void
vec4_tcs_visitor::emit_thread_end()
{
vec4_instruction *inst;
current_annotation = "thread end";
if (nir->info->tcs.vertices_out % 2) {
emit(BRW_OPCODE_ENDIF);
}
if (devinfo->gen == 7) {
struct brw_tcs_prog_data *tcs_prog_data =
(struct brw_tcs_prog_data *) prog_data;
current_annotation = "release input vertices";
/* Synchronize all threads, so we know that no one is still
* using the input URB handles.
*/
if (tcs_prog_data->instances > 1) {
dst_reg header = dst_reg(this, glsl_type::uvec4_type);
emit(TCS_OPCODE_CREATE_BARRIER_HEADER, header);
emit(SHADER_OPCODE_BARRIER, dst_null_ud(), src_reg(header));
}
/* Make thread 0 (invocations <1, 0>) release pairs of ICP handles.
* We want to compare the bottom half of invocation_id with 0, but
* use that truth value for the top half as well. Unfortunately,
* we don't have stride in the vec4 world, nor UV immediates in
* align16, so we need an opcode to get invocation_id<0,4,0>.
*/
set_condmod(BRW_CONDITIONAL_Z,
emit(TCS_OPCODE_SRC0_010_IS_ZERO, dst_null_d(),
invocation_id));
emit(IF(BRW_PREDICATE_NORMAL));
for (unsigned i = 0; i < key->input_vertices; i += 2) {
/* If we have an odd number of input vertices, the last will be
* unpaired. We don't want to use an interleaved URB write in
* that case.
*/
const bool is_unpaired = i == key->input_vertices - 1;
dst_reg header(this, glsl_type::uvec4_type);
emit(TCS_OPCODE_RELEASE_INPUT, header, brw_imm_ud(i),
brw_imm_ud(is_unpaired));
}
emit(BRW_OPCODE_ENDIF);
}
if (unlikely(INTEL_DEBUG & DEBUG_SHADER_TIME))
emit_shader_time_end();
inst = emit(TCS_OPCODE_THREAD_END);
inst->base_mrf = 14;
inst->mlen = 2;
}
void
vec4_tcs_visitor::nir_emit_intrinsic(nir_intrinsic_instr *instr)
{
switch (instr->intrinsic) {
case nir_intrinsic_load_invocation_id:
emit(MOV(get_nir_dest(instr->dest, BRW_REGISTER_TYPE_UD),
invocation_id));
break;
case nir_intrinsic_load_primitive_id:
emit(TCS_OPCODE_GET_PRIMITIVE_ID,
get_nir_dest(instr->dest, BRW_REGISTER_TYPE_UD));
break;
case nir_intrinsic_load_patch_vertices_in:
emit(MOV(get_nir_dest(instr->dest, BRW_REGISTER_TYPE_D),
brw_imm_d(key->input_vertices)));
break;
case nir_intrinsic_load_per_vertex_input: {
src_reg indirect_offset = get_indirect_offset(instr);
unsigned imm_offset = instr->const_index[0];
nir_const_value *vertex_const = nir_src_as_const_value(instr->src[0]);
src_reg vertex_index =
vertex_const ? src_reg(brw_imm_ud(vertex_const->u32[0]))
: get_nir_src(instr->src[0], BRW_REGISTER_TYPE_UD, 1);
dst_reg dst = get_nir_dest(instr->dest, BRW_REGISTER_TYPE_D);
dst.writemask = brw_writemask_for_size(instr->num_components);
emit_input_urb_read(dst, vertex_index, imm_offset,
nir_intrinsic_component(instr), indirect_offset);
break;
}
case nir_intrinsic_load_input:
unreachable("nir_lower_io should use load_per_vertex_input intrinsics");
break;
case nir_intrinsic_load_output:
case nir_intrinsic_load_per_vertex_output: {
src_reg indirect_offset = get_indirect_offset(instr);
unsigned imm_offset = instr->const_index[0];
dst_reg dst = get_nir_dest(instr->dest, BRW_REGISTER_TYPE_D);
dst.writemask = brw_writemask_for_size(instr->num_components);
if (imm_offset == 0 && indirect_offset.file == BAD_FILE) {
dst.type = BRW_REGISTER_TYPE_F;
/* This is a read of gl_TessLevelInner[], which lives in the
* Patch URB header. The layout depends on the domain.
*/
switch (key->tes_primitive_mode) {
case GL_QUADS: {
/* DWords 3-2 (reversed); use offset 0 and WZYX swizzle. */
dst_reg tmp(this, glsl_type::vec4_type);
emit_output_urb_read(tmp, 0, 0, src_reg());
emit(MOV(writemask(dst, WRITEMASK_XY),
swizzle(src_reg(tmp), BRW_SWIZZLE_WZYX)));
break;
}
case GL_TRIANGLES:
/* DWord 4; use offset 1 but normal swizzle/writemask. */
emit_output_urb_read(writemask(dst, WRITEMASK_X), 1, 0,
src_reg());
break;
case GL_ISOLINES:
/* All channels are undefined. */
return;
default:
unreachable("Bogus tessellation domain");
}
} else if (imm_offset == 1 && indirect_offset.file == BAD_FILE) {
dst.type = BRW_REGISTER_TYPE_F;
unsigned swiz = BRW_SWIZZLE_WZYX;
/* This is a read of gl_TessLevelOuter[], which lives in the
* high 4 DWords of the Patch URB header, in reverse order.
*/
switch (key->tes_primitive_mode) {
case GL_QUADS:
dst.writemask = WRITEMASK_XYZW;
break;
case GL_TRIANGLES:
dst.writemask = WRITEMASK_XYZ;
break;
case GL_ISOLINES:
/* Isolines are not reversed; swizzle .zw -> .xy */
swiz = BRW_SWIZZLE_ZWZW;
dst.writemask = WRITEMASK_XY;
return;
default:
unreachable("Bogus tessellation domain");
}
dst_reg tmp(this, glsl_type::vec4_type);
emit_output_urb_read(tmp, 1, 0, src_reg());
emit(MOV(dst, swizzle(src_reg(tmp), swiz)));
} else {
emit_output_urb_read(dst, imm_offset, nir_intrinsic_component(instr),
indirect_offset);
}
break;
}
case nir_intrinsic_store_output:
case nir_intrinsic_store_per_vertex_output: {
src_reg value = get_nir_src(instr->src[0]);
unsigned mask = instr->const_index[1];
unsigned swiz = BRW_SWIZZLE_XYZW;
src_reg indirect_offset = get_indirect_offset(instr);
unsigned imm_offset = instr->const_index[0];
/* The passthrough shader writes the whole patch header as two vec4s;
* skip all the gl_TessLevelInner/Outer swizzling.
*/
if (indirect_offset.file == BAD_FILE && !is_passthrough_shader) {
if (imm_offset == 0) {
value.type = BRW_REGISTER_TYPE_F;
mask &=
(1 << tesslevel_inner_components(key->tes_primitive_mode)) - 1;
/* This is a write to gl_TessLevelInner[], which lives in the
* Patch URB header. The layout depends on the domain.
*/
switch (key->tes_primitive_mode) {
case GL_QUADS:
/* gl_TessLevelInner[].xy lives at DWords 3-2 (reversed).
* We use an XXYX swizzle to reverse put .xy in the .wz
* channels, and use a .zw writemask.
*/
swiz = BRW_SWIZZLE4(0, 0, 1, 0);
mask = writemask_for_backwards_vector(mask);
break;
case GL_TRIANGLES:
/* gl_TessLevelInner[].x lives at DWord 4, so we set the
* writemask to X and bump the URB offset by 1.
*/
imm_offset = 1;
break;
case GL_ISOLINES:
/* Skip; gl_TessLevelInner[] doesn't exist for isolines. */
return;
default:
unreachable("Bogus tessellation domain");
}
} else if (imm_offset == 1) {
value.type = BRW_REGISTER_TYPE_F;
mask &=
(1 << tesslevel_outer_components(key->tes_primitive_mode)) - 1;
/* This is a write to gl_TessLevelOuter[] which lives in the
* Patch URB Header at DWords 4-7. However, it's reversed, so
* instead of .xyzw we have .wzyx.
