/* Emit the fragment program instructions here.
*/
void brw_wm_emit( struct brw_wm_compile *c )
{
struct brw_compile *p = &c->func;
GLuint insn;
brw_set_compression_control(p, BRW_COMPRESSION_COMPRESSED);
/* Check if any of the payload regs need to be spilled:
*/
spill_values(c, c->payload.depth, 4);
spill_values(c, c->creg, c->nr_creg);
spill_values(c, c->payload.input_interp, FRAG_ATTRIB_MAX);
for (insn = 0; insn < c->nr_insns; insn++) {
struct brw_wm_instruction *inst = &c->instruction[insn];
struct brw_reg args[3][4], dst[4];
GLuint i, dst_flags;
/* Get argument regs:
*/
for (i = 0; i < 3; i++)
get_argument_regs(c, inst->src[i], args[i]);
/* Get dest regs:
*/
for (i = 0; i < 4; i++)
if (inst->dst[i])
dst[i] = inst->dst[i]->hw_reg;
else
dst[i] = brw_null_reg();
/* Flags
*/
dst_flags = inst->writemask;
if (inst->saturate)
dst_flags |= SATURATE;
switch (inst->opcode) {
/* Generated instructions for calculating triangle interpolants:
*/
case WM_PIXELXY:
emit_pixel_xy(p, dst, dst_flags, args[0]);
break;
case WM_DELTAXY:
emit_delta_xy(p, dst, dst_flags, args[0], args[1]);
break;
case WM_WPOSXY:
emit_wpos_xy(c, dst, dst_flags, args[0]);
break;
case WM_PIXELW:
emit_pixel_w(p, dst, dst_flags, args[0], args[1]);
break;
case WM_LINTERP:
emit_linterp(p, dst, dst_flags, args[0], args[1]);
break;
case WM_PINTERP:
emit_pinterp(p, dst, dst_flags, args[0], args[1], args[2]);
break;
case WM_CINTERP:
emit_cinterp(p, dst, dst_flags, args[0]);
break;
case WM_FB_WRITE:
emit_fb_write(c, args[0], args[1], args[2], inst->target, inst->eot);
break;
/* Straightforward arithmetic:
*/
case OPCODE_ADD:
emit_alu2(p, brw_ADD, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_FRC:
emit_alu1(p, brw_FRC, dst, dst_flags, args[0]);
break;
case OPCODE_FLR:
emit_alu1(p, brw_RNDD, dst, dst_flags, args[0]);
break;
case OPCODE_DP3: /* */
emit_dp3(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_DP4:
emit_dp4(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_DPH:
emit_dph(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_LRP: /* */
emit_lrp(p, dst, dst_flags, args[0], args[1], args[2]);
break;
case OPCODE_MAD:
emit_mad(p, dst, dst_flags, args[0], args[1], args[2]);
break;
case OPCODE_MOV:
case OPCODE_SWZ:
emit_alu1(p, brw_MOV, dst, dst_flags, args[0]);
break;
case OPCODE_MUL:
emit_alu2(p, brw_MUL, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_XPD:
emit_xpd(p, dst, dst_flags, args[0], args[1]);
break;
/* Higher math functions:
*/
case OPCODE_RCP:
emit_math1(p, BRW_MATH_FUNCTION_INV, dst, dst_flags, args[0]);
break;
case OPCODE_RSQ:
emit_math1(p, BRW_MATH_FUNCTION_RSQ, dst, dst_flags, args[0]);
break;
case OPCODE_SIN:
emit_math1(p, BRW_MATH_FUNCTION_SIN, dst, dst_flags, args[0]);
break;
case OPCODE_COS:
emit_math1(p, BRW_MATH_FUNCTION_COS, dst, dst_flags, args[0]);
break;
case OPCODE_EX2:
emit_math1(p, BRW_MATH_FUNCTION_EXP, dst, dst_flags, args[0]);
break;
case OPCODE_LG2:
emit_math1(p, BRW_MATH_FUNCTION_LOG, dst, dst_flags, args[0]);
break;
case OPCODE_SCS:
/* There is an scs math function, but it would need some
* fixup for 16-element execution.
*/
if (dst_flags & WRITEMASK_X)
emit_math1(p, BRW_MATH_FUNCTION_COS, dst, (dst_flags&SATURATE)|WRITEMASK_X, args[0]);
if (dst_flags & WRITEMASK_Y)
emit_math1(p, BRW_MATH_FUNCTION_SIN, dst+1, (dst_flags&SATURATE)|WRITEMASK_X, args[0]);
break;
case OPCODE_POW:
emit_math2(p, BRW_MATH_FUNCTION_POW, dst, dst_flags, args[0], args[1]);
break;
/* Comparisons:
*/
case OPCODE_CMP:
emit_cmp(p, dst, dst_flags, args[0], args[1], args[2]);
break;
case OPCODE_MAX:
emit_max(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_MIN:
emit_min(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_SLT:
emit_slt(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_SLE:
emit_sle(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_SGT:
emit_sgt(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_SGE:
emit_sge(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_SEQ:
emit_seq(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_SNE:
emit_sne(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_LIT:
emit_lit(p, dst, dst_flags, args[0]);
break;
/* Texturing operations:
*/
case OPCODE_TEX:
emit_tex(c, inst, dst, dst_flags, args[0]);
break;
case OPCODE_TXB:
emit_txb(c, inst, dst, dst_flags, args[0]);
break;
case OPCODE_KIL:
emit_kil(c, args[0]);
break;
default:
_mesa_printf("unsupport opcode %d in fragment program\n",
inst->opcode);
}
for (i = 0; i < 4; i++)
if (inst->dst[i] && inst->dst[i]->spill_slot)
emit_spill(c,
inst->dst[i]->hw_reg,
inst->dst[i]->spill_slot);
}
}
/* Emit the fragment program instructions here.
*/
void brw_wm_emit( struct brw_wm_compile *c )
{
struct brw_compile *p = &c->func;
struct intel_context *intel = &p->brw->intel;
GLuint insn;
brw_set_compression_control(p, BRW_COMPRESSION_COMPRESSED);
if (intel->gen >= 6)
brw_set_acc_write_control(p, 1);
/* Check if any of the payload regs need to be spilled:
*/
spill_values(c, c->payload.depth, 4);
spill_values(c, c->creg, c->nr_creg);
spill_values(c, c->payload.input_interp, FRAG_ATTRIB_MAX);
for (insn = 0; insn < c->nr_insns; insn++) {
struct brw_wm_instruction *inst = &c->instruction[insn];
struct brw_reg args[3][4], dst[4];
GLuint i, dst_flags;
/* Get argument regs:
*/
for (i = 0; i < 3; i++)
get_argument_regs(c, inst->src[i], args[i]);
/* Get dest regs:
*/
for (i = 0; i < 4; i++)
if (inst->dst[i])
dst[i] = inst->dst[i]->hw_reg;
else
dst[i] = brw_null_reg();
/* Flags
*/
dst_flags = inst->writemask;
if (inst->saturate)
dst_flags |= SATURATE;
switch (inst->opcode) {
/* Generated instructions for calculating triangle interpolants:
*/
case WM_PIXELXY:
emit_pixel_xy(c, dst, dst_flags);
break;
case WM_DELTAXY:
emit_delta_xy(p, dst, dst_flags, args[0]);
break;
case WM_WPOSXY:
emit_wpos_xy(c, dst, dst_flags, args[0]);
break;
case WM_PIXELW:
emit_pixel_w(c, dst, dst_flags, args[0], args[1]);
break;
case WM_LINTERP:
emit_linterp(p, dst, dst_flags, args[0], args[1]);
break;
case WM_PINTERP:
emit_pinterp(p, dst, dst_flags, args[0], args[1], args[2]);
break;
case WM_CINTERP:
emit_cinterp(p, dst, dst_flags, args[0]);
break;
case WM_FB_WRITE:
emit_fb_write(c, args[0], args[1], args[2], inst->target, inst->eot);
break;
case WM_FRONTFACING:
emit_frontfacing(p, dst, dst_flags);
break;
/* Straightforward arithmetic:
*/
case OPCODE_ADD:
emit_alu2(p, brw_ADD, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_FRC:
