#include "brw_vec4_live_variables.h"
#include "brw_vec4_vs.h"
#include "brw_dead_control_flow.h"
-#include "common/gen_debug.h"
+#include "dev/gen_debug.h"
#include "program/prog_parameter.h"
+#include "util/u_math.h"
#define MAX_INSTRUCTION (1 << 30)
void
src_reg::init()
{
- memset(this, 0, sizeof(*this));
+ memset((void*)this, 0, sizeof(*this));
this->file = BAD_FILE;
this->type = BRW_REGISTER_TYPE_UD;
}
void
dst_reg::init()
{
- memset(this, 0, sizeof(*this));
+ memset((void*)this, 0, sizeof(*this));
this->file = BAD_FILE;
this->type = BRW_REGISTER_TYPE_UD;
this->writemask = WRITEMASK_XYZW;
switch (opcode) {
case SHADER_OPCODE_SHADER_TIME_ADD:
case VS_OPCODE_PULL_CONSTANT_LOAD_GEN7:
- case SHADER_OPCODE_UNTYPED_ATOMIC:
- case SHADER_OPCODE_UNTYPED_SURFACE_READ:
- case SHADER_OPCODE_UNTYPED_SURFACE_WRITE:
- case SHADER_OPCODE_TYPED_ATOMIC:
- case SHADER_OPCODE_TYPED_SURFACE_READ:
- case SHADER_OPCODE_TYPED_SURFACE_WRITE:
+ case VEC4_OPCODE_UNTYPED_ATOMIC:
+ case VEC4_OPCODE_UNTYPED_SURFACE_READ:
+ case VEC4_OPCODE_UNTYPED_SURFACE_WRITE:
case VEC4_OPCODE_URB_READ:
case TCS_OPCODE_URB_WRITE:
case TCS_OPCODE_RELEASE_INPUT:
{
switch (opcode) {
case SHADER_OPCODE_SHADER_TIME_ADD:
- case SHADER_OPCODE_UNTYPED_ATOMIC:
- case SHADER_OPCODE_UNTYPED_SURFACE_READ:
- case SHADER_OPCODE_UNTYPED_SURFACE_WRITE:
- case SHADER_OPCODE_TYPED_ATOMIC:
- case SHADER_OPCODE_TYPED_SURFACE_READ:
- case SHADER_OPCODE_TYPED_SURFACE_WRITE:
+ case VEC4_OPCODE_UNTYPED_ATOMIC:
+ case VEC4_OPCODE_UNTYPED_SURFACE_READ:
+ case VEC4_OPCODE_UNTYPED_SURFACE_WRITE:
case TCS_OPCODE_URB_WRITE:
if (arg == 0)
return mlen * REG_SIZE;
return true;
}
+bool
+vec4_instruction::can_do_cmod()
+{
+ if (!backend_instruction::can_do_cmod())
+ return false;
+
+ /* The accumulator result appears to get used for the conditional modifier
+ * generation. When negating a UD value, there is a 33rd bit generated for
+ * the sign in the accumulator value, so now you can't check, for example,
+ * equality with a 32-bit value. See piglit fs-op-neg-uvec4.
