#include "brw_fs.h"
#include "brw_cfg.h"
-#include "glsl/ir_print_visitor.h"
fs_generator::fs_generator(struct brw_context *brw,
struct brw_wm_compile *c,
: brw(brw), c(c), prog(prog), fp(fp), dual_source_output(dual_source_output)
{
- p = &c->func;
- intel = &brw->intel;
- ctx = &intel->ctx;
+ ctx = &brw->ctx;
shader = prog ? prog->_LinkedShaders[MESA_SHADER_FRAGMENT] : NULL;
mem_ctx = c;
+
+ p = rzalloc(mem_ctx, struct brw_compile);
+ brw_init_compile(brw, p, mem_ctx);
}
fs_generator::~fs_generator()
{
}
+void
+fs_generator::mark_surface_used(unsigned surf_index)
+{
+ assert(surf_index < BRW_MAX_WM_SURFACES);
+
+ c->prog_data.binding_table_size =
+ MAX2(c->prog_data.binding_table_size, surf_index + 1);
+}
+
+void
+fs_generator::patch_discard_jumps_to_fb_writes()
+{
+ if (brw->gen < 6 || this->discard_halt_patches.is_empty())
+ return;
+
+ /* There is a somewhat strange undocumented requirement of using
+ * HALT, according to the simulator. If some channel has HALTed to
+ * a particular UIP, then by the end of the program, every channel
+ * must have HALTed to that UIP. Furthermore, the tracking is a
+ * stack, so you can't do the final halt of a UIP after starting
+ * halting to a new UIP.
+ *
+ * Symptoms of not emitting this instruction on actual hardware
+ * included GPU hangs and sparkly rendering on the piglit discard
+ * tests.
+ */
+ struct brw_instruction *last_halt = gen6_HALT(p);
+ last_halt->bits3.break_cont.uip = 2;
+ last_halt->bits3.break_cont.jip = 2;
+
+ int ip = p->nr_insn;
+
+ foreach_list(node, &this->discard_halt_patches) {
+ ip_record *patch_ip = (ip_record *)node;
+ struct brw_instruction *patch = &p->store[patch_ip->ip];
+
+ assert(patch->header.opcode == BRW_OPCODE_HALT);
+ /* HALT takes a half-instruction distance from the pre-incremented IP. */
+ patch->bits3.break_cont.uip = (ip - patch_ip->ip) * 2;
+ }
+
+ this->discard_halt_patches.make_empty();
+}
+
void
fs_generator::generate_fb_write(fs_inst *inst)
{
brw_set_mask_control(p, BRW_MASK_DISABLE);
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
+ if (fp->UsesKill) {
+ struct brw_reg pixel_mask;
+
+ if (brw->gen >= 6)
+ pixel_mask = retype(brw_vec1_grf(1, 7), BRW_REGISTER_TYPE_UW);
+ else
+ pixel_mask = retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UW);
+
+ brw_MOV(p, pixel_mask, brw_flag_reg(0, 1));
+ }
+
if (inst->header_present) {
- if (intel->gen >= 6) {
+ if (brw->gen >= 6) {
brw_set_compression_control(p, BRW_COMPRESSION_COMPRESSED);
brw_MOV(p,
retype(brw_message_reg(inst->base_mrf), BRW_REGISTER_TYPE_UD),
retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD));
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
- if (inst->target > 0 &&
- c->key.nr_color_regions > 1 &&
- c->key.sample_alpha_to_coverage) {
+ if (inst->target > 0 && c->key.replicate_alpha) {
/* Set "Source0 Alpha Present to RenderTarget" bit in message
* header.
*/
inst->base_mrf,
implied_header,
msg_control,
- inst->target,
+ SURF_INDEX_DRAW(inst->target),
inst->mlen,
0,
eot,
inst->header_present);
+
+ mark_surface_used(SURF_INDEX_DRAW(inst->target));
}
/* Computes the integer pixel x,y values from the origin.
if (brw->has_pln &&
delta_y.nr == delta_x.nr + 1 &&
- (intel->gen >= 6 || (delta_x.nr & 1) == 0)) {
+ (brw->gen >= 6 || (delta_x.nr & 1) == 0)) {
brw_PLN(p, dst, interp, delta_x);
} else {
brw_LINE(p, brw_null_reg(), interp, delta_x);
}
}
+void
+fs_generator::generate_math_g45(fs_inst *inst,
+ struct brw_reg dst,
+ struct brw_reg src)
+{
+ if (inst->opcode == SHADER_OPCODE_POW ||
+ inst->opcode == SHADER_OPCODE_INT_QUOTIENT ||
+ inst->opcode == SHADER_OPCODE_INT_REMAINDER) {
+ generate_math_gen4(inst, dst, src);
+ return;
+ }
+
+ int op = brw_math_function(inst->opcode);
+
+ assert(inst->mlen >= 1);
+
+ brw_math(p, dst,
+ op,
+ inst->base_mrf, src,
+ BRW_MATH_DATA_VECTOR,
+ BRW_MATH_PRECISION_FULL);
+}
+
void
fs_generator::generate_tex(fs_inst *inst, struct brw_reg dst, struct brw_reg src)
{
if (dispatch_width == 16)
simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD16;
- if (intel->gen >= 5) {
+ if (brw->gen >= 5) {
switch (inst->opcode) {
case SHADER_OPCODE_TEX:
if (inst->shadow_compare) {
msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_RESINFO;
break;
case SHADER_OPCODE_TXD:
- /* There is no sample_d_c message; comparisons are done manually */
- msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_DERIVS;
+ if (inst->shadow_compare) {
+ /* Gen7.5+. Otherwise, lowered by brw_lower_texture_gradients(). */
+ assert(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_MS:
+ if (brw->gen >= 7)
+ msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_LD2DMS;
+ else
+ msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LD;
+ break;
+ case SHADER_OPCODE_LOD:
+ msg_type = GEN5_SAMPLER_MESSAGE_LOD;
+ break;
default:
assert(!"not reached");
break;
*/
if (inst->texture_offset) {
brw_push_insn_state(p);
+ brw_set_mask_control(p, BRW_MASK_DISABLE);
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
/* Explicitly set up the message header by copying g0 to the MRF. */
brw_MOV(p, retype(brw_message_reg(inst->base_mrf), BRW_REGISTER_TYPE_UD),
src,
SURF_INDEX_TEXTURE(inst->sampler),
inst->sampler,
- WRITEMASK_XYZW,
msg_type,
rlen,
inst->mlen,
inst->header_present,
simd_mode,
return_format);
+
+ mark_surface_used(SURF_INDEX_TEXTURE(inst->sampler));
}
}
void
-fs_generator::generate_discard(fs_inst *inst)
+fs_generator::generate_discard_jump(fs_inst *inst)
{
- struct brw_reg f0 = brw_flag_reg();
-
- if (intel->gen >= 6) {
- struct brw_reg g1 = retype(brw_vec1_grf(1, 7), BRW_REGISTER_TYPE_UW);
- struct brw_reg some_register;
-
- /* As of gen6, we no longer have the mask register to look at,
- * so life gets a bit more complicated.
- */
-
- /* Load the flag register with all ones. */
- brw_push_insn_state(p);
- brw_set_mask_control(p, BRW_MASK_DISABLE);
- brw_MOV(p, f0, brw_imm_uw(0xffff));
- brw_pop_insn_state(p);
-
- /* Do a comparison that should always fail, to produce 0s in the flag
- * reg where we have active channels.
- */
- some_register = retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UW);
- brw_CMP(p, retype(brw_null_reg(), BRW_REGISTER_TYPE_UD),
- BRW_CONDITIONAL_NZ, some_register, some_register);
-
- /* Undo CMP's whacking of predication*/
- brw_set_predicate_control(p, BRW_PREDICATE_NONE);
-
- brw_push_insn_state(p);
- brw_set_mask_control(p, BRW_MASK_DISABLE);
- brw_AND(p, g1, f0, g1);
- brw_pop_insn_state(p);
- } else {
- struct brw_reg g0 = retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UW);
-
- brw_push_insn_state(p);
- brw_set_mask_control(p, BRW_MASK_DISABLE);
- brw_set_compression_control(p, BRW_COMPRESSION_NONE);
+ assert(brw->gen >= 6);
- /* Unlike the 965, we have the mask reg, so we just need
- * somewhere to invert that (containing channels to be disabled)
- * so it can be ANDed with the mask of pixels still to be
- * written. Use the flag reg for consistency with gen6+.
- */
- brw_NOT(p, f0, brw_mask_reg(1)); /* IMASK */
- brw_AND(p, g0, f0, g0);
+ /* This HALT will be patched up at FB write time to point UIP at the end of
+ * the program, and at brw_uip_jip() JIP will be set to the end of the
+ * current block (or the program).
+ */
+ this->discard_halt_patches.push_tail(new(mem_ctx) ip_record(p->nr_insn));
- brw_pop_insn_state(p);
- }
+ brw_push_insn_state(p);
+ brw_set_mask_control(p, BRW_MASK_DISABLE);
+ gen6_HALT(p);
+ brw_pop_insn_state(p);
}
void
{
assert(inst->mlen != 0);
- /* Clear any post destination dependencies that would be ignored by
- * the block read. See the B-Spec for pre-gen5 send instruction.
- *
- * This could use a better solution, since texture sampling and
- * math reads could potentially run into it as well -- anywhere
- * that we have a SEND with a destination that is a register that
- * was written but not read within the last N instructions (what's
- * N? unsure). This is rare because of dead code elimination, but
- * not impossible.
- */
- if (intel->gen == 4 && !intel->is_g4x)
- brw_MOV(p, brw_null_reg(), dst);
-
brw_oword_block_read_scratch(p, dst, brw_message_reg(inst->base_mrf), 1,
inst->offset);
-
- if (intel->gen == 4 && !intel->is_g4x) {
- /* gen4 errata: destination from a send can't be used as a
- * destination until it's been read. Just read it so we don't
- * have to worry.
- */
- brw_MOV(p, brw_null_reg(), dst);
- }
}
void
-fs_generator::generate_pull_constant_load(fs_inst *inst, struct brw_reg dst,
- struct brw_reg index,
- struct brw_reg offset)
+fs_generator::generate_uniform_pull_constant_load(fs_inst *inst,
+ struct brw_reg dst,
+ struct brw_reg index,
+ struct brw_reg offset)
{
assert(inst->mlen != 0);
- /* Clear any post destination dependencies that would be ignored by
- * the block read. See the B-Spec for pre-gen5 send instruction.
- *
- * This could use a better solution, since texture sampling and
- * math reads could potentially run into it as well -- anywhere
- * that we have a SEND with a destination that is a register that
- * was written but not read within the last N instructions (what's
- * N? unsure). This is rare because of dead code elimination, but
- * not impossible.
- */
- if (intel->gen == 4 && !intel->is_g4x)
- brw_MOV(p, brw_null_reg(), dst);
-
assert(index.file == BRW_IMMEDIATE_VALUE &&
index.type == BRW_REGISTER_TYPE_UD);
uint32_t surf_index = index.dw1.ud;
brw_oword_block_read(p, dst, brw_message_reg(inst->base_mrf),
read_offset, surf_index);
- if (intel->gen == 4 && !intel->is_g4x) {
- /* gen4 errata: destination from a send can't be used as a
- * destination until it's been read. Just read it so we don't
- * have to worry.
+ mark_surface_used(surf_index);
+}
+
+void
+fs_generator::generate_uniform_pull_constant_load_gen7(fs_inst *inst,
+ struct brw_reg dst,
+ struct brw_reg index,
+ struct brw_reg offset)
+{
+ assert(inst->mlen == 0);
+
+ assert(index.file == BRW_IMMEDIATE_VALUE &&
+ index.type == BRW_REGISTER_TYPE_UD);
+ uint32_t surf_index = index.dw1.ud;
+
+ assert(offset.file == BRW_GENERAL_REGISTER_FILE);
+ /* Reference just the dword we need, to avoid angering validate_reg(). */
+ offset = brw_vec1_grf(offset.nr, 0);
+
+ brw_push_insn_state(p);
+ brw_set_compression_control(p, BRW_COMPRESSION_NONE);
+ brw_set_mask_control(p, BRW_MASK_DISABLE);
+ struct brw_instruction *send = brw_next_insn(p, BRW_OPCODE_SEND);
+ brw_pop_insn_state(p);
+
+ /* We use the SIMD4x2 mode because we want to end up with 4 components in
+ * the destination loaded consecutively from the same offset (which appears
+ * in the first component, and the rest are ignored).