emit_alu1(p, brw_FRC, dst, dst_flags, args[0]);
break;
case OPCODE_FLR:
emit_alu1(p, brw_RNDD, dst, dst_flags, args[0]);
break;
case OPCODE_DDX:
emit_ddxy(p, dst, dst_flags, true, args[0]);
break;
case OPCODE_DDY:
emit_ddxy(p, dst, dst_flags, false, args[0]);
break;
case OPCODE_DP2:
emit_dp2(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_DP3:
emit_dp3(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_DP4:
emit_dp4(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_DPH:
emit_dph(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_TRUNC:
for (i = 0; i < 4; i++) {
if (dst_flags & (1<<i)) {
brw_RNDZ(p, dst[i], args[0][i]);
}
}
break;
case OPCODE_LRP:
emit_lrp(p, dst, dst_flags, args[0], args[1], args[2]);
break;
case OPCODE_MAD:
emit_mad(p, dst, dst_flags, args[0], args[1], args[2]);
break;
case OPCODE_MOV:
case OPCODE_SWZ:
emit_alu1(p, brw_MOV, dst, dst_flags, args[0]);
break;
case OPCODE_MUL:
emit_alu2(p, brw_MUL, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_XPD:
emit_xpd(p, dst, dst_flags, args[0], args[1]);
break;
/* Higher math functions:
*/
case OPCODE_RCP:
emit_math1(c, BRW_MATH_FUNCTION_INV, dst, dst_flags, args[0]);
break;
case OPCODE_RSQ:
emit_math1(c, BRW_MATH_FUNCTION_RSQ, dst, dst_flags, args[0]);
break;
case OPCODE_SIN:
emit_math1(c, BRW_MATH_FUNCTION_SIN, dst, dst_flags, args[0]);
break;
case OPCODE_COS:
emit_math1(c, BRW_MATH_FUNCTION_COS, dst, dst_flags, args[0]);
break;
case OPCODE_EX2:
emit_math1(c, BRW_MATH_FUNCTION_EXP, dst, dst_flags, args[0]);
break;
case OPCODE_LG2:
emit_math1(c, BRW_MATH_FUNCTION_LOG, dst, dst_flags, args[0]);
break;
case OPCODE_SCS:
/* There is an scs math function, but it would need some
* fixup for 16-element execution.
*/
if (dst_flags & WRITEMASK_X)
emit_math1(c, BRW_MATH_FUNCTION_COS, dst, (dst_flags&SATURATE)|WRITEMASK_X, args[0]);
if (dst_flags & WRITEMASK_Y)
emit_math1(c, BRW_MATH_FUNCTION_SIN, dst+1, (dst_flags&SATURATE)|WRITEMASK_X, args[0]);
break;
case OPCODE_POW:
emit_math2(c, BRW_MATH_FUNCTION_POW, dst, dst_flags, args[0], args[1]);
break;
/* Comparisons:
*/
case OPCODE_CMP:
emit_cmp(p, dst, dst_flags, args[0], args[1], args[2]);
break;
case OPCODE_MAX:
emit_max(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_MIN:
emit_min(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_SLT:
emit_slt(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_SLE:
emit_sle(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_SGT:
emit_sgt(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_SGE:
emit_sge(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_SEQ:
emit_seq(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_SNE:
emit_sne(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_SSG:
emit_sign(p, dst, dst_flags, args[0]);
break;
case OPCODE_LIT:
emit_lit(c, dst, dst_flags, args[0]);
break;
/* Texturing operations:
*/
case OPCODE_TEX:
emit_tex(c, dst, dst_flags, args[0], c->payload.depth[0].hw_reg,
inst->tex_idx, inst->tex_unit,
inst->tex_shadow);
break;
case OPCODE_TXB:
emit_txb(c, dst, dst_flags, args[0], c->payload.depth[0].hw_reg,
inst->tex_idx, inst->tex_unit);
break;
case OPCODE_KIL:
emit_kil(c, args[0]);
break;
default:
printf("Unsupported opcode %i (%s) in fragment shader\n",
inst->opcode, inst->opcode < MAX_OPCODE ?
_mesa_opcode_string(inst->opcode) :
"unknown");
}
for (i = 0; i < 4; i++)
if (inst->dst[i] && inst->dst[i]->spill_slot)
emit_spill(c,
inst->dst[i]->hw_reg,
inst->dst[i]->spill_slot);
}
/* Only properly tested on ILK */
if (p->brw->intel.gen == 5) {
brw_remove_duplicate_mrf_moves(p);
if (c->dispatch_width == 16)
brw_remove_grf_to_mrf_moves(p);
}
if (unlikely(INTEL_DEBUG & DEBUG_WM)) {
int i;
printf("wm-native:\n");
for (i = 0; i < p->nr_insn; i++)
brw_disasm(stdout, &p->store[i], p->brw->intel.gen);
printf("\n");
}
}
static void brw_wm_emit_glsl(struct brw_context *brw, struct brw_wm_compile *c)
{
#define MAX_IFSN 32
#define MAX_LOOP_DEPTH 32
struct brw_instruction *if_inst[MAX_IFSN], *loop_inst[MAX_LOOP_DEPTH];
struct brw_instruction *inst0, *inst1;
int i, if_insn = 0, loop_insn = 0;
struct brw_compile *p = &c->func;
struct brw_indirect stack_index = brw_indirect(0, 0);
c->reg_index = 0;
prealloc_reg(c);
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
brw_MOV(p, get_addr_reg(stack_index), brw_address(c->stack));
for (i = 0; i < c->nr_fp_insns; i++) {
struct prog_instruction *inst = &c->prog_instructions[i];
struct prog_instruction *orig_inst;
if ((orig_inst = inst->Data) != 0)
orig_inst->Data = current_insn(p);
if (inst->CondUpdate)
brw_set_conditionalmod(p, BRW_CONDITIONAL_NZ);
else
brw_set_conditionalmod(p, BRW_CONDITIONAL_NONE);
switch (inst->Opcode) {
case WM_PIXELXY:
emit_pixel_xy(c, inst);
break;
case WM_DELTAXY:
emit_delta_xy(c, inst);
break;
case WM_PIXELW:
emit_pixel_w(c, inst);
break;
case WM_LINTERP:
emit_linterp(c, inst);
break;
case WM_PINTERP:
emit_pinterp(c, inst);
break;
case WM_CINTERP:
emit_cinterp(c, inst);
break;
case WM_WPOSXY:
emit_wpos_xy(c, inst);
break;
case WM_FB_WRITE:
emit_fb_write(c, inst);
break;
case OPCODE_ABS:
emit_abs(c, inst);
break;
case OPCODE_ADD:
emit_add(c, inst);
break;
case OPCODE_SUB:
emit_sub(c, inst);
break;
case OPCODE_FRC:
emit_frc(c, inst);
break;
case OPCODE_FLR:
emit_flr(c, inst);
break;
case OPCODE_LRP:
emit_lrp(c, inst);
break;
case OPCODE_INT:
emit_int(c, inst);
break;
case OPCODE_MOV:
emit_mov(c, inst);
break;
case OPCODE_DP3:
emit_dp3(c, inst);
break;
case OPCODE_DP4:
emit_dp4(c, inst);
break;
case OPCODE_XPD:
emit_xpd(c, inst);
break;
case OPCODE_DPH:
emit_dph(c, inst);
break;
case OPCODE_RCP:
emit_rcp(c, inst);
break;
case OPCODE_RSQ:
emit_rsq(c, inst);
break;
case OPCODE_SIN:
emit_sin(c, inst);
break;
case OPCODE_COS:
emit_cos(c, inst);
break;
case OPCODE_EX2:
emit_ex2(c, inst);
break;
case OPCODE_LG2:
emit_lg2(c, inst);
break;
case OPCODE_MAX:
emit_max(c, inst);
break;
case OPCODE_MIN:
emit_min(c, inst);
break;
case OPCODE_DDX:
emit_ddx(c, inst);
break;
case OPCODE_DDY:
emit_ddy(c, inst);
break;
case OPCODE_SLT:
emit_slt(c, inst);
break;
case OPCODE_SLE:
emit_sle(c, inst);
break;
case OPCODE_SGT:
emit_sgt(c, inst);
break;
case OPCODE_SGE:
emit_sge(c, inst);
break;
case OPCODE_SEQ:
emit_seq(c, inst);
break;
case OPCODE_SNE:
emit_sne(c, inst);
break;
case OPCODE_MUL:
emit_mul(c, inst);
break;
case OPCODE_POW:
emit_pow(c, inst);
break;
case OPCODE_MAD:
emit_mad(c, inst);
break;
case OPCODE_TEX:
emit_tex(c, inst);
break;
case OPCODE_TXB:
emit_txb(c, inst);
break;
case OPCODE_KIL_NV:
emit_kil(c);
break;
case OPCODE_IF:
assert(if_insn < MAX_IFSN);
if_inst[if_insn++] = brw_IF(p, BRW_EXECUTE_8);
break;
case OPCODE_ELSE:
if_inst[if_insn-1] = brw_ELSE(p, if_inst[if_insn-1]);
break;
case OPCODE_ENDIF:
assert(if_insn > 0);
brw_ENDIF(p, if_inst[--if_insn]);
break;
case OPCODE_BGNSUB:
case OPCODE_ENDSUB:
break;
case OPCODE_CAL:
brw_push_insn_state(p);
brw_set_mask_control(p, BRW_MASK_DISABLE);
brw_set_access_mode(p, BRW_ALIGN_1);
brw_ADD(p, deref_1ud(stack_index, 0), brw_ip_reg(), brw_imm_d(3*16));
brw_set_access_mode(p, BRW_ALIGN_16);
brw_ADD(p, get_addr_reg(stack_index),
get_addr_reg(stack_index), brw_imm_d(4));
orig_inst = inst->Data;
orig_inst->Data = &p->store[p->nr_insn];
brw_ADD(p, brw_ip_reg(), brw_ip_reg(), brw_imm_d(1*16));
brw_pop_insn_state(p);
break;
case OPCODE_RET:
brw_push_insn_state(p);
brw_set_mask_control(p, BRW_MASK_DISABLE);
brw_ADD(p, get_addr_reg(stack_index),
get_addr_reg(stack_index), brw_imm_d(-4));
brw_set_access_mode(p, BRW_ALIGN_1);
brw_MOV(p, brw_ip_reg(), deref_1ud(stack_index, 0));
brw_set_access_mode(p, BRW_ALIGN_16);
brw_pop_insn_state(p);
break;
case OPCODE_BGNLOOP:
loop_inst[loop_insn++] = brw_DO(p, BRW_EXECUTE_8);
break;
case OPCODE_BRK:
brw_BREAK(p);
brw_set_predicate_control(p, BRW_PREDICATE_NONE);
break;
case OPCODE_CONT:
brw_CONT(p);
brw_set_predicate_control(p, BRW_PREDICATE_NONE);
break;
case OPCODE_ENDLOOP:
loop_insn--;
inst0 = inst1 = brw_WHILE(p, loop_inst[loop_insn]);
/* patch all the BREAK instructions from
last BEGINLOOP */
while (inst0 > loop_inst[loop_insn]) {
inst0--;
if (inst0->header.opcode == BRW_OPCODE_BREAK) {
inst0->bits3.if_else.jump_count = inst1 - inst0 + 1;
inst0->bits3.if_else.pop_count = 0;
} else if (inst0->header.opcode == BRW_OPCODE_CONTINUE) {
inst0->bits3.if_else.jump_count = inst1 - inst0;
inst0->bits3.if_else.pop_count = 0;
}
}
break;
default:
_mesa_printf("unsupported IR in fragment shader %d\n",
inst->Opcode);
}
if (inst->CondUpdate)
brw_set_predicate_control(p, BRW_PREDICATE_NORMAL);
else
brw_set_predicate_control(p, BRW_PREDICATE_NONE);
}
post_wm_emit(c);
for (i = 0; i < c->fp->program.Base.NumInstructions; i++)
c->fp->program.Base.Instructions[i].Data = NULL;
}