+ */
+ for (unsigned i = 0; i < 3; i++) {
+ if (src[i].file != BAD_FILE &&
+ type_is_unsigned_int(src[i].type) && src[i].negate)
+ return false;
+ }
+
+ return true;
+}
+
bool
vec4_instruction::can_do_writemask(const struct gen_device_info *devinfo)
{
case SHADER_OPCODE_TG4:
case SHADER_OPCODE_TG4_OFFSET:
case SHADER_OPCODE_SAMPLEINFO:
- case VS_OPCODE_GET_BUFFER_SIZE:
+ case SHADER_OPCODE_GET_BUFFER_SIZE:
return inst->header_size;
default:
unreachable("not reached");
!reladdr && !r.reladdr);
}
+bool
+src_reg::negative_equals(const src_reg &r) const
+{
+ return this->backend_reg::negative_equals(r) &&
+ !reladdr && !r.reladdr;
+}
+
bool
vec4_visitor::opt_vector_float()
{
bool progress = false;
foreach_block(block, cfg) {
- int last_reg = -1, last_offset = -1;
+ unsigned last_reg = ~0u, last_offset = ~0u;
enum brw_reg_file last_reg_file = BAD_FILE;
uint8_t imm[4] = { 0 };
foreach_inst_in_block_safe(vec4_instruction, inst, block) {
int vf = -1;
- enum brw_reg_type need_type;
+ enum brw_reg_type need_type = BRW_REGISTER_TYPE_LAST;
/* Look for unconditional MOVs from an immediate with a partial
* writemask. Skip type-conversion MOVs other than integer 0,
need_type = BRW_REGISTER_TYPE_F;
}
} else {
- last_reg = -1;
+ last_reg = ~0u;
}
/* If this wasn't a MOV, or the destination register doesn't match,
}
inst_count = 0;
- last_reg = -1;
+ last_reg = ~0u;;
writemask = 0;
dest_type = BRW_REGISTER_TYPE_F;
* the next part of our packing algorithm.
*/
int reg = inst->src[0].nr;
- for (unsigned i = 0; i < vec4s_read; i++)
+ int channel_size = type_sz(inst->src[0].type) / 4;
+ for (unsigned i = 0; i < vec4s_read; i++) {
chans_used[reg + i] = 4;
+ channel_sizes[reg + i] = MAX2(channel_sizes[reg + i], channel_size);
+ }
}
}
break;
if (inst->saturate) {
- if (inst->dst.type != inst->src[0].type)
+ /* Full mixed-type saturates don't happen. However, we can end up
+ * with things like:
+ *
+ * mov.sat(8) g21<1>DF -1F
+ *
+ * Other mixed-size-but-same-base-type cases may also be possible.
+ */
+ if (inst->dst.type != inst->src[0].type &&
+ inst->dst.type != BRW_REGISTER_TYPE_DF &&
+ inst->src[0].type != BRW_REGISTER_TYPE_F)
assert(!"unimplemented: saturate mixed types");
- if (brw_saturate_immediate(inst->dst.type,
+ if (brw_saturate_immediate(inst->src[0].type,
&inst->src[0].as_brw_reg())) {
inst->saturate = false;
progress = true;
}
break;
+ case BRW_OPCODE_OR:
+ if (inst->src[1].is_zero()) {
+ inst->opcode = BRW_OPCODE_MOV;
+ inst->src[1] = src_reg();
+ progress = true;
+ }
+ break;
+
case VEC4_OPCODE_UNPACK_UNIFORM:
if (inst->src[0].file != UNIFORM) {
inst->opcode = BRW_OPCODE_MOV;
progress = true;
}
break;
- case BRW_OPCODE_CMP:
- if (inst->conditional_mod == BRW_CONDITIONAL_GE &&
- inst->src[0].abs &&
- inst->src[0].negate &&
- inst->src[1].is_zero()) {
- inst->src[0].abs = false;
- inst->src[0].negate = false;
- inst->conditional_mod = BRW_CONDITIONAL_Z;
- progress = true;
- break;
- }
- break;
case SHADER_OPCODE_BROADCAST:
if (is_uniform(inst->src[0]) ||
inst->src[1].is_zero()) {
}
}
+ /* VS_OPCODE_UNPACK_FLAGS_SIMD4X2 generates a bunch of mov(1)
+ * instructions, and this optimization pass is not capable of
+ * handling that. Bail on these instructions and hope that some
+ * later optimization pass can do the right thing after they are
+ * expanded.
+ */
+ if (scan_inst->opcode == VS_OPCODE_UNPACK_FLAGS_SIMD4X2)
+ break;
+
/* This doesn't handle saturation on the instruction we
* want to coalesce away if the register types do not match.
* But if scan_inst is a non type-converting 'mov', we can fix
* in the register instead.