+ */
+ dst.width = BRW_WIDTH_4;
+ brw_set_dest(p, send, dst);
+ brw_set_src0(p, send, offset);
+ brw_set_sampler_message(p, send,
+ surf_index,
+ 0, /* LD message ignores sampler unit */
+ GEN5_SAMPLER_MESSAGE_SAMPLE_LD,
+ 1, /* rlen */
+ 1, /* mlen */
+ false, /* no header */
+ BRW_SAMPLER_SIMD_MODE_SIMD4X2,
+ 0);
+
+ mark_surface_used(surf_index);
+}
+
+void
+fs_generator::generate_varying_pull_constant_load(fs_inst *inst,
+ struct brw_reg dst,
+ struct brw_reg index,
+ struct brw_reg offset)
+{
+ assert(brw->gen < 7); /* Should use the gen7 variant. */
+ assert(inst->header_present);
+ assert(inst->mlen);
+
+ assert(index.file == BRW_IMMEDIATE_VALUE &&
+ index.type == BRW_REGISTER_TYPE_UD);
+ uint32_t surf_index = index.dw1.ud;
+
+ uint32_t simd_mode, rlen, msg_type;
+ if (dispatch_width == 16) {
+ simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD16;
+ rlen = 8;
+ } else {
+ simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD8;
+ rlen = 4;
+ }
+
+ if (brw->gen >= 5)
+ msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LD;
+ else {
+ /* We always use the SIMD16 message so that we only have to load U, and
+ * not V or R.
*/
- brw_MOV(p, brw_null_reg(), dst);
+ msg_type = BRW_SAMPLER_MESSAGE_SIMD16_LD;
+ assert(inst->mlen == 3);
+ assert(inst->regs_written == 8);
+ rlen = 8;
+ simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD16;
}
+
+ struct brw_reg offset_mrf = retype(brw_message_reg(inst->base_mrf + 1),
+ BRW_REGISTER_TYPE_D);
+ brw_MOV(p, offset_mrf, offset);
+
+ struct brw_reg header = brw_vec8_grf(0, 0);
+ gen6_resolve_implied_move(p, &header, inst->base_mrf);
+
+ struct brw_instruction *send = brw_next_insn(p, BRW_OPCODE_SEND);
+ send->header.compression_control = BRW_COMPRESSION_NONE;
+ brw_set_dest(p, send, dst);
+ brw_set_src0(p, send, header);
+ if (brw->gen < 6)
+ send->header.destreg__conditionalmod = inst->base_mrf;
+
+ /* Our surface is set up as floats, regardless of what actual data is
+ * stored in it.
+ */
+ uint32_t return_format = BRW_SAMPLER_RETURN_FORMAT_FLOAT32;
+ brw_set_sampler_message(p, send,
+ surf_index,
+ 0, /* sampler (unused) */
+ msg_type,
+ rlen,
+ inst->mlen,
+ inst->header_present,
+ simd_mode,
+ return_format);
+
+ mark_surface_used(surf_index);
}
+void
+fs_generator::generate_varying_pull_constant_load_gen7(fs_inst *inst,
+ struct brw_reg dst,
+ struct brw_reg index,
+ struct brw_reg offset)
+{
+ assert(brw->gen >= 7);
+ /* Varying-offset pull constant loads are treated as a normal expression on
+ * gen7, so the fact that it's a send message is hidden at the IR level.
