radeonsi:optimizing SET_CONTEXT_REG for shaders VS
authorSonny Jiang <sonny.jiang@amd.com>
Wed, 3 Oct 2018 15:53:11 +0000 (11:53 -0400)
committerMarek Olšák <marek.olsak@amd.com>
Fri, 5 Oct 2018 23:04:13 +0000 (19:04 -0400)
Signed-off-by: Sonny Jiang <sonny.jiang@amd.com>
Signed-off-by: Marek Olšák <marek.olsak@amd.com>
src/gallium/drivers/radeonsi/si_gfx_cs.c
src/gallium/drivers/radeonsi/si_state.h
src/gallium/drivers/radeonsi/si_state_shaders.c

index 7c79bad679f064de9173ec8fc5f9e7fb79f3b1b1..1b931d71369a259043f2d332cbc34c8d7c237e98 100644 (file)
@@ -363,6 +363,12 @@ void si_begin_new_gfx_cs(struct si_context *ctx)
                ctx->tracked_regs.reg_value[SI_TRACKED_VGT_GS_INSTANCE_CNT]  = 0x00000000;
                ctx->tracked_regs.reg_value[SI_TRACKED_VGT_GS_ONCHIP_CNTL]  = 0x00000000;
                ctx->tracked_regs.reg_value[SI_TRACKED_VGT_GS_MAX_PRIMS_PER_SUBGROUP]  = 0x00000000;
+               ctx->tracked_regs.reg_value[SI_TRACKED_VGT_GS_MODE]  = 0x00000000;
+               ctx->tracked_regs.reg_value[SI_TRACKED_VGT_PRIMITIVEID_EN]  = 0x00000000;
+               ctx->tracked_regs.reg_value[SI_TRACKED_VGT_REUSE_OFF]  = 0x00000000;
+               ctx->tracked_regs.reg_value[SI_TRACKED_SPI_VS_OUT_CONFIG]  = 0x00000000;
+               ctx->tracked_regs.reg_value[SI_TRACKED_SPI_SHADER_POS_FORMAT]  = 0x00000000;
+               ctx->tracked_regs.reg_value[SI_TRACKED_PA_CL_VTE_CNTL]  = 0x00000000;
 
                /* Set all saved registers state to saved. */
                ctx->tracked_regs.reg_saved = 0xffffffffffffffff;
index 9efb4be5a584825c620a97f57e619b54f81c00b9..bf1ae9f18f86f22f57b885688947b60d0053db96 100644 (file)
@@ -295,6 +295,12 @@ enum si_tracked_reg {
        SI_TRACKED_VGT_GS_INSTANCE_CNT,
        SI_TRACKED_VGT_GS_ONCHIP_CNTL,
        SI_TRACKED_VGT_GS_MAX_PRIMS_PER_SUBGROUP,
+       SI_TRACKED_VGT_GS_MODE,
+       SI_TRACKED_VGT_PRIMITIVEID_EN,
+       SI_TRACKED_VGT_REUSE_OFF,
+       SI_TRACKED_SPI_VS_OUT_CONFIG,
+       SI_TRACKED_SPI_SHADER_POS_FORMAT,
+       SI_TRACKED_PA_CL_VTE_CNTL,
 
        SI_NUM_TRACKED_REGS,
 };
index 498deff502041412d7e142cd6e4c519eb7281c7a..25c01382a0851200f8fd211956e9a93d3825978c 100644 (file)
@@ -935,6 +935,38 @@ static void si_shader_gs(struct si_screen *sscreen, struct si_shader *shader)
        }
 }
 
+static void si_emit_shader_vs(struct si_context *sctx)
+{
+       struct si_shader *shader = sctx->queued.named.vs->shader;
+       if (!shader)
+               return;
+
+       radeon_opt_set_context_reg(sctx, R_028A40_VGT_GS_MODE,
+                                  SI_TRACKED_VGT_GS_MODE,
+                                  shader->ctx_reg.vs.vgt_gs_mode);
+       radeon_opt_set_context_reg(sctx, R_028A84_VGT_PRIMITIVEID_EN,
+                                  SI_TRACKED_VGT_PRIMITIVEID_EN,
+                                  shader->ctx_reg.vs.vgt_primitiveid_en);
+
+       if (sctx->chip_class <= VI) {
+               radeon_opt_set_context_reg(sctx, R_028AB4_VGT_REUSE_OFF,
+                                          SI_TRACKED_VGT_REUSE_OFF,
+                                          shader->ctx_reg.vs.vgt_reuse_off);
+       }
+
+       radeon_opt_set_context_reg(sctx, R_0286C4_SPI_VS_OUT_CONFIG,
+                                  SI_TRACKED_SPI_VS_OUT_CONFIG,
+                                  shader->ctx_reg.vs.spi_vs_out_config);
+
+       radeon_opt_set_context_reg(sctx, R_02870C_SPI_SHADER_POS_FORMAT,
+                                  SI_TRACKED_SPI_SHADER_POS_FORMAT,
+                                  shader->ctx_reg.vs.spi_shader_pos_format);
+
+       radeon_opt_set_context_reg(sctx, R_028818_PA_CL_VTE_CNTL,
+                                  SI_TRACKED_PA_CL_VTE_CNTL,
+                                  shader->ctx_reg.vs.pa_cl_vte_cntl);
+}
+
 /**
  * Compute the state for \p shader, which will run as a vertex shader on the
  * hardware.
@@ -947,18 +979,19 @@ static void si_shader_vs(struct si_screen *sscreen, struct si_shader *shader,
 {
        const struct tgsi_shader_info *info = &shader->selector->info;
        struct si_pm4_state *pm4;
-       unsigned num_user_sgprs;
-       unsigned nparams, vgpr_comp_cnt;
+       unsigned num_user_sgprs, vgpr_comp_cnt;
        uint64_t va;
-       unsigned oc_lds_en;
+       unsigned nparams, oc_lds_en;
        unsigned window_space =
-          info->properties[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION];
+               info->properties[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION];
        bool enable_prim_id = shader->key.mono.u.vs_export_prim_id || info->uses_primid;
 
        pm4 = si_get_shader_pm4_state(shader);
        if (!pm4)
                return;
 
+       pm4->atom.emit = si_emit_shader_vs;
+
        /* We always write VGT_GS_MODE in the VS state, because every switch
         * between different shader pipelines involving a different GS or no
         * GS at all involves a switch of the VS (different GS use different
@@ -973,19 +1006,18 @@ static void si_shader_vs(struct si_screen *sscreen, struct si_shader *shader,
                if (enable_prim_id)
                        mode = V_028A40_GS_SCENARIO_A;
 
-               si_pm4_set_reg(pm4, R_028A40_VGT_GS_MODE, S_028A40_MODE(mode));
-               si_pm4_set_reg(pm4, R_028A84_VGT_PRIMITIVEID_EN, enable_prim_id);
+               shader->ctx_reg.vs.vgt_gs_mode = S_028A40_MODE(mode);
+               shader->ctx_reg.vs.vgt_primitiveid_en = enable_prim_id;
        } else {
-               si_pm4_set_reg(pm4, R_028A40_VGT_GS_MODE,
-                              ac_vgt_gs_mode(gs->gs_max_out_vertices,
-                                             sscreen->info.chip_class));
-               si_pm4_set_reg(pm4, R_028A84_VGT_PRIMITIVEID_EN, 0);
+               shader->ctx_reg.vs.vgt_gs_mode = ac_vgt_gs_mode(gs->gs_max_out_vertices,
+                                                               sscreen->info.chip_class);
+               shader->ctx_reg.vs.vgt_primitiveid_en = 0;
        }
 
        if (sscreen->info.chip_class <= VI) {
                /* Reuse needs to be set off if we write oViewport. */
-               si_pm4_set_reg(pm4, R_028AB4_VGT_REUSE_OFF,
-                              S_028AB4_REUSE_OFF(info->writes_viewport_index));
+               shader->ctx_reg.vs.vgt_reuse_off =
+                               S_028AB4_REUSE_OFF(info->writes_viewport_index);
        }
 
        va = shader->bo->gpu_address;
@@ -1015,20 +1047,19 @@ static void si_shader_vs(struct si_screen *sscreen, struct si_shader *shader,
 
        /* VS is required to export at least one param. */
        nparams = MAX2(shader->info.nr_param_exports, 1);
-       si_pm4_set_reg(pm4, R_0286C4_SPI_VS_OUT_CONFIG,
-                      S_0286C4_VS_EXPORT_COUNT(nparams - 1));
-
-       si_pm4_set_reg(pm4, R_02870C_SPI_SHADER_POS_FORMAT,
-                      S_02870C_POS0_EXPORT_FORMAT(V_02870C_SPI_SHADER_4COMP) |
-                      S_02870C_POS1_EXPORT_FORMAT(shader->info.nr_pos_exports > 1 ?
-                                                  V_02870C_SPI_SHADER_4COMP :
-                                                  V_02870C_SPI_SHADER_NONE) |
-                      S_02870C_POS2_EXPORT_FORMAT(shader->info.nr_pos_exports > 2 ?
-                                                  V_02870C_SPI_SHADER_4COMP :
-                                                  V_02870C_SPI_SHADER_NONE) |
-                      S_02870C_POS3_EXPORT_FORMAT(shader->info.nr_pos_exports > 3 ?
-                                                  V_02870C_SPI_SHADER_4COMP :
-                                                  V_02870C_SPI_SHADER_NONE));
+       shader->ctx_reg.vs.spi_vs_out_config = S_0286C4_VS_EXPORT_COUNT(nparams - 1);
+
+       shader->ctx_reg.vs.spi_shader_pos_format =
+                       S_02870C_POS0_EXPORT_FORMAT(V_02870C_SPI_SHADER_4COMP) |
+                       S_02870C_POS1_EXPORT_FORMAT(shader->info.nr_pos_exports > 1 ?
+                                                   V_02870C_SPI_SHADER_4COMP :
+                                                   V_02870C_SPI_SHADER_NONE) |
+                       S_02870C_POS2_EXPORT_FORMAT(shader->info.nr_pos_exports > 2 ?
+                                                   V_02870C_SPI_SHADER_4COMP :
+                                                   V_02870C_SPI_SHADER_NONE) |
+                       S_02870C_POS3_EXPORT_FORMAT(shader->info.nr_pos_exports > 3 ?
+                                                   V_02870C_SPI_SHADER_4COMP :
+                                                   V_02870C_SPI_SHADER_NONE);
 
        oc_lds_en = shader->selector->type == PIPE_SHADER_TESS_EVAL ? 1 : 0;
 
@@ -1049,15 +1080,16 @@ static void si_shader_vs(struct si_screen *sscreen, struct si_shader *shader,
                       S_00B12C_SO_BASE3_EN(!!shader->selector->so.stride[3]) |
                       S_00B12C_SO_EN(!!shader->selector->so.num_outputs) |
                       S_00B12C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0));
+
        if (window_space)
-               si_pm4_set_reg(pm4, R_028818_PA_CL_VTE_CNTL,
-                              S_028818_VTX_XY_FMT(1) | S_028818_VTX_Z_FMT(1));
+               shader->ctx_reg.vs.pa_cl_vte_cntl =
+                               S_028818_VTX_XY_FMT(1) | S_028818_VTX_Z_FMT(1);
        else
-               si_pm4_set_reg(pm4, R_028818_PA_CL_VTE_CNTL,
-                              S_028818_VTX_W0_FMT(1) |
-                              S_028818_VPORT_X_SCALE_ENA(1) | S_028818_VPORT_X_OFFSET_ENA(1) |
-                              S_028818_VPORT_Y_SCALE_ENA(1) | S_028818_VPORT_Y_OFFSET_ENA(1) |
-                              S_028818_VPORT_Z_SCALE_ENA(1) | S_028818_VPORT_Z_OFFSET_ENA(1));
+               shader->ctx_reg.vs.pa_cl_vte_cntl =
+                               S_028818_VTX_W0_FMT(1) |
+                               S_028818_VPORT_X_SCALE_ENA(1) | S_028818_VPORT_X_OFFSET_ENA(1) |
+                               S_028818_VPORT_Y_SCALE_ENA(1) | S_028818_VPORT_Y_OFFSET_ENA(1) |
+                               S_028818_VPORT_Z_SCALE_ENA(1) | S_028818_VPORT_Z_OFFSET_ENA(1);
 
        if (shader->selector->type == PIPE_SHADER_TESS_EVAL)
                si_set_tesseval_regs(sscreen, shader->selector, pm4);