void
brw_init_codegen(const struct brw_device_info *devinfo,
struct brw_codegen *p, void *mem_ctx)
{
memset(p, 0, sizeof(*p));
p->devinfo = devinfo;
/*
* Set the initial instruction store array size to 1024, if found that
* isn't enough, then it will double the store size at brw_next_insn()
* until out of memory.
*/
p->store_size = 1024;
p->store = rzalloc_array(mem_ctx, brw_inst, p->store_size);
p->nr_insn = 0;
p->current = p->stack;
p->compressed = false;
memset(p->current, 0, sizeof(p->current[0]));
p->mem_ctx = mem_ctx;
/* Some defaults?
*/
brw_set_default_exec_size(p, BRW_EXECUTE_8);
brw_set_default_mask_control(p, BRW_MASK_ENABLE); /* what does this do? */
brw_set_default_saturate(p, 0);
brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
/* Set up control flow stack */
p->if_stack_depth = 0;
p->if_stack_array_size = 16;
p->if_stack = rzalloc_array(mem_ctx, int, p->if_stack_array_size);
p->loop_stack_depth = 0;
p->loop_stack_array_size = 16;
p->loop_stack = rzalloc_array(mem_ctx, int, p->loop_stack_array_size);
p->if_depth_in_loop = rzalloc_array(mem_ctx, int, p->loop_stack_array_size);
brw_init_compaction_tables(devinfo);
}
/**
* Generate the geometry shader program used on Gen6 to perform stream output
* (transform feedback).
*/
void
gen6_sol_program(struct brw_ff_gs_compile *c, struct brw_ff_gs_prog_key *key,
unsigned num_verts, bool check_edge_flags)
{
struct brw_codegen *p = &c->func;
brw_inst *inst;
c->prog_data.svbi_postincrement_value = num_verts;
brw_ff_gs_alloc_regs(c, num_verts, true);
brw_ff_gs_initialize_header(c);
if (key->num_transform_feedback_bindings > 0) {
unsigned vertex, binding;
struct brw_reg destination_indices_uw =
vec8(retype(c->reg.destination_indices, BRW_REGISTER_TYPE_UW));
/* Note: since we use the binding table to keep track of buffer offsets
* and stride, the GS doesn't need to keep track of a separate pointer
* into each buffer; it uses a single pointer which increments by 1 for
* each vertex. So we use SVBI0 for this pointer, regardless of whether
* transform feedback is in interleaved or separate attribs mode.
*
* Make sure that the buffers have enough room for all the vertices.
*/
brw_ADD(p, get_element_ud(c->reg.temp, 0),
get_element_ud(c->reg.SVBI, 0), brw_imm_ud(num_verts));
brw_CMP(p, vec1(brw_null_reg()), BRW_CONDITIONAL_LE,
get_element_ud(c->reg.temp, 0),
get_element_ud(c->reg.SVBI, 4));
brw_IF(p, BRW_EXECUTE_1);
/* Compute the destination indices to write to. Usually we use SVBI[0]
* + (0, 1, 2). However, for odd-numbered triangles in tristrips, the
* vertices come down the pipeline in reversed winding order, so we need
* to flip the order when writing to the transform feedback buffer. To
* ensure that flatshading accuracy is preserved, we need to write them
* in order SVBI[0] + (0, 2, 1) if we're using the first provoking
* vertex convention, and in order SVBI[0] + (1, 0, 2) if we're using
* the last provoking vertex convention.
*
* Note: since brw_imm_v can only be used in instructions in
* packed-word execution mode, and SVBI is a double-word, we need to
* first move the appropriate immediate constant ((0, 1, 2), (0, 2, 1),
* or (1, 0, 2)) to the destination_indices register, and then add SVBI
* using a separate instruction. Also, since the immediate constant is
* expressed as packed words, and we need to load double-words into
* destination_indices, we need to intersperse zeros to fill the upper
* halves of each double-word.
*/
brw_MOV(p, destination_indices_uw,
brw_imm_v(0x00020100)); /* (0, 1, 2) */
if (num_verts == 3) {
/* Get primitive type into temp register. */
brw_AND(p, get_element_ud(c->reg.temp, 0),
get_element_ud(c->reg.R0, 2), brw_imm_ud(0x1f));
/* Test if primitive type is TRISTRIP_REVERSE. We need to do this as
* an 8-wide comparison so that the conditional MOV that follows
* moves all 8 words correctly.
*/
brw_CMP(p, vec8(brw_null_reg()), BRW_CONDITIONAL_EQ,
get_element_ud(c->reg.temp, 0),
brw_imm_ud(_3DPRIM_TRISTRIP_REVERSE));
/* If so, then overwrite destination_indices_uw with the appropriate
* reordering.
*/
inst = brw_MOV(p, destination_indices_uw,
brw_imm_v(key->pv_first ? 0x00010200 /* (0, 2, 1) */
: 0x00020001)); /* (1, 0, 2) */
brw_inst_set_pred_control(p->devinfo, inst, BRW_PREDICATE_NORMAL);
}
assert(c->reg.destination_indices.width == BRW_EXECUTE_4);
brw_push_insn_state(p);
brw_set_default_exec_size(p, BRW_EXECUTE_4);
brw_ADD(p, c->reg.destination_indices,
c->reg.destination_indices, get_element_ud(c->reg.SVBI, 0));
brw_pop_insn_state(p);
/* For each vertex, generate code to output each varying using the
* appropriate binding table entry.
*/
for (vertex = 0; vertex < num_verts; ++vertex) {
/* Set up the correct destination index for this vertex */
brw_MOV(p, get_element_ud(c->reg.header, 5),
get_element_ud(c->reg.destination_indices, vertex));
for (binding = 0; binding < key->num_transform_feedback_bindings;
++binding) {
unsigned char varying =
key->transform_feedback_bindings[binding];
unsigned char slot = c->vue_map.varying_to_slot[varying];
/* From the Sandybridge PRM, Volume 2, Part 1, Section 4.5.1:
*
* "Prior to End of Thread with a URB_WRITE, the kernel must
* ensure that all writes are complete by sending the final
* write as a committed write."
*/
bool final_write =
binding == key->num_transform_feedback_bindings - 1 &&
vertex == num_verts - 1;
struct brw_reg vertex_slot = c->reg.vertex[vertex];
vertex_slot.nr += slot / 2;
vertex_slot.subnr = (slot % 2) * 16;
/* gl_PointSize is stored in VARYING_SLOT_PSIZ.w. */
vertex_slot.swizzle = varying == VARYING_SLOT_PSIZ
? BRW_SWIZZLE_WWWW : key->transform_feedback_swizzles[binding];
brw_set_default_access_mode(p, BRW_ALIGN_16);
brw_push_insn_state(p);
brw_set_default_exec_size(p, BRW_EXECUTE_4);
brw_MOV(p, stride(c->reg.header, 4, 4, 1),
retype(vertex_slot, BRW_REGISTER_TYPE_UD));
brw_pop_insn_state(p);
brw_set_default_access_mode(p, BRW_ALIGN_1);
brw_svb_write(p,
final_write ? c->reg.temp : brw_null_reg(), /* dest */
1, /* msg_reg_nr */
c->reg.header, /* src0 */
SURF_INDEX_GEN6_SOL_BINDING(binding), /* binding_table_index */
final_write); /* send_commit_msg */
}
}
brw_ENDIF(p);
/* Now, reinitialize the header register from R0 to restore the parts of
* the register that we overwrote while streaming out transform feedback
* data.
*/
brw_ff_gs_initialize_header(c);
/* Finally, wait for the write commit to occur so that we can proceed to
* other things safely.
*
* From the Sandybridge PRM, Volume 4, Part 1, Section 3.3:
*
* The write commit does not modify the destination register, but
* merely clears the dependency associated with the destination
* register. Thus, a simple “mov” instruction using the register as a
* source is sufficient to wait for the write commit to occur.
*/
brw_MOV(p, c->reg.temp, c->reg.temp);
}
brw_ff_gs_ff_sync(c, 1);
brw_ff_gs_overwrite_header_dw2_from_r0(c);
switch (num_verts) {
case 1:
brw_ff_gs_offset_header_dw2(c,
URB_WRITE_PRIM_START | URB_WRITE_PRIM_END);
brw_ff_gs_emit_vue(c, c->reg.vertex[0], true);
break;
case 2:
brw_ff_gs_offset_header_dw2(c, URB_WRITE_PRIM_START);
brw_ff_gs_emit_vue(c, c->reg.vertex[0], false);
brw_ff_gs_offset_header_dw2(c,
URB_WRITE_PRIM_END - URB_WRITE_PRIM_START);
brw_ff_gs_emit_vue(c, c->reg.vertex[1], true);
break;
case 3:
if (check_edge_flags) {
/* Only emit vertices 0 and 1 if this is the first triangle of the
* polygon. Otherwise they are redundant.
*/
brw_AND(p, retype(brw_null_reg(), BRW_REGISTER_TYPE_UD),
get_element_ud(c->reg.R0, 2),
brw_imm_ud(BRW_GS_EDGE_INDICATOR_0));
brw_inst_set_cond_modifier(p->devinfo, brw_last_inst, BRW_CONDITIONAL_NZ);
brw_IF(p, BRW_EXECUTE_1);
}
brw_ff_gs_offset_header_dw2(c, URB_WRITE_PRIM_START);
brw_ff_gs_emit_vue(c, c->reg.vertex[0], false);
brw_ff_gs_offset_header_dw2(c, -URB_WRITE_PRIM_START);
brw_ff_gs_emit_vue(c, c->reg.vertex[1], false);
if (check_edge_flags) {
brw_ENDIF(p);
/* Only emit vertex 2 in PRIM_END mode if this is the last triangle
* of the polygon. Otherwise leave the primitive incomplete because
* there are more polygon vertices coming.
*/
brw_AND(p, retype(brw_null_reg(), BRW_REGISTER_TYPE_UD),
get_element_ud(c->reg.R0, 2),
brw_imm_ud(BRW_GS_EDGE_INDICATOR_1));
brw_inst_set_cond_modifier(p->devinfo, brw_last_inst, BRW_CONDITIONAL_NZ);
brw_set_default_predicate_control(p, BRW_PREDICATE_NORMAL);
}
brw_ff_gs_offset_header_dw2(c, URB_WRITE_PRIM_END);
brw_set_default_predicate_control(p, BRW_PREDICATE_NONE);
brw_ff_gs_emit_vue(c, c->reg.vertex[2], true);
break;
}
}
