void
vec4_generator::generate_gs_set_vertex_count(struct brw_reg dst,
struct brw_reg src)
{
brw_push_insn_state(p);
brw_set_default_mask_control(p, BRW_MASK_DISABLE);
if (brw->gen >= 8) {
/* Move the vertex count into the second MRF for the EOT write. */
brw_MOV(p, retype(brw_message_reg(dst.nr + 1), BRW_REGISTER_TYPE_UD),
src);
} else {
/* If we think of the src and dst registers as composed of 8 DWORDs each,
* we want to pick up the contents of DWORDs 0 and 4 from src, truncate
* them to WORDs, and then pack them into DWORD 2 of dst.
*
* It's easier to get the EU to do this if we think of the src and dst
* registers as composed of 16 WORDS each; then, we want to pick up the
* contents of WORDs 0 and 8 from src, and pack them into WORDs 4 and 5
* of dst.
*
* We can do that by the following EU instruction:
*
* mov (2) dst.4<1>:uw src<8;1,0>:uw { Align1, Q1, NoMask }
*/
brw_set_default_access_mode(p, BRW_ALIGN_1);
brw_MOV(p,
suboffset(stride(retype(dst, BRW_REGISTER_TYPE_UW), 2, 2, 1), 4),
stride(retype(src, BRW_REGISTER_TYPE_UW), 8, 1, 0));
brw_set_default_access_mode(p, BRW_ALIGN_16);
}
brw_pop_insn_state(p);
}
static bool
run_tests(struct brw_context *brw)
{
bool fail = false;
for (int i = 0; i < ARRAY_SIZE(tests); i++) {
for (int align_16 = 0; align_16 <= 1; align_16++) {
struct brw_compile *p = rzalloc(NULL, struct brw_compile);
brw_init_compile(brw, p, p);
brw_set_default_predicate_control(p, BRW_PREDICATE_NONE);
if (align_16)
brw_set_default_access_mode(p, BRW_ALIGN_16);
else
brw_set_default_access_mode(p, BRW_ALIGN_1);
tests[i].func(p);
assert(p->nr_insn == 1);
if (!test_compact_instruction(p, p->store[0])) {
fail = true;
continue;
}
if (!test_fuzz_compact_instruction(p, p->store[0])) {
fail = true;
continue;
}
ralloc_free(p);
}
}
return fail;
}
void
vec4_generator::generate_gs_set_write_offset(struct brw_reg dst,
struct brw_reg src0,
struct brw_reg src1)
{
/* From p22 of volume 4 part 2 of the Ivy Bridge PRM (2.4.3.1 Message
* Header: M0.3):
*
* Slot 0 Offset. This field, after adding to the Global Offset field
* in the message descriptor, specifies the offset (in 256-bit units)
* from the start of the URB entry, as referenced by URB Handle 0, at
* which the data will be accessed.
*
* Similar text describes DWORD M0.4, which is slot 1 offset.
*
* Therefore, we want to multiply DWORDs 0 and 4 of src0 (the x components
* of the register for geometry shader invocations 0 and 1) by the
* immediate value in src1, and store the result in DWORDs 3 and 4 of dst.
*
* We can do this with the following EU instruction:
*
* mul(2) dst.3<1>UD src0<8;2,4>UD src1 { Align1 WE_all }
*/
brw_push_insn_state(p);
brw_set_default_access_mode(p, BRW_ALIGN_1);
brw_set_default_mask_control(p, BRW_MASK_DISABLE);
brw_MUL(p, suboffset(stride(dst, 2, 2, 1), 3), stride(src0, 8, 2, 4),
src1);
brw_set_default_access_mode(p, BRW_ALIGN_16);
brw_pop_insn_state(p);
}
void
vec4_generator::generate_gs_set_dword_2_immed(struct brw_reg dst,
struct brw_reg src)
{
assert(src.file == BRW_IMMEDIATE_VALUE);
brw_push_insn_state(p);
brw_set_default_access_mode(p, BRW_ALIGN_1);
brw_set_default_mask_control(p, BRW_MASK_DISABLE);
brw_MOV(p, suboffset(vec1(dst), 2), src);
brw_set_default_access_mode(p, BRW_ALIGN_16);
brw_pop_insn_state(p);
}
/* Project 'pos' to screen space (or back again), overwrite with results:
*/
void brw_clip_project_position(struct brw_clip_compile *c, struct brw_reg pos )
{
struct brw_codegen *p = &c->func;
/* calc rhw
*/
brw_math_invert(p, get_element(pos, W), get_element(pos, W));
/* value.xyz *= value.rhw
*/
brw_set_default_access_mode(p, BRW_ALIGN_16);
brw_MUL(p, brw_writemask(pos, WRITEMASK_XYZ), pos, brw_swizzle1(pos, W));
brw_set_default_access_mode(p, BRW_ALIGN_1);
}
void
vec4_generator::generate_math_gen6(vec4_instruction *inst,
struct brw_reg dst,
struct brw_reg src0,
struct brw_reg src1)
{
/* Can't do writemask because math can't be align16. */
assert(dst.dw1.bits.writemask == WRITEMASK_XYZW);
/* Source swizzles are ignored. */
check_gen6_math_src_arg(src0);
if (src1.file == BRW_GENERAL_REGISTER_FILE)
check_gen6_math_src_arg(src1);
brw_set_default_access_mode(p, BRW_ALIGN_1);
gen6_math(p, dst, brw_math_function(inst->opcode), src0, src1);
brw_set_default_access_mode(p, BRW_ALIGN_16);
}
/* This is performed against the original triangles, so no indirection
* required:
BZZZT!
*/
static void compute_tri_direction( struct brw_clip_compile *c )
{
struct brw_compile *p = &c->func;
struct brw_reg e = c->reg.tmp0;
struct brw_reg f = c->reg.tmp1;
GLuint hpos_offset = brw_varying_to_offset(&c->vue_map, VARYING_SLOT_POS);
struct brw_reg v0 = byte_offset(c->reg.vertex[0], hpos_offset);
struct brw_reg v1 = byte_offset(c->reg.vertex[1], hpos_offset);
struct brw_reg v2 = byte_offset(c->reg.vertex[2], hpos_offset);
struct brw_reg v0n = get_tmp(c);
struct brw_reg v1n = get_tmp(c);
struct brw_reg v2n = get_tmp(c);
/* Convert to NDC.
* NOTE: We can't modify the original vertex coordinates,
* as it may impact further operations.
* So, we have to keep normalized coordinates in temp registers.
*
* TBD-KC
* Try to optimize unnecessary MOV's.
*/
brw_MOV(p, v0n, v0);
brw_MOV(p, v1n, v1);
brw_MOV(p, v2n, v2);
brw_clip_project_position(c, v0n);
brw_clip_project_position(c, v1n);
brw_clip_project_position(c, v2n);
/* Calculate the vectors of two edges of the triangle:
*/
brw_ADD(p, e, v0n, negate(v2n));
brw_ADD(p, f, v1n, negate(v2n));
/* Take their crossproduct:
*/
brw_set_default_access_mode(p, BRW_ALIGN_16);
brw_MUL(p, vec4(brw_null_reg()), brw_swizzle(e, 1,2,0,3), brw_swizzle(f,2,0,1,3));
brw_MAC(p, vec4(e), negate(brw_swizzle(e, 2,0,1,3)), brw_swizzle(f,1,2,0,3));
brw_set_default_access_mode(p, BRW_ALIGN_1);
brw_MUL(p, c->reg.dir, c->reg.dir, vec4(e));
}
const unsigned *
vec4_generator::generate_assembly(const cfg_t *cfg,
unsigned *assembly_size)
{
brw_set_default_access_mode(p, BRW_ALIGN_16);
generate_code(cfg);
return brw_get_program(p, assembly_size);
}
const unsigned *
vec4_generator::generate_assembly(exec_list *instructions,
unsigned *assembly_size)
{
brw_set_default_access_mode(p, BRW_ALIGN_16);
generate_code(instructions);
return brw_get_program(p, assembly_size);
}
static bool
run_tests(const struct gen_device_info *devinfo)
{
brw_init_compaction_tables(devinfo);
bool fail = false;
for (unsigned i = 0; i < ARRAY_SIZE(tests); i++) {
for (int align_16 = 0; align_16 <= 1; align_16++) {
struct brw_codegen *p = rzalloc(NULL, struct brw_codegen);
brw_init_codegen(devinfo, p, p);
brw_set_default_predicate_control(p, BRW_PREDICATE_NONE);
if (align_16)
brw_set_default_access_mode(p, BRW_ALIGN_16);
else
brw_set_default_access_mode(p, BRW_ALIGN_1);
tests[i].func(p);
assert(p->nr_insn == 1);
if (!test_compact_instruction(p, p->store[0])) {
fail = true;
continue;
}
if (!test_fuzz_compact_instruction(p, p->store[0])) {
fail = true;
continue;
}
ralloc_free(p);
}
}
return fail;
}
void
vec4_generator::generate_gs_prepare_channel_masks(struct brw_reg dst)
{
/* We want to left shift just DWORD 4 (the x component belonging to the
* second geometry shader invocation) by 4 bits. So generate the
* instruction:
*
* shl(1) dst.4<1>UD dst.4<0,1,0>UD 4UD { align1 WE_all }
*/
dst = suboffset(vec1(dst), 4);
brw_push_insn_state(p);
brw_set_default_access_mode(p, BRW_ALIGN_1);
brw_set_default_mask_control(p, BRW_MASK_DISABLE);
brw_SHL(p, dst, dst, brw_imm_ud(4));
brw_pop_insn_state(p);
}
void
vec4_generator::generate_gs_get_instance_id(struct brw_reg dst)
{
/* We want to right shift R0.0 & R0.1 by GEN7_GS_PAYLOAD_INSTANCE_ID_SHIFT
* and store into dst.0 & dst.4. So generate the instruction:
*
* shr(8) dst<1> R0<1,4,0> GEN7_GS_PAYLOAD_INSTANCE_ID_SHIFT { align1 WE_normal 1Q }
*/
brw_push_insn_state(p);
brw_set_default_access_mode(p, BRW_ALIGN_1);
dst = retype(dst, BRW_REGISTER_TYPE_UD);
struct brw_reg r0(retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD));
brw_SHR(p, dst, stride(r0, 1, 4, 0),
brw_imm_ud(GEN7_GS_PAYLOAD_INSTANCE_ID_SHIFT));
brw_pop_insn_state(p);
}
void
vec4_generator::generate_unpack_flags(vec4_instruction *inst,
struct brw_reg dst)
{
brw_push_insn_state(p);
brw_set_default_mask_control(p, BRW_MASK_DISABLE);
brw_set_default_access_mode(p, BRW_ALIGN_1);
struct brw_reg flags = brw_flag_reg(0, 0);
struct brw_reg dst_0 = suboffset(vec1(dst), 0);
struct brw_reg dst_4 = suboffset(vec1(dst), 4);
brw_AND(p, dst_0, flags, brw_imm_ud(0x0f));
brw_AND(p, dst_4, flags, brw_imm_ud(0xf0));
brw_SHR(p, dst_4, dst_4, brw_imm_ud(4));
brw_pop_insn_state(p);
}
void
vec4_generator::generate_oword_dual_block_offsets(struct brw_reg m1,
struct brw_reg index)
{
int second_vertex_offset;
if (brw->gen >= 6)
second_vertex_offset = 1;
else
second_vertex_offset = 16;
m1 = retype(m1, BRW_REGISTER_TYPE_D);
/* Set up M1 (message payload). Only the block offsets in M1.0 and
* M1.4 are used, and the rest are ignored.
