void
gen6_gs_visitor::xfb_setup()
{
static const unsigned swizzle_for_offset[4] = {
BRW_SWIZZLE4(0, 1, 2, 3),
BRW_SWIZZLE4(1, 2, 3, 3),
BRW_SWIZZLE4(2, 3, 3, 3),
BRW_SWIZZLE4(3, 3, 3, 3)
};
struct brw_gs_prog_data *prog_data =
(struct brw_gs_prog_data *) &c->prog_data;
const struct gl_transform_feedback_info *linked_xfb_info =
&this->shader_prog->LinkedTransformFeedback;
int i;
/* Make sure that the VUE slots won't overflow the unsigned chars in
* prog_data->transform_feedback_bindings[].
*/
STATIC_ASSERT(BRW_VARYING_SLOT_COUNT <= 256);
/* Make sure that we don't need more binding table entries than we've
* set aside for use in transform feedback. (We shouldn't, since we
* set aside enough binding table entries to have one per component).
*/
assert(linked_xfb_info->NumOutputs <= BRW_MAX_SOL_BINDINGS);
prog_data->num_transform_feedback_bindings = linked_xfb_info->NumOutputs;
for (i = 0; i < prog_data->num_transform_feedback_bindings; i++) {
prog_data->transform_feedback_bindings[i] =
linked_xfb_info->Outputs[i].OutputRegister;
prog_data->transform_feedback_swizzles[i] =
swizzle_for_offset[linked_xfb_info->Outputs[i].ComponentOffset];
}
}
static bool
try_copy_propagate(const struct brw_device_info *devinfo,
vec4_instruction *inst,
int arg, struct copy_entry *entry)
{
/* For constant propagation, we only handle the same constant
* across all 4 channels. Some day, we should handle the 8-bit
* float vector format, which would let us constant propagate
* vectors better.
*/
src_reg value = *entry->value[0];
for (int i = 1; i < 4; i++) {
/* This is equals() except we don't care about the swizzle. */
if (value.file != entry->value[i]->file ||
value.reg != entry->value[i]->reg ||
value.reg_offset != entry->value[i]->reg_offset ||
value.type != entry->value[i]->type ||
value.negate != entry->value[i]->negate ||
value.abs != entry->value[i]->abs) {
return false;
}
}
/* Compute the swizzle of the original register by swizzling the
* component loaded from each value according to the swizzle of
* operand we're going to change.
*/
int s[4];
for (int i = 0; i < 4; i++) {
s[i] = BRW_GET_SWZ(entry->value[i]->swizzle, i);
}
value.swizzle = brw_compose_swizzle(inst->src[arg].swizzle,
BRW_SWIZZLE4(s[0], s[1], s[2], s[3]));
if (value.file != UNIFORM &&
value.file != GRF &&
value.file != ATTR)
return false;
if (devinfo->gen >= 8 && (value.negate || value.abs) &&
is_logic_op(inst->opcode)) {
return false;
}
if (inst->src[arg].abs) {
value.negate = false;
value.abs = true;
}
if (inst->src[arg].negate)
value.negate = !value.negate;
bool has_source_modifiers = value.negate || value.abs;
/* gen6 math and gen7+ SENDs from GRFs ignore source modifiers on
* instructions.
*/
if ((has_source_modifiers || value.file == UNIFORM ||
value.swizzle != BRW_SWIZZLE_XYZW) && !inst->can_do_source_mods(devinfo))
return false;
if (has_source_modifiers && value.type != inst->src[arg].type)
return false;
if (has_source_modifiers &&
inst->opcode == SHADER_OPCODE_GEN4_SCRATCH_WRITE)
return false;
if (inst->is_3src() && value.file == UNIFORM)
return false;
if (inst->is_send_from_grf())
return false;
/* we can't generally copy-propagate UD negations becuse we
* end up accessing the resulting values as signed integers
* instead. See also resolve_ud_negate().
*/
if (value.negate &&
value.type == BRW_REGISTER_TYPE_UD)
return false;
/* Don't report progress if this is a noop. */
if (value.equals(inst->src[arg]))
return false;
const unsigned dst_saturate_mask = inst->dst.writemask &
brw_apply_swizzle_to_mask(inst->src[arg].swizzle, entry->saturatemask);
if (dst_saturate_mask) {
/* We either saturate all or nothing. */
if (dst_saturate_mask != inst->dst.writemask)
return false;
/* Limit saturate propagation only to SEL with src1 bounded within 0.0
* and 1.0, otherwise skip copy propagate altogether.
