radeonsi: move 2nd-shader descriptor pointers into s[0:1]
authorMarek Olšák <marek.olsak@amd.com>
Fri, 2 Feb 2018 20:35:20 +0000 (21:35 +0100)
committerMarek Olšák <marek.olsak@amd.com>
Sat, 24 Feb 2018 22:08:29 +0000 (23:08 +0100)
If 32-bit pointers are supported, both pointers can be moved into s[0:1]
and then ESGS has exactly the same user data SGPR declarations as VS.

If 32-bit pointers are not supported, only one pointer can be moved into
s[0:1]. In that case, the 2nd pointer is moved before TCS constants,
so that the location is the same in HS and GS.

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
src/gallium/drivers/radeonsi/si_descriptors.c
src/gallium/drivers/radeonsi/si_shader.c
src/gallium/drivers/radeonsi/si_shader.h

index 65b0d3ac194ed9ffdc2a6367c590df49fc03b5fa..fac71bd602cc2f9bdbc5c5a3889a98a018e508e1 100644 (file)
@@ -2053,6 +2053,24 @@ static void si_emit_consecutive_shader_pointers(struct si_context *sctx,
        }
 }
 
+static void si_emit_disjoint_shader_pointers(struct si_context *sctx,
+                                            unsigned pointer_mask,
+                                            unsigned sh_base)
+{
+       if (!sh_base)
+               return;
+
+       struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
+       unsigned mask = sctx->shader_pointers_dirty & pointer_mask;
+
+       while (mask) {
+               struct si_descriptors *descs = &sctx->descriptors[u_bit_scan(&mask)];
+
+               si_emit_shader_pointer_head(cs, descs, sh_base, 1);
+               si_emit_shader_pointer_body(sctx->screen, cs, descs);
+       }
+}
+
 static void si_emit_global_shader_pointers(struct si_context *sctx,
                                           struct si_descriptors *descs)
 {
@@ -2089,14 +2107,21 @@ void si_emit_graphics_shader_pointers(struct si_context *sctx,
 
        si_emit_consecutive_shader_pointers(sctx, SI_DESCS_SHADER_MASK(VERTEX),
                                            sh_base[PIPE_SHADER_VERTEX]);
-       si_emit_consecutive_shader_pointers(sctx, SI_DESCS_SHADER_MASK(TESS_CTRL),
-                                           sh_base[PIPE_SHADER_TESS_CTRL]);
        si_emit_consecutive_shader_pointers(sctx, SI_DESCS_SHADER_MASK(TESS_EVAL),
                                            sh_base[PIPE_SHADER_TESS_EVAL]);
-       si_emit_consecutive_shader_pointers(sctx, SI_DESCS_SHADER_MASK(GEOMETRY),
-                                           sh_base[PIPE_SHADER_GEOMETRY]);
        si_emit_consecutive_shader_pointers(sctx, SI_DESCS_SHADER_MASK(FRAGMENT),
                                            sh_base[PIPE_SHADER_FRAGMENT]);
+       if (HAVE_32BIT_POINTERS || sctx->b.chip_class <= VI) {
+               si_emit_consecutive_shader_pointers(sctx, SI_DESCS_SHADER_MASK(TESS_CTRL),
+                                                   sh_base[PIPE_SHADER_TESS_CTRL]);
+               si_emit_consecutive_shader_pointers(sctx, SI_DESCS_SHADER_MASK(GEOMETRY),
+                                                   sh_base[PIPE_SHADER_GEOMETRY]);
+       } else {
+               si_emit_disjoint_shader_pointers(sctx, SI_DESCS_SHADER_MASK(TESS_CTRL),
+                                                sh_base[PIPE_SHADER_TESS_CTRL]);
+               si_emit_disjoint_shader_pointers(sctx, SI_DESCS_SHADER_MASK(GEOMETRY),
+                                                sh_base[PIPE_SHADER_GEOMETRY]);
+       }
 
        sctx->shader_pointers_dirty &=
                ~u_bit_consecutive(SI_DESCS_RW_BUFFERS, SI_DESCS_FIRST_COMPUTE);
@@ -2572,40 +2597,56 @@ void si_init_all_descriptors(struct si_context *sctx)
        int i;
 