*/
struct brw_reg m1_0 = suboffset(vec1(m1), 0);
struct brw_reg m1_4 = suboffset(vec1(m1), 4);
struct brw_reg index_0 = suboffset(vec1(index), 0);
struct brw_reg index_4 = suboffset(vec1(index), 4);
brw_push_insn_state(p);
brw_set_default_mask_control(p, BRW_MASK_DISABLE);
brw_set_default_access_mode(p, BRW_ALIGN_1);
brw_MOV(p, m1_0, index_0);
if (index.file == BRW_IMMEDIATE_VALUE) {
index_4.dw1.ud += second_vertex_offset;
brw_MOV(p, m1_4, index_4);
} else {
brw_ADD(p, m1_4, index_4, brw_imm_d(second_vertex_offset));
}
brw_pop_insn_state(p);
}
/* Interpolate between two vertices and put the result into a0.0.
* Increment a0.0 accordingly.
*
* Beware that dest_ptr can be equal to v0_ptr!
*/
void brw_clip_interp_vertex( struct brw_clip_compile *c,
struct brw_indirect dest_ptr,
struct brw_indirect v0_ptr, /* from */
struct brw_indirect v1_ptr, /* to */
struct brw_reg t0,
bool force_edgeflag)
{
struct brw_codegen *p = &c->func;
struct brw_reg t_nopersp, v0_ndc_copy;
GLuint slot;
/* Just copy the vertex header:
*/
/*
* After CLIP stage, only first 256 bits of the VUE are read
* back on Ironlake, so needn't change it
*/
brw_copy_indirect_to_indirect(p, dest_ptr, v0_ptr, 1);
/* First handle the 3D and NDC interpolation, in case we
* need noperspective interpolation. Doing it early has no
* performance impact in any case.
*/
/* Take a copy of the v0 NDC coordinates, in case dest == v0. */
if (c->has_noperspective_shading) {
GLuint offset = brw_varying_to_offset(&c->vue_map,
BRW_VARYING_SLOT_NDC);
v0_ndc_copy = get_tmp(c);
brw_MOV(p, v0_ndc_copy, deref_4f(v0_ptr, offset));
}
/* Compute the new 3D position
*
* dest_hpos = v0_hpos * (1 - t0) + v1_hpos * t0
*/
{
GLuint delta = brw_varying_to_offset(&c->vue_map, VARYING_SLOT_POS);
struct brw_reg tmp = get_tmp(c);
brw_MUL(p, vec4(brw_null_reg()), deref_4f(v1_ptr, delta), t0);
brw_MAC(p, tmp, negate(deref_4f(v0_ptr, delta)), t0);
brw_ADD(p, deref_4f(dest_ptr, delta), deref_4f(v0_ptr, delta), tmp);
release_tmp(c, tmp);
}
/* Recreate the projected (NDC) coordinate in the new vertex header */
brw_clip_project_vertex(c, dest_ptr);
/* If we have noperspective attributes,
* we need to compute the screen-space t
*/
if (c->has_noperspective_shading) {
GLuint delta = brw_varying_to_offset(&c->vue_map,
BRW_VARYING_SLOT_NDC);
struct brw_reg tmp = get_tmp(c);
t_nopersp = get_tmp(c);
/* t_nopersp = vec4(v1.xy, dest.xy) */
brw_MOV(p, t_nopersp, deref_4f(v1_ptr, delta));
brw_MOV(p, tmp, deref_4f(dest_ptr, delta));
brw_set_default_access_mode(p, BRW_ALIGN_16);
brw_MOV(p,
brw_writemask(t_nopersp, WRITEMASK_ZW),
brw_swizzle(tmp, 0, 1, 0, 1));
/* t_nopersp = vec4(v1.xy, dest.xy) - v0.xyxy */
brw_ADD(p, t_nopersp, t_nopersp,
negate(brw_swizzle(v0_ndc_copy, 0, 1, 0, 1)));
/* Add the absolute values of the X and Y deltas so that if
* the points aren't in the same place on the screen we get
* nonzero values to divide.
*
* After that, we have vert1 - vert0 in t_nopersp.x and
* vertnew - vert0 in t_nopersp.y
*
* t_nopersp = vec2(|v1.x -v0.x| + |v1.y -v0.y|,
* |dest.x-v0.x| + |dest.y-v0.y|)
*/
brw_ADD(p,
brw_writemask(t_nopersp, WRITEMASK_XY),
brw_abs(brw_swizzle(t_nopersp, 0, 2, 0, 0)),
brw_abs(brw_swizzle(t_nopersp, 1, 3, 0, 0)));
brw_set_default_access_mode(p, BRW_ALIGN_1);
/* If the points are in the same place, just substitute a
* value to avoid divide-by-zero
*/
brw_CMP(p, vec1(brw_null_reg()), BRW_CONDITIONAL_EQ,
vec1(t_nopersp),
brw_imm_f(0));
brw_IF(p, BRW_EXECUTE_1);
brw_MOV(p, t_nopersp, brw_imm_vf4(brw_float_to_vf(1.0),
brw_float_to_vf(0.0),
brw_float_to_vf(0.0),
brw_float_to_vf(0.0)));
brw_ENDIF(p);
/* Now compute t_nopersp = t_nopersp.y/t_nopersp.x and broadcast it. */
brw_math_invert(p, get_element(t_nopersp, 0), get_element(t_nopersp, 0));
brw_MUL(p, vec1(t_nopersp), vec1(t_nopersp),
vec1(suboffset(t_nopersp, 1)));
brw_set_default_access_mode(p, BRW_ALIGN_16);
brw_MOV(p, t_nopersp, brw_swizzle(t_nopersp, 0, 0, 0, 0));
brw_set_default_access_mode(p, BRW_ALIGN_1);
release_tmp(c, tmp);
release_tmp(c, v0_ndc_copy);
}
/* Now we can iterate over each attribute
* (could be done in pairs?)
*/
for (slot = 0; slot < c->vue_map.num_slots; slot++) {
int varying = c->vue_map.slot_to_varying[slot];
GLuint delta = brw_vue_slot_to_offset(slot);
/* HPOS, NDC already handled above */
if (varying == VARYING_SLOT_POS || varying == BRW_VARYING_SLOT_NDC)
continue;
if (varying == VARYING_SLOT_EDGE) {
if (force_edgeflag)
brw_MOV(p, deref_4f(dest_ptr, delta), brw_imm_f(1));
else
brw_MOV(p, deref_4f(dest_ptr, delta), deref_4f(v0_ptr, delta));
} else if (varying == VARYING_SLOT_PSIZ) {
/* PSIZ doesn't need interpolation because it isn't used by the
* fragment shader.
*/
} else if (varying < VARYING_SLOT_MAX) {
/* This is a true vertex result (and not a special value for the VUE
* header), so interpolate:
*
* New = attr0 + t*attr1 - t*attr0
*
* Unless the attribute is flat shaded -- in which case just copy
* from one of the sources (doesn't matter which; already copied from pv)
*/
GLuint interp = c->key.interpolation_mode.mode[slot];
if (interp != INTERP_QUALIFIER_FLAT) {
struct brw_reg tmp = get_tmp(c);
struct brw_reg t =
interp == INTERP_QUALIFIER_NOPERSPECTIVE ? t_nopersp : t0;
brw_MUL(p,
vec4(brw_null_reg()),
deref_4f(v1_ptr, delta),
t);
brw_MAC(p,
tmp,
negate(deref_4f(v0_ptr, delta)),
t);
brw_ADD(p,
deref_4f(dest_ptr, delta),
deref_4f(v0_ptr, delta),
tmp);
release_tmp(c, tmp);
}
else {
brw_MOV(p,
deref_4f(dest_ptr, delta),
deref_4f(v0_ptr, delta));
}
}
}
if (c->vue_map.num_slots % 2) {
GLuint delta = brw_vue_slot_to_offset(c->vue_map.num_slots);
brw_MOV(p, deref_4f(dest_ptr, delta), brw_imm_f(0));
}
if (c->has_noperspective_shading)
release_tmp(c, t_nopersp);
}
/**
* 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;
}
}
void
vec4_generator::generate_pull_constant_load_gen7(vec4_instruction *inst,
struct brw_reg dst,
struct brw_reg surf_index,
struct brw_reg offset)
{
assert(surf_index.type == BRW_REGISTER_TYPE_UD);
if (surf_index.file == BRW_IMMEDIATE_VALUE) {
brw_inst *insn = brw_next_insn(p, BRW_OPCODE_SEND);
brw_set_dest(p, insn, dst);
brw_set_src0(p, insn, offset);
brw_set_sampler_message(p, insn,
surf_index.dw1.ud,
0, /* LD message ignores sampler unit */
GEN5_SAMPLER_MESSAGE_SAMPLE_LD,
1, /* rlen */
1, /* mlen */
false, /* no header */
BRW_SAMPLER_SIMD_MODE_SIMD4X2,
0);
brw_mark_surface_used(&prog_data->base, surf_index.dw1.ud);
} else {
struct brw_reg addr = vec1(retype(brw_address_reg(0), BRW_REGISTER_TYPE_UD));
brw_push_insn_state(p);
brw_set_default_mask_control(p, BRW_MASK_DISABLE);
brw_set_default_access_mode(p, BRW_ALIGN_1);
/* a0.0 = surf_index & 0xff */
brw_inst *insn_and = brw_next_insn(p, BRW_OPCODE_AND);
brw_inst_set_exec_size(p->brw, insn_and, BRW_EXECUTE_1);
brw_set_dest(p, insn_and, addr);
brw_set_src0(p, insn_and, vec1(retype(surf_index, BRW_REGISTER_TYPE_UD)));
brw_set_src1(p, insn_and, brw_imm_ud(0x0ff));
/* a0.0 |= <descriptor> */
brw_inst *insn_or = brw_next_insn(p, BRW_OPCODE_OR);
brw_set_sampler_message(p, insn_or,
0 /* surface */,
0 /* sampler */,
GEN5_SAMPLER_MESSAGE_SAMPLE_LD,
1 /* rlen */,
1 /* mlen */,
false /* header */,
BRW_SAMPLER_SIMD_MODE_SIMD4X2,
0);
brw_inst_set_exec_size(p->brw, insn_or, BRW_EXECUTE_1);
brw_inst_set_src1_reg_type(p->brw, insn_or, BRW_REGISTER_TYPE_UD);
brw_set_src0(p, insn_or, addr);
brw_set_dest(p, insn_or, addr);
/* dst = send(offset, a0.0) */
brw_inst *insn_send = brw_next_insn(p, BRW_OPCODE_SEND);
brw_set_dest(p, insn_send, dst);
brw_set_src0(p, insn_send, offset);
brw_set_indirect_send_descriptor(p, insn_send, BRW_SFID_SAMPLER, addr);
brw_pop_insn_state(p);
/* visitor knows more than we do about the surface limit required,
* so has already done marking.
*/
}
}
void
vec4_generator::generate_tex(vec4_instruction *inst,
struct brw_reg dst,
struct brw_reg src)
{
int msg_type = -1;
if (brw->gen >= 5) {
switch (inst->opcode) {
case SHADER_OPCODE_TEX:
case SHADER_OPCODE_TXL:
if (inst->shadow_compare) {
msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LOD_COMPARE;
} else {
msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LOD;
}
break;
case SHADER_OPCODE_TXD:
if (inst->shadow_compare) {
/* Gen7.5+. Otherwise, lowered by brw_lower_texture_gradients(). */
assert(brw->gen >= 8 || brw->is_haswell);
msg_type = HSW_SAMPLER_MESSAGE_SAMPLE_DERIV_COMPARE;
} else {
msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_DERIVS;
}
break;
case SHADER_OPCODE_TXF:
msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LD;
break;
case SHADER_OPCODE_TXF_CMS:
if (brw->gen >= 7)
msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_LD2DMS;
else
msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LD;
break;
case SHADER_OPCODE_TXF_MCS:
assert(brw->gen >= 7);
msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_LD_MCS;
break;
case SHADER_OPCODE_TXS:
msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_RESINFO;
break;
case SHADER_OPCODE_TG4:
if (inst->shadow_compare) {
msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4_C;
} else {
msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4;
}
break;
case SHADER_OPCODE_TG4_OFFSET:
if (inst->shadow_compare) {
msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4_PO_C;
} else {
msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4_PO;
}
break;
default:
unreachable("should not get here: invalid vec4 texture opcode");
}
} else {
switch (inst->opcode) {
case SHADER_OPCODE_TEX:
case SHADER_OPCODE_TXL:
if (inst->shadow_compare) {
msg_type = BRW_SAMPLER_MESSAGE_SIMD4X2_SAMPLE_LOD_COMPARE;
assert(inst->mlen == 3);
} else {
msg_type = BRW_SAMPLER_MESSAGE_SIMD4X2_SAMPLE_LOD;
assert(inst->mlen == 2);
}
break;
case SHADER_OPCODE_TXD:
/* There is no sample_d_c message; comparisons are done manually. */
msg_type = BRW_SAMPLER_MESSAGE_SIMD4X2_SAMPLE_GRADIENTS;
assert(inst->mlen == 4);
break;
case SHADER_OPCODE_TXF:
msg_type = BRW_SAMPLER_MESSAGE_SIMD4X2_LD;
assert(inst->mlen == 2);
break;
case SHADER_OPCODE_TXS:
msg_type = BRW_SAMPLER_MESSAGE_SIMD4X2_RESINFO;
assert(inst->mlen == 2);
break;
default:
unreachable("should not get here: invalid vec4 texture opcode");
}
}
assert(msg_type != -1);
/* Load the message header if present. If there's a texture offset, we need
* to set it up explicitly and load the offset bitfield. Otherwise, we can
* use an implied move from g0 to the first message register.
*/
if (inst->header_present) {
if (brw->gen < 6 && !inst->texture_offset) {
/* Set up an implied move from g0 to the MRF. */
src = brw_vec8_grf(0, 0);
} else {
struct brw_reg header =
retype(brw_message_reg(inst->base_mrf), BRW_REGISTER_TYPE_UD);
/* Explicitly set up the message header by copying g0 to the MRF. */
brw_push_insn_state(p);
brw_set_default_mask_control(p, BRW_MASK_DISABLE);
brw_MOV(p, header, retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD));
brw_set_default_access_mode(p, BRW_ALIGN_1);
if (inst->texture_offset) {
/* Set the texel offset bits in DWord 2. */
brw_MOV(p, get_element_ud(header, 2),
brw_imm_ud(inst->texture_offset));
}
if (inst->sampler >= 16) {
/* The "Sampler Index" field can only store values between 0 and 15.
* However, we can add an offset to the "Sampler State Pointer"
* field, effectively selecting a different set of 16 samplers.
*
* The "Sampler State Pointer" needs to be aligned to a 32-byte
* offset, and each sampler state is only 16-bytes, so we can't
* exclusively use the offset - we have to use both.
*/
assert(brw->gen >= 8 || brw->is_haswell);
brw_ADD(p,
get_element_ud(header, 3),
get_element_ud(brw_vec8_grf(0, 0), 3),
brw_imm_ud(16 * (inst->sampler / 16) *
sizeof(gen7_sampler_state)));
}
brw_pop_insn_state(p);
}
}
uint32_t return_format;
switch (dst.type) {
case BRW_REGISTER_TYPE_D:
return_format = BRW_SAMPLER_RETURN_FORMAT_SINT32;
break;
case BRW_REGISTER_TYPE_UD:
return_format = BRW_SAMPLER_RETURN_FORMAT_UINT32;
break;
default:
return_format = BRW_SAMPLER_RETURN_FORMAT_FLOAT32;
break;
}
uint32_t surface_index = ((inst->opcode == SHADER_OPCODE_TG4 ||
inst->opcode == SHADER_OPCODE_TG4_OFFSET)
? prog_data->base.binding_table.gather_texture_start
: prog_data->base.binding_table.texture_start) + inst->sampler;
brw_SAMPLE(p,
dst,
inst->base_mrf,
src,
surface_index,
inst->sampler % 16,
msg_type,
1, /* response length */
inst->mlen,
inst->header_present,
BRW_SAMPLER_SIMD_MODE_SIMD4X2,
return_format);
brw_mark_surface_used(&prog_data->base, surface_index);
}
void
vec4_generator::generate_tex(vec4_instruction *inst,
struct brw_reg dst,
struct brw_reg src,
struct brw_reg sampler_index)
{
int msg_type = -1;
if (brw->gen >= 5) {
switch (inst->opcode) {
case SHADER_OPCODE_TEX:
case SHADER_OPCODE_TXL:
if (inst->shadow_compare) {
msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LOD_COMPARE;
} else {
msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LOD;
}
break;
case SHADER_OPCODE_TXD:
if (inst->shadow_compare) {
/* Gen7.5+. Otherwise, lowered by brw_lower_texture_gradients(). */
assert(brw->gen >= 8 || brw->is_haswell);
msg_type = HSW_SAMPLER_MESSAGE_SAMPLE_DERIV_COMPARE;
} else {
msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_DERIVS;
}
break;
case SHADER_OPCODE_TXF:
msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LD;
break;
case SHADER_OPCODE_TXF_CMS:
if (brw->gen >= 7)
msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_LD2DMS;
else
msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LD;
break;
case SHADER_OPCODE_TXF_MCS:
assert(brw->gen >= 7);
msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_LD_MCS;
break;
case SHADER_OPCODE_TXS:
msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_RESINFO;
break;
case SHADER_OPCODE_TG4:
if (inst->shadow_compare) {
msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4_C;
} else {
msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4;
}
break;
case SHADER_OPCODE_TG4_OFFSET:
if (inst->shadow_compare) {
msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4_PO_C;
} else {
msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4_PO;
}
break;
default:
unreachable("should not get here: invalid vec4 texture opcode");
}
} else {
switch (inst->opcode) {
case SHADER_OPCODE_TEX:
case SHADER_OPCODE_TXL:
if (inst->shadow_compare) {
msg_type = BRW_SAMPLER_MESSAGE_SIMD4X2_SAMPLE_LOD_COMPARE;
assert(inst->mlen == 3);
} else {
msg_type = BRW_SAMPLER_MESSAGE_SIMD4X2_SAMPLE_LOD;
assert(inst->mlen == 2);
}
break;
case SHADER_OPCODE_TXD:
/* There is no sample_d_c message; comparisons are done manually. */
msg_type = BRW_SAMPLER_MESSAGE_SIMD4X2_SAMPLE_GRADIENTS;
assert(inst->mlen == 4);
break;
case SHADER_OPCODE_TXF:
msg_type = BRW_SAMPLER_MESSAGE_SIMD4X2_LD;
assert(inst->mlen == 2);
break;
case SHADER_OPCODE_TXS:
msg_type = BRW_SAMPLER_MESSAGE_SIMD4X2_RESINFO;
assert(inst->mlen == 2);
break;
default:
unreachable("should not get here: invalid vec4 texture opcode");
}
}
assert(msg_type != -1);
assert(sampler_index.type == BRW_REGISTER_TYPE_UD);
/* Load the message header if present. If there's a texture offset, we need
* to set it up explicitly and load the offset bitfield. Otherwise, we can
* use an implied move from g0 to the first message register.