*/
if (to_mrf && scan_inst->mlen > 0) {
- if (inst->dst.nr >= scan_inst->base_mrf &&
- inst->dst.nr < scan_inst->base_mrf + scan_inst->mlen) {
+ unsigned start = scan_inst->base_mrf;
+ unsigned end = scan_inst->base_mrf + scan_inst->mlen;
+
+ if (inst->dst.nr >= start && inst->dst.nr < end) {
break;
}
} else {
vec4_instruction *inst = (vec4_instruction *)be_inst;
if (inst->predicate) {
- fprintf(file, "(%cf0.%d%s) ",
+ fprintf(file, "(%cf%d.%d%s) ",
inst->predicate_inverse ? '-' : '+',
- inst->flag_subreg,
+ inst->flag_subreg / 2,
+ inst->flag_subreg % 2,
pred_ctrl_align16[inst->predicate]);
}
fprintf(file, "%s", conditional_modifier[inst->conditional_mod]);
if (!inst->predicate &&
(devinfo->gen < 5 || (inst->opcode != BRW_OPCODE_SEL &&
+ inst->opcode != BRW_OPCODE_CSEL &&
inst->opcode != BRW_OPCODE_IF &&
inst->opcode != BRW_OPCODE_WHILE))) {
- fprintf(file, ".f0.%d", inst->flag_subreg);
+ fprintf(file, ".f%d.%d", inst->flag_subreg / 2, inst->flag_subreg % 2);
}
}
fprintf(file, " ");
* matter what, or the GPU would hang.
*/
if (devinfo->gen < 6 && this->uniforms == 0) {
- stage_prog_data->param =
- reralloc(NULL, stage_prog_data->param, uint32_t, 4);
+ brw_stage_prog_data_add_params(stage_prog_data, 4);
for (unsigned int i = 0; i < 4; i++) {
unsigned int slot = this->uniforms * 4 + i;
stage_prog_data->param[slot] = BRW_PARAM_BUILTIN_ZERO;
}
}
+/**
+ * Three source instruction must have a GRF/MRF destination register.
+ * ARF NULL is not allowed. Fix that up by allocating a temporary GRF.
+ */
+void
+vec4_visitor::fixup_3src_null_dest()
+{
+ bool progress = false;
+
+ foreach_block_and_inst_safe (block, vec4_instruction, inst, cfg) {
+ if (inst->is_3src(devinfo) && inst->dst.is_null()) {
+ const unsigned size_written = type_sz(inst->dst.type);
+ const unsigned num_regs = DIV_ROUND_UP(size_written, REG_SIZE);
+
+ inst->dst = retype(dst_reg(VGRF, alloc.allocate(num_regs)),
+ inst->dst.type);
+ progress = true;
+ }
+ }
+
+ if (progress)
+ invalidate_live_intervals();
+}
+
void
vec4_visitor::convert_to_hw_regs()
{
if (linst->src[i].file == BAD_FILE)
continue;
- if (!is_uniform(linst->src[i]))
+ bool is_interleaved_attr =
+ linst->src[i].file == ATTR &&
+ stage_uses_interleaved_attributes(stage,
+ prog_data->dispatch_mode);
+
+ if (!is_uniform(linst->src[i]) && !is_interleaved_attr)
linst->src[i] = horiz_offset(linst->src[i], channel_offset);
}
OPT(scalarize_df);
}
+ fixup_3src_null_dest();
+
bool allocated_without_spills = reg_allocate();
if (!allocated_without_spills) {
void *mem_ctx,
const struct brw_vs_prog_key *key,
struct brw_vs_prog_data *prog_data,
- const nir_shader *src_shader,
- bool use_legacy_snorm_formula,
+ nir_shader *shader,
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, &key->tex, is_scalar);
const unsigned *assembly = NULL;
}
prog_data->inputs_read = shader->info.inputs_read;
- prog_data->double_inputs_read = shader->info.double_inputs_read;
+ prog_data->double_inputs_read = shader->info.vs.double_inputs;
- brw_nir_lower_vs_inputs(shader, use_legacy_snorm_formula,
- key->gl_attrib_wa_flags);
- brw_nir_lower_vue_outputs(shader, is_scalar);
+ brw_nir_lower_vs_inputs(shader, key->gl_attrib_wa_flags);
+ brw_nir_lower_vue_outputs(shader);
shader = brw_postprocess_nir(shader, compiler, is_scalar);
prog_data->base.clip_distance_mask =
((1 << shader->info.cull_distance_array_size) - 1) <<
shader->info.clip_distance_array_size;
- unsigned nr_attribute_slots = _mesa_bitcount_64(prog_data->inputs_read);
+ unsigned nr_attribute_slots = util_bitcount64(prog_data->inputs_read);
/* gl_VertexID and gl_InstanceID are system values, but arrive via an
* incoming vertex attribute. So, add an extra slot.
*/
if (shader->info.system_values_read &
- (BITFIELD64_BIT(SYSTEM_VALUE_BASE_VERTEX) |
+ (BITFIELD64_BIT(SYSTEM_VALUE_FIRST_VERTEX) |
BITFIELD64_BIT(SYSTEM_VALUE_BASE_INSTANCE) |
BITFIELD64_BIT(SYSTEM_VALUE_VERTEX_ID_ZERO_BASE) |
BITFIELD64_BIT(SYSTEM_VALUE_INSTANCE_ID))) {
nr_attribute_slots++;
}
+ /* gl_DrawID and IsIndexedDraw share its very own vec4 */
+ if (shader->info.system_values_read &
+ (BITFIELD64_BIT(SYSTEM_VALUE_DRAW_ID) |
+ BITFIELD64_BIT(SYSTEM_VALUE_IS_INDEXED_DRAW))) {
+ nr_attribute_slots++;
+ }
+
+ if (shader->info.system_values_read &
+ BITFIELD64_BIT(SYSTEM_VALUE_IS_INDEXED_DRAW))
+ prog_data->uses_is_indexed_draw = true;
+
if (shader->info.system_values_read &
- BITFIELD64_BIT(SYSTEM_VALUE_BASE_VERTEX))
- prog_data->uses_basevertex = true;
+ BITFIELD64_BIT(SYSTEM_VALUE_FIRST_VERTEX))
+ prog_data->uses_firstvertex = true;
if (shader->info.system_values_read &
BITFIELD64_BIT(SYSTEM_VALUE_BASE_INSTANCE))
BITFIELD64_BIT(SYSTEM_VALUE_INSTANCE_ID))
prog_data->uses_instanceid = true;
- /* gl_DrawID has its very own vec4 */
if (shader->info.system_values_read &
- BITFIELD64_BIT(SYSTEM_VALUE_DRAW_ID)) {
- prog_data->uses_drawid = true;
- nr_attribute_slots++;
- }
-
- unsigned nr_attributes = nr_attribute_slots -
- DIV_ROUND_UP(_mesa_bitcount_64(shader->info.double_inputs_read), 2);
+ BITFIELD64_BIT(SYSTEM_VALUE_DRAW_ID))
+ prog_data->uses_drawid = true;
/* 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
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
prog_data->base.base.dispatch_grf_start_reg = v.payload.num_regs;
- fs_generator g(compiler, log_data, mem_ctx, (void *) key,
+ fs_generator g(compiler, log_data, mem_ctx,
&prog_data->base.base, v.promoted_constants,
v.runtime_check_aads_emit, MESA_SHADER_VERTEX);
if (INTEL_DEBUG & DEBUG_VS) {
g.enable_debug(debug_name);
}
g.generate_code(v.cfg, 8);
- assembly = g.get_assembly(final_assembly_size);
+ assembly = g.get_assembly();
}
if (!assembly) {
prog_data->base.dispatch_mode = DISPATCH_MODE_4X2_DUAL_OBJECT;
vec4_vs_visitor v(compiler, log_data, key, prog_data,
- shader, mem_ctx,
- shader_time_index, use_legacy_snorm_formula);
+ shader, mem_ctx, shader_time_index);
if (!v.run()) {
if (error_str)
*error_str = ralloc_strdup(mem_ctx, v.fail_msg);
}
assembly = brw_vec4_generate_assembly(compiler, log_data, mem_ctx,
- shader, &prog_data->base, v.cfg,
- final_assembly_size);
+ shader, &prog_data->base, v.cfg);
}
return assembly;
projects / mesa.git / blobdiff