*/
if (key->tes_primitive_mode == GL_ISOLINES) {
/* Isolines .xy should be stored in .zw, in order. */
swiz = BRW_SWIZZLE4(0, 0, 0, 1);
mask <<= 2;
} else {
/* Other domains are reversed; store .wzyx instead of .xyzw. */
swiz = BRW_SWIZZLE_WZYX;
mask = writemask_for_backwards_vector(mask);
}
}
}
unsigned first_component = nir_intrinsic_component(instr);
if (first_component) {
assert(swiz == BRW_SWIZZLE_XYZW);
swiz = BRW_SWZ_COMP_OUTPUT(first_component);
mask = mask << first_component;
}
emit_urb_write(swizzle(value, swiz), mask,
imm_offset, indirect_offset);
break;
}
case nir_intrinsic_barrier: {
dst_reg header = dst_reg(this, glsl_type::uvec4_type);
emit(TCS_OPCODE_CREATE_BARRIER_HEADER, header);
emit(SHADER_OPCODE_BARRIER, dst_null_ud(), src_reg(header));
break;
}
default:
vec4_visitor::nir_emit_intrinsic(instr);
}
}
void
vec4_gs_visitor::gs_emit_vertex(int stream_id)
{
this->current_annotation = "emit vertex: safety check";
/* Haswell and later hardware ignores the "Render Stream Select" bits
* from the 3DSTATE_STREAMOUT packet when the SOL stage is disabled,
* and instead sends all primitives down the pipeline for rasterization.
* If the SOL stage is enabled, "Render Stream Select" is honored and
* primitives bound to non-zero streams are discarded after stream output.
*
* Since the only purpose of primives sent to non-zero streams is to
* be recorded by transform feedback, we can simply discard all geometry
* bound to these streams when transform feedback is disabled.
*/
if (stream_id > 0 && !nir->info->has_transform_feedback_varyings)
return;
/* If we're outputting 32 control data bits or less, then we can wait
* until the shader is over to output them all. Otherwise we need to
* output them as we go. Now is the time to do it, since we're about to
* output the vertex_count'th vertex, so it's guaranteed that the
* control data bits associated with the (vertex_count - 1)th vertex are
* correct.
*/
if (c->control_data_header_size_bits > 32) {
this->current_annotation = "emit vertex: emit control data bits";
/* Only emit control data bits if we've finished accumulating a batch
* of 32 bits. This is the case when:
*
* (vertex_count * bits_per_vertex) % 32 == 0
*
* (in other words, when the last 5 bits of vertex_count *
* bits_per_vertex are 0). Assuming bits_per_vertex == 2^n for some
* integer n (which is always the case, since bits_per_vertex is
* always 1 or 2), this is equivalent to requiring that the last 5-n
* bits of vertex_count are 0:
*
* vertex_count & (2^(5-n) - 1) == 0
*
* 2^(5-n) == 2^5 / 2^n == 32 / bits_per_vertex, so this is
* equivalent to:
*
* vertex_count & (32 / bits_per_vertex - 1) == 0
*/
vec4_instruction *inst =
emit(AND(dst_null_ud(), this->vertex_count,
brw_imm_ud(32 / c->control_data_bits_per_vertex - 1)));
inst->conditional_mod = BRW_CONDITIONAL_Z;
emit(IF(BRW_PREDICATE_NORMAL));
{
/* If vertex_count is 0, then no control data bits have been
* accumulated yet, so we skip emitting them.
*/
emit(CMP(dst_null_ud(), this->vertex_count, brw_imm_ud(0u),
BRW_CONDITIONAL_NEQ));
emit(IF(BRW_PREDICATE_NORMAL));
emit_control_data_bits();
emit(BRW_OPCODE_ENDIF);
/* Reset control_data_bits to 0 so we can start accumulating a new
* batch.
*
* Note: in the case where vertex_count == 0, this neutralizes the
* effect of any call to EndPrimitive() that the shader may have
* made before outputting its first vertex.
*/
inst = emit(MOV(dst_reg(this->control_data_bits), brw_imm_ud(0u)));
inst->force_writemask_all = true;
}
emit(BRW_OPCODE_ENDIF);
}
this->current_annotation = "emit vertex: vertex data";
emit_vertex();
/* In stream mode we have to set control data bits for all vertices
* unless we have disabled control data bits completely (which we do
* do for GL_POINTS outputs that don't use streams).
*/
if (c->control_data_header_size_bits > 0 &&
gs_prog_data->control_data_format ==
GEN7_GS_CONTROL_DATA_FORMAT_GSCTL_SID) {
this->current_annotation = "emit vertex: Stream control data bits";
set_stream_control_data_bits(stream_id);
}
this->current_annotation = NULL;
}