radeonsi: calculate optimal GS ring sizes to fix GS hangs on Tonga
[mesa.git] / src / gallium / drivers / radeonsi / si_state_shaders.c
index f673388b121a8c8007311ac66cb622e360a6aa7f..007a0ecb47628472a43be7b95270c011ea39e71f 100644 (file)
@@ -33,6 +33,7 @@
 #include "tgsi/tgsi_parse.h"
 #include "tgsi/tgsi_ureg.h"
 #include "util/u_memory.h"
+#include "util/u_prim.h"
 #include "util/u_simple_shaders.h"
 
 static void si_set_tesseval_regs(struct si_shader *shader,
@@ -122,7 +123,8 @@ static void si_shader_ls(struct si_shader *shader)
 
        shader->ls_rsrc1 = S_00B528_VGPRS((shader->num_vgprs - 1) / 4) |
                           S_00B528_SGPRS((num_sgprs - 1) / 8) |
-                          S_00B528_VGPR_COMP_CNT(vgpr_comp_cnt);
+                          S_00B528_VGPR_COMP_CNT(vgpr_comp_cnt) |
+                          S_00B528_DX10_CLAMP(shader->dx10_clamp_mode);
        shader->ls_rsrc2 = S_00B52C_USER_SGPR(num_user_sgprs) |
                           S_00B52C_SCRATCH_EN(shader->scratch_bytes_per_wave > 0);
 }
@@ -154,7 +156,8 @@ static void si_shader_hs(struct si_shader *shader)
        si_pm4_set_reg(pm4, R_00B424_SPI_SHADER_PGM_HI_HS, va >> 40);
        si_pm4_set_reg(pm4, R_00B428_SPI_SHADER_PGM_RSRC1_HS,
                       S_00B428_VGPRS((shader->num_vgprs - 1) / 4) |
-                      S_00B428_SGPRS((num_sgprs - 1) / 8));
+                      S_00B428_SGPRS((num_sgprs - 1) / 8) |
+                      S_00B428_DX10_CLAMP(shader->dx10_clamp_mode));
        si_pm4_set_reg(pm4, R_00B42C_SPI_SHADER_PGM_RSRC2_HS,
                       S_00B42C_USER_SGPR(num_user_sgprs) |
                       S_00B42C_SCRATCH_EN(shader->scratch_bytes_per_wave > 0));
@@ -177,7 +180,7 @@ static void si_shader_es(struct si_shader *shader)
 
        if (shader->selector->type == PIPE_SHADER_VERTEX) {
                vgpr_comp_cnt = shader->uses_instanceid ? 3 : 0;
-               num_user_sgprs = SI_VS_NUM_USER_SGPR;
+               num_user_sgprs = SI_ES_NUM_USER_SGPR;
        } else if (shader->selector->type == PIPE_SHADER_TESS_EVAL) {
                vgpr_comp_cnt = 3; /* all components are needed for TES */
                num_user_sgprs = SI_TES_NUM_USER_SGPR;
@@ -207,32 +210,17 @@ static void si_shader_es(struct si_shader *shader)
                si_set_tesseval_regs(shader, pm4);
 }
 
-static unsigned si_gs_get_max_stream(struct si_shader *shader)
-{
-       struct pipe_stream_output_info *so = &shader->selector->so;
-       unsigned max_stream = 0, i;
-
-       if (so->num_outputs == 0)
-               return 0;
-
-       for (i = 0; i < so->num_outputs; i++) {
-               if (so->output[i].stream > max_stream)
-                       max_stream = so->output[i].stream;
-       }
-       return max_stream;
-}
-
 static void si_shader_gs(struct si_shader *shader)
 {
-       unsigned gs_vert_itemsize = shader->selector->info.num_outputs * 16;
+       unsigned gs_vert_itemsize = shader->selector->gsvs_vertex_size;
        unsigned gs_max_vert_out = shader->selector->gs_max_out_vertices;
-       unsigned gsvs_itemsize = (gs_vert_itemsize * gs_max_vert_out) >> 2;
+       unsigned gsvs_itemsize = shader->selector->max_gsvs_emit_size >> 2;
        unsigned gs_num_invocations = shader->selector->gs_num_invocations;
        unsigned cut_mode;
        struct si_pm4_state *pm4;
        unsigned num_sgprs, num_user_sgprs;
        uint64_t va;
-       unsigned max_stream = si_gs_get_max_stream(shader);
+       unsigned max_stream = shader->selector->max_gs_stream;
 
        /* The GSVS_RING_ITEMSIZE register takes 15 bits */
        assert(gsvs_itemsize < (1 << 15));
@@ -264,7 +252,7 @@ static void si_shader_gs(struct si_shader *shader)
        si_pm4_set_reg(pm4, R_028A68_VGT_GSVS_RING_OFFSET_3, gsvs_itemsize * ((max_stream >= 3) ? 3 : 1));
 
        si_pm4_set_reg(pm4, R_028AAC_VGT_ESGS_RING_ITEMSIZE,
-                      util_bitcount64(shader->selector->inputs_read) * (16 >> 2));
+                      shader->selector->esgs_itemsize / 4);
        si_pm4_set_reg(pm4, R_028AB0_VGT_GSVS_RING_ITEMSIZE, gsvs_itemsize * (max_stream + 1));
 
        si_pm4_set_reg(pm4, R_028B38_VGT_GS_MAX_VERT_OUT, gs_max_vert_out);
@@ -402,6 +390,7 @@ static void si_shader_ps(struct si_shader *shader)
        unsigned num_sgprs, num_user_sgprs;
        unsigned spi_baryc_cntl = 0;
        uint64_t va;
+       bool has_centroid;
 
        pm4 = shader->pm4 = CALLOC_STRUCT(si_pm4_state);
 
@@ -433,8 +422,11 @@ static void si_shader_ps(struct si_shader *shader)
                }
        }
 
+       has_centroid = G_0286CC_PERSP_CENTROID_ENA(shader->spi_ps_input_ena) ||
+                      G_0286CC_LINEAR_CENTROID_ENA(shader->spi_ps_input_ena);
+
        spi_ps_in_control = S_0286D8_NUM_INTERP(shader->nparam) |
-               S_0286D8_BC_OPTIMIZE_DISABLE(1);
+                           S_0286D8_BC_OPTIMIZE_DISABLE(has_centroid);
 
        si_pm4_set_reg(pm4, R_0286E0_SPI_BARYC_CNTL, spi_baryc_cntl);
        si_pm4_set_reg(pm4, R_0286D8_SPI_PS_IN_CONTROL, spi_ps_in_control);
@@ -521,26 +513,26 @@ static inline void si_shader_selector_key(struct pipe_context *ctx,
                                key->vs.instance_divisors[i] =
                                        sctx->vertex_elements->elements[i].instance_divisor;
 
-               if (sctx->tes_shader)
+               if (sctx->tes_shader.cso)
                        key->vs.as_ls = 1;
-               else if (sctx->gs_shader) {
+               else if (sctx->gs_shader.cso) {
                        key->vs.as_es = 1;
-                       key->vs.es_enabled_outputs = sctx->gs_shader->inputs_read;
+                       key->vs.es_enabled_outputs = sctx->gs_shader.cso->inputs_read;
                }
 
-               if (!sctx->gs_shader && sctx->ps_shader &&
-                   sctx->ps_shader->info.uses_primid)
+               if (!sctx->gs_shader.cso && sctx->ps_shader.cso &&
+                   sctx->ps_shader.cso->info.uses_primid)
                        key->vs.export_prim_id = 1;
                break;
        case PIPE_SHADER_TESS_CTRL:
                key->tcs.prim_mode =
-                       sctx->tes_shader->info.properties[TGSI_PROPERTY_TES_PRIM_MODE];
+                       sctx->tes_shader.cso->info.properties[TGSI_PROPERTY_TES_PRIM_MODE];
                break;
        case PIPE_SHADER_TESS_EVAL:
-               if (sctx->gs_shader) {
+               if (sctx->gs_shader.cso) {
                        key->tes.as_es = 1;
-                       key->tes.es_enabled_outputs = sctx->gs_shader->inputs_read;
-               } else if (sctx->ps_shader && sctx->ps_shader->info.uses_primid)
+                       key->tes.es_enabled_outputs = sctx->gs_shader.cso->inputs_read;
+               } else if (sctx->ps_shader.cso && sctx->ps_shader.cso->info.uses_primid)
                        key->tes.export_prim_id = 1;
                break;
        case PIPE_SHADER_GEOMETRY:
@@ -570,6 +562,7 @@ static inline void si_shader_selector_key(struct pipe_context *ctx,
                        key->ps.poly_line_smoothing = ((is_poly && rs->poly_smooth) ||
                                                       (is_line && rs->line_smooth)) &&
                                                      sctx->framebuffer.nr_samples <= 1;
+                       key->ps.clamp_color = rs->clamp_fragment_color;
                }
 
                key->ps.alpha_func = PIPE_FUNC_ALWAYS;
@@ -586,11 +579,13 @@ static inline void si_shader_selector_key(struct pipe_context *ctx,
 
 /* Select the hw shader variant depending on the current state. */
 static int si_shader_select(struct pipe_context *ctx,
-                           struct si_shader_selector *sel)
+                           struct si_shader_ctx_state *state)
 {
        struct si_context *sctx = (struct si_context *)ctx;
+       struct si_shader_selector *sel = state->cso;
+       struct si_shader *current = state->current;
        union si_shader_key key;
-       struct si_shader * shader = NULL;
+       struct si_shader *iter, *shader = NULL;
        int r;
 
        si_shader_selector_key(ctx, sel, &key);
@@ -599,55 +594,56 @@ static int si_shader_select(struct pipe_context *ctx,
         * This path is also used for most shaders that don't need multiple
         * variants, it will cost just a computation of the key and this
         * test. */
-       if (likely(sel->current && memcmp(&sel->current->key, &key, sizeof(key)) == 0)) {
+       if (likely(current && memcmp(&current->key, &key, sizeof(key)) == 0))
                return 0;
-       }
 
-       /* lookup if we have other variants in the list */
-       if (sel->num_shaders > 1) {
-               struct si_shader *p = sel->current, *c = p->next_variant;
+       pipe_mutex_lock(sel->mutex);
 
-               while (c && memcmp(&c->key, &key, sizeof(key)) != 0) {
-                       p = c;
-                       c = c->next_variant;
+       /* Find the shader variant. */
+       for (iter = sel->first_variant; iter; iter = iter->next_variant) {
+               /* Don't check the "current" shader. We checked it above. */
+               if (current != iter &&
+                   memcmp(&iter->key, &key, sizeof(key)) == 0) {
+                       state->current = iter;
+                       pipe_mutex_unlock(sel->mutex);
+                       return 0;
                }
+       }
 
-               if (c) {
-                       p->next_variant = c->next_variant;
-                       shader = c;
-               }
+       /* Build a new shader. */
+       shader = CALLOC_STRUCT(si_shader);
+       if (!shader) {
+               pipe_mutex_unlock(sel->mutex);
+               return -ENOMEM;
        }
+       shader->selector = sel;
+       shader->key = key;
+
+       r = si_shader_create(sctx->screen, sctx->tm, shader);
+       if (unlikely(r)) {
+               R600_ERR("Failed to build shader variant (type=%u) %d\n",
+                        sel->type, r);
+               FREE(shader);
+               pipe_mutex_unlock(sel->mutex);
+               return r;
+       }
+       si_shader_init_pm4_state(shader);
 
-       if (shader) {
-               shader->next_variant = sel->current;
-               sel->current = shader;
+       if (!sel->last_variant) {
+               sel->first_variant = shader;
+               sel->last_variant = shader;
        } else {
-               shader = CALLOC(1, sizeof(struct si_shader));
-               shader->selector = sel;
-               shader->key = key;
-
-               shader->next_variant = sel->current;
-               sel->current = shader;
-               r = si_shader_create((struct si_screen*)ctx->screen, sctx->tm,
-                                    shader);
-               if (unlikely(r)) {
-                       R600_ERR("Failed to build shader variant (type=%u) %d\n",
-                                sel->type, r);
-                       sel->current = NULL;
-                       FREE(shader);
-                       return r;
-               }
-               si_shader_init_pm4_state(shader);
-               sel->num_shaders++;
-               p_atomic_inc(&sctx->screen->b.num_compilations);
+               sel->last_variant->next_variant = shader;
+               sel->last_variant = shader;
        }
-
+       state->current = shader;
+       p_atomic_inc(&sctx->screen->b.num_compilations);
+       pipe_mutex_unlock(sel->mutex);
        return 0;
 }
 
-static void *si_create_shader_state(struct pipe_context *ctx,
-                                   const struct pipe_shader_state *state,
-                                   unsigned pipe_shader_type)
+static void *si_create_shader_selector(struct pipe_context *ctx,
+                                      const struct pipe_shader_state *state)
 {
        struct si_screen *sscreen = (struct si_screen *)ctx->screen;
        struct si_shader_selector *sel = CALLOC_STRUCT(si_shader_selector);
@@ -656,7 +652,6 @@ static void *si_create_shader_state(struct pipe_context *ctx,
        if (!sel)
                return NULL;
 
-       sel->type = pipe_shader_type;
        sel->tokens = tgsi_dup_tokens(state->tokens);
        if (!sel->tokens) {
                FREE(sel);
@@ -665,6 +660,7 @@ static void *si_create_shader_state(struct pipe_context *ctx,
 
        sel->so = state->stream_output;
        tgsi_scan_shader(state->tokens, &sel->info);
+       sel->type = util_pipe_shader_from_tgsi_processor(sel->info.processor);
        p_atomic_inc(&sscreen->b.num_shaders_created);
 
        /* First set which opcode uses which (i,j) pair. */
@@ -695,7 +691,7 @@ static void *si_create_shader_state(struct pipe_context *ctx,
                sel->info.uses_linear_centroid +
                sel->info.uses_linear_sample >= 2;
 
-       switch (pipe_shader_type) {
+       switch (sel->type) {
        case PIPE_SHADER_GEOMETRY:
                sel->gs_output_prim =
                        sel->info.properties[TGSI_PROPERTY_GS_OUTPUT_PRIM];
@@ -703,8 +699,17 @@ static void *si_create_shader_state(struct pipe_context *ctx,
                        sel->info.properties[TGSI_PROPERTY_GS_MAX_OUTPUT_VERTICES];
                sel->gs_num_invocations =
                        sel->info.properties[TGSI_PROPERTY_GS_INVOCATIONS];
-               sel->gsvs_itemsize = sel->info.num_outputs * 16 *
-                                    sel->gs_max_out_vertices;
+               sel->gsvs_vertex_size = sel->info.num_outputs * 16;
+               sel->max_gsvs_emit_size = sel->gsvs_vertex_size *
+                                         sel->gs_max_out_vertices;
+
+               sel->max_gs_stream = 0;
+               for (i = 0; i < sel->so.num_outputs; i++)
+                       sel->max_gs_stream = MAX2(sel->max_gs_stream,
+                                                 sel->so.output[i].stream);
+
+               sel->gs_input_verts_per_prim =
+                       u_vertices_per_prim(sel->info.properties[TGSI_PROPERTY_GS_INPUT_PRIM]);
 
                for (i = 0; i < sel->info.num_inputs; i++) {
                        unsigned name = sel->info.input_semantic_name[i];
@@ -722,6 +727,7 @@ static void *si_create_shader_state(struct pipe_context *ctx,
 
        case PIPE_SHADER_VERTEX:
        case PIPE_SHADER_TESS_CTRL:
+       case PIPE_SHADER_TESS_EVAL:
                for (i = 0; i < sel->info.num_outputs; i++) {
                        unsigned name = sel->info.output_semantic_name[i];
                        unsigned index = sel->info.output_semantic_index[i];
@@ -738,6 +744,7 @@ static void *si_create_shader_state(struct pipe_context *ctx,
                                        1llu << si_shader_io_get_unique_index(name, index);
                        }
                }
+               sel->esgs_itemsize = util_last_bit64(sel->outputs_written) * 16;
                break;
        case PIPE_SHADER_FRAGMENT:
                for (i = 0; i < sel->info.num_outputs; i++) {
@@ -750,47 +757,21 @@ static void *si_create_shader_state(struct pipe_context *ctx,
                break;
        }
 
-       if (sscreen->b.debug_flags & DBG_PRECOMPILE)
-               if (si_shader_select(ctx, sel)) {
+       if (sscreen->b.debug_flags & DBG_PRECOMPILE) {
+               struct si_shader_ctx_state state = {sel};
+
+               if (si_shader_select(ctx, &state)) {
                        fprintf(stderr, "radeonsi: can't create a shader\n");
                        tgsi_free_tokens(sel->tokens);
                        FREE(sel);
                        return NULL;
                }
+       }
 
+       pipe_mutex_init(sel->mutex);
        return sel;
 }
 
-static void *si_create_fs_state(struct pipe_context *ctx,
-                               const struct pipe_shader_state *state)
-{
-       return si_create_shader_state(ctx, state, PIPE_SHADER_FRAGMENT);
-}
-
-static void *si_create_gs_state(struct pipe_context *ctx,
-                               const struct pipe_shader_state *state)
-{
-       return si_create_shader_state(ctx, state, PIPE_SHADER_GEOMETRY);
-}
-
-static void *si_create_vs_state(struct pipe_context *ctx,
-                               const struct pipe_shader_state *state)
-{
-       return si_create_shader_state(ctx, state, PIPE_SHADER_VERTEX);
-}
-
-static void *si_create_tcs_state(struct pipe_context *ctx,
-                                const struct pipe_shader_state *state)
-{
-       return si_create_shader_state(ctx, state, PIPE_SHADER_TESS_CTRL);
-}
-
-static void *si_create_tes_state(struct pipe_context *ctx,
-                                const struct pipe_shader_state *state)
-{
-       return si_create_shader_state(ctx, state, PIPE_SHADER_TESS_EVAL);
-}
-
 /**
  * Normally, we only emit 1 viewport and 1 scissor if no shader is using
  * the VIEWPORT_INDEX output, and emitting the other viewports and scissors
@@ -815,10 +796,11 @@ static void si_bind_vs_shader(struct pipe_context *ctx, void *state)
        struct si_context *sctx = (struct si_context *)ctx;
        struct si_shader_selector *sel = state;
 
-       if (sctx->vs_shader == sel || !sel)
+       if (sctx->vs_shader.cso == sel)
                return;
 
-       sctx->vs_shader = sel;
+       sctx->vs_shader.cso = sel;
+       sctx->vs_shader.current = sel ? sel->first_variant : NULL;
        si_mark_atom_dirty(sctx, &sctx->clip_regs);
        si_update_viewports_and_scissors(sctx);
 }
@@ -827,12 +809,13 @@ static void si_bind_gs_shader(struct pipe_context *ctx, void *state)
 {
        struct si_context *sctx = (struct si_context *)ctx;
        struct si_shader_selector *sel = state;
-       bool enable_changed = !!sctx->gs_shader != !!sel;
+       bool enable_changed = !!sctx->gs_shader.cso != !!sel;
 
-       if (sctx->gs_shader == sel)
+       if (sctx->gs_shader.cso == sel)
                return;
 
-       sctx->gs_shader = sel;
+       sctx->gs_shader.cso = sel;
+       sctx->gs_shader.current = sel ? sel->first_variant : NULL;
        si_mark_atom_dirty(sctx, &sctx->clip_regs);
        sctx->last_rast_prim = -1; /* reset this so that it gets updated */
 
@@ -845,12 +828,13 @@ static void si_bind_tcs_shader(struct pipe_context *ctx, void *state)
 {
        struct si_context *sctx = (struct si_context *)ctx;
        struct si_shader_selector *sel = state;
-       bool enable_changed = !!sctx->tcs_shader != !!sel;
+       bool enable_changed = !!sctx->tcs_shader.cso != !!sel;
 
-       if (sctx->tcs_shader == sel)
+       if (sctx->tcs_shader.cso == sel)
                return;
 
-       sctx->tcs_shader = sel;
+       sctx->tcs_shader.cso = sel;
+       sctx->tcs_shader.current = sel ? sel->first_variant : NULL;
 
        if (enable_changed)
                sctx->last_tcs = NULL; /* invalidate derived tess state */
@@ -860,12 +844,13 @@ static void si_bind_tes_shader(struct pipe_context *ctx, void *state)
 {
        struct si_context *sctx = (struct si_context *)ctx;
        struct si_shader_selector *sel = state;
-       bool enable_changed = !!sctx->tes_shader != !!sel;
+       bool enable_changed = !!sctx->tes_shader.cso != !!sel;
 
-       if (sctx->tes_shader == sel)
+       if (sctx->tes_shader.cso == sel)
                return;
 
-       sctx->tes_shader = sel;
+       sctx->tes_shader.cso = sel;
+       sctx->tes_shader.current = sel ? sel->first_variant : NULL;
        si_mark_atom_dirty(sctx, &sctx->clip_regs);
        sctx->last_rast_prim = -1; /* reset this so that it gets updated */
 
@@ -876,40 +861,37 @@ static void si_bind_tes_shader(struct pipe_context *ctx, void *state)
        si_update_viewports_and_scissors(sctx);
 }
 
-static void si_make_dummy_ps(struct si_context *sctx)
-{
-       if (!sctx->dummy_pixel_shader) {
-               sctx->dummy_pixel_shader =
-                       util_make_fragment_cloneinput_shader(&sctx->b.b, 0,
-                                                            TGSI_SEMANTIC_GENERIC,
-                                                            TGSI_INTERPOLATE_CONSTANT);
-       }
-}
-
 static void si_bind_ps_shader(struct pipe_context *ctx, void *state)
 {
        struct si_context *sctx = (struct si_context *)ctx;
        struct si_shader_selector *sel = state;
 
        /* skip if supplied shader is one already in use */
-       if (sctx->ps_shader == sel)
+       if (sctx->ps_shader.cso == sel)
                return;
 
-       /* use a dummy shader if binding a NULL shader */
-       if (!sel) {
-               si_make_dummy_ps(sctx);
-               sel = sctx->dummy_pixel_shader;
-       }
-
-       sctx->ps_shader = sel;
+       sctx->ps_shader.cso = sel;
+       sctx->ps_shader.current = sel ? sel->first_variant : NULL;
        si_mark_atom_dirty(sctx, &sctx->cb_target_mask);
 }
 
-static void si_delete_shader_selector(struct pipe_context *ctx,
-                                     struct si_shader_selector *sel)
+static void si_delete_shader_selector(struct pipe_context *ctx, void *state)
 {
        struct si_context *sctx = (struct si_context *)ctx;
-       struct si_shader *p = sel->current, *c;
+       struct si_shader_selector *sel = (struct si_shader_selector *)state;
+       struct si_shader *p = sel->first_variant, *c;
+       struct si_shader_ctx_state *current_shader[SI_NUM_SHADERS] = {
+               [PIPE_SHADER_VERTEX] = &sctx->vs_shader,
+               [PIPE_SHADER_TESS_CTRL] = &sctx->tcs_shader,
+               [PIPE_SHADER_TESS_EVAL] = &sctx->tes_shader,
+               [PIPE_SHADER_GEOMETRY] = &sctx->gs_shader,
+               [PIPE_SHADER_FRAGMENT] = &sctx->ps_shader,
+       };
+
+       if (current_shader[sel->type]->cso == sel) {
+               current_shader[sel->type]->cso = NULL;
+               current_shader[sel->type]->current = NULL;
+       }
 
        while (p) {
                c = p->next_variant;
@@ -940,87 +922,30 @@ static void si_delete_shader_selector(struct pipe_context *ctx,
                        break;
                }
 
-               si_shader_destroy(ctx, p);
+               si_shader_destroy(p);
                free(p);
                p = c;
        }
 
+       pipe_mutex_destroy(sel->mutex);
        free(sel->tokens);
        free(sel);
 }
 
-static void si_delete_vs_shader(struct pipe_context *ctx, void *state)
-{
-       struct si_context *sctx = (struct si_context *)ctx;
-       struct si_shader_selector *sel = (struct si_shader_selector *)state;
-
-       if (sctx->vs_shader == sel) {
-               sctx->vs_shader = NULL;
-       }
-
-       si_delete_shader_selector(ctx, sel);
-}
-
-static void si_delete_gs_shader(struct pipe_context *ctx, void *state)
-{
-       struct si_context *sctx = (struct si_context *)ctx;
-       struct si_shader_selector *sel = (struct si_shader_selector *)state;
-
-       if (sctx->gs_shader == sel) {
-               sctx->gs_shader = NULL;
-       }
-
-       si_delete_shader_selector(ctx, sel);
-}
-
-static void si_delete_ps_shader(struct pipe_context *ctx, void *state)
-{
-       struct si_context *sctx = (struct si_context *)ctx;
-       struct si_shader_selector *sel = (struct si_shader_selector *)state;
-
-       if (sctx->ps_shader == sel) {
-               sctx->ps_shader = NULL;
-       }
-
-       si_delete_shader_selector(ctx, sel);
-}
-
-static void si_delete_tcs_shader(struct pipe_context *ctx, void *state)
-{
-       struct si_context *sctx = (struct si_context *)ctx;
-       struct si_shader_selector *sel = (struct si_shader_selector *)state;
-
-       if (sctx->tcs_shader == sel) {
-               sctx->tcs_shader = NULL;
-       }
-
-       si_delete_shader_selector(ctx, sel);
-}
-
-static void si_delete_tes_shader(struct pipe_context *ctx, void *state)
-{
-       struct si_context *sctx = (struct si_context *)ctx;
-       struct si_shader_selector *sel = (struct si_shader_selector *)state;
-
-       if (sctx->tes_shader == sel) {
-               sctx->tes_shader = NULL;
-       }
-
-       si_delete_shader_selector(ctx, sel);
-}
-
 static void si_emit_spi_map(struct si_context *sctx, struct r600_atom *atom)
 {
-       struct radeon_winsys_cs *cs = sctx->b.rings.gfx.cs;
-       struct si_shader *ps = sctx->ps_shader->current;
+       struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
+       struct si_shader *ps = sctx->ps_shader.current;
        struct si_shader *vs = si_get_vs_state(sctx);
-       struct tgsi_shader_info *psinfo = &ps->selector->info;
+       struct tgsi_shader_info *psinfo;
        struct tgsi_shader_info *vsinfo = &vs->selector->info;
        unsigned i, j, tmp, num_written = 0;
 
-       if (!ps->nparam)
+       if (!ps || !ps->nparam)
                return;
 
+       psinfo = &ps->selector->info;
+
        radeon_set_context_reg_seq(cs, R_028644_SPI_PS_INPUT_CNTL_0, ps->nparam);
 
        for (i = 0; i < psinfo->num_inputs; i++) {
@@ -1081,9 +1006,14 @@ bcolor:
 
 static void si_emit_spi_ps_input(struct si_context *sctx, struct r600_atom *atom)
 {
-       struct radeon_winsys_cs *cs = sctx->b.rings.gfx.cs;
-       struct si_shader *ps = sctx->ps_shader->current;
-       unsigned input_ena = ps->spi_ps_input_ena;
+       struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
+       struct si_shader *ps = sctx->ps_shader.current;
+       unsigned input_ena;
+
+       if (!ps)
+               return;
+
+       input_ena = ps->spi_ps_input_ena;
 
        /* we need to enable at least one of them, otherwise we hang the GPU */
        assert(G_0286CC_PERSP_SAMPLE_ENA(input_ena) ||
@@ -1144,6 +1074,7 @@ static void si_init_config_add_vgt_flush(struct si_context *sctx)
        if (sctx->init_config_has_vgt_flush)
                return;
 
+       /* VGT_FLUSH is required even if VGT is idle. It resets VGT pointers. */
        si_pm4_cmd_begin(sctx->init_config, PKT3_EVENT_WRITE);
        si_pm4_cmd_add(sctx->init_config, EVENT_TYPE(V_028A90_VGT_FLUSH) | EVENT_INDEX(0));
        si_pm4_cmd_end(sctx->init_config, false);
@@ -1151,70 +1082,127 @@ static void si_init_config_add_vgt_flush(struct si_context *sctx)
 }
 
 /* Initialize state related to ESGS / GSVS ring buffers */
-static void si_init_gs_rings(struct si_context *sctx)
+static bool si_update_gs_ring_buffers(struct si_context *sctx)
 {
-       unsigned esgs_ring_size = 128 * 1024;
-       unsigned gsvs_ring_size = 60 * 1024 * 1024;
+       struct si_shader_selector *es =
+               sctx->tes_shader.cso ? sctx->tes_shader.cso : sctx->vs_shader.cso;
+       struct si_shader_selector *gs = sctx->gs_shader.cso;
+       struct si_pm4_state *pm4;
 
-       assert(!sctx->esgs_ring && !sctx->gsvs_ring);
+       /* Chip constants. */
+       unsigned num_se = sctx->screen->b.info.max_se;
+       unsigned wave_size = 64;
+       unsigned max_gs_waves = 32 * num_se; /* max 32 per SE on GCN */
+       unsigned gs_vertex_reuse = 16 * num_se; /* GS_VERTEX_REUSE register (per SE) */
+       unsigned alignment = 256 * num_se;
+       /* The maximum size is 63.999 MB per SE. */
+       unsigned max_size = ((unsigned)(63.999 * 1024 * 1024) & ~255) * num_se;
+
+       /* Calculate the minimum size. */
+       unsigned min_esgs_ring_size = align(es->esgs_itemsize * gs_vertex_reuse *
+                                           wave_size, alignment);
+
+       /* These are recommended sizes, not minimum sizes. */
+       unsigned esgs_ring_size = max_gs_waves * 2 * wave_size *
+                                 es->esgs_itemsize * gs->gs_input_verts_per_prim;
+       unsigned gsvs_ring_size = max_gs_waves * 2 * wave_size *
+                                 gs->max_gsvs_emit_size * (gs->max_gs_stream + 1);
+
+       min_esgs_ring_size = align(min_esgs_ring_size, alignment);
+       esgs_ring_size = align(esgs_ring_size, alignment);
+       gsvs_ring_size = align(gsvs_ring_size, alignment);
+
+       esgs_ring_size = CLAMP(esgs_ring_size, min_esgs_ring_size, max_size);
+       gsvs_ring_size = MIN2(gsvs_ring_size, max_size);
+
+       /* Some rings don't have to be allocated if shaders don't use them.
+        * (e.g. no varyings between ES and GS or GS and VS)
+        */
+       bool update_esgs = esgs_ring_size &&
+                          (!sctx->esgs_ring ||
+                           sctx->esgs_ring->width0 < esgs_ring_size);
+       bool update_gsvs = gsvs_ring_size &&
+                          (!sctx->gsvs_ring ||
+                           sctx->gsvs_ring->width0 < gsvs_ring_size);
 
-       sctx->esgs_ring = pipe_buffer_create(sctx->b.b.screen, PIPE_BIND_CUSTOM,
-                                      PIPE_USAGE_DEFAULT, esgs_ring_size);
-       if (!sctx->esgs_ring)
-               return;
+       if (!update_esgs && !update_gsvs)
+               return true;
 
-       sctx->gsvs_ring = pipe_buffer_create(sctx->b.b.screen, PIPE_BIND_CUSTOM,
-                                            PIPE_USAGE_DEFAULT, gsvs_ring_size);
-       if (!sctx->gsvs_ring) {
+       if (update_esgs) {
                pipe_resource_reference(&sctx->esgs_ring, NULL);
-               return;
+               sctx->esgs_ring = pipe_buffer_create(sctx->b.b.screen, PIPE_BIND_CUSTOM,
+                                                    PIPE_USAGE_DEFAULT,
+                                                    esgs_ring_size);
+               if (!sctx->esgs_ring)
+                       return false;
        }
 
-       si_init_config_add_vgt_flush(sctx);
+       if (update_gsvs) {
+               pipe_resource_reference(&sctx->gsvs_ring, NULL);
+               sctx->gsvs_ring = pipe_buffer_create(sctx->b.b.screen, PIPE_BIND_CUSTOM,
+                                                    PIPE_USAGE_DEFAULT,
+                                                    gsvs_ring_size);
+               if (!sctx->gsvs_ring)
+                       return false;
+       }
+
+       /* Create the "init_config_gs_rings" state. */
+       pm4 = CALLOC_STRUCT(si_pm4_state);
+       if (!pm4)
+               return false;
 
-       /* Append these registers to the init config state. */
        if (sctx->b.chip_class >= CIK) {
-               if (sctx->b.chip_class >= VI) {
-                       /* The maximum sizes are 63.999 MB on VI, because
-                        * the register fields only have 18 bits. */
-                       assert(esgs_ring_size / 256 < (1 << 18));
-                       assert(gsvs_ring_size / 256 < (1 << 18));
-               }
-               si_pm4_set_reg(sctx->init_config, R_030900_VGT_ESGS_RING_SIZE,
-                              esgs_ring_size / 256);
-               si_pm4_set_reg(sctx->init_config, R_030904_VGT_GSVS_RING_SIZE,
-                              gsvs_ring_size / 256);
+               if (sctx->esgs_ring)
+                       si_pm4_set_reg(pm4, R_030900_VGT_ESGS_RING_SIZE,
+                                      sctx->esgs_ring->width0 / 256);
+               if (sctx->gsvs_ring)
+                       si_pm4_set_reg(pm4, R_030904_VGT_GSVS_RING_SIZE,
+                                      sctx->gsvs_ring->width0 / 256);
        } else {
-               si_pm4_set_reg(sctx->init_config, R_0088C8_VGT_ESGS_RING_SIZE,
-                              esgs_ring_size / 256);
-               si_pm4_set_reg(sctx->init_config, R_0088CC_VGT_GSVS_RING_SIZE,
-                              gsvs_ring_size / 256);
+               if (sctx->esgs_ring)
+                       si_pm4_set_reg(pm4, R_0088C8_VGT_ESGS_RING_SIZE,
+                                      sctx->esgs_ring->width0 / 256);
+               if (sctx->gsvs_ring)
+                       si_pm4_set_reg(pm4, R_0088CC_VGT_GSVS_RING_SIZE,
+                                      sctx->gsvs_ring->width0 / 256);
        }
 
-       /* Flush the context to re-emit the init_config state.
-        * This is done only once in a lifetime of a context.
-        */
-       si_pm4_upload_indirect_buffer(sctx, sctx->init_config);
+       /* Set the state. */
+       if (sctx->init_config_gs_rings)
+               si_pm4_free_state(sctx, sctx->init_config_gs_rings, ~0);
+       sctx->init_config_gs_rings = pm4;
+
+       if (!sctx->init_config_has_vgt_flush) {
+               si_init_config_add_vgt_flush(sctx);
+               si_pm4_upload_indirect_buffer(sctx, sctx->init_config);
+       }
+
+       /* Flush the context to re-emit both init_config states. */
        sctx->b.initial_gfx_cs_size = 0; /* force flush */
        si_context_gfx_flush(sctx, RADEON_FLUSH_ASYNC, NULL);
 
-       si_set_ring_buffer(&sctx->b.b, PIPE_SHADER_VERTEX, SI_RING_ESGS,
-                          sctx->esgs_ring, 0, esgs_ring_size,
-                          true, true, 4, 64, 0);
-       si_set_ring_buffer(&sctx->b.b, PIPE_SHADER_GEOMETRY, SI_RING_ESGS,
-                          sctx->esgs_ring, 0, esgs_ring_size,
-                          false, false, 0, 0, 0);
-       si_set_ring_buffer(&sctx->b.b, PIPE_SHADER_VERTEX, SI_RING_GSVS,
-                          sctx->gsvs_ring, 0, gsvs_ring_size,
-                          false, false, 0, 0, 0);
+       /* Set ring bindings. */
+       if (sctx->esgs_ring) {
+               si_set_ring_buffer(&sctx->b.b, PIPE_SHADER_VERTEX, SI_RING_ESGS,
+                                  sctx->esgs_ring, 0, sctx->esgs_ring->width0,
+                                  true, true, 4, 64, 0);
+               si_set_ring_buffer(&sctx->b.b, PIPE_SHADER_GEOMETRY, SI_RING_ESGS,
+                                  sctx->esgs_ring, 0, sctx->esgs_ring->width0,
+                                  false, false, 0, 0, 0);
+       }
+       if (sctx->gsvs_ring)
+               si_set_ring_buffer(&sctx->b.b, PIPE_SHADER_VERTEX, SI_RING_GSVS,
+                                  sctx->gsvs_ring, 0, sctx->gsvs_ring->width0,
+                                  false, false, 0, 0, 0);
+       return true;
 }
 
-static void si_update_gs_rings(struct si_context *sctx)
+static void si_update_gsvs_ring_bindings(struct si_context *sctx)
 {
-       unsigned gsvs_itemsize = sctx->gs_shader->gsvs_itemsize;
+       unsigned gsvs_itemsize = sctx->gs_shader.cso->max_gsvs_emit_size;
        uint64_t offset;
 
-       if (gsvs_itemsize == sctx->last_gsvs_itemsize)
+       if (!sctx->gsvs_ring || gsvs_itemsize == sctx->last_gsvs_itemsize)
                return;
 
        sctx->last_gsvs_itemsize = gsvs_itemsize;
@@ -1245,17 +1233,14 @@ static void si_update_gs_rings(struct si_context *sctx)
  *          < 0 if there was a failure
  */
 static int si_update_scratch_buffer(struct si_context *sctx,
-                                   struct si_shader_selector *sel)
+                                   struct si_shader *shader)
 {
-       struct si_shader *shader;
        uint64_t scratch_va = sctx->scratch_buffer->gpu_address;
        int r;
 
-       if (!sel)
+       if (!shader)
                return 0;
 
-       shader = sel->current;
-
        /* This shader doesn't need a scratch buffer */
        if (shader->scratch_bytes_per_wave == 0)
                return 0;
@@ -1284,30 +1269,23 @@ static int si_update_scratch_buffer(struct si_context *sctx,
 
 static unsigned si_get_current_scratch_buffer_size(struct si_context *sctx)
 {
-       if (!sctx->scratch_buffer)
-               return 0;
-
-       return sctx->scratch_buffer->b.b.width0;
+       return sctx->scratch_buffer ? sctx->scratch_buffer->b.b.width0 : 0;
 }
 
-static unsigned si_get_scratch_buffer_bytes_per_wave(struct si_context *sctx,
-                                       struct si_shader_selector *sel)
+static unsigned si_get_scratch_buffer_bytes_per_wave(struct si_shader *shader)
 {
-       if (!sel)
-               return 0;
-
-       return sel->current->scratch_bytes_per_wave;
+       return shader ? shader->scratch_bytes_per_wave : 0;
 }
 
 static unsigned si_get_max_scratch_bytes_per_wave(struct si_context *sctx)
 {
        unsigned bytes = 0;
 
-       bytes = MAX2(bytes, si_get_scratch_buffer_bytes_per_wave(sctx, sctx->ps_shader));
-       bytes = MAX2(bytes, si_get_scratch_buffer_bytes_per_wave(sctx, sctx->gs_shader));
-       bytes = MAX2(bytes, si_get_scratch_buffer_bytes_per_wave(sctx, sctx->vs_shader));
-       bytes = MAX2(bytes, si_get_scratch_buffer_bytes_per_wave(sctx, sctx->tcs_shader));
-       bytes = MAX2(bytes, si_get_scratch_buffer_bytes_per_wave(sctx, sctx->tes_shader));
+       bytes = MAX2(bytes, si_get_scratch_buffer_bytes_per_wave(sctx->ps_shader.current));
+       bytes = MAX2(bytes, si_get_scratch_buffer_bytes_per_wave(sctx->gs_shader.current));
+       bytes = MAX2(bytes, si_get_scratch_buffer_bytes_per_wave(sctx->vs_shader.current));
+       bytes = MAX2(bytes, si_get_scratch_buffer_bytes_per_wave(sctx->tcs_shader.current));
+       bytes = MAX2(bytes, si_get_scratch_buffer_bytes_per_wave(sctx->tes_shader.current));
        return bytes;
 }
 
@@ -1322,7 +1300,6 @@ static bool si_update_spi_tmpring_size(struct si_context *sctx)
        int r;
 
        if (scratch_needed_size > 0) {
-
                if (scratch_needed_size > current_scratch_buffer_size) {
                        /* Create a bigger scratch buffer */
                        pipe_resource_reference(
@@ -1342,58 +1319,46 @@ static bool si_update_spi_tmpring_size(struct si_context *sctx)
                 * last used, so we still need to try to update them, even if
                 * they require scratch buffers smaller than the current size.
                 */
-               r = si_update_scratch_buffer(sctx, sctx->ps_shader);
+               r = si_update_scratch_buffer(sctx, sctx->ps_shader.current);
                if (r < 0)
                        return false;
                if (r == 1)
-                       si_pm4_bind_state(sctx, ps, sctx->ps_shader->current->pm4);
+                       si_pm4_bind_state(sctx, ps, sctx->ps_shader.current->pm4);
 
-               r = si_update_scratch_buffer(sctx, sctx->gs_shader);
+               r = si_update_scratch_buffer(sctx, sctx->gs_shader.current);
                if (r < 0)
                        return false;
                if (r == 1)
-                       si_pm4_bind_state(sctx, gs, sctx->gs_shader->current->pm4);
+                       si_pm4_bind_state(sctx, gs, sctx->gs_shader.current->pm4);
 
-               r = si_update_scratch_buffer(sctx, sctx->tcs_shader);
+               r = si_update_scratch_buffer(sctx, sctx->tcs_shader.current);
                if (r < 0)
                        return false;
                if (r == 1)
-                       si_pm4_bind_state(sctx, hs, sctx->tcs_shader->current->pm4);
+                       si_pm4_bind_state(sctx, hs, sctx->tcs_shader.current->pm4);
 
                /* VS can be bound as LS, ES, or VS. */
-               if (sctx->tes_shader) {
-                       r = si_update_scratch_buffer(sctx, sctx->vs_shader);
-                       if (r < 0)
-                               return false;
-                       if (r == 1)
-                               si_pm4_bind_state(sctx, ls, sctx->vs_shader->current->pm4);
-               } else if (sctx->gs_shader) {
-                       r = si_update_scratch_buffer(sctx, sctx->vs_shader);
-                       if (r < 0)
-                               return false;
-                       if (r == 1)
-                               si_pm4_bind_state(sctx, es, sctx->vs_shader->current->pm4);
-               } else {
-                       r = si_update_scratch_buffer(sctx, sctx->vs_shader);
-                       if (r < 0)
-                               return false;
-                       if (r == 1)
-                               si_pm4_bind_state(sctx, vs, sctx->vs_shader->current->pm4);
+               r = si_update_scratch_buffer(sctx, sctx->vs_shader.current);
+               if (r < 0)
+                       return false;
+               if (r == 1) {
+                       if (sctx->tes_shader.current)
+                               si_pm4_bind_state(sctx, ls, sctx->vs_shader.current->pm4);
+                       else if (sctx->gs_shader.current)
+                               si_pm4_bind_state(sctx, es, sctx->vs_shader.current->pm4);
+                       else
+                               si_pm4_bind_state(sctx, vs, sctx->vs_shader.current->pm4);
                }
 
                /* TES can be bound as ES or VS. */
-               if (sctx->gs_shader) {
-                       r = si_update_scratch_buffer(sctx, sctx->tes_shader);
-                       if (r < 0)
-                               return false;
-                       if (r == 1)
-                               si_pm4_bind_state(sctx, es, sctx->tes_shader->current->pm4);
-               } else {
-                       r = si_update_scratch_buffer(sctx, sctx->tes_shader);
-                       if (r < 0)
-                               return false;
-                       if (r == 1)
-                               si_pm4_bind_state(sctx, vs, sctx->tes_shader->current->pm4);
+               r = si_update_scratch_buffer(sctx, sctx->tes_shader.current);
+               if (r < 0)
+                       return false;
+               if (r == 1) {
+                       if (sctx->gs_shader.current)
+                               si_pm4_bind_state(sctx, es, sctx->tes_shader.current->pm4);
+                       else
+                               si_pm4_bind_state(sctx, vs, sctx->tes_shader.current->pm4);
                }
        }
 
@@ -1459,7 +1424,7 @@ static void si_generate_fixed_func_tcs(struct si_context *sctx)
        if (!ureg)
                return; /* if we get here, we're screwed */
 
-       assert(!sctx->fixed_func_tcs_shader);
+       assert(!sctx->fixed_func_tcs_shader.cso);
 
        ureg_DECL_constant2D(ureg, 0, 1, SI_DRIVER_STATE_CONST_BUF);
        const0 = ureg_src_dimension(ureg_src_register(TGSI_FILE_CONSTANT, 0),
@@ -1474,7 +1439,7 @@ static void si_generate_fixed_func_tcs(struct si_context *sctx)
        ureg_MOV(ureg, tessinner, const1);
        ureg_END(ureg);
 
-       sctx->fixed_func_tcs_shader =
+       sctx->fixed_func_tcs_shader.cso =
                ureg_create_shader_and_destroy(ureg, &sctx->b.b);
 }
 
@@ -1482,7 +1447,7 @@ static void si_update_vgt_shader_config(struct si_context *sctx)
 {
        /* Calculate the index of the config.
         * 0 = VS, 1 = VS+GS, 2 = VS+Tess, 3 = VS+Tess+GS */
-       unsigned index = 2*!!sctx->tes_shader + !!sctx->gs_shader;
+       unsigned index = 2*!!sctx->tes_shader.cso + !!sctx->gs_shader.cso;
        struct si_pm4_state **pm4 = &sctx->vgt_shader_config[index];
 
        if (!*pm4) {
@@ -1490,17 +1455,17 @@ static void si_update_vgt_shader_config(struct si_context *sctx)
 
                *pm4 = CALLOC_STRUCT(si_pm4_state);
 
-               if (sctx->tes_shader) {
+               if (sctx->tes_shader.cso) {
                        stages |= S_028B54_LS_EN(V_028B54_LS_STAGE_ON) |
                                  S_028B54_HS_EN(1);
 
-                       if (sctx->gs_shader)
+                       if (sctx->gs_shader.cso)
                                stages |= S_028B54_ES_EN(V_028B54_ES_STAGE_DS) |
                                          S_028B54_GS_EN(1) |
                                          S_028B54_VS_EN(V_028B54_VS_STAGE_COPY_SHADER);
                        else
                                stages |= S_028B54_VS_EN(V_028B54_VS_STAGE_DS);
-               } else if (sctx->gs_shader) {
+               } else if (sctx->gs_shader.cso) {
                        stages |= S_028B54_ES_EN(V_028B54_ES_STAGE_REAL) |
                                  S_028B54_GS_EN(1) |
                                  S_028B54_VS_EN(V_028B54_VS_STAGE_COPY_SHADER);
@@ -1530,7 +1495,7 @@ bool si_update_shaders(struct si_context *sctx)
        int r;
 
        /* Update stages before GS. */
-       if (sctx->tes_shader) {
+       if (sctx->tes_shader.cso) {
                if (!sctx->tf_ring) {
                        si_init_tess_factor_ring(sctx);
                        if (!sctx->tf_ring)
@@ -1538,73 +1503,70 @@ bool si_update_shaders(struct si_context *sctx)
                }
 
                /* VS as LS */
-               r = si_shader_select(ctx, sctx->vs_shader);
+               r = si_shader_select(ctx, &sctx->vs_shader);
                if (r)
                        return false;
-               si_pm4_bind_state(sctx, ls, sctx->vs_shader->current->pm4);
+               si_pm4_bind_state(sctx, ls, sctx->vs_shader.current->pm4);
 
-               if (sctx->tcs_shader) {
-                       r = si_shader_select(ctx, sctx->tcs_shader);
+               if (sctx->tcs_shader.cso) {
+                       r = si_shader_select(ctx, &sctx->tcs_shader);
                        if (r)
                                return false;
-                       si_pm4_bind_state(sctx, hs, sctx->tcs_shader->current->pm4);
+                       si_pm4_bind_state(sctx, hs, sctx->tcs_shader.current->pm4);
                } else {
-                       if (!sctx->fixed_func_tcs_shader) {
+                       if (!sctx->fixed_func_tcs_shader.cso) {
                                si_generate_fixed_func_tcs(sctx);
-                               if (!sctx->fixed_func_tcs_shader)
+                               if (!sctx->fixed_func_tcs_shader.cso)
                                        return false;
                        }
 
-                       r = si_shader_select(ctx, sctx->fixed_func_tcs_shader);
+                       r = si_shader_select(ctx, &sctx->fixed_func_tcs_shader);
                        if (r)
                                return false;
                        si_pm4_bind_state(sctx, hs,
-                                         sctx->fixed_func_tcs_shader->current->pm4);
+                                         sctx->fixed_func_tcs_shader.current->pm4);
                }
 
-               r = si_shader_select(ctx, sctx->tes_shader);
+               r = si_shader_select(ctx, &sctx->tes_shader);
                if (r)
                        return false;
 
-               if (sctx->gs_shader) {
+               if (sctx->gs_shader.cso) {
                        /* TES as ES */
-                       si_pm4_bind_state(sctx, es, sctx->tes_shader->current->pm4);
+                       si_pm4_bind_state(sctx, es, sctx->tes_shader.current->pm4);
                } else {
                        /* TES as VS */
-                       si_pm4_bind_state(sctx, vs, sctx->tes_shader->current->pm4);
-                       si_update_so(sctx, sctx->tes_shader);
+                       si_pm4_bind_state(sctx, vs, sctx->tes_shader.current->pm4);
+                       si_update_so(sctx, sctx->tes_shader.cso);
                }
-       } else if (sctx->gs_shader) {
+       } else if (sctx->gs_shader.cso) {
                /* VS as ES */
-               r = si_shader_select(ctx, sctx->vs_shader);
+               r = si_shader_select(ctx, &sctx->vs_shader);
                if (r)
                        return false;
-               si_pm4_bind_state(sctx, es, sctx->vs_shader->current->pm4);
+               si_pm4_bind_state(sctx, es, sctx->vs_shader.current->pm4);
        } else {
                /* VS as VS */
-               r = si_shader_select(ctx, sctx->vs_shader);
+               r = si_shader_select(ctx, &sctx->vs_shader);
                if (r)
                        return false;
-               si_pm4_bind_state(sctx, vs, sctx->vs_shader->current->pm4);
-               si_update_so(sctx, sctx->vs_shader);
+               si_pm4_bind_state(sctx, vs, sctx->vs_shader.current->pm4);
+               si_update_so(sctx, sctx->vs_shader.cso);
        }
 
        /* Update GS. */
-       if (sctx->gs_shader) {
-               r = si_shader_select(ctx, sctx->gs_shader);
+       if (sctx->gs_shader.cso) {
+               r = si_shader_select(ctx, &sctx->gs_shader);
                if (r)
                        return false;
-               si_pm4_bind_state(sctx, gs, sctx->gs_shader->current->pm4);
-               si_pm4_bind_state(sctx, vs, sctx->gs_shader->current->gs_copy_shader->pm4);
-               si_update_so(sctx, sctx->gs_shader);
+               si_pm4_bind_state(sctx, gs, sctx->gs_shader.current->pm4);
+               si_pm4_bind_state(sctx, vs, sctx->gs_shader.current->gs_copy_shader->pm4);
+               si_update_so(sctx, sctx->gs_shader.cso);
 
-               if (!sctx->gsvs_ring) {
-                       si_init_gs_rings(sctx);
-                       if (!sctx->gsvs_ring)
-                               return false;
-               }
+               if (!si_update_gs_ring_buffers(sctx))
+                       return false;
 
-               si_update_gs_rings(sctx);
+               si_update_gsvs_ring_bindings(sctx);
        } else {
                si_pm4_bind_state(sctx, gs, NULL);
                si_pm4_bind_state(sctx, es, NULL);
@@ -1612,23 +1574,38 @@ bool si_update_shaders(struct si_context *sctx)
 
        si_update_vgt_shader_config(sctx);
 
-       r = si_shader_select(ctx, sctx->ps_shader);
-       if (r)
-               return false;
-       si_pm4_bind_state(sctx, ps, sctx->ps_shader->current->pm4);
-
-       if (si_pm4_state_changed(sctx, ps) || si_pm4_state_changed(sctx, vs) ||
-           sctx->sprite_coord_enable != rs->sprite_coord_enable ||
-           sctx->flatshade != rs->flatshade) {
-               sctx->sprite_coord_enable = rs->sprite_coord_enable;
-               sctx->flatshade = rs->flatshade;
-               si_mark_atom_dirty(sctx, &sctx->spi_map);
-       }
+       if (sctx->ps_shader.cso) {
+               r = si_shader_select(ctx, &sctx->ps_shader);
+               if (r)
+                       return false;
+               si_pm4_bind_state(sctx, ps, sctx->ps_shader.current->pm4);
+
+               if (si_pm4_state_changed(sctx, ps) || si_pm4_state_changed(sctx, vs) ||
+                   sctx->sprite_coord_enable != rs->sprite_coord_enable ||
+                   sctx->flatshade != rs->flatshade) {
+                       sctx->sprite_coord_enable = rs->sprite_coord_enable;
+                       sctx->flatshade = rs->flatshade;
+                       si_mark_atom_dirty(sctx, &sctx->spi_map);
+               }
+
+               if (si_pm4_state_changed(sctx, ps) ||
+                   sctx->force_persample_interp != rs->force_persample_interp) {
+                       sctx->force_persample_interp = rs->force_persample_interp;
+                       si_mark_atom_dirty(sctx, &sctx->spi_ps_input);
+               }
 
-       if (si_pm4_state_changed(sctx, ps) ||
-           sctx->force_persample_interp != rs->force_persample_interp) {
-               sctx->force_persample_interp = rs->force_persample_interp;
-               si_mark_atom_dirty(sctx, &sctx->spi_ps_input);
+               if (sctx->ps_db_shader_control != sctx->ps_shader.current->db_shader_control) {
+                       sctx->ps_db_shader_control = sctx->ps_shader.current->db_shader_control;
+                       si_mark_atom_dirty(sctx, &sctx->db_render_state);
+               }
+
+               if (sctx->smoothing_enabled != sctx->ps_shader.current->key.ps.poly_line_smoothing) {
+                       sctx->smoothing_enabled = sctx->ps_shader.current->key.ps.poly_line_smoothing;
+                       si_mark_atom_dirty(sctx, &sctx->msaa_config);
+
+                       if (sctx->b.chip_class == SI)
+                               si_mark_atom_dirty(sctx, &sctx->db_render_state);
+               }
        }
 
        if (si_pm4_state_changed(sctx, ls) ||
@@ -1640,19 +1617,6 @@ bool si_update_shaders(struct si_context *sctx)
                if (!si_update_spi_tmpring_size(sctx))
                        return false;
        }
-
-       if (sctx->ps_db_shader_control != sctx->ps_shader->current->db_shader_control) {
-               sctx->ps_db_shader_control = sctx->ps_shader->current->db_shader_control;
-               si_mark_atom_dirty(sctx, &sctx->db_render_state);
-       }
-
-       if (sctx->smoothing_enabled != sctx->ps_shader->current->key.ps.poly_line_smoothing) {
-               sctx->smoothing_enabled = sctx->ps_shader->current->key.ps.poly_line_smoothing;
-               si_mark_atom_dirty(sctx, &sctx->msaa_config);
-
-               if (sctx->b.chip_class == SI)
-                       si_mark_atom_dirty(sctx, &sctx->db_render_state);
-       }
        return true;
 }
 
@@ -1661,11 +1625,11 @@ void si_init_shader_functions(struct si_context *sctx)
        si_init_atom(sctx, &sctx->spi_map, &sctx->atoms.s.spi_map, si_emit_spi_map);
        si_init_atom(sctx, &sctx->spi_ps_input, &sctx->atoms.s.spi_ps_input, si_emit_spi_ps_input);
 
-       sctx->b.b.create_vs_state = si_create_vs_state;
-       sctx->b.b.create_tcs_state = si_create_tcs_state;
-       sctx->b.b.create_tes_state = si_create_tes_state;
-       sctx->b.b.create_gs_state = si_create_gs_state;
-       sctx->b.b.create_fs_state = si_create_fs_state;
+       sctx->b.b.create_vs_state = si_create_shader_selector;
+       sctx->b.b.create_tcs_state = si_create_shader_selector;
+       sctx->b.b.create_tes_state = si_create_shader_selector;
+       sctx->b.b.create_gs_state = si_create_shader_selector;
+       sctx->b.b.create_fs_state = si_create_shader_selector;
 
        sctx->b.b.bind_vs_state = si_bind_vs_shader;
        sctx->b.b.bind_tcs_state = si_bind_tcs_shader;
@@ -1673,9 +1637,9 @@ void si_init_shader_functions(struct si_context *sctx)
        sctx->b.b.bind_gs_state = si_bind_gs_shader;
        sctx->b.b.bind_fs_state = si_bind_ps_shader;
 
-       sctx->b.b.delete_vs_state = si_delete_vs_shader;
-       sctx->b.b.delete_tcs_state = si_delete_tcs_shader;
-       sctx->b.b.delete_tes_state = si_delete_tes_shader;
-       sctx->b.b.delete_gs_state = si_delete_gs_shader;
-       sctx->b.b.delete_fs_state = si_delete_ps_shader;
+       sctx->b.b.delete_vs_state = si_delete_shader_selector;
+       sctx->b.b.delete_tcs_state = si_delete_shader_selector;
+       sctx->b.b.delete_tes_state = si_delete_shader_selector;
+       sctx->b.b.delete_gs_state = si_delete_shader_selector;
+       sctx->b.b.delete_fs_state = si_delete_shader_selector;
 }