*/
switch(inst->opcode) {
case BRW_OPCODE_SEL:
if (arg != 0 ||
inst->src[0].type != BRW_REGISTER_TYPE_F ||
inst->src[1].file != IMM ||
inst->src[1].type != BRW_REGISTER_TYPE_F ||
inst->src[1].fixed_hw_reg.dw1.f < 0.0 ||
inst->src[1].fixed_hw_reg.dw1.f > 1.0) {
return false;
}
if (!inst->saturate)
inst->saturate = true;
break;
default:
return false;
}
}
value.type = inst->src[arg].type;
inst->src[arg] = value;
return true;
}
static void populate_key(struct brw_context *brw,
struct brw_ff_gs_prog_key *key)
{
static const unsigned swizzle_for_offset[4] = {
BRW_SWIZZLE4(0, 1, 2, 3),
BRW_SWIZZLE4(1, 2, 3, 3),
BRW_SWIZZLE4(2, 3, 3, 3),
BRW_SWIZZLE4(3, 3, 3, 3)
};
struct gl_context *ctx = &brw->ctx;
memset(key, 0, sizeof(*key));
/* CACHE_NEW_VS_PROG (part of VUE map) */
key->attrs = brw->vs.prog_data->base.vue_map.slots_valid;
/* BRW_NEW_PRIMITIVE */
key->primitive = brw->primitive;
/* _NEW_LIGHT */
key->pv_first = (ctx->Light.ProvokingVertex == GL_FIRST_VERTEX_CONVENTION);
if (key->primitive == _3DPRIM_QUADLIST && ctx->Light.ShadeModel != GL_FLAT) {
/* Provide consistent primitive order with brw_set_prim's
* optimization of single quads to trifans.
*/
key->pv_first = true;
}
if (brw->gen >= 7) {
/* On Gen7 and later, we don't use GS (yet). */
key->need_gs_prog = false;
} else if (brw->gen == 6) {
/* On Gen6, GS is used for transform feedback. */
/* BRW_NEW_TRANSFORM_FEEDBACK */
if (_mesa_is_xfb_active_and_unpaused(ctx)) {
const struct gl_shader_program *shaderprog =
ctx->Shader.CurrentProgram[MESA_SHADER_VERTEX];
const struct gl_transform_feedback_info *linked_xfb_info =
&shaderprog->LinkedTransformFeedback;
int i;
/* Make sure that the VUE slots won't overflow the unsigned chars in
* key->transform_feedback_bindings[].
*/
STATIC_ASSERT(BRW_VARYING_SLOT_COUNT <= 256);
/* Make sure that we don't need more binding table entries than we've
* set aside for use in transform feedback. (We shouldn't, since we
* set aside enough binding table entries to have one per component).
*/
assert(linked_xfb_info->NumOutputs <= BRW_MAX_SOL_BINDINGS);
key->need_gs_prog = true;
key->num_transform_feedback_bindings = linked_xfb_info->NumOutputs;
for (i = 0; i < key->num_transform_feedback_bindings; ++i) {
key->transform_feedback_bindings[i] =
linked_xfb_info->Outputs[i].OutputRegister;
key->transform_feedback_swizzles[i] =
swizzle_for_offset[linked_xfb_info->Outputs[i].ComponentOffset];
}
}
} else {
/* Pre-gen6, GS is used to transform QUADLIST, QUADSTRIP, and LINELOOP
* into simpler primitives.
*/
key->need_gs_prog = (brw->primitive == _3DPRIM_QUADLIST ||
brw->primitive == _3DPRIM_QUADSTRIP ||
brw->primitive == _3DPRIM_LINELOOP);
}
}
void
vec4_vs_visitor::emit_prolog()
{
dst_reg sign_recovery_shift;
dst_reg normalize_factor;
dst_reg es3_normalize_factor;
for (int i = 0; i < VERT_ATTRIB_MAX; i++) {
if (vs_prog_data->inputs_read & BITFIELD64_BIT(i)) {
uint8_t wa_flags = key->gl_attrib_wa_flags[i];
dst_reg reg(ATTR, i);
dst_reg reg_d = reg;
reg_d.type = BRW_REGISTER_TYPE_D;
dst_reg reg_ud = reg;
reg_ud.type = BRW_REGISTER_TYPE_UD;
/* Do GL_FIXED rescaling for GLES2.0. Our GL_FIXED attributes
* come in as floating point conversions of the integer values.