*/
if (inst->header_present) {
if (brw->gen < 6 && !inst->texture_offset) {
/* Set up an implied move from g0 to the MRF. */
src = brw_vec8_grf(0, 0);
} else {
struct brw_reg header =
retype(brw_message_reg(inst->base_mrf), BRW_REGISTER_TYPE_UD);
/* Explicitly set up the message header by copying g0 to the MRF. */
brw_push_insn_state(p);
brw_set_default_mask_control(p, BRW_MASK_DISABLE);
brw_MOV(p, header, retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD));
brw_set_default_access_mode(p, BRW_ALIGN_1);
if (inst->texture_offset) {
/* Set the texel offset bits in DWord 2. */
brw_MOV(p, get_element_ud(header, 2),
brw_imm_ud(inst->texture_offset));
}
brw_adjust_sampler_state_pointer(p, header, sampler_index, dst);
brw_pop_insn_state(p);
}
}
uint32_t return_format;
switch (dst.type) {
case BRW_REGISTER_TYPE_D:
return_format = BRW_SAMPLER_RETURN_FORMAT_SINT32;
break;
case BRW_REGISTER_TYPE_UD:
return_format = BRW_SAMPLER_RETURN_FORMAT_UINT32;
break;
default:
return_format = BRW_SAMPLER_RETURN_FORMAT_FLOAT32;
break;
}
uint32_t base_binding_table_index = (inst->opcode == SHADER_OPCODE_TG4 ||
inst->opcode == SHADER_OPCODE_TG4_OFFSET)
? prog_data->base.binding_table.gather_texture_start
: prog_data->base.binding_table.texture_start;
if (sampler_index.file == BRW_IMMEDIATE_VALUE) {
uint32_t sampler = sampler_index.dw1.ud;
brw_SAMPLE(p,
dst,
inst->base_mrf,
src,
sampler + base_binding_table_index,
sampler % 16,
msg_type,
1, /* response length */
inst->mlen,
inst->header_present,
BRW_SAMPLER_SIMD_MODE_SIMD4X2,
return_format);
brw_mark_surface_used(&prog_data->base, sampler + base_binding_table_index);
} else {
/* Non-constant sampler index. */
/* Note: this clobbers `dst` as a temporary before emitting the send */
struct brw_reg addr = vec1(retype(brw_address_reg(0), BRW_REGISTER_TYPE_UD));
struct brw_reg temp = vec1(retype(dst, BRW_REGISTER_TYPE_UD));
struct brw_reg sampler_reg = vec1(retype(sampler_index, BRW_REGISTER_TYPE_UD));
brw_push_insn_state(p);
brw_set_default_mask_control(p, BRW_MASK_DISABLE);
brw_set_default_access_mode(p, BRW_ALIGN_1);
/* Some care required: `sampler` and `temp` may alias:
* addr = sampler & 0xff
* temp = (sampler << 8) & 0xf00
* addr = addr | temp
*/
brw_ADD(p, addr, sampler_reg, brw_imm_ud(base_binding_table_index));
brw_SHL(p, temp, sampler_reg, brw_imm_ud(8u));
brw_AND(p, temp, temp, brw_imm_ud(0x0f00));
brw_AND(p, addr, addr, brw_imm_ud(0x0ff));
brw_OR(p, addr, addr, temp);
/* a0.0 |= <descriptor> */
brw_inst *insn_or = brw_next_insn(p, BRW_OPCODE_OR);
brw_set_sampler_message(p, insn_or,
0 /* surface */,
0 /* sampler */,
msg_type,
1 /* rlen */,
inst->mlen /* mlen */,
inst->header_present /* header */,
BRW_SAMPLER_SIMD_MODE_SIMD4X2,
return_format);
brw_inst_set_exec_size(p->brw, insn_or, BRW_EXECUTE_1);
brw_inst_set_src1_reg_type(p->brw, insn_or, BRW_REGISTER_TYPE_UD);
brw_set_src0(p, insn_or, addr);
brw_set_dest(p, insn_or, addr);
/* dst = send(offset, a0.0) */
brw_inst *insn_send = brw_next_insn(p, BRW_OPCODE_SEND);
brw_set_dest(p, insn_send, dst);
brw_set_src0(p, insn_send, src);
brw_set_indirect_send_descriptor(p, insn_send, BRW_SFID_SAMPLER, addr);
brw_pop_insn_state(p);
/* visitor knows more than we do about the surface limit required,
* so has already done marking.
*/
}
}
void brw_emit_point_sprite_setup(struct brw_sf_compile *c, bool allocate)
{
struct brw_compile *p = &c->func;
GLuint i;
c->flag_value = 0xff;
c->nr_verts = 1;
if (allocate)
alloc_regs(c);
copy_z_inv_w(c);
for (i = 0; i < c->nr_setup_regs; i++)
{
struct brw_reg a0 = offset(c->vert[0], i);
GLushort pc, pc_persp, pc_linear, pc_coord_replace;
bool last = calculate_masks(c, i, &pc, &pc_persp, &pc_linear);
pc_coord_replace = calculate_point_sprite_mask(c, i);
pc_persp &= ~pc_coord_replace;
if (pc_persp) {
set_predicate_control_flag_value(p, c, pc_persp);
brw_MUL(p, a0, a0, c->inv_w[0]);
}
/* Point sprite coordinate replacement: A texcoord with this
* enabled gets replaced with the value (x, y, 0, 1) where x and
* y vary from 0 to 1 across the horizontal and vertical of the
* point.
*/
if (pc_coord_replace) {
set_predicate_control_flag_value(p, c, pc_coord_replace);
/* Caculate 1.0/PointWidth */
gen4_math(&c->func,
c->tmp,
BRW_MATH_FUNCTION_INV,
0,
c->dx0,
BRW_MATH_DATA_SCALAR,
BRW_MATH_PRECISION_FULL);
brw_set_default_access_mode(p, BRW_ALIGN_16);
/* dA/dx, dA/dy */
brw_MOV(p, c->m1Cx, brw_imm_f(0.0));
brw_MOV(p, c->m2Cy, brw_imm_f(0.0));
brw_MOV(p, brw_writemask(c->m1Cx, WRITEMASK_X), c->tmp);
if (c->key.sprite_origin_lower_left) {
brw_MOV(p, brw_writemask(c->m2Cy, WRITEMASK_Y), negate(c->tmp));
} else {
brw_MOV(p, brw_writemask(c->m2Cy, WRITEMASK_Y), c->tmp);
}
/* attribute constant offset */
brw_MOV(p, c->m3C0, brw_imm_f(0.0));
if (c->key.sprite_origin_lower_left) {
brw_MOV(p, brw_writemask(c->m3C0, WRITEMASK_YW), brw_imm_f(1.0));
} else {
brw_MOV(p, brw_writemask(c->m3C0, WRITEMASK_W), brw_imm_f(1.0));
}
brw_set_default_access_mode(p, BRW_ALIGN_1);
}
if (pc & ~pc_coord_replace) {
set_predicate_control_flag_value(p, c, pc & ~pc_coord_replace);
brw_MOV(p, c->m1Cx, brw_imm_ud(0));
brw_MOV(p, c->m2Cy, brw_imm_ud(0));
brw_MOV(p, c->m3C0, a0); /* constant value */
}
set_predicate_control_flag_value(p, c, pc);
/* Copy m0..m3 to URB. */
brw_urb_WRITE(p,
brw_null_reg(),
0,
brw_vec8_grf(0, 0),
last ? BRW_URB_WRITE_EOT_COMPLETE
: BRW_URB_WRITE_NO_FLAGS,
4, /* msg len */
0, /* response len */
i*4, /* urb destination offset */
BRW_URB_SWIZZLE_TRANSPOSE);
}
brw_set_default_predicate_control(p, BRW_PREDICATE_NONE);
}
void
vec4_generator::generate_gs_set_channel_masks(struct brw_reg dst,
struct brw_reg src)
{
/* From p21 of volume 4 part 2 of the Ivy Bridge PRM (2.4.3.1 Message
* Header: M0.5):
*
* 15 Vertex 1 DATA [3] / Vertex 0 DATA[7] Channel Mask
*
* When Swizzle Control = URB_INTERLEAVED this bit controls Vertex 1
* DATA[3], when Swizzle Control = URB_NOSWIZZLE this bit controls
* Vertex 0 DATA[7]. This bit is ANDed with the corresponding
* channel enable to determine the final channel enable. For the
* URB_READ_OWORD & URB_READ_HWORD messages, when final channel
* enable is 1 it indicates that Vertex 1 DATA [3] will be included
* in the writeback message. For the URB_WRITE_OWORD &
* URB_WRITE_HWORD messages, when final channel enable is 1 it
* indicates that Vertex 1 DATA [3] will be written to the surface.
*
* 0: Vertex 1 DATA [3] / Vertex 0 DATA[7] channel not included
* 1: Vertex DATA [3] / Vertex 0 DATA[7] channel included
*
* 14 Vertex 1 DATA [2] Channel Mask
* 13 Vertex 1 DATA [1] Channel Mask
* 12 Vertex 1 DATA [0] Channel Mask
* 11 Vertex 0 DATA [3] Channel Mask
* 10 Vertex 0 DATA [2] Channel Mask
* 9 Vertex 0 DATA [1] Channel Mask
* 8 Vertex 0 DATA [0] Channel Mask
*
* (This is from a section of the PRM that is agnostic to the particular
* type of shader being executed, so "Vertex 0" and "Vertex 1" refer to
* geometry shader invocations 0 and 1, respectively). Since we have the
* enable flags for geometry shader invocation 0 in bits 3:0 of DWORD 0,
* and the enable flags for geometry shader invocation 1 in bits 7:0 of
* DWORD 4, we just need to OR them together and store the result in bits
* 15:8 of DWORD 5.
*
* It's easier to get the EU to do this if we think of the src and dst
* registers as composed of 32 bytes each; then, we want to pick up the
* contents of bytes 0 and 16 from src, OR them together, and store them in
* byte 21.
*
* We can do that by the following EU instruction:
*
* or(1) dst.21<1>UB src<0,1,0>UB src.16<0,1,0>UB { align1 WE_all }
*
* Note: this relies on the source register having zeros in (a) bits 7:4 of
* DWORD 0 and (b) bits 3:0 of DWORD 4. We can rely on (b) because the
* source register was prepared by GS_OPCODE_PREPARE_CHANNEL_MASKS (which
* shifts DWORD 4 left by 4 bits), and we can rely on (a) because prior to
* the execution of GS_OPCODE_PREPARE_CHANNEL_MASKS, DWORDs 0 and 4 need to
* contain valid channel mask values (which are in the range 0x0-0xf).
*/
dst = retype(dst, BRW_REGISTER_TYPE_UB);
src = retype(src, BRW_REGISTER_TYPE_UB);
brw_push_insn_state(p);
brw_set_default_access_mode(p, BRW_ALIGN_1);
brw_set_default_mask_control(p, BRW_MASK_DISABLE);
brw_OR(p, suboffset(vec1(dst), 21), vec1(src), suboffset(vec1(src), 16));
brw_pop_insn_state(p);
}