+ */
+ assert(!inst->header_present);
+ assert(!inst->mlen);
+
+ assert(index.file == BRW_IMMEDIATE_VALUE &&
+ index.type == BRW_REGISTER_TYPE_UD);
+ uint32_t surf_index = index.dw1.ud;
+
+ uint32_t simd_mode, rlen, mlen;
+ if (dispatch_width == 16) {
+ mlen = 2;
+ rlen = 8;
+ simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD16;
+ } else {
+ mlen = 1;
+ rlen = 4;
+ simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD8;
+ }
+
+ struct brw_instruction *send = brw_next_insn(p, BRW_OPCODE_SEND);
+ brw_set_dest(p, send, dst);
+ brw_set_src0(p, send, offset);
+ brw_set_sampler_message(p, send,
+ surf_index,
+ 0, /* LD message ignores sampler unit */
+ GEN5_SAMPLER_MESSAGE_SAMPLE_LD,
+ rlen,
+ mlen,
+ false, /* no header */
+ simd_mode,
+ 0);
+
+ mark_surface_used(surf_index);
+}
/**
* Cause the current pixel/sample mask (from R1.7 bits 15:0) to be transferred
* Used only on Gen6 and above.
*/
void
-fs_generator::generate_mov_dispatch_to_flags()
+fs_generator::generate_mov_dispatch_to_flags(fs_inst *inst)
{
- struct brw_reg f0 = brw_flag_reg();
- struct brw_reg g1 = retype(brw_vec1_grf(1, 7), BRW_REGISTER_TYPE_UW);
+ struct brw_reg flags = brw_flag_reg(0, inst->flag_subreg);
+ struct brw_reg dispatch_mask;
+
+ if (brw->gen >= 6)
+ dispatch_mask = retype(brw_vec1_grf(1, 7), BRW_REGISTER_TYPE_UW);
+ else
+ dispatch_mask = retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UW);
- assert (intel->gen >= 6);
brw_push_insn_state(p);
brw_set_mask_control(p, BRW_MASK_DISABLE);
- brw_MOV(p, f0, g1);
+ brw_MOV(p, flags, dispatch_mask);
brw_pop_insn_state(p);
}
break;
}
break;
- case FIXED_HW_REG:
+ case HW_REG:
brw_reg = reg->fixed_hw_reg;
break;
case BAD_FILE:
return brw_reg;
}
+/**
+ * Sets the first word of a vgrf for gen7+ simd4x2 uniform pull constant
+ * sampler LD messages.
+ *
+ * We don't want to bake it into the send message's code generation because
+ * that means we don't get a chance to schedule the instructions.
+ */
+void
+fs_generator::generate_set_simd4x2_offset(fs_inst *inst,
+ struct brw_reg dst,
+ struct brw_reg value)
+{
+ assert(value.file == BRW_IMMEDIATE_VALUE);
+
+ brw_push_insn_state(p);
+ brw_set_compression_control(p, BRW_COMPRESSION_NONE);
+ brw_set_mask_control(p, BRW_MASK_DISABLE);
+ brw_MOV(p, retype(brw_vec1_reg(dst.file, dst.nr, 0), value.type), value);
+ brw_pop_insn_state(p);
+}
+
+/**
+ * Change the register's data type from UD to W, doubling the strides in order
+ * to compensate for halving the data type width.
+ */
+static struct brw_reg
+ud_reg_to_w(struct brw_reg r)
+{
+ assert(r.type == BRW_REGISTER_TYPE_UD);
+ r.type = BRW_REGISTER_TYPE_W;
+
+ /* The BRW_*_STRIDE enums are defined so that incrementing the field
+ * doubles the real stride.
+ */
+ if (r.hstride != 0)
+ ++r.hstride;
+ if (r.vstride != 0)
+ ++r.vstride;
+
+ return r;
+}
+
+void
+fs_generator::generate_pack_half_2x16_split(fs_inst *inst,
+ struct brw_reg dst,
+ struct brw_reg x,
+ struct brw_reg y)
+{
+ assert(brw->gen >= 7);
+ assert(dst.type == BRW_REGISTER_TYPE_UD);
+ assert(x.type == BRW_REGISTER_TYPE_F);
+ assert(y.type == BRW_REGISTER_TYPE_F);
+
+ /* From the Ivybridge PRM, Vol4, Part3, Section 6.27 f32to16:
+ *
+ * Because this instruction does not have a 16-bit floating-point type,
+ * the destination data type must be Word (W).
+ *
+ * The destination must be DWord-aligned and specify a horizontal stride
+ * (HorzStride) of 2. The 16-bit result is stored in the lower word of
+ * each destination channel and the upper word is not modified.