*/
if (wa_flags & BRW_ATTRIB_WA_COMPONENT_MASK) {
dst_reg dst = reg;
dst.type = brw_type_for_base_type(glsl_type::vec4_type);
dst.writemask = (1 << (wa_flags & BRW_ATTRIB_WA_COMPONENT_MASK)) - 1;
emit(MUL(dst, src_reg(dst), src_reg(1.0f / 65536.0f)));
}
/* Do sign recovery for 2101010 formats if required. */
if (wa_flags & BRW_ATTRIB_WA_SIGN) {
if (sign_recovery_shift.file == BAD_FILE) {
/* shift constant: <22,22,22,30> */
sign_recovery_shift = dst_reg(this, glsl_type::uvec4_type);
emit(MOV(writemask(sign_recovery_shift, WRITEMASK_XYZ), src_reg(22u)));
emit(MOV(writemask(sign_recovery_shift, WRITEMASK_W), src_reg(30u)));
}
emit(SHL(reg_ud, src_reg(reg_ud), src_reg(sign_recovery_shift)));
emit(ASR(reg_d, src_reg(reg_d), src_reg(sign_recovery_shift)));
}
/* Apply BGRA swizzle if required. */
if (wa_flags & BRW_ATTRIB_WA_BGRA) {
src_reg temp = src_reg(reg);
temp.swizzle = BRW_SWIZZLE4(2,1,0,3);
emit(MOV(reg, temp));
}
if (wa_flags & BRW_ATTRIB_WA_NORMALIZE) {
/* ES 3.0 has different rules for converting signed normalized
* fixed-point numbers than desktop GL.
*/
if ((wa_flags & BRW_ATTRIB_WA_SIGN) && !use_legacy_snorm_formula) {
/* According to equation 2.2 of the ES 3.0 specification,
* signed normalization conversion is done by:
*
* f = c / (2^(b-1)-1)
*/
if (es3_normalize_factor.file == BAD_FILE) {
/* mul constant: 1 / (2^(b-1) - 1) */
es3_normalize_factor = dst_reg(this, glsl_type::vec4_type);
emit(MOV(writemask(es3_normalize_factor, WRITEMASK_XYZ),
src_reg(1.0f / ((1<<9) - 1))));
emit(MOV(writemask(es3_normalize_factor, WRITEMASK_W),
src_reg(1.0f / ((1<<1) - 1))));
}
dst_reg dst = reg;
dst.type = brw_type_for_base_type(glsl_type::vec4_type);
emit(MOV(dst, src_reg(reg_d)));
emit(MUL(dst, src_reg(dst), src_reg(es3_normalize_factor)));
emit_minmax(BRW_CONDITIONAL_GE, dst, src_reg(dst), src_reg(-1.0f));
} else {
/* The following equations are from the OpenGL 3.2 specification:
*
* 2.1 unsigned normalization
* f = c/(2^n-1)
*
* 2.2 signed normalization
* f = (2c+1)/(2^n-1)
*
* Both of these share a common divisor, which is represented by
* "normalize_factor" in the code below.
*/
if (normalize_factor.file == BAD_FILE) {
/* 1 / (2^b - 1) for b=<10,10,10,2> */
normalize_factor = dst_reg(this, glsl_type::vec4_type);
emit(MOV(writemask(normalize_factor, WRITEMASK_XYZ),
src_reg(1.0f / ((1<<10) - 1))));
emit(MOV(writemask(normalize_factor, WRITEMASK_W),
src_reg(1.0f / ((1<<2) - 1))));
}
dst_reg dst = reg;
dst.type = brw_type_for_base_type(glsl_type::vec4_type);
emit(MOV(dst, src_reg((wa_flags & BRW_ATTRIB_WA_SIGN) ? reg_d : reg_ud)));
/* For signed normalization, we want the numerator to be 2c+1. */
if (wa_flags & BRW_ATTRIB_WA_SIGN) {
emit(MUL(dst, src_reg(dst), src_reg(2.0f)));
emit(ADD(dst, src_reg(dst), src_reg(1.0f)));
}
emit(MUL(dst, src_reg(dst), src_reg(normalize_factor)));
}
}
if (wa_flags & BRW_ATTRIB_WA_SCALE) {
dst_reg dst = reg;
dst.type = brw_type_for_base_type(glsl_type::vec4_type);
emit(MOV(dst, src_reg((wa_flags & BRW_ATTRIB_WA_SIGN) ? reg_d : reg_ud)));
}
}
}
}
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);
}
}
static void populate_key( struct brw_context *brw,
struct brw_gs_prog_key *key )
{
static const unsigned swizzle_for_offset[4] = {
BRW_SWIZZLE4(0, 1, 2, 3),
BRW_SWIZZLE4(1, 2, 3, 3),
BRW_SWIZZLE4(2, 3, 3, 3),
BRW_SWIZZLE4(3, 3, 3, 3)
};
struct gl_context *ctx = &brw->intel.ctx;
struct intel_context *intel = &brw->intel;
memset(key, 0, sizeof(*key));
/* CACHE_NEW_VS_PROG */
key->attrs = brw->vs.prog_data->outputs_written;
/* BRW_NEW_PRIMITIVE */
key->primitive = brw->primitive;
/* _NEW_LIGHT */
key->pv_first = (ctx->Light.ProvokingVertex == GL_FIRST_VERTEX_CONVENTION);
if (key->primitive == _3DPRIM_QUADLIST && ctx->Light.ShadeModel != GL_FLAT) {
/* Provide consistent primitive order with brw_set_prim's
* optimization of single quads to trifans.
*/
key->pv_first = true;
}
/* _NEW_TRANSFORM */
key->userclip_active = (ctx->Transform.ClipPlanesEnabled != 0);
if (intel->gen >= 7) {
/* On Gen7 and later, we don't use GS (yet). */
key->need_gs_prog = false;
} else if (intel->gen == 6) {
/* On Gen6, GS is used for transform feedback. */
/* _NEW_TRANSFORM_FEEDBACK */
if (ctx->TransformFeedback.CurrentObject->Active &&
!ctx->TransformFeedback.CurrentObject->Paused) {
const struct gl_shader_program *shaderprog =
ctx->Shader.CurrentVertexProgram;
const struct gl_transform_feedback_info *linked_xfb_info =
&shaderprog->LinkedTransformFeedback;
int i;
/* Make sure that the VUE slots won't overflow the unsigned chars in
* key->transform_feedback_bindings[].
*/
STATIC_ASSERT(BRW_VERT_RESULT_MAX <= 256);
/* Make sure that we don't need more binding table entries than we've
* set aside for use in transform feedback. (We shouldn't, since we
* set aside enough binding table entries to have one per component).
*/
assert(linked_xfb_info->NumOutputs <= BRW_MAX_SOL_BINDINGS);
key->need_gs_prog = true;
key->num_transform_feedback_bindings = linked_xfb_info->NumOutputs;
for (i = 0; i < key->num_transform_feedback_bindings; ++i) {
key->transform_feedback_bindings[i] =
linked_xfb_info->Outputs[i].OutputRegister;
key->transform_feedback_swizzles[i] =
swizzle_for_offset[linked_xfb_info->Outputs[i].ComponentOffset];
}
}
/* On Gen6, GS is also used for rasterizer discard. */
/* _NEW_RASTERIZER_DISCARD */
if (ctx->RasterDiscard) {
key->need_gs_prog = true;
key->rasterizer_discard = true;
}
} else {
/* Pre-gen6, GS is used to transform QUADLIST, QUADSTRIP, and LINELOOP
* into simpler primitives.
*/
key->need_gs_prog = (brw->primitive == _3DPRIM_QUADLIST ||
brw->primitive == _3DPRIM_QUADSTRIP ||
brw->primitive == _3DPRIM_LINELOOP);
}
/* For testing, the environment variable INTEL_FORCE_GS can be used to
* force a GS program to be used, even if it's not necessary.
*/
if (getenv("INTEL_FORCE_GS"))
key->need_gs_prog = true;
}
