turnip: rework streamout state and add missing counter buffer read/writes

Rework the streamout state and at the same time fix some issues, the
biggest one being to actually use the counter buffers instead of ignoring
them completely.

(note it appears the dEQP tests are bad and able to pass with the previous
broken behavior of not ever reading/writing from the counter buffers)

Signed-off-by: Jonathan Marek <[email protected]>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5579>

4 files changed, 188 insertions, 168 deletions

diff --git a/src/freedreno/vulkan/tu_clear_blit.c b/src/freedreno/vulkan/tu_clear_blit.c
index d17b990f491..9b729a562f5 100644
--- a/src/freedreno/vulkan/tu_clear_blit.c
+++ b/src/freedreno/vulkan/tu_clear_blit.c
@@ -517,7 +517,7 @@ r3d_common(struct tu_cmd_buffer *cmd, struct tu_cs *cs, bool blit, uint32_t num_
    tu_cs_emit_regs(cs, A6XX_PC_PRIMITIVE_CNTL_0());
    tu_cs_emit_regs(cs, A6XX_VFD_CONTROL_0());
 
-   tu6_emit_vpc(cs, &vs, NULL, NULL, gs, &fs, NULL);
+   tu6_emit_vpc(cs, &vs, NULL, NULL, gs, &fs);
 
    /* REPL_MODE for varying with RECTLIST (2 vertices only) */
    tu_cs_emit_regs(cs, A6XX_VPC_VARYING_INTERP_MODE(0, 0));
diff --git a/src/freedreno/vulkan/tu_cmd_buffer.c b/src/freedreno/vulkan/tu_cmd_buffer.c
index 41d6bfcdedd..e3f625d583b 100644
--- a/src/freedreno/vulkan/tu_cmd_buffer.c
+++ b/src/freedreno/vulkan/tu_cmd_buffer.c
@@ -708,6 +708,13 @@ use_hw_binning(struct tu_cmd_buffer *cmd)
 {
    const struct tu_tiling_config *tiling = &cmd->state.tiling_config;
 
+   /* XFB commands are emitted for BINNING || SYSMEM, which makes it incompatible
+    * with non-hw binning GMEM rendering. this is required because some of the
+    * XFB commands need to only be executed once
+    */
+   if (cmd->state.xfb_used)
+      return true;
+
    if (unlikely(cmd->device->physical_device->instance->debug_flags & TU_DEBUG_NOBIN))
       return false;
 
@@ -989,13 +996,6 @@ tu6_init_hw(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
    tu_cs_emit(cs, CP_SET_DRAW_STATE__1_ADDR_LO(0));
    tu_cs_emit(cs, CP_SET_DRAW_STATE__2_ADDR_HI(0));
 
-   /* Set not to use streamout by default, */
-   tu_cs_emit_pkt7(cs, CP_CONTEXT_REG_BUNCH, 4);
-   tu_cs_emit(cs, REG_A6XX_VPC_SO_CNTL);
-   tu_cs_emit(cs, 0);
-   tu_cs_emit(cs, REG_A6XX_VPC_SO_BUF_CNTL);
-   tu_cs_emit(cs, 0);
-
    tu_cs_emit_regs(cs,
                    A6XX_SP_HS_CTRL_REG0(0));
 
@@ -1441,6 +1441,9 @@ tu_cmd_render_tiles(struct tu_cmd_buffer *cmd)
 {
    const struct tu_tiling_config *tiling = &cmd->state.tiling_config;
 
+   if (use_hw_binning(cmd))
+      cmd->use_vsc_data = true;
+
    tu6_tile_render_begin(cmd, &cmd->cs);
 
    for (uint32_t y = 0; y < tiling->tile_count.height; y++) {
@@ -1925,33 +1928,86 @@ void tu_CmdBindTransformFeedbackBuffersEXT(VkCommandBuffer commandBuffer,
                                            const VkDeviceSize *pSizes)
 {
    TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-   assert(firstBinding + bindingCount <= IR3_MAX_SO_BUFFERS);
+   struct tu_cs *cs = &cmd->draw_cs;
+
+   /* using COND_REG_EXEC for xfb commands matches the blob behavior
+    * presumably there isn't any benefit using a draw state when the
+    * condition is (SYSMEM | BINNING)
+    */
+   tu_cond_exec_start(cs, CP_COND_REG_EXEC_0_MODE(RENDER_MODE) |
+                          CP_COND_REG_EXEC_0_SYSMEM |
+                          CP_COND_REG_EXEC_0_BINNING);
 
    for (uint32_t i = 0; i < bindingCount; i++) {
-      uint32_t idx = firstBinding + i;
       TU_FROM_HANDLE(tu_buffer, buf, pBuffers[i]);
+      uint64_t iova = buf->bo->iova + pOffsets[i];
+      uint32_t size = buf->bo->size - pOffsets[i];
+      uint32_t idx = i + firstBinding;
 
-      if (pOffsets[i] != 0)
-         cmd->state.streamout_reset |= 1 << idx;
+      if (pSizes && pSizes[i] != VK_WHOLE_SIZE)
+         size = pSizes[i];
 
-      cmd->state.streamout_buf.buffers[idx] = buf;
-      cmd->state.streamout_buf.offsets[idx] = pOffsets[i];
-      cmd->state.streamout_buf.sizes[idx] = pSizes[i];
+      /* BUFFER_BASE is 32-byte aligned, add remaining offset to BUFFER_OFFSET */
+      uint32_t offset = iova & 0x1f;
+      iova &= ~(uint64_t) 0x1f;
 
-      cmd->state.streamout_enabled |= 1 << idx;
+      tu_cs_emit_pkt4(cs, REG_A6XX_VPC_SO_BUFFER_BASE(idx), 3);
+      tu_cs_emit_qw(cs, iova);
+      tu_cs_emit(cs, size + offset);
+
+      cmd->state.streamout_offset[idx] = offset;
+
+      tu_bo_list_add(&cmd->bo_list, buf->bo, MSM_SUBMIT_BO_WRITE);
    }
 
-   cmd->state.dirty |= TU_CMD_DIRTY_STREAMOUT_BUFFERS;
+   tu_cond_exec_end(cs);
 }
 
-void tu_CmdBeginTransformFeedbackEXT(VkCommandBuffer commandBuffer,
-                                       uint32_t firstCounterBuffer,
-                                       uint32_t counterBufferCount,
-                                       const VkBuffer *pCounterBuffers,
-                                       const VkDeviceSize *pCounterBufferOffsets)
+void
+tu_CmdBeginTransformFeedbackEXT(VkCommandBuffer commandBuffer,
+                                uint32_t firstCounterBuffer,
+                                uint32_t counterBufferCount,
+                                const VkBuffer *pCounterBuffers,
+                                const VkDeviceSize *pCounterBufferOffsets)
 {
-   assert(firstCounterBuffer + counterBufferCount <= IR3_MAX_SO_BUFFERS);
-   /* TODO do something with counter buffer? */
+   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
+   struct tu_cs *cs = &cmd->draw_cs;
+
+   tu_cond_exec_start(cs, CP_COND_REG_EXEC_0_MODE(RENDER_MODE) |
+                          CP_COND_REG_EXEC_0_SYSMEM |
+                          CP_COND_REG_EXEC_0_BINNING);
+
+   /* TODO: only update offset for active buffers */
+   for (uint32_t i = 0; i < IR3_MAX_SO_BUFFERS; i++)
+      tu_cs_emit_regs(cs, A6XX_VPC_SO_BUFFER_OFFSET(i, cmd->state.streamout_offset[i]));
+
+   for (uint32_t i = 0; i < counterBufferCount; i++) {
+      uint32_t idx = firstCounterBuffer + i;
+      uint32_t offset = cmd->state.streamout_offset[idx];
+
+      if (!pCounterBuffers[i])
+         continue;
+
+      TU_FROM_HANDLE(tu_buffer, buf, pCounterBuffers[i]);
+
+      tu_bo_list_add(&cmd->bo_list, buf->bo, MSM_SUBMIT_BO_READ);
+
+      tu_cs_emit_pkt7(cs, CP_MEM_TO_REG, 3);
+      tu_cs_emit(cs, CP_MEM_TO_REG_0_REG(REG_A6XX_VPC_SO_BUFFER_OFFSET(idx)) |
+                     CP_MEM_TO_REG_0_UNK31 |
+                     CP_MEM_TO_REG_0_CNT(1));
+      tu_cs_emit_qw(cs, buf->bo->iova + pCounterBufferOffsets[i]);
+
+      if (offset) {
+         tu_cs_emit_pkt7(cs, CP_REG_RMW, 3);
+         tu_cs_emit(cs, CP_REG_RMW_0_DST_REG(REG_A6XX_VPC_SO_BUFFER_OFFSET(idx)) |
+                        CP_REG_RMW_0_SRC1_ADD);
+         tu_cs_emit_qw(cs, 0xffffffff);
+         tu_cs_emit_qw(cs, offset);
+      }
+   }
+
+   tu_cond_exec_end(cs);
 }
 
 void tu_CmdEndTransformFeedbackEXT(VkCommandBuffer commandBuffer,
@@ -1960,11 +2016,58 @@ void tu_CmdEndTransformFeedbackEXT(VkCommandBuffer commandBuffer,
                                        const VkBuffer *pCounterBuffers,
                                        const VkDeviceSize *pCounterBufferOffsets)
 {
-   assert(firstCounterBuffer + counterBufferCount <= IR3_MAX_SO_BUFFERS);
-   /* TODO do something with counter buffer? */
-
    TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-   cmd->state.streamout_enabled = 0;
+   struct tu_cs *cs = &cmd->draw_cs;
+
+   tu_cond_exec_start(cs, CP_COND_REG_EXEC_0_MODE(RENDER_MODE) |
+                          CP_COND_REG_EXEC_0_SYSMEM |
+                          CP_COND_REG_EXEC_0_BINNING);
+
+   /* TODO: only flush buffers that need to be flushed */
+   for (uint32_t i = 0; i < IR3_MAX_SO_BUFFERS; i++) {
+      /* note: FLUSH_BASE is always the same, so it could go in init_hw()? */
+      tu_cs_emit_pkt4(cs, REG_A6XX_VPC_SO_FLUSH_BASE(i), 2);
+      tu_cs_emit_qw(cs, cmd->scratch_bo.iova + ctrl_offset(flush_base[i]));
+      tu6_emit_event_write(cmd, cs, FLUSH_SO_0 + i);
+   }
+
+   for (uint32_t i = 0; i < counterBufferCount; i++) {
+      uint32_t idx = firstCounterBuffer + i;
+      uint32_t offset = cmd->state.streamout_offset[idx];
+
+      if (!pCounterBuffers[i])
+         continue;
+
+      TU_FROM_HANDLE(tu_buffer, buf, pCounterBuffers[i]);
+
+      tu_bo_list_add(&cmd->bo_list, buf->bo, MSM_SUBMIT_BO_WRITE);
+
+      /* VPC_SO_FLUSH_BASE has dwords counter, but counter should be in bytes */
+      tu_cs_emit_pkt7(cs, CP_MEM_TO_REG, 3);
+      tu_cs_emit(cs, CP_MEM_TO_REG_0_REG(REG_A6XX_CP_SCRATCH_REG(0)) |
+                     CP_MEM_TO_REG_0_SHIFT_BY_2 |
+                     0x40000 | /* ??? */
+                     CP_MEM_TO_REG_0_UNK31 |
+                     CP_MEM_TO_REG_0_CNT(1));
+      tu_cs_emit_qw(cs, cmd->scratch_bo.iova + ctrl_offset(flush_base[idx]));
+
+      if (offset) {
+         tu_cs_emit_pkt7(cs, CP_REG_RMW, 3);
+         tu_cs_emit(cs, CP_REG_RMW_0_DST_REG(REG_A6XX_CP_SCRATCH_REG(0)) |
+                        CP_REG_RMW_0_SRC1_ADD);
+         tu_cs_emit_qw(cs, 0xffffffff);
+         tu_cs_emit_qw(cs, -offset);
+      }
+
+      tu_cs_emit_pkt7(cs, CP_REG_TO_MEM, 3);
+      tu_cs_emit(cs, CP_REG_TO_MEM_0_REG(REG_A6XX_CP_SCRATCH_REG(0)) |
+                     CP_REG_TO_MEM_0_CNT(1));
+      tu_cs_emit_qw(cs, buf->bo->iova + pCounterBufferOffsets[i]);
+   }
+
+   tu_cond_exec_end(cs);
+
+   cmd->state.xfb_used = true;
 }
 
 void
@@ -2694,10 +2797,6 @@ tu_CmdBeginRenderPass(VkCommandBuffer commandBuffer,
 
    tu_set_input_attachments(cmd, cmd->state.subpass);
 
-   /* note: use_hw_binning only checks tiling config */
-   if (use_hw_binning(cmd))
-      cmd->use_vsc_data = true;
-
    for (uint32_t i = 0; i < fb->attachment_count; ++i) {
       const struct tu_image_view *iview = fb->attachments[i].attachment;
       tu_bo_list_add(&cmd->bo_list, iview->image->bo,
@@ -2972,67 +3071,6 @@ tu6_emit_vertex_buffers(struct tu_cmd_buffer *cmd,
    return tu_cs_end_sub_stream(&cmd->sub_cs, &cs);
 }
 
-static void
-tu6_emit_streamout(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
-{
-   struct tu_streamout_state *tf = &cmd->state.pipeline->streamout;
-
-   for (unsigned i = 0; i < IR3_MAX_SO_BUFFERS; i++) {
-      struct tu_buffer *buf = cmd->state.streamout_buf.buffers[i];
-      if (!buf)
-         continue;
-
-      uint32_t offset;
-      offset = cmd->state.streamout_buf.offsets[i];
-
-      tu_cs_emit_regs(cs, A6XX_VPC_SO_BUFFER_BASE(i, .bo = buf->bo,
-                                                     .bo_offset = buf->bo_offset));
-      tu_cs_emit_regs(cs, A6XX_VPC_SO_BUFFER_SIZE(i, buf->size));
-
-      if (cmd->state.streamout_reset & (1 << i)) {
-         tu_cs_emit_regs(cs, A6XX_VPC_SO_BUFFER_OFFSET(i, offset));
-         cmd->state.streamout_reset &= ~(1  << i);
-      } else {
-         tu_cs_emit_pkt7(cs, CP_MEM_TO_REG, 3);
-         tu_cs_emit(cs, CP_MEM_TO_REG_0_REG(REG_A6XX_VPC_SO_BUFFER_OFFSET(i)) |
-                        CP_MEM_TO_REG_0_SHIFT_BY_2 | CP_MEM_TO_REG_0_UNK31 |
-                        CP_MEM_TO_REG_0_CNT(0));
-         tu_cs_emit_qw(cs, cmd->scratch_bo.iova +
-                           ctrl_offset(flush_base[i].offset));
-      }
-
-      tu_cs_emit_regs(cs, A6XX_VPC_SO_FLUSH_BASE(i, .bo = &cmd->scratch_bo,
-                                                    .bo_offset =
-                                                       ctrl_offset(flush_base[i])));
-   }
-
-   if (cmd->state.streamout_enabled) {
-      tu_cs_emit_pkt7(cs, CP_CONTEXT_REG_BUNCH, 12 + (2 * tf->prog_count));
-      tu_cs_emit(cs, REG_A6XX_VPC_SO_BUF_CNTL);
-      tu_cs_emit(cs, tf->vpc_so_buf_cntl);
-      tu_cs_emit(cs, REG_A6XX_VPC_SO_NCOMP(0));
-      tu_cs_emit(cs, tf->ncomp[0]);
-      tu_cs_emit(cs, REG_A6XX_VPC_SO_NCOMP(1));
-      tu_cs_emit(cs, tf->ncomp[1]);
-      tu_cs_emit(cs, REG_A6XX_VPC_SO_NCOMP(2));
-      tu_cs_emit(cs, tf->ncomp[2]);
-      tu_cs_emit(cs, REG_A6XX_VPC_SO_NCOMP(3));
-      tu_cs_emit(cs, tf->ncomp[3]);
-      tu_cs_emit(cs, REG_A6XX_VPC_SO_CNTL);
-      tu_cs_emit(cs, A6XX_VPC_SO_CNTL_ENABLE);
-      for (unsigned i = 0; i < tf->prog_count; i++) {
-         tu_cs_emit(cs, REG_A6XX_VPC_SO_PROG);
-         tu_cs_emit(cs, tf->prog[i]);
-      }
-   } else {
-      tu_cs_emit_pkt7(cs, CP_CONTEXT_REG_BUNCH, 4);
-      tu_cs_emit(cs, REG_A6XX_VPC_SO_CNTL);
-      tu_cs_emit(cs, 0);
-      tu_cs_emit(cs, REG_A6XX_VPC_SO_BUF_CNTL);
-      tu_cs_emit(cs, 0);
-   }
-}
-
 static uint64_t
 get_tess_param_bo_size(const struct tu_pipeline *pipeline,
                        const struct tu_draw_info *draw_info)
@@ -3180,9 +3218,6 @@ tu6_bind_draw_states(struct tu_cmd_buffer *cmd,
          tu6_emit_consts(cmd, pipeline, descriptors_state, MESA_SHADER_FRAGMENT);
    }
 
-   if (cmd->state.dirty & TU_CMD_DIRTY_STREAMOUT_BUFFERS)
-      tu6_emit_streamout(cmd, cs);
-
    if (cmd->state.dirty & TU_CMD_DIRTY_DESCRIPTOR_SETS) {
       /* We need to reload the descriptors every time the descriptor sets
        * change. However, the commands we send only depend on the pipeline
@@ -3300,17 +3335,6 @@ tu6_bind_draw_states(struct tu_cmd_buffer *cmd,
 
    tu_cs_sanity_check(cs);
 
-   /* track BOs */
-   if (cmd->state.dirty & TU_CMD_DIRTY_STREAMOUT_BUFFERS) {
-      for (unsigned i = 0; i < IR3_MAX_SO_BUFFERS; i++) {
-         const struct tu_buffer *buf = cmd->state.streamout_buf.buffers[i];
-         if (buf) {
-            tu_bo_list_add(&cmd->bo_list, buf->bo,
-                              MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_WRITE);
-         }
-      }
-   }
-
    /* There are too many graphics dirty bits to list here, so just list the
     * bits to preserve instead. The only things not emitted here are
     * compute-related state.
@@ -3470,13 +3494,6 @@ tu_draw(struct tu_cmd_buffer *cmd, const struct tu_draw_info *draw)
    else
       tu6_emit_draw_direct(cmd, cs, draw);
 
-   if (cmd->state.streamout_enabled) {
-      for (unsigned i = 0; i < IR3_MAX_SO_BUFFERS; i++) {
-         if (cmd->state.streamout_enabled & (1 << i))
-            tu6_emit_event_write(cmd, cs, FLUSH_SO_0 + i);
-      }
-   }
-
    tu_cs_sanity_check(cs);
 }
 
diff --git a/src/freedreno/vulkan/tu_pipeline.c b/src/freedreno/vulkan/tu_pipeline.c
index fe6c346c2d1..a3251621e8f 100644
--- a/src/freedreno/vulkan/tu_pipeline.c
+++ b/src/freedreno/vulkan/tu_pipeline.c
@@ -591,20 +591,28 @@ tu6_link_streamout(struct ir3_shader_linkage *l,
 }
 
 static void
-tu6_setup_streamout(const struct ir3_shader_variant *v,
-            struct ir3_shader_linkage *l, struct tu_streamout_state *tf)
+tu6_setup_streamout(struct tu_cs *cs,
+                    const struct ir3_shader_variant *v,
+                    struct ir3_shader_linkage *l)
 {
    const struct ir3_stream_output_info *info = &v->shader->stream_output;
+   uint32_t prog[IR3_MAX_SO_OUTPUTS * 2] = {};
+   uint32_t ncomp[IR3_MAX_SO_BUFFERS] = {};
+   uint32_t prog_count = align(l->max_loc, 2) / 2;
 
-   memset(tf, 0, sizeof(*tf));
+   /* TODO: streamout state should be in a non-GMEM draw state */
 
-   tf->prog_count = align(l->max_loc, 2) / 2;
-
-   debug_assert(tf->prog_count < ARRAY_SIZE(tf->prog));
+   /* no streamout: */
+   if (info->num_outputs == 0) {
+      tu_cs_emit_pkt7(cs, CP_CONTEXT_REG_BUNCH, 4);
+      tu_cs_emit(cs, REG_A6XX_VPC_SO_CNTL);
+      tu_cs_emit(cs, 0);
+      tu_cs_emit(cs, REG_A6XX_VPC_SO_BUF_CNTL);
+      tu_cs_emit(cs, 0);
+      return;
+   }
 
-   /* set stride info to the streamout state */
-   for (unsigned i = 0; i < IR3_MAX_SO_BUFFERS; i++)
-      tf->stride[i] = info->stride[i];
+   /* is there something to do with info->stride[i]? */
 
    for (unsigned i = 0; i < info->num_outputs; i++) {
       const struct ir3_stream_output *out = &info->output[i];
@@ -615,7 +623,7 @@ tu6_setup_streamout(const struct ir3_shader_variant *v,
       if (v->outputs[k].regid == INVALID_REG)
          continue;
 
-      tf->ncomp[out->output_buffer] += out->num_components;
+      ncomp[out->output_buffer] += out->num_components;
 
       /* linkage map sorted by order frag shader wants things, so
        * a bit less ideal here..
@@ -632,22 +640,35 @@ tu6_setup_streamout(const struct ir3_shader_variant *v,
          unsigned off = j + out->dst_offset;  /* in dwords */
 
          if (loc & 1) {
-            tf->prog[loc/2] |= A6XX_VPC_SO_PROG_B_EN |
-                        A6XX_VPC_SO_PROG_B_BUF(out->output_buffer) |
-                        A6XX_VPC_SO_PROG_B_OFF(off * 4);
+            prog[loc/2] |= A6XX_VPC_SO_PROG_B_EN |
+                           A6XX_VPC_SO_PROG_B_BUF(out->output_buffer) |
+                           A6XX_VPC_SO_PROG_B_OFF(off * 4);
          } else {
-            tf->prog[loc/2] |= A6XX_VPC_SO_PROG_A_EN |
-                        A6XX_VPC_SO_PROG_A_BUF(out->output_buffer) |
-                        A6XX_VPC_SO_PROG_A_OFF(off * 4);
+            prog[loc/2] |= A6XX_VPC_SO_PROG_A_EN |
+                           A6XX_VPC_SO_PROG_A_BUF(out->output_buffer) |
+                           A6XX_VPC_SO_PROG_A_OFF(off * 4);
          }
       }
    }
 
-   tf->vpc_so_buf_cntl = A6XX_VPC_SO_BUF_CNTL_ENABLE |
-               COND(tf->ncomp[0] > 0, A6XX_VPC_SO_BUF_CNTL_BUF0) |
-               COND(tf->ncomp[1] > 0, A6XX_VPC_SO_BUF_CNTL_BUF1) |
-               COND(tf->ncomp[2] > 0, A6XX_VPC_SO_BUF_CNTL_BUF2) |
-               COND(tf->ncomp[3] > 0, A6XX_VPC_SO_BUF_CNTL_BUF3);
+   tu_cs_emit_pkt7(cs, CP_CONTEXT_REG_BUNCH, 12 + 2 * prog_count);
+   tu_cs_emit(cs, REG_A6XX_VPC_SO_BUF_CNTL);
+   tu_cs_emit(cs, A6XX_VPC_SO_BUF_CNTL_ENABLE |
+                  COND(ncomp[0] > 0, A6XX_VPC_SO_BUF_CNTL_BUF0) |
+                  COND(ncomp[1] > 0, A6XX_VPC_SO_BUF_CNTL_BUF1) |
+                  COND(ncomp[2] > 0, A6XX_VPC_SO_BUF_CNTL_BUF2) |
+                  COND(ncomp[3] > 0, A6XX_VPC_SO_BUF_CNTL_BUF3));
+   for (uint32_t i = 0; i < 4; i++) {
+      tu_cs_emit(cs, REG_A6XX_VPC_SO_NCOMP(i));
+      tu_cs_emit(cs, ncomp[i]);
+   }
+   /* note: "VPC_SO_CNTL" write seems to be responsible for resetting the SO_PROG */
+   tu_cs_emit(cs, REG_A6XX_VPC_SO_CNTL);
+   tu_cs_emit(cs, A6XX_VPC_SO_CNTL_ENABLE);
+   for (uint32_t i = 0; i < prog_count; i++) {
+      tu_cs_emit(cs, REG_A6XX_VPC_SO_PROG);
+      tu_cs_emit(cs, prog[i]);
+   }
 }
 
 static void
@@ -710,8 +731,7 @@ tu6_emit_vpc(struct tu_cs *cs,
              const struct ir3_shader_variant *hs,
              const struct ir3_shader_variant *ds,
              const struct ir3_shader_variant *gs,
-             const struct ir3_shader_variant *fs,
-             struct tu_streamout_state *tf)
+             const struct ir3_shader_variant *fs)
 {
    const struct ir3_shader_variant *last_shader;
    if (gs) {
@@ -762,8 +782,7 @@ tu6_emit_vpc(struct tu_cs *cs,
       ir3_link_add(&linkage, pointsize_regid, 0x1, linkage.max_loc);
    }
 
-   if (last_shader->shader->stream_output.num_outputs)
-      tu6_setup_streamout(last_shader, &linkage, tf);
+   tu6_setup_streamout(cs, last_shader, &linkage);
 
    /* map outputs of the last shader to VPC */
    assert(linkage.cnt <= 32);
@@ -1316,8 +1335,7 @@ static void
 tu6_emit_program(struct tu_cs *cs,
                  struct tu_pipeline_builder *builder,
                  const struct tu_bo *binary_bo,
-                 bool binning_pass,
-                 struct tu_streamout_state *tf)
+                 bool binning_pass)
 {
    const struct ir3_shader_variant *vs = builder->variants[MESA_SHADER_VERTEX];
    const struct ir3_shader_variant *bs = builder->binning_variant;
@@ -1355,7 +1373,7 @@ tu6_emit_program(struct tu_cs *cs,
    tu_cs_emit_pkt4(cs, REG_A6XX_SP_HS_UNKNOWN_A831, 1);
    tu_cs_emit(cs, 0);
 
-   tu6_emit_vpc(cs, vs, hs, ds, gs, fs, tf);
+   tu6_emit_vpc(cs, vs, hs, ds, gs, fs);
    tu6_emit_vpc_varying_modes(cs, fs);
 
    if (fs) {
@@ -2040,11 +2058,11 @@ tu_pipeline_builder_parse_shader_stages(struct tu_pipeline_builder *builder,
 {
    struct tu_cs prog_cs;
    tu_cs_begin_sub_stream(&pipeline->cs, 512, &prog_cs);
-   tu6_emit_program(&prog_cs, builder, &pipeline->program.binary_bo, false, &pipeline->streamout);
+   tu6_emit_program(&prog_cs, builder, &pipeline->program.binary_bo, false);
    pipeline->program.state_ib = tu_cs_end_sub_stream(&pipeline->cs, &prog_cs);
 
    tu_cs_begin_sub_stream(&pipeline->cs, 512, &prog_cs);
-   tu6_emit_program(&prog_cs, builder, &pipeline->program.binary_bo, true, &pipeline->streamout);
+   tu6_emit_program(&prog_cs, builder, &pipeline->program.binary_bo, true);
    pipeline->program.binning_state_ib = tu_cs_end_sub_stream(&pipeline->cs, &prog_cs);
 
    VkShaderStageFlags stages = 0;
diff --git a/src/freedreno/vulkan/tu_private.h b/src/freedreno/vulkan/tu_private.h
index 532ab64362c..0109b5c2ba5 100644
--- a/src/freedreno/vulkan/tu_private.h
+++ b/src/freedreno/vulkan/tu_private.h
@@ -674,19 +674,10 @@ enum tu_cmd_dirty_bits
    TU_CMD_DIRTY_DESCRIPTOR_SETS = 1 << 3,
    TU_CMD_DIRTY_COMPUTE_DESCRIPTOR_SETS = 1 << 4,
    TU_CMD_DIRTY_SHADER_CONSTS = 1 << 5,
-   TU_CMD_DIRTY_STREAMOUT_BUFFERS = 1 << 6,
    /* all draw states were disabled and need to be re-enabled: */
    TU_CMD_DIRTY_DRAW_STATE = 1 << 7,
 };
 
-struct tu_streamout_state {
-   uint16_t stride[IR3_MAX_SO_BUFFERS];
-   uint32_t ncomp[IR3_MAX_SO_BUFFERS];
-   uint32_t prog[IR3_MAX_SO_OUTPUTS * 2];
-   uint32_t prog_count;
-   uint32_t vpc_so_buf_cntl;
-};
-
 /* There are only three cache domains we have to care about: the CCU, or
  * color cache unit, which is used for color and depth/stencil attachments
  * and copy/blit destinations, and is split conceptually into color and depth,
@@ -824,17 +815,6 @@ struct tu_cmd_state
    struct tu_cs_entry desc_sets_ib, desc_sets_load_ib;
    struct tu_cs_entry ia_gmem_ib, ia_sysmem_ib;
 
-   /* Stream output buffers */
-   struct
-   {
-      struct tu_buffer *buffers[IR3_MAX_SO_BUFFERS];
-      VkDeviceSize offsets[IR3_MAX_SO_BUFFERS];
-      VkDeviceSize sizes[IR3_MAX_SO_BUFFERS];
-   } streamout_buf;
-
-   uint8_t streamout_reset;
-   uint8_t streamout_enabled;
-
    /* Index buffer */
    struct tu_buffer *index_buffer;
    uint64_t index_offset;
@@ -842,6 +822,12 @@ struct tu_cmd_state
    uint32_t max_index_count;
    uint64_t index_va;
 
+   /* because streamout base has to be 32-byte aligned
+    * there is an extra offset to deal with when it is
+    * unaligned
+    */
+   uint8_t streamout_offset[IR3_MAX_SO_BUFFERS];
+
    /* Renderpasses are tricky, because we may need to flush differently if
     * using sysmem vs. gmem and therefore we have to delay any flushing that
     * happens before a renderpass. So we have to have two copies of the flush
@@ -860,6 +846,8 @@ struct tu_cmd_state
    struct tu_tiling_config tiling_config;
 
    struct tu_cs_entry tile_store_ib;
+
+   bool xfb_used;
 };
 
 struct tu_cmd_pool
@@ -1067,8 +1055,6 @@ struct tu_pipeline
    VkShaderStageFlags active_stages;
    uint32_t active_desc_sets;
 
-   struct tu_streamout_state streamout;
-
    /* mask of enabled dynamic states
     * if BIT(i) is set, pipeline->dynamic_state[i] is *NOT* used
     */
@@ -1169,8 +1155,7 @@ tu6_emit_vpc(struct tu_cs *cs,
              const struct ir3_shader_variant *hs,
              const struct ir3_shader_variant *ds,
              const struct ir3_shader_variant *gs,
-             const struct ir3_shader_variant *fs,
-             struct tu_streamout_state *tf);
+             const struct ir3_shader_variant *fs);
 
 void
 tu6_emit_fs_inputs(struct tu_cs *cs, const struct ir3_shader_variant *fs);