+ */
+ struct brw_reg dst_w = ud_reg_to_w(dst);
+
+ /* Give each 32-bit channel of dst the form below , where "." means
+ * unchanged.
+ * 0x....hhhh
+ */
+ brw_F32TO16(p, dst_w, y);
+
+ /* Now the form:
+ * 0xhhhh0000
+ */
+ brw_SHL(p, dst, dst, brw_imm_ud(16u));
+
+ /* And, finally the form of packHalf2x16's output:
+ * 0xhhhhllll
+ */
+ brw_F32TO16(p, dst_w, x);
+}
+
+void
+fs_generator::generate_unpack_half_2x16_split(fs_inst *inst,
+ struct brw_reg dst,
+ struct brw_reg src)
+{
+ assert(brw->gen >= 7);
+ assert(dst.type == BRW_REGISTER_TYPE_F);
+ assert(src.type == BRW_REGISTER_TYPE_UD);
+
+ /* From the Ivybridge PRM, Vol4, Part3, Section 6.26 f16to32:
+ *
+ * Because this instruction does not have a 16-bit floating-point type,
+ * the source data type must be Word (W). The destination type must be
+ * F (Float).
+ */
+ struct brw_reg src_w = ud_reg_to_w(src);
+
+ /* Each channel of src has the form of unpackHalf2x16's input: 0xhhhhllll.
+ * For the Y case, we wish to access only the upper word; therefore
+ * a 16-bit subregister offset is needed.
+ */
+ assert(inst->opcode == FS_OPCODE_UNPACK_HALF_2x16_SPLIT_X ||
+ inst->opcode == FS_OPCODE_UNPACK_HALF_2x16_SPLIT_Y);
+ if (inst->opcode == FS_OPCODE_UNPACK_HALF_2x16_SPLIT_Y)
+ src_w.subnr += 2;
+
+ brw_F16TO32(p, dst, src_w);
+}
+
+void
+fs_generator::generate_shader_time_add(fs_inst *inst,
+ struct brw_reg payload,
+ struct brw_reg offset,
+ struct brw_reg value)
+{
+ assert(brw->gen >= 7);
+ brw_push_insn_state(p);
+ brw_set_mask_control(p, true);
+
+ assert(payload.file == BRW_GENERAL_REGISTER_FILE);
+ struct brw_reg payload_offset = retype(brw_vec1_grf(payload.nr, 0),
+ offset.type);
+ struct brw_reg payload_value = retype(brw_vec1_grf(payload.nr + 1, 0),
+ value.type);
+
+ assert(offset.file == BRW_IMMEDIATE_VALUE);
+ if (value.file == BRW_GENERAL_REGISTER_FILE) {
+ value.width = BRW_WIDTH_1;
+ value.hstride = BRW_HORIZONTAL_STRIDE_0;
+ value.vstride = BRW_VERTICAL_STRIDE_0;
+ } else {
+ assert(value.file == BRW_IMMEDIATE_VALUE);
+ }
+
+ /* Trying to deal with setup of the params from the IR is crazy in the FS8
+ * case, and we don't really care about squeezing every bit of performance
+ * out of this path, so we just emit the MOVs from here.
+ */
+ brw_MOV(p, payload_offset, offset);
+ brw_MOV(p, payload_value, value);
+ brw_shader_time_add(p, payload, SURF_INDEX_WM_SHADER_TIME);
+ brw_pop_insn_state(p);
+
+ mark_surface_used(SURF_INDEX_WM_SHADER_TIME);
+}
+
void
fs_generator::generate_code(exec_list *instructions)
{
brw_set_conditionalmod(p, inst->conditional_mod);
brw_set_predicate_control(p, inst->predicate);
brw_set_predicate_inverse(p, inst->predicate_inverse);
+ brw_set_flag_reg(p, 0, inst->flag_subreg);
brw_set_saturate(p, inst->saturate);
+ brw_set_mask_control(p, inst->force_writemask_all);
if (inst->force_uncompressed || dispatch_width == 8) {
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
brw_set_access_mode(p, BRW_ALIGN_1);
break;
+ case BRW_OPCODE_LRP:
+ brw_set_access_mode(p, BRW_ALIGN_16);
+ if (dispatch_width == 16) {
+ brw_set_compression_control(p, BRW_COMPRESSION_NONE);
+ brw_LRP(p, dst, src[0], src[1], src[2]);
+ brw_set_compression_control(p, BRW_COMPRESSION_2NDHALF);
+ brw_LRP(p, sechalf(dst), sechalf(src[0]), sechalf(src[1]), sechalf(src[2]));
+ brw_set_compression_control(p, BRW_COMPRESSION_COMPRESSED);
+ } else {
+ brw_LRP(p, dst, src[0], src[1], src[2]);
+ }
+ brw_set_access_mode(p, BRW_ALIGN_1);
+ break;
+
case BRW_OPCODE_FRC:
brw_FRC(p, dst, src[0]);
break;
case BRW_OPCODE_SHL:
brw_SHL(p, dst, src[0], src[1]);
break;
-
+ case BRW_OPCODE_F32TO16:
+ brw_F32TO16(p, dst, src[0]);
+ break;
+ case BRW_OPCODE_F16TO32:
+ brw_F16TO32(p, dst, src[0]);
+ break;
case BRW_OPCODE_CMP:
brw_CMP(p, dst, inst->conditional_mod, src[0], src[1]);
break;
case BRW_OPCODE_SEL:
brw_SEL(p, dst, src[0], src[1]);
break;
+ case BRW_OPCODE_BFREV:
+ /* BFREV only supports UD type for src and dst. */
+ brw_BFREV(p, retype(dst, BRW_REGISTER_TYPE_UD),
+ retype(src[0], BRW_REGISTER_TYPE_UD));
+ break;
+ case BRW_OPCODE_FBH:
+ /* FBH only supports UD type for dst. */
+ brw_FBH(p, retype(dst, BRW_REGISTER_TYPE_UD), src[0]);
+ break;
+ case BRW_OPCODE_FBL:
+ /* FBL only supports UD type for dst. */
+ brw_FBL(p, retype(dst, BRW_REGISTER_TYPE_UD), src[0]);
+ break;
+ case BRW_OPCODE_CBIT:
+ /* CBIT only supports UD type for dst. */
+ brw_CBIT(p, retype(dst, BRW_REGISTER_TYPE_UD), src[0]);
+ break;
+
+ case BRW_OPCODE_BFE:
+ brw_set_access_mode(p, BRW_ALIGN_16);
+ if (dispatch_width == 16) {
+ brw_set_compression_control(p, BRW_COMPRESSION_NONE);
+ brw_BFE(p, dst, src[0], src[1], src[2]);
+ brw_set_compression_control(p, BRW_COMPRESSION_2NDHALF);
+ brw_BFE(p, sechalf(dst), sechalf(src[0]), sechalf(src[1]), sechalf(src[2]));
+ brw_set_compression_control(p, BRW_COMPRESSION_COMPRESSED);
+ } else {
+ brw_BFE(p, dst, src[0], src[1], src[2]);
+ }
+ brw_set_access_mode(p, BRW_ALIGN_1);
+ break;
+
+ case BRW_OPCODE_BFI1:
+ brw_BFI1(p, dst, src[0], src[1]);
+ break;
+ case BRW_OPCODE_BFI2:
+ brw_set_access_mode(p, BRW_ALIGN_16);
+ if (dispatch_width == 16) {
+ brw_set_compression_control(p, BRW_COMPRESSION_NONE);
+ brw_BFI2(p, dst, src[0], src[1], src[2]);
+ brw_set_compression_control(p, BRW_COMPRESSION_2NDHALF);
+ brw_BFI2(p, sechalf(dst), sechalf(src[0]), sechalf(src[1]), sechalf(src[2]));
+ brw_set_compression_control(p, BRW_COMPRESSION_COMPRESSED);
+ } else {
+ brw_BFI2(p, dst, src[0], src[1], src[2]);
+ }
+ brw_set_access_mode(p, BRW_ALIGN_1);
+ break;
case BRW_OPCODE_IF:
if (inst->src[0].file != BAD_FILE) {
/* The instruction has an embedded compare (only allowed on gen6) */
- assert(intel->gen == 6);
+ assert(brw->gen == 6);
gen6_IF(p, inst->conditional_mod, src[0], src[1]);
} else {
brw_IF(p, dispatch_width == 16 ? BRW_EXECUTE_16 : BRW_EXECUTE_8);
break;
case BRW_OPCODE_CONTINUE:
/* FINISHME: We need to write the loop instruction support still. */
- if (intel->gen >= 6)
+ if (brw->gen >= 6)
gen6_CONT(p);
else
brw_CONT(p);
case SHADER_OPCODE_LOG2:
case SHADER_OPCODE_SIN:
case SHADER_OPCODE_COS:
- if (intel->gen >= 7) {
+ if (brw->gen >= 7) {
generate_math1_gen7(inst, dst, src[0]);
- } else if (intel->gen == 6) {
+ } else if (brw->gen == 6) {
generate_math1_gen6(inst, dst, src[0]);
+ } else if (brw->gen == 5 || brw->is_g4x) {
+ generate_math_g45(inst, dst, src[0]);
} else {
generate_math_gen4(inst, dst, src[0]);
}
case SHADER_OPCODE_INT_QUOTIENT:
case SHADER_OPCODE_INT_REMAINDER:
case SHADER_OPCODE_POW:
- if (intel->gen >= 7) {
+ if (brw->gen >= 7) {
generate_math2_gen7(inst, dst, src[0], src[1]);
- } else if (intel->gen == 6) {
+ } else if (brw->gen == 6) {
generate_math2_gen6(inst, dst, src[0], src[1]);
} else {
generate_math_gen4(inst, dst, src[0]);
case FS_OPCODE_TXB:
case SHADER_OPCODE_TXD:
case SHADER_OPCODE_TXF:
+ case SHADER_OPCODE_TXF_MS:
case SHADER_OPCODE_TXL:
case SHADER_OPCODE_TXS:
+ case SHADER_OPCODE_LOD:
generate_tex(inst, dst, src[0]);
break;
- case FS_OPCODE_DISCARD:
- generate_discard(inst);
- break;
case FS_OPCODE_DDX:
generate_ddx(inst, dst, src[0]);
break;
generate_unspill(inst, dst);
break;
- case FS_OPCODE_PULL_CONSTANT_LOAD:
- generate_pull_constant_load(inst, dst, src[0], src[1]);
+ case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD:
+ generate_uniform_pull_constant_load(inst, dst, src[0], src[1]);
+ break;
+
+ case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD_GEN7:
+ generate_uniform_pull_constant_load_gen7(inst, dst, src[0], src[1]);
+ break;
+
+ case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD:
+ generate_varying_pull_constant_load(inst, dst, src[0], src[1]);
+ break;
+
+ case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GEN7:
+ generate_varying_pull_constant_load_gen7(inst, dst, src[0], src[1]);
break;
case FS_OPCODE_FB_WRITE:
break;
case FS_OPCODE_MOV_DISPATCH_TO_FLAGS:
- generate_mov_dispatch_to_flags();
+ generate_mov_dispatch_to_flags(inst);
+ break;
+
+ case FS_OPCODE_DISCARD_JUMP:
+ generate_discard_jump(inst);
+ break;
+
+ case SHADER_OPCODE_SHADER_TIME_ADD:
+ generate_shader_time_add(inst, src[0], src[1], src[2]);
+ break;
+
+ case FS_OPCODE_SET_SIMD4X2_OFFSET:
+ generate_set_simd4x2_offset(inst, dst, src[0]);
+ break;
+
+ case FS_OPCODE_PACK_HALF_2x16_SPLIT:
+ generate_pack_half_2x16_split(inst, dst, src[0], src[1]);
+ break;
+
+ case FS_OPCODE_UNPACK_HALF_2x16_SPLIT_X:
+ case FS_OPCODE_UNPACK_HALF_2x16_SPLIT_Y:
+ generate_unpack_half_2x16_split(inst, dst, src[0]);
+ break;
+
+ case FS_OPCODE_PLACEHOLDER_HALT:
+ /* This is the place where the final HALT needs to be inserted if
+ * we've emitted any discards. If not, this will emit no code.
+ */
+ patch_discard_jumps_to_fb_writes();
break;
default: