path: root/src/gallium/drivers/freedreno/a5xx/fd5_program.c

/*
 * Copyright (C) 2016 Rob Clark <robclark@freedesktop.org>
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
 * copy of this software and associated documentation files (the "Software"),
 * to deal in the Software without restriction, including without limitation
 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
 * and/or sell copies of the Software, and to permit persons to whom the
 * Software is furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice (including the next
 * paragraph) shall be included in all copies or substantial portions of the
 * Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 * SOFTWARE.
 *
 * Authors:
 *    Rob Clark <robclark@freedesktop.org>
 */

#include "pipe/p_state.h"
#include "util/bitset.h"
#include "util/format/u_format.h"
#include "util/u_inlines.h"
#include "util/u_memory.h"
#include "util/u_string.h"

#include "freedreno_program.h"

#include "fd5_emit.h"
#include "fd5_format.h"
#include "fd5_program.h"
#include "fd5_texture.h"

#include "ir3_cache.h"

void
fd5_emit_shader(struct fd_ringbuffer *ring, const struct ir3_shader_variant *so)
{
   const struct ir3_info *si = &so->info;
   enum a4xx_state_block sb = fd4_stage2shadersb(so->type);
   enum a4xx_state_src src;
   uint32_t i, sz, *bin;

   if (FD_DBG(DIRECT)) {
      sz = si->sizedwords;
      src = SS4_DIRECT;
      bin = fd_bo_map(so->bo);
   } else {
      sz = 0;
      src = SS4_INDIRECT;
      bin = NULL;
   }

   OUT_PKT7(ring, CP_LOAD_STATE4, 3 + sz);
   OUT_RING(ring, CP_LOAD_STATE4_0_DST_OFF(0) |
                     CP_LOAD_STATE4_0_STATE_SRC(src) |
                     CP_LOAD_STATE4_0_STATE_BLOCK(sb) |
                     CP_LOAD_STATE4_0_NUM_UNIT(so->instrlen));
   if (bin) {
      OUT_RING(ring, CP_LOAD_STATE4_1_EXT_SRC_ADDR(0) |
                        CP_LOAD_STATE4_1_STATE_TYPE(ST4_SHADER));
      OUT_RING(ring, CP_LOAD_STATE4_2_EXT_SRC_ADDR_HI(0));
   } else {
      OUT_RELOC(ring, so->bo, 0, CP_LOAD_STATE4_1_STATE_TYPE(ST4_SHADER), 0);
   }

   /* for how clever coverity is, it is sometimes rather dull, and
    * doesn't realize that the only case where bin==NULL, sz==0:
    */
   assume(bin || (sz == 0));

   for (i = 0; i < sz; i++) {
      OUT_RING(ring, bin[i]);
   }
}

/* TODO maybe some of this we could pre-compute once rather than having
 * so much draw-time logic?
 */
static void
emit_stream_out(struct fd_ringbuffer *ring, const struct ir3_shader_variant *v,
                struct ir3_shader_linkage *l)
{
   const struct ir3_stream_output_info *strmout = &v->stream_output;
   unsigned ncomp[PIPE_MAX_SO_BUFFERS] = {0};
   unsigned prog[align(l->max_loc, 2) / 2];

   memset(prog, 0, sizeof(prog));

   for (unsigned i = 0; i < strmout->num_outputs; i++) {
      const struct ir3_stream_output *out = &strmout->output[i];
      unsigned k = out->register_index;
      unsigned idx;

      ncomp[out->output_buffer] += out->num_components;

      /* linkage map sorted by order frag shader wants things, so
       * a bit less ideal here..
       */
      for (idx = 0; idx < l->cnt; idx++)
         if (l->var[idx].slot == v->outputs[k].slot)
            break;

      assert(idx < l->cnt);

      for (unsigned j = 0; j < out->num_components; j++) {
         unsigned c = j + out->start_component;
         unsigned loc = l->var[idx].loc + c;
         unsigned off = j + out->dst_offset; /* in dwords */

         if (loc & 1) {
            prog[loc / 2] |= A5XX_VPC_SO_PROG_B_EN |
                             A5XX_VPC_SO_PROG_B_BUF(out->output_buffer) |
                             A5XX_VPC_SO_PROG_B_OFF(off * 4);
         } else {
            prog[loc / 2] |= A5XX_VPC_SO_PROG_A_EN |
                             A5XX_VPC_SO_PROG_A_BUF(out->output_buffer) |
                             A5XX_VPC_SO_PROG_A_OFF(off * 4);
         }
      }
   }

   OUT_PKT7(ring, CP_CONTEXT_REG_BUNCH, 12 + (2 * ARRAY_SIZE(prog)));
   OUT_RING(ring, REG_A5XX_VPC_SO_BUF_CNTL);
   OUT_RING(ring, A5XX_VPC_SO_BUF_CNTL_ENABLE |
                     COND(ncomp[0] > 0, A5XX_VPC_SO_BUF_CNTL_BUF0) |
                     COND(ncomp[1] > 0, A5XX_VPC_SO_BUF_CNTL_BUF1) |
                     COND(ncomp[2] > 0, A5XX_VPC_SO_BUF_CNTL_BUF2) |
                     COND(ncomp[3] > 0, A5XX_VPC_SO_BUF_CNTL_BUF3));
   OUT_RING(ring, REG_A5XX_VPC_SO_NCOMP(0));
   OUT_RING(ring, ncomp[0]);
   OUT_RING(ring, REG_A5XX_VPC_SO_NCOMP(1));
   OUT_RING(ring, ncomp[1]);
   OUT_RING(ring, REG_A5XX_VPC_SO_NCOMP(2));
   OUT_RING(ring, ncomp[2]);
   OUT_RING(ring, REG_A5XX_VPC_SO_NCOMP(3));
   OUT_RING(ring, ncomp[3]);
   OUT_RING(ring, REG_A5XX_VPC_SO_CNTL);
   OUT_RING(ring, A5XX_VPC_SO_CNTL_ENABLE);
   for (unsigned i = 0; i < ARRAY_SIZE(prog); i++) {
      OUT_RING(ring, REG_A5XX_VPC_SO_PROG);
      OUT_RING(ring, prog[i]);
   }
}

struct stage {
   const struct ir3_shader_variant *v;
   const struct ir3_info *i;
   /* const sizes are in units of 4 * vec4 */
   uint8_t constoff;
   uint8_t constlen;
   /* instr sizes are in units of 16 instructions */
   uint8_t instroff;
   uint8_t instrlen;
};

enum { VS = 0, FS = 1, HS = 2, DS = 3, GS = 4, MAX_STAGES };

static void
setup_stages(struct fd5_emit *emit, struct stage *s)
{
   unsigned i;

   s[VS].v = fd5_emit_get_vp(emit);
   s[FS].v = fd5_emit_get_fp(emit);

   s[HS].v = s[DS].v = s[GS].v = NULL; /* for now */

   for (i = 0; i < MAX_STAGES; i++) {
      if (s[i].v) {
         s[i].i = &s[i].v->info;
         /* constlen is in units of 4 * vec4: */
         assert(s[i].v->constlen % 4 == 0);
         s[i].constlen = s[i].v->constlen / 4;
         /* instrlen is already in units of 16 instr.. although
          * probably we should ditch that and not make the compiler
          * care about instruction group size of a3xx vs a5xx
          */
         s[i].instrlen = s[i].v->instrlen;
      } else {
         s[i].i = NULL;
         s[i].constlen = 0;
         s[i].instrlen = 0;
      }
   }

   /* NOTE: at least for gles2, blob partitions VS at bottom of const
    * space and FS taking entire remaining space.  We probably don't
    * need to do that the same way, but for now mimic what the blob
    * does to make it easier to diff against register values from blob
    *
    * NOTE: if VS.instrlen + FS.instrlen > 64, then one or both shaders
    * is run from external memory.
    */
   if ((s[VS].instrlen + s[FS].instrlen) > 64) {
      /* prioritize FS for internal memory: */
      if (s[FS].instrlen < 64) {
         /* if FS can fit, kick VS out to external memory: */
         s[VS].instrlen = 0;
      } else if (s[VS].instrlen < 64) {
         /* otherwise if VS can fit, kick out FS: */
         s[FS].instrlen = 0;
      } else {
         /* neither can fit, run both from external memory: */
         s[VS].instrlen = 0;
         s[FS].instrlen = 0;
      }
   }

   unsigned constoff = 0;
   for (i = 0; i < MAX_STAGES; i++) {
      s[i].constoff = constoff;
      constoff += s[i].constlen;
   }

   s[VS].instroff = 0;
   s[FS].instroff = 64 - s[FS].instrlen;
   s[HS].instroff = s[DS].instroff = s[GS].instroff = s[FS].instroff;
}

static inline uint32_t
next_regid(uint32_t reg, uint32_t increment)
{
   if (VALIDREG(reg))
      return reg + increment;
   else
      return regid(63, 0);
}
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, samp_id_regid,
      samp_mask_regid;
   uint32_t ij_regid[IJ_COUNT], vertex_regid, instance_regid, clip0_regid,
      clip1_regid;
   enum a3xx_threadsize fssz;
   uint8_t psize_loc = ~0;
   int i, j;

   setup_stages(emit, s);

   bool do_streamout = (s[VS].v->stream_output.num_outputs > 0);
   uint8_t clip_mask = s[VS].v->clip_mask,
           cull_mask = s[VS].v->cull_mask;
   uint8_t clip_cull_mask = clip_mask | cull_mask;

   /* Unlike a6xx, we don't factor the rasterizer's clip enables in here.  It's
    * already handled by the frontend by storing 0.0 to the clipdist in the
    * shader variant (using either nir_lower_clip_disable for clip distances
    * from the source shader, or nir_lower_clip_vs for user clip planes).
    * Masking the disabled clipdists off causes GPU hangs in tests like
    * spec@glsl-1.20@execution@clipping@vs-clip-vertex-enables.
    */

   fssz = (s[FS].i->double_threadsize) ? FOUR_QUADS : TWO_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);
   clip0_regid = ir3_find_output_regid(s[VS].v, VARYING_SLOT_CLIP_DIST0);
   clip1_regid = ir3_find_output_regid(s[VS].v, VARYING_SLOT_CLIP_DIST1);
   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);
   }

   samp_id_regid = ir3_find_sysval_regid(s[FS].v, SYSTEM_VALUE_SAMPLE_ID);
   samp_mask_regid =
      ir3_find_sysval_regid(s[FS].v, SYSTEM_VALUE_SAMPLE_MASK_IN);
   face_regid = ir3_find_sysval_regid(s[FS].v, SYSTEM_VALUE_FRONT_FACE);
   coord_regid = ir3_find_sysval_regid(s[FS].v, SYSTEM_VALUE_FRAG_COORD);
   zwcoord_regid = next_regid(coord_regid, 2);
   for (unsigned i = 0; i < ARRAY_SIZE(ij_regid); i++)
      ij_regid[i] = ir3_find_sysval_regid(
         s[FS].v, SYSTEM_VALUE_BARYCENTRIC_PERSP_PIXEL + i);

   /* 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) |
         COND(s[VS].instrlen != 0, A5XX_SP_VS_CTRL_REG0_BUFFER) |
         /* XXX: 0x2 is only unset in
          * dEQP-GLES3.functional.ubo.single_nested_struct_array.single_buffer.packed_instance_array_vertex
          * on a collection of blob traces.  That shader is 1091 instrs, 0
          * half, 3 full, 108 constlen.  Other >1091 instr non-VS shaders don't
          * unset it, so that's not the trick.
          */
         0x2 |
         A5XX_SP_VS_CTRL_REG0_BRANCHSTACK(ir3_shader_branchstack_hw(s[VS].v)) |
         COND(s[VS].v->need_pixlod, A5XX_SP_VS_CTRL_REG0_PIXLODENABLE));

   /* If we have streamout, link against the real FS in the binning program,
    * rather than the dummy FS used for binning pass state, to ensure the
    * OUTLOC's match.  Depending on whether we end up doing sysmem or gmem, the
    * actual streamout could happen with either the binning pass or draw pass
    * program, but the same streamout stateobj is used in either case:
    */
   const struct ir3_shader_variant *link_fs = s[FS].v;
   if (do_streamout && emit->binning_pass)
      link_fs = emit->prog->fs;
   struct ir3_shader_linkage l = {0};
   ir3_link_shaders(&l, s[VS].v, link_fs, true);

   uint8_t clip0_loc = l.clip0_loc;
   uint8_t clip1_loc = l.clip1_loc;

   OUT_PKT4(ring, REG_A5XX_VPC_VAR_DISABLE(0), 4);
   OUT_RING(ring, ~l.varmask[0]); /* VPC_VAR[0].DISABLE */
   OUT_RING(ring, ~l.varmask[1]); /* VPC_VAR[1].DISABLE */
   OUT_RING(ring, ~l.varmask[2]); /* VPC_VAR[2].DISABLE */
   OUT_RING(ring, ~l.varmask[3]); /* VPC_VAR[3].DISABLE */

   /* Add stream out outputs after computing the VPC_VAR_DISABLE bitmask. */
   ir3_link_stream_out(&l, s[VS].v);

   /* a5xx appends pos/psize to end of the linkage map: */
   if (VALIDREG(pos_regid))
      ir3_link_add(&l, VARYING_SLOT_POS, pos_regid, 0xf, l.max_loc);

   if (VALIDREG(psize_regid)) {
      psize_loc = l.max_loc;
      ir3_link_add(&l, VARYING_SLOT_PSIZ, psize_regid, 0x1, l.max_loc);
   }

   /* Handle the case where clip/cull distances aren't read by the FS. Make
    * sure to avoid adding an output with an empty writemask if the user
    * disables all the clip distances in the API so that the slot is unused.
    */
   if (clip0_loc == 0xff && VALIDREG(clip0_regid) &&
       (clip_cull_mask & 0xf) != 0) {
      clip0_loc = l.max_loc;
      ir3_link_add(&l, VARYING_SLOT_CLIP_DIST0, clip0_regid,
                   clip_cull_mask & 0xf, l.max_loc);
   }

   if (clip1_loc == 0xff && VALIDREG(clip1_regid) &&
       (clip_cull_mask >> 4) != 0) {
      clip1_loc = l.max_loc;
      ir3_link_add(&l, VARYING_SLOT_CLIP_DIST1, clip1_regid,
                   clip_cull_mask >> 4, l.max_loc);
   }

   /* If we have stream-out, we use the full shader for binning
    * pass, rather than the optimized binning pass one, so that we
    * have all the varying outputs available for xfb.  So streamout
    * state should always be derived from the non-binning pass
    * program:
    */
   if (do_streamout && !emit->binning_pass)
      emit_stream_out(ring, s[VS].v, &l);

   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) |
                     0x10000); // XXX

   fd5_context(ctx)->max_loc = l.max_loc;

   if (emit->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) |
                     A5XX_HLSQ_CONTROL_2_REG_SAMPLEID(samp_id_regid) |
                     A5XX_HLSQ_CONTROL_2_REG_SAMPLEMASK(samp_mask_regid) |
                     A5XX_HLSQ_CONTROL_2_REG_CENTERRHW(ij_regid[IJ_PERSP_CENTER_RHW]));
   OUT_RING(
      ring,
      A5XX_HLSQ_CONTROL_3_REG_IJ_PERSP_PIXEL(ij_regid[IJ_PERSP_PIXEL]) |
         A5XX_HLSQ_CONTROL_3_REG_IJ_LINEAR_PIXEL(ij_regid[IJ_LINEAR_PIXEL]) |
         A5XX_HLSQ_CONTROL_3_REG_IJ_PERSP_CENTROID(
            ij_regid[IJ_PERSP_CENTROID]) |
         A5XX_HLSQ_CONTROL_3_REG_IJ_PERSP_CENTROID(
            ij_regid[IJ_LINEAR_CENTROID]));
   OUT_RING(
      ring,
      A5XX_HLSQ_CONTROL_4_REG_XYCOORDREGID(coord_regid) |
         A5XX_HLSQ_CONTROL_4_REG_ZWCOORDREGID(zwcoord_regid) |
         A5XX_HLSQ_CONTROL_4_REG_IJ_PERSP_SAMPLE(ij_regid[IJ_PERSP_SAMPLE]) |
         A5XX_HLSQ_CONTROL_4_REG_IJ_LINEAR_SAMPLE(ij_regid[IJ_LINEAR_SAMPLE]));

   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) |
         0x40002 | /* XXX set pretty much everywhere */
         COND(s[FS].instrlen != 0, A5XX_SP_FS_CTRL_REG0_BUFFER) |
         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(ir3_shader_branchstack_hw(s[FS].v)) |
         COND(s[FS].v->need_pixlod, 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,
            CONDREG(ij_regid[IJ_PERSP_PIXEL], A5XX_GRAS_CNTL_IJ_PERSP_PIXEL) |
               CONDREG(ij_regid[IJ_PERSP_CENTROID],
                       A5XX_GRAS_CNTL_IJ_PERSP_CENTROID) |
               CONDREG(ij_regid[IJ_PERSP_SAMPLE],
                       A5XX_GRAS_CNTL_IJ_PERSP_SAMPLE) |
               CONDREG(ij_regid[IJ_LINEAR_PIXEL], A5XX_GRAS_CNTL_IJ_LINEAR_PIXEL) |
               CONDREG(ij_regid[IJ_LINEAR_CENTROID],
                       A5XX_GRAS_CNTL_IJ_LINEAR_CENTROID) |
               CONDREG(ij_regid[IJ_LINEAR_SAMPLE],
                       A5XX_GRAS_CNTL_IJ_LINEAR_SAMPLE) |
               COND(s[FS].v->fragcoord_compmask != 0,
                    A5XX_GRAS_CNTL_COORD_MASK(s[FS].v->fragcoord_compmask) |
                       A5XX_GRAS_CNTL_IJ_LINEAR_PIXEL) |
               COND(s[FS].v->frag_face, A5XX_GRAS_CNTL_IJ_LINEAR_PIXEL) |
               CONDREG(ij_regid[IJ_LINEAR_PIXEL], A5XX_GRAS_CNTL_IJ_LINEAR_PIXEL));

   OUT_PKT4(ring, REG_A5XX_RB_RENDER_CONTROL0, 2);
   OUT_RING(
      ring,
      CONDREG(ij_regid[IJ_PERSP_PIXEL],
              A5XX_RB_RENDER_CONTROL0_IJ_PERSP_PIXEL) |
         CONDREG(ij_regid[IJ_PERSP_CENTROID],
                 A5XX_RB_RENDER_CONTROL0_IJ_PERSP_CENTROID) |
         CONDREG(ij_regid[IJ_PERSP_SAMPLE],
                 A5XX_RB_RENDER_CONTROL0_IJ_PERSP_SAMPLE) |
         CONDREG(ij_regid[IJ_LINEAR_PIXEL],
              A5XX_RB_RENDER_CONTROL0_IJ_LINEAR_PIXEL) |
         CONDREG(ij_regid[IJ_LINEAR_CENTROID],
                 A5XX_RB_RENDER_CONTROL0_IJ_LINEAR_CENTROID) |
         CONDREG(ij_regid[IJ_LINEAR_SAMPLE],
                 A5XX_RB_RENDER_CONTROL0_IJ_LINEAR_SAMPLE) |
         COND(s[FS].v->fragcoord_compmask != 0,
              A5XX_RB_RENDER_CONTROL0_COORD_MASK(s[FS].v->fragcoord_compmask) |
                 A5XX_RB_RENDER_CONTROL0_IJ_LINEAR_PIXEL) |
         COND(s[FS].v->frag_face, A5XX_RB_RENDER_CONTROL0_IJ_LINEAR_PIXEL) |
         CONDREG(ij_regid[IJ_LINEAR_PIXEL], A5XX_RB_RENDER_CONTROL0_IJ_LINEAR_PIXEL));
   OUT_RING(ring,
            CONDREG(samp_mask_regid, A5XX_RB_RENDER_CONTROL1_SAMPLEMASK) |
               COND(s[FS].v->frag_face, A5XX_RB_RENDER_CONTROL1_FACENESS) |
               CONDREG(samp_id_regid, A5XX_RB_RENDER_CONTROL1_SAMPLEID));

   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(color_regid[i] & HALF_REG_ID,
                             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->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].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++;
               }
            }
         }

         bool coord_mode = emit->sprite_coord_mode;
         if (ir3_point_sprite(s[FS].v, j, emit->sprite_coord_enable,
                              &coord_mode)) {
            /* mask is two 2-bit fields, where:
             *   '01' -> S
             *   '10' -> T
             *   '11' -> 1 - T  (flip mode)
             */
            unsigned mask = 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++;
            }
         }
      }

      OUT_PKT4(ring, REG_A5XX_VPC_VARYING_INTERP_MODE(0), 8);
      for (i = 0; i < 8; i++)
         OUT_RING(ring, vinterp[i]); /* VPC_VARYING_INTERP[i].MODE */

      OUT_PKT4(ring, REG_A5XX_VPC_VARYING_PS_REPL_MODE(0), 8);
      for (i = 0; i < 8; i++)
         OUT_RING(ring, vpsrepl[i]); /* VPC_VARYING_PS_REPL[i] */
   }

   OUT_PKT4(ring, REG_A5XX_GRAS_VS_CL_CNTL, 1);
   OUT_RING(ring, A5XX_GRAS_VS_CL_CNTL_CLIP_MASK(clip_mask) |
                     A5XX_GRAS_VS_CL_CNTL_CULL_MASK(cull_mask));

   OUT_PKT4(ring, REG_A5XX_VPC_CLIP_CNTL, 1);
   OUT_RING(ring, A5XX_VPC_CLIP_CNTL_CLIP_MASK(clip_cull_mask) |
                     A5XX_VPC_CLIP_CNTL_CLIP_DIST_03_LOC(clip0_loc) |
                     A5XX_VPC_CLIP_CNTL_CLIP_DIST_47_LOC(clip1_loc));

   OUT_PKT4(ring, REG_A5XX_PC_CLIP_CNTL, 1);
   OUT_RING(ring, A5XX_PC_CLIP_CNTL_CLIP_MASK(clip_mask));

   if (!emit->binning_pass)
      if (s[FS].instrlen)
         fd5_emit_shader(ring, s[FS].v);

   OUT_PKT4(ring, REG_A5XX_VFD_CONTROL_1, 5);
   OUT_RING(ring, A5XX_VFD_CONTROL_1_REGID4VTX(vertex_regid) |
                     A5XX_VFD_CONTROL_1_REGID4INST(instance_regid) | 0xfc0000);
   OUT_RING(ring, 0x0000fcfc); /* VFD_CONTROL_2 */
   OUT_RING(ring, 0x0000fcfc); /* VFD_CONTROL_3 */
   OUT_RING(ring, 0x000000fc); /* VFD_CONTROL_4 */
   OUT_RING(ring, 0x00000000); /* VFD_CONTROL_5 */
}

static struct ir3_program_state *
fd5_program_create(void *data, struct ir3_shader_variant *bs,
                   struct ir3_shader_variant *vs, struct ir3_shader_variant *hs,
                   struct ir3_shader_variant *ds, struct ir3_shader_variant *gs,
                   struct ir3_shader_variant *fs,
                   const struct ir3_cache_key *key) in_dt
{
   struct fd_context *ctx = fd_context(data);
   struct fd5_program_state *state = CALLOC_STRUCT(fd5_program_state);

   tc_assert_driver_thread(ctx->tc);

   state->bs = bs;
   state->vs = vs;
   state->fs = fs;

   return &state->base;
}

static void
fd5_program_destroy(void *data, struct ir3_program_state *state)
{
   struct fd5_program_state *so = fd5_program_state(state);
   free(so);
}

static const struct ir3_cache_funcs cache_funcs = {
   .create_state = fd5_program_create,
   .destroy_state = fd5_program_destroy,
};

void
fd5_prog_init(struct pipe_context *pctx)
{
   struct fd_context *ctx = fd_context(pctx);

   ctx->shader_cache = ir3_cache_create(&cache_funcs, ctx);
   ir3_prog_init(pctx);
   fd_prog_init(pctx);
}