radeonsi/gfx9: set registers and shader key for merged ES-GS
authorMarek Olšák <marek.olsak@amd.com>
Wed, 19 Apr 2017 01:21:16 +0000 (03:21 +0200)
committerMarek Olšák <marek.olsak@amd.com>
Fri, 28 Apr 2017 19:47:35 +0000 (21:47 +0200)
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
src/gallium/drivers/radeonsi/si_shader.c
src/gallium/drivers/radeonsi/si_shader.h
src/gallium/drivers/radeonsi/si_state.c
src/gallium/drivers/radeonsi/si_state_draw.c
src/gallium/drivers/radeonsi/si_state_shaders.c

index 50efb13cbaf5d5cb0e6ed62a40ea0c91c6117d9f..44f3da034e4331a59c14075566443534001cafb6 100644 (file)
@@ -6914,6 +6914,14 @@ static void si_dump_shader_key(unsigned processor, struct si_shader *shader,
                break;
 
        case PIPE_SHADER_GEOMETRY:
+               if (shader->is_gs_copy_shader)
+                       break;
+
+               if (shader->selector->screen->b.chip_class >= GFX9 &&
+                   key->part.gs.es->type == PIPE_SHADER_VERTEX) {
+                       si_dump_shader_key_vs(key, &key->part.gs.vs_prolog,
+                                             "part.gs.vs_prolog", f);
+               }
                fprintf(f, "  part.gs.prolog.tri_strip_adj_fix = %u\n", key->part.gs.prolog.tri_strip_adj_fix);
                break;
 
index 8a779aedb0f56d2033bce9d62f5d1051cada49db..0f74a906e7672402ebda5644b030eb1d98932dd5 100644 (file)
@@ -419,6 +419,8 @@ struct si_shader_key {
                        struct si_vs_epilog_bits epilog; /* same as VS */
                } tes; /* tessellation evaluation shader */
                struct {
+                       struct si_vs_prolog_bits vs_prolog; /* for merged ES-GS */
+                       struct si_shader_selector *es;   /* for merged ES-GS */
                        struct si_gs_prolog_bits prolog;
                } gs;
                struct {
index ec99326dda40cedc1b33c899916ed260b92d16c1..39494cc1a17a882073302a64e7e4bfa43e4742b9 100644 (file)
@@ -4331,8 +4331,10 @@ static void si_init_config(struct si_context *sctx)
        si_pm4_set_reg(pm4, R_028A1C_VGT_HOS_MIN_TESS_LEVEL, fui(0));
 
        /* FIXME calculate these values somehow ??? */
-       si_pm4_set_reg(pm4, R_028A54_VGT_GS_PER_ES, SI_GS_PER_ES);
-       si_pm4_set_reg(pm4, R_028A58_VGT_ES_PER_GS, 0x40);
+       if (sctx->b.chip_class <= VI) {
+               si_pm4_set_reg(pm4, R_028A54_VGT_GS_PER_ES, SI_GS_PER_ES);
+               si_pm4_set_reg(pm4, R_028A58_VGT_ES_PER_GS, 0x40);
+       }
        si_pm4_set_reg(pm4, R_028A5C_VGT_GS_PER_VS, 0x2);
 
        si_pm4_set_reg(pm4, R_028A8C_VGT_PRIMITIVEID_RESET, 0x0);
@@ -4482,20 +4484,20 @@ static void si_init_config(struct si_context *sctx)
        }
 
        if (sctx->b.chip_class >= CIK) {
-               /* If this is 0, Bonaire can hang even if GS isn't being used.
-                * Other chips are unaffected. These are suboptimal values,
-                * but we don't use on-chip GS.
-                */
-               si_pm4_set_reg(pm4, R_028A44_VGT_GS_ONCHIP_CNTL,
-                              S_028A44_ES_VERTS_PER_SUBGRP(64) |
-                              S_028A44_GS_PRIMS_PER_SUBGRP(4));
-
                if (sctx->b.chip_class >= GFX9) {
                        si_pm4_set_reg(pm4, R_00B41C_SPI_SHADER_PGM_RSRC3_HS, S_00B41C_CU_EN(0xffff));
                } else {
                        si_pm4_set_reg(pm4, R_00B51C_SPI_SHADER_PGM_RSRC3_LS, S_00B51C_CU_EN(0xffff));
                        si_pm4_set_reg(pm4, R_00B41C_SPI_SHADER_PGM_RSRC3_HS, 0);
                        si_pm4_set_reg(pm4, R_00B31C_SPI_SHADER_PGM_RSRC3_ES, S_00B31C_CU_EN(0xffff));
+
+                       /* If this is 0, Bonaire can hang even if GS isn't being used.
+                        * Other chips are unaffected. These are suboptimal values,
+                        * but we don't use on-chip GS.
+                        */
+                       si_pm4_set_reg(pm4, R_028A44_VGT_GS_ONCHIP_CNTL,
+                                      S_028A44_ES_VERTS_PER_SUBGRP(64) |
+                                      S_028A44_GS_PRIMS_PER_SUBGRP(4));
                }
                si_pm4_set_reg(pm4, R_00B21C_SPI_SHADER_PGM_RSRC3_GS, S_00B21C_CU_EN(0xffff));
 
index de97c0e8f0607789477a5da82bf609cc4bb93c23..393f64f372fed06af5a310bbbbb0160ea1496172 100644 (file)
@@ -479,7 +479,8 @@ static unsigned si_get_ia_multi_vgt_param(struct si_context *sctx,
 
        if (sctx->gs_shader.cso) {
                /* GS requirement. */
-               if (SI_GS_PER_ES / primgroup_size >= sctx->screen->gs_table_depth - 3)
+               if (sctx->b.chip_class <= VI &&
+                   SI_GS_PER_ES / primgroup_size >= sctx->screen->gs_table_depth - 3)
                        ia_multi_vgt_param |= S_028AA8_PARTIAL_ES_WAVE_ON(1);
 
                /* GS hw bug with single-primitive instances and SWITCH_ON_EOI.
index cb65c0a8cb231290fb8113a83f58752af48856a6..6cacbc7a33aace00c1a5056a03bc63cb7c6c30a8 100644 (file)
@@ -573,6 +573,7 @@ static void si_shader_es(struct si_screen *sscreen, struct si_shader *shader)
  */
 static uint32_t si_vgt_gs_mode(struct si_shader_selector *sel)
 {
+       enum chip_class chip_class = sel->screen->b.chip_class;
        unsigned gs_max_vert_out = sel->gs_max_out_vertices;
        unsigned cut_mode;
 
@@ -589,11 +590,118 @@ static uint32_t si_vgt_gs_mode(struct si_shader_selector *sel)
 
        return S_028A40_MODE(V_028A40_GS_SCENARIO_G) |
               S_028A40_CUT_MODE(cut_mode)|
-              S_028A40_ES_WRITE_OPTIMIZE(1) |
-              S_028A40_GS_WRITE_OPTIMIZE(1);
+              S_028A40_ES_WRITE_OPTIMIZE(chip_class <= VI) |
+              S_028A40_GS_WRITE_OPTIMIZE(1) |
+              S_028A40_ONCHIP(chip_class >= GFX9 ? 1 : 0);
 }
 
-static void si_shader_gs(struct si_shader *shader)
+struct gfx9_gs_info {
+       unsigned es_verts_per_subgroup;
+       unsigned gs_prims_per_subgroup;
+       unsigned gs_inst_prims_in_subgroup;
+       unsigned max_prims_per_subgroup;
+       unsigned lds_size;
+};
+
+static void gfx9_get_gs_info(struct si_shader_selector *es,
+                                  struct si_shader_selector *gs,
+                                  struct gfx9_gs_info *out)
+{
+       unsigned gs_num_invocations = MAX2(gs->gs_num_invocations, 1);
+       unsigned input_prim = gs->info.properties[TGSI_PROPERTY_GS_INPUT_PRIM];
+       bool uses_adjacency = input_prim >= PIPE_PRIM_LINES_ADJACENCY &&
+                             input_prim <= PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY;
+
+       /* All these are in dwords: */
+       /* We can't allow using the whole LDS, because GS waves compete with
+        * other shader stages for LDS space. */
+       const unsigned max_lds_size = 8 * 1024;
+       const unsigned esgs_itemsize = es->esgs_itemsize / 4;
+       unsigned esgs_lds_size;
+
+       /* All these are per subgroup: */
+       const unsigned max_out_prims = 32 * 1024;
+       const unsigned max_es_verts = 255;
+       const unsigned ideal_gs_prims = 64;
+       unsigned max_gs_prims, gs_prims;
+       unsigned min_es_verts, es_verts, worst_case_es_verts;
+
+       assert(gs_num_invocations <= 32); /* GL maximum */
+
+       if (uses_adjacency || gs_num_invocations > 1)
+               max_gs_prims = 127 / gs_num_invocations;
+       else
+               max_gs_prims = 255;
+
+       /* MAX_PRIMS_PER_SUBGROUP = gs_prims * max_vert_out * gs_invocations.
+        * Make sure we don't go over the maximum value.
+        */
+       max_gs_prims = MIN2(max_gs_prims,
+                           max_out_prims /
+                           (gs->gs_max_out_vertices * gs_num_invocations));
+       assert(max_gs_prims > 0);
+
+       /* If the primitive has adjacency, halve the number of vertices
+        * that will be reused in multiple primitives.
+        */
+       min_es_verts = gs->gs_input_verts_per_prim / (uses_adjacency ? 2 : 1);
+
+       gs_prims = MIN2(ideal_gs_prims, max_gs_prims);
+       worst_case_es_verts = MIN2(min_es_verts * gs_prims, max_es_verts);
+
+       /* Compute ESGS LDS size based on the worst case number of ES vertices
+        * needed to create the target number of GS prims per subgroup.
+        */
+       esgs_lds_size = esgs_itemsize * worst_case_es_verts;
+
+       /* If total LDS usage is too big, refactor partitions based on ratio
+        * of ESGS item sizes.
+        */
+       if (esgs_lds_size > max_lds_size) {
+               /* Our target GS Prims Per Subgroup was too large. Calculate
+                * the maximum number of GS Prims Per Subgroup that will fit
+                * into LDS, capped by the maximum that the hardware can support.
+                */
+               gs_prims = MIN2((max_lds_size / (esgs_itemsize * min_es_verts)),
+                               max_gs_prims);
+               assert(gs_prims > 0);
+               worst_case_es_verts = MIN2(min_es_verts * gs_prims,
+                                          max_es_verts);
+
+               esgs_lds_size = esgs_itemsize * worst_case_es_verts;
+               assert(esgs_lds_size <= max_lds_size);
+       }
+
+       /* Now calculate remaining ESGS information. */
+       if (esgs_lds_size)
+               es_verts = MIN2(esgs_lds_size / esgs_itemsize, max_es_verts);
+       else
+               es_verts = max_es_verts;
+
+       /* Vertices for adjacency primitives are not always reused, so restore
+        * it for ES_VERTS_PER_SUBGRP.
+        */
+       min_es_verts = gs->gs_input_verts_per_prim;
+
+       /* For normal primitives, the VGT only checks if they are past the ES
+        * verts per subgroup after allocating a full GS primitive and if they
+        * are, kick off a new subgroup.  But if those additional ES verts are
+        * unique (e.g. not reused) we need to make sure there is enough LDS
+        * space to account for those ES verts beyond ES_VERTS_PER_SUBGRP.
+        */
+       es_verts -= min_es_verts - 1;
+
+       out->es_verts_per_subgroup = es_verts;
+       out->gs_prims_per_subgroup = gs_prims;
+       out->gs_inst_prims_in_subgroup = gs_prims * gs_num_invocations;
+       out->max_prims_per_subgroup = out->gs_inst_prims_in_subgroup *
+                                     gs->gs_max_out_vertices;
+       out->lds_size = align(esgs_lds_size, 128) / 128;
+
+       assert(out->max_prims_per_subgroup <= max_out_prims);
+}
+
+static void si_shader_gs(struct si_screen *sscreen, struct si_shader *shader)
 {
        struct si_shader_selector *sel = shader->selector;
        const ubyte *num_components = sel->info.num_stream_output_components;
@@ -622,7 +730,7 @@ static void si_shader_gs(struct si_shader *shader)
        /* The GSVS_RING_ITEMSIZE register takes 15 bits */
        assert(offset < (1 << 15));
 
-       si_pm4_set_reg(pm4, R_028B38_VGT_GS_MAX_VERT_OUT, shader->selector->gs_max_out_vertices);
+       si_pm4_set_reg(pm4, R_028B38_VGT_GS_MAX_VERT_OUT, sel->gs_max_out_vertices);
 
        si_pm4_set_reg(pm4, R_028B5C_VGT_GS_VERT_ITEMSIZE, num_components[0]);
        si_pm4_set_reg(pm4, R_028B60_VGT_GS_VERT_ITEMSIZE_1, (max_stream >= 1) ? num_components[1] : 0);
@@ -635,17 +743,72 @@ static void si_shader_gs(struct si_shader *shader)
 
        va = shader->bo->gpu_address;
        si_pm4_add_bo(pm4, shader->bo, RADEON_USAGE_READ, RADEON_PRIO_SHADER_BINARY);
-       si_pm4_set_reg(pm4, R_00B220_SPI_SHADER_PGM_LO_GS, va >> 8);
-       si_pm4_set_reg(pm4, R_00B224_SPI_SHADER_PGM_HI_GS, va >> 40);
-
-       si_pm4_set_reg(pm4, R_00B228_SPI_SHADER_PGM_RSRC1_GS,
-                      S_00B228_VGPRS((shader->config.num_vgprs - 1) / 4) |
-                      S_00B228_SGPRS((shader->config.num_sgprs - 1) / 8) |
-                      S_00B228_DX10_CLAMP(1) |
-                      S_00B228_FLOAT_MODE(shader->config.float_mode));
-       si_pm4_set_reg(pm4, R_00B22C_SPI_SHADER_PGM_RSRC2_GS,
-                      S_00B22C_USER_SGPR(GFX6_GS_NUM_USER_SGPR) |
-                      S_00B22C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0));
+
+       if (sscreen->b.chip_class >= GFX9) {
+               unsigned input_prim = sel->info.properties[TGSI_PROPERTY_GS_INPUT_PRIM];
+               unsigned es_type = shader->key.part.gs.es->type;
+               unsigned es_vgpr_comp_cnt, gs_vgpr_comp_cnt;
+               struct gfx9_gs_info gs_info;
+
+               if (es_type == PIPE_SHADER_VERTEX)
+                       es_vgpr_comp_cnt = shader->info.uses_instanceid ? 3 : 0;
+               else if (es_type == PIPE_SHADER_TESS_EVAL)
+                       es_vgpr_comp_cnt = 3; /* all components are needed for TES */
+               else
+                       unreachable("invalid shader selector type");
+
+               /* If offsets 4, 5 are used, GS_VGPR_COMP_CNT is ignored and
+                * VGPR[0:4] are always loaded.
+                */
+               if (sel->info.uses_invocationid)
+                       gs_vgpr_comp_cnt = 3; /* VGPR3 contains InvocationID. */
+               else if (sel->info.uses_primid)
+                       gs_vgpr_comp_cnt = 2; /* VGPR2 contains PrimitiveID. */
+               else if (input_prim >= PIPE_PRIM_TRIANGLES)
+                       gs_vgpr_comp_cnt = 1; /* VGPR1 contains offsets 2, 3 */
+               else
+                       gs_vgpr_comp_cnt = 0; /* VGPR0 contains offsets 0, 1 */
+
+               gfx9_get_gs_info(shader->key.part.gs.es, sel, &gs_info);
+
+               si_pm4_set_reg(pm4, R_00B210_SPI_SHADER_PGM_LO_ES, va >> 8);
+               si_pm4_set_reg(pm4, R_00B214_SPI_SHADER_PGM_HI_ES, va >> 40);
+
+               si_pm4_set_reg(pm4, R_00B228_SPI_SHADER_PGM_RSRC1_GS,
+                              S_00B228_VGPRS((shader->config.num_vgprs - 1) / 4) |
+                              S_00B228_SGPRS((shader->config.num_sgprs - 1) / 8) |
+                              S_00B228_DX10_CLAMP(1) |
+                              S_00B228_FLOAT_MODE(shader->config.float_mode) |
+                              S_00B228_GS_VGPR_COMP_CNT(gs_vgpr_comp_cnt));
+               si_pm4_set_reg(pm4, R_00B22C_SPI_SHADER_PGM_RSRC2_GS,
+                              S_00B22C_USER_SGPR(GFX9_GS_NUM_USER_SGPR) |
+                              S_00B22C_USER_SGPR_MSB(GFX9_GS_NUM_USER_SGPR >> 5) |
+                              S_00B22C_ES_VGPR_COMP_CNT(es_vgpr_comp_cnt) |
+                              S_00B22C_OC_LDS_EN(es_type == PIPE_SHADER_TESS_EVAL) |
+                              S_00B22C_LDS_SIZE(gs_info.lds_size) |
+                              S_00B22C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0));
+
+               si_pm4_set_reg(pm4, R_028A44_VGT_GS_ONCHIP_CNTL,
+                              S_028A44_ES_VERTS_PER_SUBGRP(gs_info.es_verts_per_subgroup) |
+                              S_028A44_GS_PRIMS_PER_SUBGRP(gs_info.gs_prims_per_subgroup) |
+                              S_028A44_GS_INST_PRIMS_IN_SUBGRP(gs_info.gs_inst_prims_in_subgroup));
+               si_pm4_set_reg(pm4, R_028A94_VGT_GS_MAX_PRIMS_PER_SUBGROUP,
+                              S_028A94_MAX_PRIMS_PER_SUBGROUP(gs_info.max_prims_per_subgroup));
+               si_pm4_set_reg(pm4, R_028AAC_VGT_ESGS_RING_ITEMSIZE,
+                              shader->key.part.gs.es->esgs_itemsize / 4);
+       } else {
+               si_pm4_set_reg(pm4, R_00B220_SPI_SHADER_PGM_LO_GS, va >> 8);
+               si_pm4_set_reg(pm4, R_00B224_SPI_SHADER_PGM_HI_GS, va >> 40);
+
+               si_pm4_set_reg(pm4, R_00B228_SPI_SHADER_PGM_RSRC1_GS,
+                              S_00B228_VGPRS((shader->config.num_vgprs - 1) / 4) |
+                              S_00B228_SGPRS((shader->config.num_sgprs - 1) / 8) |
+                              S_00B228_DX10_CLAMP(1) |
+                              S_00B228_FLOAT_MODE(shader->config.float_mode));
+               si_pm4_set_reg(pm4, R_00B22C_SPI_SHADER_PGM_RSRC2_GS,
+                              S_00B22C_USER_SGPR(GFX6_GS_NUM_USER_SGPR) |
+                              S_00B22C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0));
+       }
 }
 
 /**
@@ -969,7 +1132,7 @@ static void si_shader_init_pm4_state(struct si_screen *sscreen,
                        si_shader_vs(sscreen, shader, NULL);
                break;
        case PIPE_SHADER_GEOMETRY:
-               si_shader_gs(shader);
+               si_shader_gs(sscreen, shader);
                break;
        case PIPE_SHADER_FRAGMENT:
                si_shader_ps(shader);
@@ -1108,6 +1271,15 @@ static inline void si_shader_selector_key(struct pipe_context *ctx,
                }
                break;
        case PIPE_SHADER_GEOMETRY:
+               if (sctx->b.chip_class >= GFX9) {
+                       if (sctx->tes_shader.cso) {
+                               key->part.gs.es = sctx->tes_shader.cso;
+                       } else {
+                               si_shader_selector_key_vs(sctx, sctx->vs_shader.cso,
+                                                         key, &key->part.gs.vs_prolog);
+                               key->part.gs.es = sctx->vs_shader.cso;
+                       }
+               }
                key->part.gs.prolog.tri_strip_adj_fix = sctx->gs_tri_strip_adj_fix;
                break;
        case PIPE_SHADER_FRAGMENT: {
@@ -1729,6 +1901,12 @@ static void *si_create_shader_selector(struct pipe_context *ctx,
                        }
                }
                sel->esgs_itemsize = util_last_bit64(sel->outputs_written) * 16;
+
+               /* For the ESGS ring in LDS, add 1 dword to reduce LDS bank
+                * conflicts, i.e. each vertex will start at a different bank.
+                */
+               if (sctx->b.chip_class >= GFX9)
+                       sel->esgs_itemsize += 4;
                break;
 
        case PIPE_SHADER_FRAGMENT: