void fd5_program_emit(struct fd_context *ctx, struct fd_ringbuffer *ring, struct fd5_emit *emit) { struct stage s[MAX_STAGES]; uint32_t pos_regid, psize_regid, color_regid[8]; uint32_t face_regid, coord_regid, zwcoord_regid; uint32_t vcoord_regid, vertex_regid, instance_regid; enum a3xx_threadsize fssz; uint8_t psize_loc = ~0; int i, j; setup_stages(emit, s); fssz = (s[FS].i->max_reg >= 24) ? TWO_QUADS : FOUR_QUADS; pos_regid = ir3_find_output_regid(s[VS].v, VARYING_SLOT_POS); psize_regid = ir3_find_output_regid(s[VS].v, VARYING_SLOT_PSIZ); vertex_regid = ir3_find_sysval_regid(s[VS].v, SYSTEM_VALUE_VERTEX_ID_ZERO_BASE); instance_regid = ir3_find_sysval_regid(s[VS].v, SYSTEM_VALUE_INSTANCE_ID); if (s[FS].v->color0_mrt) { color_regid[0] = color_regid[1] = color_regid[2] = color_regid[3] = color_regid[4] = color_regid[5] = color_regid[6] = color_regid[7] = ir3_find_output_regid(s[FS].v, FRAG_RESULT_COLOR); } else { color_regid[0] = ir3_find_output_regid(s[FS].v, FRAG_RESULT_DATA0); color_regid[1] = ir3_find_output_regid(s[FS].v, FRAG_RESULT_DATA1); color_regid[2] = ir3_find_output_regid(s[FS].v, FRAG_RESULT_DATA2); color_regid[3] = ir3_find_output_regid(s[FS].v, FRAG_RESULT_DATA3); color_regid[4] = ir3_find_output_regid(s[FS].v, FRAG_RESULT_DATA4); color_regid[5] = ir3_find_output_regid(s[FS].v, FRAG_RESULT_DATA5); color_regid[6] = ir3_find_output_regid(s[FS].v, FRAG_RESULT_DATA6); color_regid[7] = ir3_find_output_regid(s[FS].v, FRAG_RESULT_DATA7); } /* TODO get these dynamically: */ face_regid = s[FS].v->frag_face ? regid(0,0) : regid(63,0); coord_regid = s[FS].v->frag_coord ? regid(0,0) : regid(63,0); zwcoord_regid = s[FS].v->frag_coord ? regid(0,2) : regid(63,0); vcoord_regid = (s[FS].v->total_in > 0) ? s[FS].v->pos_regid : regid(63,0); /* we could probably divide this up into things that need to be * emitted if frag-prog is dirty vs if vert-prog is dirty.. */ OUT_PKT4(ring, REG_A5XX_HLSQ_VS_CONFIG, 5); OUT_RING(ring, A5XX_HLSQ_VS_CONFIG_CONSTOBJECTOFFSET(s[VS].constoff) | A5XX_HLSQ_VS_CONFIG_SHADEROBJOFFSET(s[VS].instroff) | COND(s[VS].v, A5XX_HLSQ_VS_CONFIG_ENABLED)); OUT_RING(ring, A5XX_HLSQ_FS_CONFIG_CONSTOBJECTOFFSET(s[FS].constoff) | A5XX_HLSQ_FS_CONFIG_SHADEROBJOFFSET(s[FS].instroff) | COND(s[FS].v, A5XX_HLSQ_FS_CONFIG_ENABLED)); OUT_RING(ring, A5XX_HLSQ_HS_CONFIG_CONSTOBJECTOFFSET(s[HS].constoff) | A5XX_HLSQ_HS_CONFIG_SHADEROBJOFFSET(s[HS].instroff) | COND(s[HS].v, A5XX_HLSQ_HS_CONFIG_ENABLED)); OUT_RING(ring, A5XX_HLSQ_DS_CONFIG_CONSTOBJECTOFFSET(s[DS].constoff) | A5XX_HLSQ_DS_CONFIG_SHADEROBJOFFSET(s[DS].instroff) | COND(s[DS].v, A5XX_HLSQ_DS_CONFIG_ENABLED)); OUT_RING(ring, A5XX_HLSQ_GS_CONFIG_CONSTOBJECTOFFSET(s[GS].constoff) | A5XX_HLSQ_GS_CONFIG_SHADEROBJOFFSET(s[GS].instroff) | COND(s[GS].v, A5XX_HLSQ_GS_CONFIG_ENABLED)); OUT_PKT4(ring, REG_A5XX_HLSQ_CS_CONFIG, 1); OUT_RING(ring, 0x00000000); OUT_PKT4(ring, REG_A5XX_HLSQ_VS_CNTL, 5); OUT_RING(ring, A5XX_HLSQ_VS_CNTL_INSTRLEN(s[VS].instrlen) | COND(s[VS].v && s[VS].v->has_ssbo, A5XX_HLSQ_VS_CNTL_SSBO_ENABLE)); OUT_RING(ring, A5XX_HLSQ_FS_CNTL_INSTRLEN(s[FS].instrlen) | COND(s[FS].v && s[FS].v->has_ssbo, A5XX_HLSQ_FS_CNTL_SSBO_ENABLE)); OUT_RING(ring, A5XX_HLSQ_HS_CNTL_INSTRLEN(s[HS].instrlen) | COND(s[HS].v && s[HS].v->has_ssbo, A5XX_HLSQ_HS_CNTL_SSBO_ENABLE)); OUT_RING(ring, A5XX_HLSQ_DS_CNTL_INSTRLEN(s[DS].instrlen) | COND(s[DS].v && s[DS].v->has_ssbo, A5XX_HLSQ_DS_CNTL_SSBO_ENABLE)); OUT_RING(ring, A5XX_HLSQ_GS_CNTL_INSTRLEN(s[GS].instrlen) | COND(s[GS].v && s[GS].v->has_ssbo, A5XX_HLSQ_GS_CNTL_SSBO_ENABLE)); OUT_PKT4(ring, REG_A5XX_SP_VS_CONFIG, 5); OUT_RING(ring, A5XX_SP_VS_CONFIG_CONSTOBJECTOFFSET(s[VS].constoff) | A5XX_SP_VS_CONFIG_SHADEROBJOFFSET(s[VS].instroff) | COND(s[VS].v, A5XX_SP_VS_CONFIG_ENABLED)); OUT_RING(ring, A5XX_SP_FS_CONFIG_CONSTOBJECTOFFSET(s[FS].constoff) | A5XX_SP_FS_CONFIG_SHADEROBJOFFSET(s[FS].instroff) | COND(s[FS].v, A5XX_SP_FS_CONFIG_ENABLED)); OUT_RING(ring, A5XX_SP_HS_CONFIG_CONSTOBJECTOFFSET(s[HS].constoff) | A5XX_SP_HS_CONFIG_SHADEROBJOFFSET(s[HS].instroff) | COND(s[HS].v, A5XX_SP_HS_CONFIG_ENABLED)); OUT_RING(ring, A5XX_SP_DS_CONFIG_CONSTOBJECTOFFSET(s[DS].constoff) | A5XX_SP_DS_CONFIG_SHADEROBJOFFSET(s[DS].instroff) | COND(s[DS].v, A5XX_SP_DS_CONFIG_ENABLED)); OUT_RING(ring, A5XX_SP_GS_CONFIG_CONSTOBJECTOFFSET(s[GS].constoff) | A5XX_SP_GS_CONFIG_SHADEROBJOFFSET(s[GS].instroff) | COND(s[GS].v, A5XX_SP_GS_CONFIG_ENABLED)); OUT_PKT4(ring, REG_A5XX_SP_CS_CONFIG, 1); OUT_RING(ring, 0x00000000); OUT_PKT4(ring, REG_A5XX_HLSQ_VS_CONSTLEN, 2); OUT_RING(ring, s[VS].constlen); /* HLSQ_VS_CONSTLEN */ OUT_RING(ring, s[VS].instrlen); /* HLSQ_VS_INSTRLEN */ OUT_PKT4(ring, REG_A5XX_HLSQ_FS_CONSTLEN, 2); OUT_RING(ring, s[FS].constlen); /* HLSQ_FS_CONSTLEN */ OUT_RING(ring, s[FS].instrlen); /* HLSQ_FS_INSTRLEN */ OUT_PKT4(ring, REG_A5XX_HLSQ_HS_CONSTLEN, 2); OUT_RING(ring, s[HS].constlen); /* HLSQ_HS_CONSTLEN */ OUT_RING(ring, s[HS].instrlen); /* HLSQ_HS_INSTRLEN */ OUT_PKT4(ring, REG_A5XX_HLSQ_DS_CONSTLEN, 2); OUT_RING(ring, s[DS].constlen); /* HLSQ_DS_CONSTLEN */ OUT_RING(ring, s[DS].instrlen); /* HLSQ_DS_INSTRLEN */ OUT_PKT4(ring, REG_A5XX_HLSQ_GS_CONSTLEN, 2); OUT_RING(ring, s[GS].constlen); /* HLSQ_GS_CONSTLEN */ OUT_RING(ring, s[GS].instrlen); /* HLSQ_GS_INSTRLEN */ OUT_PKT4(ring, REG_A5XX_HLSQ_CS_CONSTLEN, 2); OUT_RING(ring, 0x00000000); /* HLSQ_CS_CONSTLEN */ OUT_RING(ring, 0x00000000); /* HLSQ_CS_INSTRLEN */ OUT_PKT4(ring, REG_A5XX_SP_VS_CTRL_REG0, 1); OUT_RING(ring, A5XX_SP_VS_CTRL_REG0_HALFREGFOOTPRINT(s[VS].i->max_half_reg + 1) | A5XX_SP_VS_CTRL_REG0_FULLREGFOOTPRINT(s[VS].i->max_reg + 1) | 0x6 | /* XXX seems to be always set? */ A5XX_SP_VS_CTRL_REG0_BRANCHSTACK(0x3) | // XXX need to figure this out somehow.. COND(s[VS].v->has_samp, A5XX_SP_VS_CTRL_REG0_PIXLODENABLE)); struct ir3_shader_linkage l = {0}; ir3_link_shaders(&l, s[VS].v, s[FS].v); if ((s[VS].v->shader->stream_output.num_outputs > 0) && !emit->key.binning_pass) link_stream_out(&l, s[VS].v); BITSET_DECLARE(varbs, 128) = {0}; uint32_t *varmask = (uint32_t *)varbs; for (i = 0; i < l.cnt; i++) for (j = 0; j < util_last_bit(l.var[i].compmask); j++) BITSET_SET(varbs, l.var[i].loc + j); OUT_PKT4(ring, REG_A5XX_VPC_VAR_DISABLE(0), 4); OUT_RING(ring, ~varmask[0]); /* VPC_VAR[0].DISABLE */ OUT_RING(ring, ~varmask[1]); /* VPC_VAR[1].DISABLE */ OUT_RING(ring, ~varmask[2]); /* VPC_VAR[2].DISABLE */ OUT_RING(ring, ~varmask[3]); /* VPC_VAR[3].DISABLE */ /* a5xx appends pos/psize to end of the linkage map: */ if (pos_regid != regid(63,0)) ir3_link_add(&l, pos_regid, 0xf, l.max_loc); if (psize_regid != regid(63,0)) { psize_loc = l.max_loc; ir3_link_add(&l, psize_regid, 0x1, l.max_loc); } if ((s[VS].v->shader->stream_output.num_outputs > 0) && !emit->key.binning_pass) { emit_stream_out(ring, s[VS].v, &l); OUT_PKT4(ring, REG_A5XX_VPC_SO_OVERRIDE, 1); OUT_RING(ring, 0x00000000); } else { OUT_PKT4(ring, REG_A5XX_VPC_SO_OVERRIDE, 1); OUT_RING(ring, A5XX_VPC_SO_OVERRIDE_SO_DISABLE); } for (i = 0, j = 0; (i < 16) && (j < l.cnt); i++) { uint32_t reg = 0; OUT_PKT4(ring, REG_A5XX_SP_VS_OUT_REG(i), 1); reg |= A5XX_SP_VS_OUT_REG_A_REGID(l.var[j].regid); reg |= A5XX_SP_VS_OUT_REG_A_COMPMASK(l.var[j].compmask); j++; reg |= A5XX_SP_VS_OUT_REG_B_REGID(l.var[j].regid); reg |= A5XX_SP_VS_OUT_REG_B_COMPMASK(l.var[j].compmask); j++; OUT_RING(ring, reg); } for (i = 0, j = 0; (i < 8) && (j < l.cnt); i++) { uint32_t reg = 0; OUT_PKT4(ring, REG_A5XX_SP_VS_VPC_DST_REG(i), 1); reg |= A5XX_SP_VS_VPC_DST_REG_OUTLOC0(l.var[j++].loc); reg |= A5XX_SP_VS_VPC_DST_REG_OUTLOC1(l.var[j++].loc); reg |= A5XX_SP_VS_VPC_DST_REG_OUTLOC2(l.var[j++].loc); reg |= A5XX_SP_VS_VPC_DST_REG_OUTLOC3(l.var[j++].loc); OUT_RING(ring, reg); } OUT_PKT4(ring, REG_A5XX_SP_VS_OBJ_START_LO, 2); OUT_RELOC(ring, s[VS].v->bo, 0, 0, 0); /* SP_VS_OBJ_START_LO/HI */ if (s[VS].instrlen) fd5_emit_shader(ring, s[VS].v); // TODO depending on other bits in this reg (if any) set somewhere else? OUT_PKT4(ring, REG_A5XX_PC_PRIM_VTX_CNTL, 1); OUT_RING(ring, COND(s[VS].v->writes_psize, A5XX_PC_PRIM_VTX_CNTL_PSIZE)); OUT_PKT4(ring, REG_A5XX_SP_PRIMITIVE_CNTL, 1); OUT_RING(ring, A5XX_SP_PRIMITIVE_CNTL_VSOUT(l.cnt)); OUT_PKT4(ring, REG_A5XX_VPC_CNTL_0, 1); OUT_RING(ring, A5XX_VPC_CNTL_0_STRIDE_IN_VPC(l.max_loc) | COND(s[FS].v->total_in > 0, A5XX_VPC_CNTL_0_VARYING) | COND(s[FS].v->frag_coord, A5XX_VPC_CNTL_0_VARYING) | 0x10000); // XXX fd5_context(ctx)->max_loc = l.max_loc; if (emit->key.binning_pass) { OUT_PKT4(ring, REG_A5XX_SP_FS_OBJ_START_LO, 2); OUT_RING(ring, 0x00000000); /* SP_FS_OBJ_START_LO */ OUT_RING(ring, 0x00000000); /* SP_FS_OBJ_START_HI */ } else { OUT_PKT4(ring, REG_A5XX_SP_FS_OBJ_START_LO, 2); OUT_RELOC(ring, s[FS].v->bo, 0, 0, 0); /* SP_FS_OBJ_START_LO/HI */ } OUT_PKT4(ring, REG_A5XX_HLSQ_CONTROL_0_REG, 5); OUT_RING(ring, A5XX_HLSQ_CONTROL_0_REG_FSTHREADSIZE(fssz) | A5XX_HLSQ_CONTROL_0_REG_CSTHREADSIZE(TWO_QUADS) | 0x00000880); /* XXX HLSQ_CONTROL_0 */ OUT_RING(ring, A5XX_HLSQ_CONTROL_1_REG_PRIMALLOCTHRESHOLD(63)); OUT_RING(ring, A5XX_HLSQ_CONTROL_2_REG_FACEREGID(face_regid) | 0xfcfcfc00); /* XXX */ OUT_RING(ring, A5XX_HLSQ_CONTROL_3_REG_FRAGCOORDXYREGID(vcoord_regid) | 0xfcfcfc00); /* XXX */ OUT_RING(ring, A5XX_HLSQ_CONTROL_4_REG_XYCOORDREGID(coord_regid) | A5XX_HLSQ_CONTROL_4_REG_ZWCOORDREGID(zwcoord_regid) | 0x0000fcfc); /* XXX */ OUT_PKT4(ring, REG_A5XX_SP_FS_CTRL_REG0, 1); OUT_RING(ring, COND(s[FS].v->total_in > 0, A5XX_SP_FS_CTRL_REG0_VARYING) | COND(s[FS].v->frag_coord, A5XX_SP_FS_CTRL_REG0_VARYING) | 0x40006 | /* XXX set pretty much everywhere */ A5XX_SP_FS_CTRL_REG0_THREADSIZE(fssz) | A5XX_SP_FS_CTRL_REG0_HALFREGFOOTPRINT(s[FS].i->max_half_reg + 1) | A5XX_SP_FS_CTRL_REG0_FULLREGFOOTPRINT(s[FS].i->max_reg + 1) | A5XX_SP_FS_CTRL_REG0_BRANCHSTACK(0x3) | // XXX need to figure this out somehow.. COND(s[FS].v->has_samp, A5XX_SP_FS_CTRL_REG0_PIXLODENABLE)); OUT_PKT4(ring, REG_A5XX_HLSQ_UPDATE_CNTL, 1); OUT_RING(ring, 0x020fffff); /* XXX */ OUT_PKT4(ring, REG_A5XX_VPC_GS_SIV_CNTL, 1); OUT_RING(ring, 0x0000ffff); /* XXX */ OUT_PKT4(ring, REG_A5XX_SP_SP_CNTL, 1); OUT_RING(ring, 0x00000010); /* XXX */ OUT_PKT4(ring, REG_A5XX_GRAS_CNTL, 1); OUT_RING(ring, COND(s[FS].v->total_in > 0, A5XX_GRAS_CNTL_VARYING) | COND(s[FS].v->frag_coord, A5XX_GRAS_CNTL_XCOORD | A5XX_GRAS_CNTL_YCOORD | A5XX_GRAS_CNTL_ZCOORD | A5XX_GRAS_CNTL_WCOORD | A5XX_GRAS_CNTL_UNK3) | COND(s[FS].v->frag_face, A5XX_GRAS_CNTL_UNK3)); OUT_PKT4(ring, REG_A5XX_RB_RENDER_CONTROL0, 2); OUT_RING(ring, COND(s[FS].v->total_in > 0, A5XX_RB_RENDER_CONTROL0_VARYING) | COND(s[FS].v->frag_coord, A5XX_RB_RENDER_CONTROL0_XCOORD | A5XX_RB_RENDER_CONTROL0_YCOORD | A5XX_RB_RENDER_CONTROL0_ZCOORD | A5XX_RB_RENDER_CONTROL0_WCOORD | A5XX_RB_RENDER_CONTROL0_UNK3) | COND(s[FS].v->frag_face, A5XX_RB_RENDER_CONTROL0_UNK3)); OUT_RING(ring, COND(s[FS].v->frag_face, A5XX_RB_RENDER_CONTROL1_FACENESS)); OUT_PKT4(ring, REG_A5XX_SP_FS_OUTPUT_REG(0), 8); for (i = 0; i < 8; i++) { OUT_RING(ring, A5XX_SP_FS_OUTPUT_REG_REGID(color_regid[i]) | COND(emit->key.half_precision, A5XX_SP_FS_OUTPUT_REG_HALF_PRECISION)); } OUT_PKT4(ring, REG_A5XX_VPC_PACK, 1); OUT_RING(ring, A5XX_VPC_PACK_NUMNONPOSVAR(s[FS].v->total_in) | A5XX_VPC_PACK_PSIZELOC(psize_loc)); if (!emit->key.binning_pass) { uint32_t vinterp[8], vpsrepl[8]; memset(vinterp, 0, sizeof(vinterp)); memset(vpsrepl, 0, sizeof(vpsrepl)); /* looks like we need to do int varyings in the frag * shader on a5xx (no flatshad reg? or a420.0 bug?): * * (sy)(ss)nop * (sy)ldlv.u32 r0.x,l[r0.x], 1 * ldlv.u32 r0.y,l[r0.x+1], 1 * (ss)bary.f (ei)r63.x, 0, r0.x * (ss)(rpt1)cov.s32f16 hr0.x, (r)r0.x * (rpt5)nop * sam (f16)(xyzw)hr0.x, hr0.x, s#0, t#0 * * Possibly on later a5xx variants we'll be able to use * something like the code below instead of workaround * in the shader: */ /* figure out VARYING_INTERP / VARYING_PS_REPL register values: */ for (j = -1; (j = ir3_next_varying(s[FS].v, j)) < (int)s[FS].v->inputs_count; ) { /* NOTE: varyings are packed, so if compmask is 0xb * then first, third, and fourth component occupy * three consecutive varying slots: */ unsigned compmask = s[FS].v->inputs[j].compmask; uint32_t inloc = s[FS].v->inputs[j].inloc; if ((s[FS].v->inputs[j].interpolate == INTERP_MODE_FLAT) || (s[FS].v->inputs[j].rasterflat && emit->rasterflat)) { uint32_t loc = inloc; for (i = 0; i < 4; i++) { if (compmask & (1 << i)) { vinterp[loc / 16] |= 1 << ((loc % 16) * 2); //flatshade[loc / 32] |= 1 << (loc % 32); loc++; } } } gl_varying_slot slot = s[FS].v->inputs[j].slot; /* since we don't enable PIPE_CAP_TGSI_TEXCOORD: */ if (slot >= VARYING_SLOT_VAR0) { unsigned texmask = 1 << (slot - VARYING_SLOT_VAR0); /* Replace the .xy coordinates with S/T from the point sprite. Set * interpolation bits for .zw such that they become .01 */ if (emit->sprite_coord_enable & texmask) { /* mask is two 2-bit fields, where: * '01' -> S * '10' -> T * '11' -> 1 - T (flip mode) */ unsigned mask = emit->sprite_coord_mode ? 0b1101 : 0b1001; uint32_t loc = inloc; if (compmask & 0x1) { vpsrepl[loc / 16] |= ((mask >> 0) & 0x3) << ((loc % 16) * 2); loc++; } if (compmask & 0x2) { vpsrepl[loc / 16] |= ((mask >> 2) & 0x3) << ((loc % 16) * 2); loc++; }
void fd4_program_emit(struct fd_ringbuffer *ring, struct fd4_emit *emit) { struct stage s[MAX_STAGES]; uint32_t pos_regid, posz_regid, psize_regid, color_regid; uint32_t face_regid, coord_regid, zwcoord_regid; int constmode; int i, j, k; setup_stages(emit, s); /* blob seems to always use constmode currently: */ constmode = 1; pos_regid = ir3_find_output_regid(s[VS].v, ir3_semantic_name(TGSI_SEMANTIC_POSITION, 0)); posz_regid = ir3_find_output_regid(s[FS].v, ir3_semantic_name(TGSI_SEMANTIC_POSITION, 0)); psize_regid = ir3_find_output_regid(s[VS].v, ir3_semantic_name(TGSI_SEMANTIC_PSIZE, 0)); color_regid = ir3_find_output_regid(s[FS].v, ir3_semantic_name(TGSI_SEMANTIC_COLOR, 0)); if (util_format_is_alpha(emit->pformat)) color_regid += 3; /* TODO get these dynamically: */ face_regid = s[FS].v->frag_face ? regid(0,0) : regid(63,0); coord_regid = s[FS].v->frag_coord ? regid(0,0) : regid(63,0); zwcoord_regid = s[FS].v->frag_coord ? regid(0,2) : regid(63,0); /* we could probably divide this up into things that need to be * emitted if frag-prog is dirty vs if vert-prog is dirty.. */ OUT_PKT0(ring, REG_A4XX_HLSQ_UPDATE_CONTROL, 1); OUT_RING(ring, 0x00000003); OUT_PKT0(ring, REG_A4XX_HLSQ_CONTROL_0_REG, 5); OUT_RING(ring, A4XX_HLSQ_CONTROL_0_REG_FSTHREADSIZE(FOUR_QUADS) | A4XX_HLSQ_CONTROL_0_REG_CONSTMODE(constmode) | A4XX_HLSQ_CONTROL_0_REG_FSSUPERTHREADENABLE | /* NOTE: I guess SHADERRESTART and CONSTFULLUPDATE maybe * flush some caches? I think we only need to set those * bits if we have updated const or shader.. */ A4XX_HLSQ_CONTROL_0_REG_SPSHADERRESTART | A4XX_HLSQ_CONTROL_0_REG_SPCONSTFULLUPDATE); OUT_RING(ring, A4XX_HLSQ_CONTROL_1_REG_VSTHREADSIZE(TWO_QUADS) | A4XX_HLSQ_CONTROL_1_REG_VSSUPERTHREADENABLE | A4XX_HLSQ_CONTROL_1_REG_COORDREGID(coord_regid) | A4XX_HLSQ_CONTROL_1_REG_ZWCOORDREGID(zwcoord_regid)); OUT_RING(ring, A4XX_HLSQ_CONTROL_2_REG_PRIMALLOCTHRESHOLD(63) | 0x3f3f000 | /* XXX */ A4XX_HLSQ_CONTROL_2_REG_FACEREGID(face_regid)); OUT_RING(ring, A4XX_HLSQ_CONTROL_3_REG_REGID(s[FS].v->pos_regid) | 0xfcfcfc00); OUT_RING(ring, 0x00fcfcfc); /* XXX HLSQ_CONTROL_4 */ OUT_PKT0(ring, REG_A4XX_HLSQ_VS_CONTROL_REG, 5); OUT_RING(ring, A4XX_HLSQ_VS_CONTROL_REG_CONSTLENGTH(s[VS].constlen) | A4XX_HLSQ_VS_CONTROL_REG_CONSTOBJECTOFFSET(s[VS].constoff) | A4XX_HLSQ_VS_CONTROL_REG_INSTRLENGTH(s[VS].instrlen) | A4XX_HLSQ_VS_CONTROL_REG_SHADEROBJOFFSET(s[VS].instroff)); OUT_RING(ring, A4XX_HLSQ_FS_CONTROL_REG_CONSTLENGTH(s[FS].constlen) | A4XX_HLSQ_FS_CONTROL_REG_CONSTOBJECTOFFSET(s[FS].constoff) | A4XX_HLSQ_FS_CONTROL_REG_INSTRLENGTH(s[FS].instrlen) | A4XX_HLSQ_FS_CONTROL_REG_SHADEROBJOFFSET(s[FS].instroff)); OUT_RING(ring, A4XX_HLSQ_HS_CONTROL_REG_CONSTLENGTH(s[HS].constlen) | A4XX_HLSQ_HS_CONTROL_REG_CONSTOBJECTOFFSET(s[HS].constoff) | A4XX_HLSQ_HS_CONTROL_REG_INSTRLENGTH(s[HS].instrlen) | A4XX_HLSQ_HS_CONTROL_REG_SHADEROBJOFFSET(s[HS].instroff)); OUT_RING(ring, A4XX_HLSQ_DS_CONTROL_REG_CONSTLENGTH(s[DS].constlen) | A4XX_HLSQ_DS_CONTROL_REG_CONSTOBJECTOFFSET(s[DS].constoff) | A4XX_HLSQ_DS_CONTROL_REG_INSTRLENGTH(s[DS].instrlen) | A4XX_HLSQ_DS_CONTROL_REG_SHADEROBJOFFSET(s[DS].instroff)); OUT_RING(ring, A4XX_HLSQ_GS_CONTROL_REG_CONSTLENGTH(s[GS].constlen) | A4XX_HLSQ_GS_CONTROL_REG_CONSTOBJECTOFFSET(s[GS].constoff) | A4XX_HLSQ_GS_CONTROL_REG_INSTRLENGTH(s[GS].instrlen) | A4XX_HLSQ_GS_CONTROL_REG_SHADEROBJOFFSET(s[GS].instroff)); OUT_PKT0(ring, REG_A4XX_SP_SP_CTRL_REG, 1); OUT_RING(ring, 0x140010 | /* XXX */ COND(emit->key.binning_pass, A4XX_SP_SP_CTRL_REG_BINNING_PASS)); OUT_PKT0(ring, REG_A4XX_SP_INSTR_CACHE_CTRL, 1); OUT_RING(ring, 0x7f | /* XXX */ COND(s[VS].instrlen, A4XX_SP_INSTR_CACHE_CTRL_VS_BUFFER) | COND(s[FS].instrlen, A4XX_SP_INSTR_CACHE_CTRL_FS_BUFFER) | COND(s[VS].instrlen && s[FS].instrlen, A4XX_SP_INSTR_CACHE_CTRL_INSTR_BUFFER)); OUT_PKT0(ring, REG_A4XX_SP_VS_LENGTH_REG, 1); OUT_RING(ring, s[VS].v->instrlen); /* SP_VS_LENGTH_REG */ OUT_PKT0(ring, REG_A4XX_SP_VS_CTRL_REG0, 3); OUT_RING(ring, A4XX_SP_VS_CTRL_REG0_THREADMODE(MULTI) | A4XX_SP_VS_CTRL_REG0_HALFREGFOOTPRINT(s[VS].i->max_half_reg + 1) | A4XX_SP_VS_CTRL_REG0_FULLREGFOOTPRINT(s[VS].i->max_reg + 1) | A4XX_SP_VS_CTRL_REG0_INOUTREGOVERLAP(0) | A4XX_SP_VS_CTRL_REG0_THREADSIZE(TWO_QUADS) | A4XX_SP_VS_CTRL_REG0_SUPERTHREADMODE | COND(s[VS].v->has_samp, A4XX_SP_VS_CTRL_REG0_PIXLODENABLE)); OUT_RING(ring, A4XX_SP_VS_CTRL_REG1_CONSTLENGTH(s[VS].constlen) | A4XX_SP_VS_CTRL_REG1_INITIALOUTSTANDING(s[VS].v->total_in)); OUT_RING(ring, A4XX_SP_VS_PARAM_REG_POSREGID(pos_regid) | A4XX_SP_VS_PARAM_REG_PSIZEREGID(psize_regid) | A4XX_SP_VS_PARAM_REG_TOTALVSOUTVAR(align(s[FS].v->total_in, 4) / 4)); for (i = 0, j = -1; (i < 16) && (j < (int)s[FS].v->inputs_count); i++) { uint32_t reg = 0; OUT_PKT0(ring, REG_A4XX_SP_VS_OUT_REG(i), 1); j = ir3_next_varying(s[FS].v, j); if (j < s[FS].v->inputs_count) { k = ir3_find_output(s[VS].v, s[FS].v->inputs[j].semantic); reg |= A4XX_SP_VS_OUT_REG_A_REGID(s[VS].v->outputs[k].regid); reg |= A4XX_SP_VS_OUT_REG_A_COMPMASK(s[FS].v->inputs[j].compmask); } j = ir3_next_varying(s[FS].v, j); if (j < s[FS].v->inputs_count) { k = ir3_find_output(s[VS].v, s[FS].v->inputs[j].semantic); reg |= A4XX_SP_VS_OUT_REG_B_REGID(s[VS].v->outputs[k].regid); reg |= A4XX_SP_VS_OUT_REG_B_COMPMASK(s[FS].v->inputs[j].compmask); } OUT_RING(ring, reg); } for (i = 0, j = -1; (i < 8) && (j < (int)s[FS].v->inputs_count); i++) { uint32_t reg = 0; OUT_PKT0(ring, REG_A4XX_SP_VS_VPC_DST_REG(i), 1); j = ir3_next_varying(s[FS].v, j); if (j < s[FS].v->inputs_count) reg |= A4XX_SP_VS_VPC_DST_REG_OUTLOC0(s[FS].v->inputs[j].inloc); j = ir3_next_varying(s[FS].v, j); if (j < s[FS].v->inputs_count) reg |= A4XX_SP_VS_VPC_DST_REG_OUTLOC1(s[FS].v->inputs[j].inloc); j = ir3_next_varying(s[FS].v, j); if (j < s[FS].v->inputs_count) reg |= A4XX_SP_VS_VPC_DST_REG_OUTLOC2(s[FS].v->inputs[j].inloc); j = ir3_next_varying(s[FS].v, j); if (j < s[FS].v->inputs_count) reg |= A4XX_SP_VS_VPC_DST_REG_OUTLOC3(s[FS].v->inputs[j].inloc); OUT_RING(ring, reg); } OUT_PKT0(ring, REG_A4XX_SP_VS_OBJ_OFFSET_REG, 2); OUT_RING(ring, A4XX_SP_VS_OBJ_OFFSET_REG_CONSTOBJECTOFFSET(s[VS].constoff) | A4XX_SP_VS_OBJ_OFFSET_REG_SHADEROBJOFFSET(s[VS].instroff)); OUT_RELOC(ring, s[VS].v->bo, 0, 0, 0); /* SP_VS_OBJ_START_REG */ OUT_PKT0(ring, REG_A4XX_SP_FS_LENGTH_REG, 1); OUT_RING(ring, s[FS].v->instrlen); /* SP_FS_LENGTH_REG */ OUT_PKT0(ring, REG_A4XX_SP_FS_CTRL_REG0, 2); OUT_RING(ring, A4XX_SP_FS_CTRL_REG0_THREADMODE(MULTI) | COND(s[FS].v->total_in > 0, A4XX_SP_FS_CTRL_REG0_VARYING) | A4XX_SP_FS_CTRL_REG0_HALFREGFOOTPRINT(s[FS].i->max_half_reg + 1) | A4XX_SP_FS_CTRL_REG0_FULLREGFOOTPRINT(s[FS].i->max_reg + 1) | A4XX_SP_FS_CTRL_REG0_INOUTREGOVERLAP(1) | A4XX_SP_FS_CTRL_REG0_THREADSIZE(FOUR_QUADS) | A4XX_SP_FS_CTRL_REG0_SUPERTHREADMODE | COND(s[FS].v->has_samp, A4XX_SP_FS_CTRL_REG0_PIXLODENABLE)); OUT_RING(ring, A4XX_SP_FS_CTRL_REG1_CONSTLENGTH(s[FS].constlen) | 0x80000000 | /* XXX */ COND(s[FS].v->frag_face, A4XX_SP_FS_CTRL_REG1_FACENESS) | COND(s[FS].v->total_in > 0, A4XX_SP_FS_CTRL_REG1_VARYING) | COND(s[FS].v->frag_coord, A4XX_SP_FS_CTRL_REG1_FRAGCOORD)); OUT_PKT0(ring, REG_A4XX_SP_FS_OBJ_OFFSET_REG, 2); OUT_RING(ring, A4XX_SP_FS_OBJ_OFFSET_REG_CONSTOBJECTOFFSET(s[FS].constoff) | A4XX_SP_FS_OBJ_OFFSET_REG_SHADEROBJOFFSET(s[FS].instroff)); if (emit->key.binning_pass) OUT_RING(ring, 0x00000000); else OUT_RELOC(ring, s[FS].v->bo, 0, 0, 0); /* SP_FS_OBJ_START_REG */ OUT_PKT0(ring, REG_A4XX_SP_HS_OBJ_OFFSET_REG, 1); OUT_RING(ring, A4XX_SP_HS_OBJ_OFFSET_REG_CONSTOBJECTOFFSET(s[HS].constoff) | A4XX_SP_HS_OBJ_OFFSET_REG_SHADEROBJOFFSET(s[HS].instroff)); OUT_PKT0(ring, REG_A4XX_SP_DS_OBJ_OFFSET_REG, 1); OUT_RING(ring, A4XX_SP_DS_OBJ_OFFSET_REG_CONSTOBJECTOFFSET(s[DS].constoff) | A4XX_SP_DS_OBJ_OFFSET_REG_SHADEROBJOFFSET(s[DS].instroff)); OUT_PKT0(ring, REG_A4XX_SP_GS_OBJ_OFFSET_REG, 1); OUT_RING(ring, A4XX_SP_GS_OBJ_OFFSET_REG_CONSTOBJECTOFFSET(s[GS].constoff) | A4XX_SP_GS_OBJ_OFFSET_REG_SHADEROBJOFFSET(s[GS].instroff)); OUT_PKT0(ring, REG_A4XX_RB_RENDER_CONTROL2, 1); OUT_RING(ring, A4XX_RB_RENDER_CONTROL2_MSAA_SAMPLES(0) | COND(s[FS].v->total_in > 0, A4XX_RB_RENDER_CONTROL2_VARYING) | COND(s[FS].v->frag_face, A4XX_RB_RENDER_CONTROL2_FACENESS) | COND(s[FS].v->frag_coord, A4XX_RB_RENDER_CONTROL2_XCOORD | A4XX_RB_RENDER_CONTROL2_YCOORD | // TODO enabling gl_FragCoord.z is causing lockups on 0ad (but seems // to work everywhere else). // A4XX_RB_RENDER_CONTROL2_ZCOORD | A4XX_RB_RENDER_CONTROL2_WCOORD)); OUT_PKT0(ring, REG_A4XX_RB_FS_OUTPUT_REG, 1); OUT_RING(ring, A4XX_RB_FS_OUTPUT_REG_MRT(1) | COND(s[FS].v->writes_pos, A4XX_RB_FS_OUTPUT_REG_FRAG_WRITES_Z)); OUT_PKT0(ring, REG_A4XX_SP_FS_OUTPUT_REG, 1); if (s[FS].v->writes_pos) { OUT_RING(ring, 0x00000001 | A4XX_SP_FS_OUTPUT_REG_DEPTH_ENABLE | A4XX_SP_FS_OUTPUT_REG_DEPTH_REGID(posz_regid)); } else { OUT_RING(ring, 0x00000001); } OUT_PKT0(ring, REG_A4XX_SP_FS_MRT_REG(0), 8); OUT_RING(ring, A4XX_SP_FS_MRT_REG_REGID(color_regid) | A4XX_SP_FS_MRT_REG_MRTFORMAT(emit->format) | COND(emit->key.half_precision, A4XX_SP_FS_MRT_REG_HALF_PRECISION)); OUT_RING(ring, A4XX_SP_FS_MRT_REG_REGID(0)); OUT_RING(ring, A4XX_SP_FS_MRT_REG_REGID(0)); OUT_RING(ring, A4XX_SP_FS_MRT_REG_REGID(0)); OUT_RING(ring, A4XX_SP_FS_MRT_REG_REGID(0)); OUT_RING(ring, A4XX_SP_FS_MRT_REG_REGID(0)); OUT_RING(ring, A4XX_SP_FS_MRT_REG_REGID(0)); OUT_RING(ring, A4XX_SP_FS_MRT_REG_REGID(0)); if (emit->key.binning_pass) { OUT_PKT0(ring, REG_A4XX_VPC_ATTR, 2); OUT_RING(ring, A4XX_VPC_ATTR_THRDASSIGN(1) | 0x40000000 | /* XXX */ COND(s[VS].v->writes_psize, A4XX_VPC_ATTR_PSIZE)); OUT_RING(ring, 0x00000000); } else { uint32_t vinterp[8], flatshade[2]; memset(vinterp, 0, sizeof(vinterp)); memset(flatshade, 0, sizeof(flatshade)); /* looks like we need to do int varyings in the frag * shader on a4xx (no flatshad reg? or a420.0 bug?): * * (sy)(ss)nop * (sy)ldlv.u32 r0.x,l[r0.x], 1 * ldlv.u32 r0.y,l[r0.x+1], 1 * (ss)bary.f (ei)r63.x, 0, r0.x * (ss)(rpt1)cov.s32f16 hr0.x, (r)r0.x * (rpt5)nop * sam (f16)(xyzw)hr0.x, hr0.x, s#0, t#0 * * Possibly on later a4xx variants we'll be able to use * something like the code below instead of workaround * in the shader: */ #if 0 /* figure out VARYING_INTERP / FLAT_SHAD register values: */ for (j = -1; (j = ir3_next_varying(s[FS].v, j)) < (int)s[FS].v->inputs_count; ) { uint32_t interp = s[FS].v->inputs[j].interpolate; if ((interp == TGSI_INTERPOLATE_CONSTANT) || ((interp == TGSI_INTERPOLATE_COLOR) && emit->rasterflat)) { /* TODO might be cleaner to just +8 in SP_VS_VPC_DST_REG * instead.. rather than -8 everywhere else.. */ uint32_t loc = s[FS].v->inputs[j].inloc - 8; /* currently assuming varyings aligned to 4 (not * packed): */ debug_assert((loc % 4) == 0); for (i = 0; i < 4; i++, loc++) { vinterp[loc / 16] |= 1 << ((loc % 16) * 2); flatshade[loc / 32] |= 1 << (loc % 32); } } } #endif OUT_PKT0(ring, REG_A4XX_VPC_ATTR, 2); OUT_RING(ring, A4XX_VPC_ATTR_TOTALATTR(s[FS].v->total_in) | A4XX_VPC_ATTR_THRDASSIGN(1) | COND(s[FS].v->total_in > 0, A4XX_VPC_ATTR_ENABLE) | 0x40000000 | /* XXX */ COND(s[VS].v->writes_psize, A4XX_VPC_ATTR_PSIZE)); OUT_RING(ring, A4XX_VPC_PACK_NUMFPNONPOSVAR(s[FS].v->total_in) | A4XX_VPC_PACK_NUMNONPOSVSVAR(s[FS].v->total_in)); OUT_PKT0(ring, REG_A4XX_VPC_VARYING_INTERP_MODE(0), 8); for (i = 0; i < 8; i++) OUT_RING(ring, vinterp[i]); /* VPC_VARYING_INTERP[i].MODE */ OUT_PKT0(ring, REG_A4XX_VPC_VARYING_PS_REPL_MODE(0), 8); for (i = 0; i < 8; i++) OUT_RING(ring, s[FS].v->shader->vpsrepl[i]); /* VPC_VARYING_PS_REPL[i] */ } if (s[VS].instrlen) emit_shader(ring, s[VS].v); if (!emit->key.binning_pass) if (s[FS].instrlen) emit_shader(ring, s[FS].v); }
void fd4_program_emit(struct fd_ringbuffer *ring, struct fd4_emit *emit, int nr, struct pipe_surface **bufs) { struct stage s[MAX_STAGES]; uint32_t pos_regid, posz_regid, psize_regid, color_regid[8]; uint32_t face_regid, coord_regid, zwcoord_regid; enum a3xx_threadsize fssz; int constmode; int i, j, k; debug_assert(nr <= ARRAY_SIZE(color_regid)); if (emit->key.binning_pass) nr = 0; setup_stages(emit, s); fssz = (s[FS].i->max_reg >= 24) ? TWO_QUADS : FOUR_QUADS; /* blob seems to always use constmode currently: */ constmode = 1; pos_regid = ir3_find_output_regid(s[VS].v, VARYING_SLOT_POS); if (pos_regid == regid(63, 0)) { /* hw dislikes when there is no position output, which can * happen for transform-feedback vertex shaders. Just tell * the hw to use r0.x, with whatever random value is there: */ pos_regid = regid(0, 0); } posz_regid = ir3_find_output_regid(s[FS].v, FRAG_RESULT_DEPTH); psize_regid = ir3_find_output_regid(s[VS].v, VARYING_SLOT_PSIZ); if (s[FS].v->color0_mrt) { color_regid[0] = color_regid[1] = color_regid[2] = color_regid[3] = color_regid[4] = color_regid[5] = color_regid[6] = color_regid[7] = ir3_find_output_regid(s[FS].v, FRAG_RESULT_COLOR); } else { color_regid[0] = ir3_find_output_regid(s[FS].v, FRAG_RESULT_DATA0); color_regid[1] = ir3_find_output_regid(s[FS].v, FRAG_RESULT_DATA1); color_regid[2] = ir3_find_output_regid(s[FS].v, FRAG_RESULT_DATA2); color_regid[3] = ir3_find_output_regid(s[FS].v, FRAG_RESULT_DATA3); color_regid[4] = ir3_find_output_regid(s[FS].v, FRAG_RESULT_DATA4); color_regid[5] = ir3_find_output_regid(s[FS].v, FRAG_RESULT_DATA5); color_regid[6] = ir3_find_output_regid(s[FS].v, FRAG_RESULT_DATA6); color_regid[7] = ir3_find_output_regid(s[FS].v, FRAG_RESULT_DATA7); } /* TODO get these dynamically: */ face_regid = s[FS].v->frag_face ? regid(0,0) : regid(63,0); coord_regid = s[FS].v->frag_coord ? regid(0,0) : regid(63,0); zwcoord_regid = s[FS].v->frag_coord ? regid(0,2) : regid(63,0); /* we could probably divide this up into things that need to be * emitted if frag-prog is dirty vs if vert-prog is dirty.. */ OUT_PKT0(ring, REG_A4XX_HLSQ_UPDATE_CONTROL, 1); OUT_RING(ring, 0x00000003); OUT_PKT0(ring, REG_A4XX_HLSQ_CONTROL_0_REG, 5); OUT_RING(ring, A4XX_HLSQ_CONTROL_0_REG_FSTHREADSIZE(fssz) | A4XX_HLSQ_CONTROL_0_REG_CONSTMODE(constmode) | A4XX_HLSQ_CONTROL_0_REG_FSSUPERTHREADENABLE | /* NOTE: I guess SHADERRESTART and CONSTFULLUPDATE maybe * flush some caches? I think we only need to set those * bits if we have updated const or shader.. */ A4XX_HLSQ_CONTROL_0_REG_SPSHADERRESTART | A4XX_HLSQ_CONTROL_0_REG_SPCONSTFULLUPDATE); OUT_RING(ring, A4XX_HLSQ_CONTROL_1_REG_VSTHREADSIZE(TWO_QUADS) | A4XX_HLSQ_CONTROL_1_REG_VSSUPERTHREADENABLE | A4XX_HLSQ_CONTROL_1_REG_COORDREGID(coord_regid) | A4XX_HLSQ_CONTROL_1_REG_ZWCOORDREGID(zwcoord_regid)); OUT_RING(ring, A4XX_HLSQ_CONTROL_2_REG_PRIMALLOCTHRESHOLD(63) | 0x3f3f000 | /* XXX */ A4XX_HLSQ_CONTROL_2_REG_FACEREGID(face_regid)); OUT_RING(ring, A4XX_HLSQ_CONTROL_3_REG_REGID(s[FS].v->pos_regid) | 0xfcfcfc00); OUT_RING(ring, 0x00fcfcfc); /* XXX HLSQ_CONTROL_4 */ OUT_PKT0(ring, REG_A4XX_HLSQ_VS_CONTROL_REG, 5); OUT_RING(ring, A4XX_HLSQ_VS_CONTROL_REG_CONSTLENGTH(s[VS].constlen) | A4XX_HLSQ_VS_CONTROL_REG_CONSTOBJECTOFFSET(s[VS].constoff) | A4XX_HLSQ_VS_CONTROL_REG_INSTRLENGTH(s[VS].instrlen) | A4XX_HLSQ_VS_CONTROL_REG_SHADEROBJOFFSET(s[VS].instroff)); OUT_RING(ring, A4XX_HLSQ_FS_CONTROL_REG_CONSTLENGTH(s[FS].constlen) | A4XX_HLSQ_FS_CONTROL_REG_CONSTOBJECTOFFSET(s[FS].constoff) | A4XX_HLSQ_FS_CONTROL_REG_INSTRLENGTH(s[FS].instrlen) | A4XX_HLSQ_FS_CONTROL_REG_SHADEROBJOFFSET(s[FS].instroff)); OUT_RING(ring, A4XX_HLSQ_HS_CONTROL_REG_CONSTLENGTH(s[HS].constlen) | A4XX_HLSQ_HS_CONTROL_REG_CONSTOBJECTOFFSET(s[HS].constoff) | A4XX_HLSQ_HS_CONTROL_REG_INSTRLENGTH(s[HS].instrlen) | A4XX_HLSQ_HS_CONTROL_REG_SHADEROBJOFFSET(s[HS].instroff)); OUT_RING(ring, A4XX_HLSQ_DS_CONTROL_REG_CONSTLENGTH(s[DS].constlen) | A4XX_HLSQ_DS_CONTROL_REG_CONSTOBJECTOFFSET(s[DS].constoff) | A4XX_HLSQ_DS_CONTROL_REG_INSTRLENGTH(s[DS].instrlen) | A4XX_HLSQ_DS_CONTROL_REG_SHADEROBJOFFSET(s[DS].instroff)); OUT_RING(ring, A4XX_HLSQ_GS_CONTROL_REG_CONSTLENGTH(s[GS].constlen) | A4XX_HLSQ_GS_CONTROL_REG_CONSTOBJECTOFFSET(s[GS].constoff) | A4XX_HLSQ_GS_CONTROL_REG_INSTRLENGTH(s[GS].instrlen) | A4XX_HLSQ_GS_CONTROL_REG_SHADEROBJOFFSET(s[GS].instroff)); OUT_PKT0(ring, REG_A4XX_SP_SP_CTRL_REG, 1); OUT_RING(ring, 0x140010 | /* XXX */ COND(emit->key.binning_pass, A4XX_SP_SP_CTRL_REG_BINNING_PASS)); OUT_PKT0(ring, REG_A4XX_SP_INSTR_CACHE_CTRL, 1); OUT_RING(ring, 0x7f | /* XXX */ COND(s[VS].instrlen, A4XX_SP_INSTR_CACHE_CTRL_VS_BUFFER) | COND(s[FS].instrlen, A4XX_SP_INSTR_CACHE_CTRL_FS_BUFFER) | COND(s[VS].instrlen && s[FS].instrlen, A4XX_SP_INSTR_CACHE_CTRL_INSTR_BUFFER)); OUT_PKT0(ring, REG_A4XX_SP_VS_LENGTH_REG, 1); OUT_RING(ring, s[VS].v->instrlen); /* SP_VS_LENGTH_REG */ OUT_PKT0(ring, REG_A4XX_SP_VS_CTRL_REG0, 3); OUT_RING(ring, A4XX_SP_VS_CTRL_REG0_THREADMODE(MULTI) | A4XX_SP_VS_CTRL_REG0_HALFREGFOOTPRINT(s[VS].i->max_half_reg + 1) | A4XX_SP_VS_CTRL_REG0_FULLREGFOOTPRINT(s[VS].i->max_reg + 1) | A4XX_SP_VS_CTRL_REG0_INOUTREGOVERLAP(0) | A4XX_SP_VS_CTRL_REG0_THREADSIZE(TWO_QUADS) | A4XX_SP_VS_CTRL_REG0_SUPERTHREADMODE | COND(s[VS].v->has_samp, A4XX_SP_VS_CTRL_REG0_PIXLODENABLE)); OUT_RING(ring, A4XX_SP_VS_CTRL_REG1_CONSTLENGTH(s[VS].constlen) | A4XX_SP_VS_CTRL_REG1_INITIALOUTSTANDING(s[VS].v->total_in)); OUT_RING(ring, A4XX_SP_VS_PARAM_REG_POSREGID(pos_regid) | A4XX_SP_VS_PARAM_REG_PSIZEREGID(psize_regid) | A4XX_SP_VS_PARAM_REG_TOTALVSOUTVAR(s[FS].v->varying_in)); for (i = 0, j = -1; (i < 16) && (j < (int)s[FS].v->inputs_count); i++) { uint32_t reg = 0; OUT_PKT0(ring, REG_A4XX_SP_VS_OUT_REG(i), 1); j = ir3_next_varying(s[FS].v, j); if (j < s[FS].v->inputs_count) { k = ir3_find_output(s[VS].v, s[FS].v->inputs[j].slot); reg |= A4XX_SP_VS_OUT_REG_A_REGID(s[VS].v->outputs[k].regid); reg |= A4XX_SP_VS_OUT_REG_A_COMPMASK(s[FS].v->inputs[j].compmask); } j = ir3_next_varying(s[FS].v, j); if (j < s[FS].v->inputs_count) { k = ir3_find_output(s[VS].v, s[FS].v->inputs[j].slot); reg |= A4XX_SP_VS_OUT_REG_B_REGID(s[VS].v->outputs[k].regid); reg |= A4XX_SP_VS_OUT_REG_B_COMPMASK(s[FS].v->inputs[j].compmask); } OUT_RING(ring, reg); } for (i = 0, j = -1; (i < 8) && (j < (int)s[FS].v->inputs_count); i++) { uint32_t reg = 0; OUT_PKT0(ring, REG_A4XX_SP_VS_VPC_DST_REG(i), 1); j = ir3_next_varying(s[FS].v, j); if (j < s[FS].v->inputs_count) reg |= A4XX_SP_VS_VPC_DST_REG_OUTLOC0(s[FS].v->inputs[j].inloc); j = ir3_next_varying(s[FS].v, j); if (j < s[FS].v->inputs_count) reg |= A4XX_SP_VS_VPC_DST_REG_OUTLOC1(s[FS].v->inputs[j].inloc); j = ir3_next_varying(s[FS].v, j); if (j < s[FS].v->inputs_count) reg |= A4XX_SP_VS_VPC_DST_REG_OUTLOC2(s[FS].v->inputs[j].inloc); j = ir3_next_varying(s[FS].v, j); if (j < s[FS].v->inputs_count) reg |= A4XX_SP_VS_VPC_DST_REG_OUTLOC3(s[FS].v->inputs[j].inloc); OUT_RING(ring, reg); } OUT_PKT0(ring, REG_A4XX_SP_VS_OBJ_OFFSET_REG, 2); OUT_RING(ring, A4XX_SP_VS_OBJ_OFFSET_REG_CONSTOBJECTOFFSET(s[VS].constoff) | A4XX_SP_VS_OBJ_OFFSET_REG_SHADEROBJOFFSET(s[VS].instroff)); OUT_RELOC(ring, s[VS].v->bo, 0, 0, 0); /* SP_VS_OBJ_START_REG */ if (emit->key.binning_pass) { OUT_PKT0(ring, REG_A4XX_SP_FS_LENGTH_REG, 1); OUT_RING(ring, 0x00000000); /* SP_FS_LENGTH_REG */ OUT_PKT0(ring, REG_A4XX_SP_FS_CTRL_REG0, 2); OUT_RING(ring, A4XX_SP_FS_CTRL_REG0_THREADMODE(MULTI) | COND(s[FS].v->total_in > 0, A4XX_SP_FS_CTRL_REG0_VARYING) | A4XX_SP_FS_CTRL_REG0_HALFREGFOOTPRINT(0) | A4XX_SP_FS_CTRL_REG0_FULLREGFOOTPRINT(0) | A4XX_SP_FS_CTRL_REG0_INOUTREGOVERLAP(1) | A4XX_SP_FS_CTRL_REG0_THREADSIZE(fssz) | A4XX_SP_FS_CTRL_REG0_SUPERTHREADMODE); OUT_RING(ring, A4XX_SP_FS_CTRL_REG1_CONSTLENGTH(s[FS].constlen) | 0x80000000); OUT_PKT0(ring, REG_A4XX_SP_FS_OBJ_OFFSET_REG, 2); OUT_RING(ring, A4XX_SP_FS_OBJ_OFFSET_REG_CONSTOBJECTOFFSET(s[FS].constoff) | A4XX_SP_FS_OBJ_OFFSET_REG_SHADEROBJOFFSET(s[FS].instroff)); OUT_RING(ring, 0x00000000); } else { OUT_PKT0(ring, REG_A4XX_SP_FS_LENGTH_REG, 1); OUT_RING(ring, s[FS].v->instrlen); /* SP_FS_LENGTH_REG */ OUT_PKT0(ring, REG_A4XX_SP_FS_CTRL_REG0, 2); OUT_RING(ring, A4XX_SP_FS_CTRL_REG0_THREADMODE(MULTI) | COND(s[FS].v->total_in > 0, A4XX_SP_FS_CTRL_REG0_VARYING) | A4XX_SP_FS_CTRL_REG0_HALFREGFOOTPRINT(s[FS].i->max_half_reg + 1) | A4XX_SP_FS_CTRL_REG0_FULLREGFOOTPRINT(s[FS].i->max_reg + 1) | A4XX_SP_FS_CTRL_REG0_INOUTREGOVERLAP(1) | A4XX_SP_FS_CTRL_REG0_THREADSIZE(fssz) | A4XX_SP_FS_CTRL_REG0_SUPERTHREADMODE | COND(s[FS].v->has_samp, A4XX_SP_FS_CTRL_REG0_PIXLODENABLE)); OUT_RING(ring, A4XX_SP_FS_CTRL_REG1_CONSTLENGTH(s[FS].constlen) | 0x80000000 | /* XXX */ COND(s[FS].v->frag_face, A4XX_SP_FS_CTRL_REG1_FACENESS) | COND(s[FS].v->total_in > 0, A4XX_SP_FS_CTRL_REG1_VARYING) | COND(s[FS].v->frag_coord, A4XX_SP_FS_CTRL_REG1_FRAGCOORD)); OUT_PKT0(ring, REG_A4XX_SP_FS_OBJ_OFFSET_REG, 2); OUT_RING(ring, A4XX_SP_FS_OBJ_OFFSET_REG_CONSTOBJECTOFFSET(s[FS].constoff) | A4XX_SP_FS_OBJ_OFFSET_REG_SHADEROBJOFFSET(s[FS].instroff)); OUT_RELOC(ring, s[FS].v->bo, 0, 0, 0); /* SP_FS_OBJ_START_REG */ } OUT_PKT0(ring, REG_A4XX_SP_HS_OBJ_OFFSET_REG, 1); OUT_RING(ring, A4XX_SP_HS_OBJ_OFFSET_REG_CONSTOBJECTOFFSET(s[HS].constoff) | A4XX_SP_HS_OBJ_OFFSET_REG_SHADEROBJOFFSET(s[HS].instroff)); OUT_PKT0(ring, REG_A4XX_SP_DS_OBJ_OFFSET_REG, 1); OUT_RING(ring, A4XX_SP_DS_OBJ_OFFSET_REG_CONSTOBJECTOFFSET(s[DS].constoff) | A4XX_SP_DS_OBJ_OFFSET_REG_SHADEROBJOFFSET(s[DS].instroff)); OUT_PKT0(ring, REG_A4XX_SP_GS_OBJ_OFFSET_REG, 1); OUT_RING(ring, A4XX_SP_GS_OBJ_OFFSET_REG_CONSTOBJECTOFFSET(s[GS].constoff) | A4XX_SP_GS_OBJ_OFFSET_REG_SHADEROBJOFFSET(s[GS].instroff)); OUT_PKT0(ring, REG_A4XX_RB_RENDER_CONTROL2, 1); OUT_RING(ring, A4XX_RB_RENDER_CONTROL2_MSAA_SAMPLES(0) | COND(s[FS].v->total_in > 0, A4XX_RB_RENDER_CONTROL2_VARYING) | COND(s[FS].v->frag_face, A4XX_RB_RENDER_CONTROL2_FACENESS) | COND(s[FS].v->frag_coord, A4XX_RB_RENDER_CONTROL2_XCOORD | A4XX_RB_RENDER_CONTROL2_YCOORD | A4XX_RB_RENDER_CONTROL2_ZCOORD | A4XX_RB_RENDER_CONTROL2_WCOORD)); OUT_PKT0(ring, REG_A4XX_RB_FS_OUTPUT_REG, 1); OUT_RING(ring, A4XX_RB_FS_OUTPUT_REG_MRT(nr) | COND(s[FS].v->writes_pos, A4XX_RB_FS_OUTPUT_REG_FRAG_WRITES_Z)); OUT_PKT0(ring, REG_A4XX_SP_FS_OUTPUT_REG, 1); OUT_RING(ring, A4XX_SP_FS_OUTPUT_REG_MRT(nr) | COND(s[FS].v->writes_pos, A4XX_SP_FS_OUTPUT_REG_DEPTH_ENABLE) | A4XX_SP_FS_OUTPUT_REG_DEPTH_REGID(posz_regid)); OUT_PKT0(ring, REG_A4XX_SP_FS_MRT_REG(0), 8); for (i = 0; i < 8; i++) { enum a4xx_color_fmt format = 0; bool srgb = false; if (i < nr) { format = fd4_emit_format(bufs[i]); if (bufs[i] && !emit->no_decode_srgb) srgb = util_format_is_srgb(bufs[i]->format); } OUT_RING(ring, A4XX_SP_FS_MRT_REG_REGID(color_regid[i]) | A4XX_SP_FS_MRT_REG_MRTFORMAT(format) | COND(srgb, A4XX_SP_FS_MRT_REG_COLOR_SRGB) | COND(emit->key.half_precision, A4XX_SP_FS_MRT_REG_HALF_PRECISION)); } if (emit->key.binning_pass) { OUT_PKT0(ring, REG_A4XX_VPC_ATTR, 2); OUT_RING(ring, A4XX_VPC_ATTR_THRDASSIGN(1) | 0x40000000 | /* XXX */ COND(s[VS].v->writes_psize, A4XX_VPC_ATTR_PSIZE)); OUT_RING(ring, 0x00000000); } else { uint32_t vinterp[8], vpsrepl[8]; memset(vinterp, 0, sizeof(vinterp)); memset(vpsrepl, 0, sizeof(vpsrepl)); /* looks like we need to do int varyings in the frag * shader on a4xx (no flatshad reg? or a420.0 bug?): * * (sy)(ss)nop * (sy)ldlv.u32 r0.x,l[r0.x], 1 * ldlv.u32 r0.y,l[r0.x+1], 1 * (ss)bary.f (ei)r63.x, 0, r0.x * (ss)(rpt1)cov.s32f16 hr0.x, (r)r0.x * (rpt5)nop * sam (f16)(xyzw)hr0.x, hr0.x, s#0, t#0 * * Possibly on later a4xx variants we'll be able to use * something like the code below instead of workaround * in the shader: */ /* figure out VARYING_INTERP / VARYING_PS_REPL register values: */ for (j = -1; (j = ir3_next_varying(s[FS].v, j)) < (int)s[FS].v->inputs_count; ) { /* NOTE: varyings are packed, so if compmask is 0xb * then first, third, and fourth component occupy * three consecutive varying slots: */ unsigned compmask = s[FS].v->inputs[j].compmask; /* TODO might be cleaner to just +8 in SP_VS_VPC_DST_REG * instead.. rather than -8 everywhere else.. */ uint32_t inloc = s[FS].v->inputs[j].inloc - 8; if ((s[FS].v->inputs[j].interpolate == INTERP_MODE_FLAT) || (s[FS].v->inputs[j].rasterflat && emit->rasterflat)) { uint32_t loc = inloc; for (i = 0; i < 4; i++) { if (compmask & (1 << i)) { vinterp[loc / 16] |= 1 << ((loc % 16) * 2); //flatshade[loc / 32] |= 1 << (loc % 32); loc++; } } } gl_varying_slot slot = s[FS].v->inputs[j].slot; /* since we don't enable PIPE_CAP_TGSI_TEXCOORD: */ if (slot >= VARYING_SLOT_VAR0) { unsigned texmask = 1 << (slot - VARYING_SLOT_VAR0); /* Replace the .xy coordinates with S/T from the point sprite. Set * interpolation bits for .zw such that they become .01 */ if (emit->sprite_coord_enable & texmask) { /* mask is two 2-bit fields, where: * '01' -> S * '10' -> T * '11' -> 1 - T (flip mode) */ unsigned mask = emit->sprite_coord_mode ? 0b1101 : 0b1001; uint32_t loc = inloc; if (compmask & 0x1) { vpsrepl[loc / 16] |= ((mask >> 0) & 0x3) << ((loc % 16) * 2); loc++; } if (compmask & 0x2) { vpsrepl[loc / 16] |= ((mask >> 2) & 0x3) << ((loc % 16) * 2); loc++; } if (compmask & 0x4) { /* .z <- 0.0f */ vinterp[loc / 16] |= 0b10 << ((loc % 16) * 2); loc++; } if (compmask & 0x8) { /* .w <- 1.0f */ vinterp[loc / 16] |= 0b11 << ((loc % 16) * 2); loc++; } }