->inputs_read);
* incoming vertex attribute. So, add an extra slot.
*/
if (shader->info.system_values_read &
- (BITFIELD64_BIT(SYSTEM_VALUE_VERTEX_ID_ZERO_BASE) |
+ (BITFIELD64_BIT(SYSTEM_VALUE_BASE_VERTEX) |
+ BITFIELD64_BIT(SYSTEM_VALUE_BASE_INSTANCE) |
+ BITFIELD64_BIT(SYSTEM_VALUE_VERTEX_ID_ZERO_BASE) |
BITFIELD64_BIT(SYSTEM_VALUE_INSTANCE_ID))) {
nr_attributes++;
}
+ /* gl_DrawID has its very own vec4 */
+ if (shader->info.system_values_read & BITFIELD64_BIT(SYSTEM_VALUE_DRAW_ID)) {
+ nr_attributes++;
+ }
+
+ unsigned nr_attribute_slots =
+ nr_attributes +
+ _mesa_bitcount_64(shader->info.double_inputs_read);
+
/* The 3DSTATE_VS documentation lists the lower bound on "Vertex URB Entry
* Read Length" as 1 in vec4 mode, and 0 in SIMD8 mode. Empirically, in
* vec4 mode, the hardware appears to wedge unless we read something.
*/
- if (compiler->scalar_stage[MESA_SHADER_VERTEX])
- prog_data->base.urb_read_length = DIV_ROUND_UP(nr_attributes, 2);
+ if (is_scalar)
+ prog_data->base.urb_read_length =
+ DIV_ROUND_UP(nr_attribute_slots, 2);
else
- prog_data->base.urb_read_length = DIV_ROUND_UP(MAX2(nr_attributes, 1), 2);
+ prog_data->base.urb_read_length =
+ DIV_ROUND_UP(MAX2(nr_attribute_slots, 1), 2);
prog_data->nr_attributes = nr_attributes;
+ prog_data->nr_attribute_slots = nr_attribute_slots;
/* Since vertex shaders reuse the same VUE entry for inputs and outputs
* (overwriting the original contents), we need to make sure the size is
* the larger of the two.
*/
const unsigned vue_entries =
- MAX2(nr_attributes, (unsigned)prog_data->base.vue_map.num_slots);
+ MAX2(nr_attribute_slots, (unsigned)prog_data->base.vue_map.num_slots);
if (compiler->devinfo->gen == 6)
prog_data->base.urb_entry_size = DIV_ROUND_UP(vue_entries, 8);
else
prog_data->base.urb_entry_size = DIV_ROUND_UP(vue_entries, 4);
- if (compiler->scalar_stage[MESA_SHADER_VERTEX]) {
+ if (is_scalar) {
prog_data->base.dispatch_mode = DISPATCH_MODE_SIMD8;
fs_visitor v(compiler, log_data, mem_ctx, key, &prog_data->base.base,
return NULL;
}
+ prog_data->base.base.dispatch_grf_start_reg = v.payload.num_regs;
+
fs_generator g(compiler, log_data, mem_ctx, (void *) key,
&prog_data->base.base, v.promoted_constants,
- v.runtime_check_aads_emit, "VS");
+ v.runtime_check_aads_emit, MESA_SHADER_VERTEX);
if (INTEL_DEBUG & DEBUG_VS) {
const char *debug_name =
ralloc_asprintf(mem_ctx, "%s vertex shader %s",