 #if !HAVE_32BIT_POINTERS
-       STATIC_ASSERT(GFX9_SGPR_TCS_CONST_AND_SHADER_BUFFERS % 2 == 0);
-       STATIC_ASSERT(GFX9_SGPR_GS_CONST_AND_SHADER_BUFFERS % 2 == 0);
+       STATIC_ASSERT(GFX9_SGPR_2ND_SAMPLERS_AND_IMAGES % 2 == 0);
 #endif
 
        for (i = 0; i < SI_NUM_SHADERS; i++) {
-               bool gfx9_tcs = false;
-               bool gfx9_gs = false;
+               bool is_2nd = sctx->b.chip_class >= GFX9 &&
+                                    (i == PIPE_SHADER_TESS_CTRL ||
+                                     i == PIPE_SHADER_GEOMETRY);
                unsigned num_sampler_slots = SI_NUM_IMAGES / 2 + SI_NUM_SAMPLERS;
                unsigned num_buffer_slots = SI_NUM_SHADER_BUFFERS + SI_NUM_CONST_BUFFERS;
+               int rel_dw_offset;
                struct si_descriptors *desc;
 
-               if (sctx->b.chip_class >= GFX9) {
-                       gfx9_tcs = i == PIPE_SHADER_TESS_CTRL;
-                       gfx9_gs = i == PIPE_SHADER_GEOMETRY;
+               if (is_2nd) {
+                       if (i == PIPE_SHADER_TESS_CTRL) {
+                               rel_dw_offset = (R_00B408_SPI_SHADER_USER_DATA_ADDR_LO_HS -
+                                                R_00B430_SPI_SHADER_USER_DATA_LS_0) / 4;
+                       } else { /* PIPE_SHADER_GEOMETRY */
+                               rel_dw_offset = (R_00B208_SPI_SHADER_USER_DATA_ADDR_LO_GS -
+                                                R_00B330_SPI_SHADER_USER_DATA_ES_0) / 4;
+                       }
+               } else {
+                       rel_dw_offset = SI_SGPR_CONST_AND_SHADER_BUFFERS;
                }
-
                desc = si_const_and_shader_buffer_descriptors(sctx, i);
                si_init_buffer_resources(&sctx->const_and_shader_buffers[i], desc,
-                                        num_buffer_slots,
-                                        gfx9_tcs ? GFX9_SGPR_TCS_CONST_AND_SHADER_BUFFERS :
-                                        gfx9_gs ? GFX9_SGPR_GS_CONST_AND_SHADER_BUFFERS :
-                                                  SI_SGPR_CONST_AND_SHADER_BUFFERS,
+                                        num_buffer_slots, rel_dw_offset,
                                         RADEON_USAGE_READWRITE,
                                         RADEON_USAGE_READ,
                                         RADEON_PRIO_SHADER_RW_BUFFER,
                                         RADEON_PRIO_CONST_BUFFER);
                desc->slot_index_to_bind_directly = si_get_constbuf_slot(0);
 
+               if (is_2nd) {
+#if HAVE_32BIT_POINTERS
+                       if (i == PIPE_SHADER_TESS_CTRL) {
+                               rel_dw_offset = (R_00B40C_SPI_SHADER_USER_DATA_ADDR_HI_HS -
+                                                R_00B430_SPI_SHADER_USER_DATA_LS_0) / 4;
+                       } else { /* PIPE_SHADER_GEOMETRY */
+                               rel_dw_offset = (R_00B20C_SPI_SHADER_USER_DATA_ADDR_HI_GS -
+                                                R_00B330_SPI_SHADER_USER_DATA_ES_0) / 4;
+                       }
+#else
+                       rel_dw_offset = GFX9_SGPR_2ND_SAMPLERS_AND_IMAGES;
+#endif
+               } else {
+                       rel_dw_offset = SI_SGPR_SAMPLERS_AND_IMAGES;
+               }
+
                desc = si_sampler_and_image_descriptors(sctx, i);
-               si_init_descriptors(desc,
-                                   gfx9_tcs ? GFX9_SGPR_TCS_SAMPLERS_AND_IMAGES :
-                                   gfx9_gs ? GFX9_SGPR_GS_SAMPLERS_AND_IMAGES :
-                                             SI_SGPR_SAMPLERS_AND_IMAGES,
-                                   16, num_sampler_slots);
+               si_init_descriptors(desc, rel_dw_offset, 16, num_sampler_slots);
 
                int j;
                for (j = 0; j < SI_NUM_IMAGES; j++)
index efdf80b961ec01c996ec724ce4b120235e505d37..8e975a597c6e781ec1c9bc4210d727f14d61eb76 100644 (file)
@@ -3351,6 +3351,9 @@ static void si_set_ls_return_value_for_tcs(struct si_shader_context *ctx)
 {
        LLVMValueRef ret = ctx->return_value;
 
+       ret = si_insert_input_ptr(ctx, ret, 0, 0);
+       if (HAVE_32BIT_POINTERS)
+               ret = si_insert_input_ptr(ctx, ret, 1, 1);
        ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_offchip_offset, 2);
        ret = si_insert_input_ret(ctx, ret, ctx->param_merged_wave_info, 3);
        ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_factor_offset, 4);
@@ -3364,6 +3367,12 @@ static void si_set_ls_return_value_for_tcs(struct si_shader_context *ctx)
 
        ret = si_insert_input_ret(ctx, ret, ctx->param_vs_state_bits,
                                  8 + SI_SGPR_VS_STATE_BITS);
+
+#if !HAVE_32BIT_POINTERS
+       ret = si_insert_input_ptr(ctx, ret, ctx->param_vs_state_bits + 1,
+                                 8 + GFX9_SGPR_2ND_SAMPLERS_AND_IMAGES);
+#endif
+
        ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_offchip_layout,
                                  8 + GFX9_SGPR_TCS_OFFCHIP_LAYOUT);
        ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_out_lds_offsets,
@@ -3375,13 +3384,6 @@ static void si_set_ls_return_value_for_tcs(struct si_shader_context *ctx)
        ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_factor_addr_base64k,
                                  8 + GFX9_SGPR_TCS_FACTOR_ADDR_BASE64K);
 
-       unsigned desc_param = ctx->param_tcs_factor_addr_base64k +
-                             (HAVE_32BIT_POINTERS ? 1 : 2);
-       ret = si_insert_input_ptr(ctx, ret, desc_param,
-                                 8 + GFX9_SGPR_TCS_CONST_AND_SHADER_BUFFERS);
-       ret = si_insert_input_ptr(ctx, ret, desc_param + 1,
-                                 8 + GFX9_SGPR_TCS_SAMPLERS_AND_IMAGES);
-
        unsigned vgpr = 8 + GFX9_TCS_NUM_USER_SGPR;
        ret = LLVMBuildInsertValue(ctx->ac.builder, ret,
                                   ac_to_float(&ctx->ac, ctx->abi.tcs_patch_id),
@@ -3397,6 +3399,9 @@ static void si_set_es_return_value_for_gs(struct si_shader_context *ctx)
 {
        LLVMValueRef ret = ctx->return_value;
 
+       ret = si_insert_input_ptr(ctx, ret, 0, 0);
+       if (HAVE_32BIT_POINTERS)
+               ret = si_insert_input_ptr(ctx, ret, 1, 1);
        ret = si_insert_input_ret(ctx, ret, ctx->param_gs2vs_offset, 2);
        ret = si_insert_input_ret(ctx, ret, ctx->param_merged_wave_info, 3);
        ret = si_insert_input_ret(ctx, ret, ctx->param_merged_scratch_offset, 5);
@@ -3407,11 +3412,10 @@ static void si_set_es_return_value_for_gs(struct si_shader_context *ctx)
                                  ctx->param_bindless_samplers_and_images,
                                  8 + SI_SGPR_BINDLESS_SAMPLERS_AND_IMAGES);
 
-       unsigned desc_param = ctx->param_vs_state_bits + 1;
-       ret = si_insert_input_ptr(ctx, ret, desc_param,
-                                 8 + GFX9_SGPR_GS_CONST_AND_SHADER_BUFFERS);
-       ret = si_insert_input_ptr(ctx, ret, desc_param + 1,
-                                 8 + GFX9_SGPR_GS_SAMPLERS_AND_IMAGES);
+#if !HAVE_32BIT_POINTERS
+       ret = si_insert_input_ptr(ctx, ret, ctx->param_vs_state_bits + 1,
+                                 8 + GFX9_SGPR_2ND_SAMPLERS_AND_IMAGES);
+#endif
 
        unsigned vgpr = 8 + GFX9_GS_NUM_USER_SGPR;
        for (unsigned i = 0; i < 5; i++) {
@@ -4489,9 +4493,9 @@ static unsigned si_get_max_workgroup_size(const struct si_shader *shader)
        return max_work_group_size;
 }
 
-static void declare_per_stage_desc_pointers(struct si_shader_context *ctx,
-                                           struct si_function_info *fninfo,
-                                           bool assign_params)
+static void declare_const_and_shader_buffers(struct si_shader_context *ctx,
+                                            struct si_function_info *fninfo,
+                                            bool assign_params)
 {
        LLVMTypeRef const_shader_buf_type;
 
@@ -4505,14 +4509,28 @@ static void declare_per_stage_desc_pointers(struct si_shader_context *ctx,
                add_arg(fninfo, ARG_SGPR,
                        ac_array_in_const32_addr_space(const_shader_buf_type));
 
+       if (assign_params)
+               ctx->param_const_and_shader_buffers = const_and_shader_buffers;
+}
+
+static void declare_samplers_and_images(struct si_shader_context *ctx,
+                                       struct si_function_info *fninfo,
+                                       bool assign_params)
+{
        unsigned samplers_and_images =
                add_arg(fninfo, ARG_SGPR,
                        ac_array_in_const32_addr_space(ctx->v8i32));
 
-       if (assign_params) {
-               ctx->param_const_and_shader_buffers = const_and_shader_buffers;
+       if (assign_params)
                ctx->param_samplers_and_images = samplers_and_images;
-       }
+}
+
+static void declare_per_stage_desc_pointers(struct si_shader_context *ctx,
+                                           struct si_function_info *fninfo,
+                                           bool assign_params)
+{
+       declare_const_and_shader_buffers(ctx, fninfo, assign_params);
+       declare_samplers_and_images(ctx, fninfo, assign_params);
 }
 
 static void declare_global_desc_pointers(struct si_shader_context *ctx,
@@ -4677,8 +4695,14 @@ static void create_function(struct si_shader_context *ctx)
 
        case SI_SHADER_MERGED_VERTEX_TESSCTRL:
                /* Merged stages have 8 system SGPRs at the beginning. */
-               add_arg(&fninfo, ARG_SGPR, ctx->i32); /* SPI_SHADER_USER_DATA_ADDR_LO_HS */
-               add_arg(&fninfo, ARG_SGPR, ctx->i32); /* SPI_SHADER_USER_DATA_ADDR_HI_HS */
+               /* SPI_SHADER_USER_DATA_ADDR_LO/HI_HS */
+               if (HAVE_32BIT_POINTERS) {
+                       declare_per_stage_desc_pointers(ctx, &fninfo,
+                                                       ctx->type == PIPE_SHADER_TESS_CTRL);
+               } else {
+                       declare_const_and_shader_buffers(ctx, &fninfo,
+                                                        ctx->type == PIPE_SHADER_TESS_CTRL);
+               }
                ctx->param_tcs_offchip_offset = add_arg(&fninfo, ARG_SGPR, ctx->i32);
                ctx->param_merged_wave_info = add_arg(&fninfo, ARG_SGPR, ctx->i32);
                ctx->param_tcs_factor_offset = add_arg(&fninfo, ARG_SGPR, ctx->i32);
@@ -4691,16 +4715,15 @@ static void create_function(struct si_shader_context *ctx)
                                                ctx->type == PIPE_SHADER_VERTEX);
                declare_vs_specific_input_sgprs(ctx, &fninfo);
 
+               if (!HAVE_32BIT_POINTERS) {
+                       declare_samplers_and_images(ctx, &fninfo,
+                                                   ctx->type == PIPE_SHADER_TESS_CTRL);
+               }
                ctx->param_tcs_offchip_layout = add_arg(&fninfo, ARG_SGPR, ctx->i32);
                ctx->param_tcs_out_lds_offsets = add_arg(&fninfo, ARG_SGPR, ctx->i32);
                ctx->param_tcs_out_lds_layout = add_arg(&fninfo, ARG_SGPR, ctx->i32);
                ctx->param_tcs_offchip_addr_base64k = add_arg(&fninfo, ARG_SGPR, ctx->i32);
                ctx->param_tcs_factor_addr_base64k = add_arg(&fninfo, ARG_SGPR, ctx->i32);
-               if (!HAVE_32BIT_POINTERS)
-                       add_arg(&fninfo, ARG_SGPR, ctx->i32); /* unused */
-
-               declare_per_stage_desc_pointers(ctx, &fninfo,
-                                               ctx->type == PIPE_SHADER_TESS_CTRL);
 
                /* VGPRs (first TCS, then VS) */
                add_arg_assign(&fninfo, ARG_VGPR, ctx->i32, &ctx->abi.tcs_patch_id);
@@ -4731,8 +4754,14 @@ static void create_function(struct si_shader_context *ctx)
 
        case SI_SHADER_MERGED_VERTEX_OR_TESSEVAL_GEOMETRY:
                /* Merged stages have 8 system SGPRs at the beginning. */
-               add_arg(&fninfo, ARG_SGPR, ctx->i32); /* unused (SPI_SHADER_USER_DATA_ADDR_LO_GS) */
-               add_arg(&fninfo, ARG_SGPR, ctx->i32); /* unused (SPI_SHADER_USER_DATA_ADDR_HI_GS) */
+               /* SPI_SHADER_USER_DATA_ADDR_LO/HI_GS */
+               if (HAVE_32BIT_POINTERS) {
+                       declare_per_stage_desc_pointers(ctx, &fninfo,
+                                                       ctx->type == PIPE_SHADER_GEOMETRY);
+               } else {
+                       declare_const_and_shader_buffers(ctx, &fninfo,
+                                                        ctx->type == PIPE_SHADER_GEOMETRY);
+               }
                ctx->param_gs2vs_offset = add_arg(&fninfo, ARG_SGPR, ctx->i32);
                ctx->param_merged_wave_info = add_arg(&fninfo, ARG_SGPR, ctx->i32);
                ctx->param_tcs_offchip_offset = add_arg(&fninfo, ARG_SGPR, ctx->i32);
@@ -4758,8 +4787,10 @@ static void create_function(struct si_shader_context *ctx)
                        ctx->param_vs_state_bits = add_arg(&fninfo, ARG_SGPR, ctx->i32); /* unused */
                }
 
-               declare_per_stage_desc_pointers(ctx, &fninfo,
-                                               ctx->type == PIPE_SHADER_GEOMETRY);
+               if (!HAVE_32BIT_POINTERS) {
+                       declare_samplers_and_images(ctx, &fninfo,
+                                                   ctx->type == PIPE_SHADER_GEOMETRY);
+               }
 
                /* VGPRs (first GS, then VS/TES) */
                ctx->param_gs_vtx01_offset = add_arg(&fninfo, ARG_VGPR, ctx->i32);
@@ -7266,7 +7297,8 @@ static void si_build_tcs_epilog_function(struct si_shader_context *ctx,
        si_init_function_info(&fninfo);
 
        if (ctx->screen->info.chip_class >= GFX9) {
-               add_arg(&fninfo, ARG_SGPR, ctx->i64);
+               add_arg(&fninfo, ARG_SGPR, ctx->i32);
+               add_arg(&fninfo, ARG_SGPR, ctx->i32);
                ctx->param_tcs_offchip_offset = add_arg(&fninfo, ARG_SGPR, ctx->i32);
                add_arg(&fninfo, ARG_SGPR, ctx->i32); /* wave info */
                ctx->param_tcs_factor_offset = add_arg(&fninfo, ARG_SGPR, ctx->i32);
@@ -7282,6 +7314,8 @@ static void si_build_tcs_epilog_function(struct si_shader_context *ctx,
                add_arg(&fninfo, ARG_SGPR, ctx->i32);
                add_arg(&fninfo, ARG_SGPR, ctx->i32);
                add_arg(&fninfo, ARG_SGPR, ctx->i32);
+               if (!HAVE_32BIT_POINTERS)
+                       add_arg(&fninfo, ARG_SGPR, ctx->ac.intptr);
                ctx->param_tcs_offchip_layout = add_arg(&fninfo, ARG_SGPR, ctx->i32);
                add_arg(&fninfo, ARG_SGPR, ctx->i32);
                add_arg(&fninfo, ARG_SGPR, ctx->i32);
index ef4472ba9944923dd89a81dd6753c1838f40603f..e0d6f701483ee93c4b5d749ad49a523e7ffabd83 100644 (file)
@@ -195,38 +195,29 @@ enum {
        GFX6_SGPR_TCS_FACTOR_ADDR_BASE64K,
        GFX6_TCS_NUM_USER_SGPR,
 
+       /* GFX9: Merged shaders. */
+#if HAVE_32BIT_POINTERS
+       /* 2ND_CONST_AND_SHADER_BUFFERS is set in USER_DATA_ADDR_LO (SGPR0). */
+       /* 2ND_SAMPLERS_AND_IMAGES is set in USER_DATA_ADDR_HI (SGPR1). */
+       GFX9_MERGED_NUM_USER_SGPR = SI_VS_NUM_USER_SGPR,
+#else
+       /* 2ND_CONST_AND_SHADER_BUFFERS is set in USER_DATA_ADDR_LO/HI (SGPR[0:1]). */
+       GFX9_SGPR_2ND_SAMPLERS_AND_IMAGES = SI_VS_NUM_USER_SGPR,
+       GFX9_SGPR_2ND_SAMPLERS_AND_IMAGES_HI,
+       GFX9_MERGED_NUM_USER_SGPR,
+#endif
+
        /* GFX9: Merged LS-HS (VS-TCS) only. */
-       GFX9_SGPR_TCS_OFFCHIP_LAYOUT = SI_VS_NUM_USER_SGPR,
+       GFX9_SGPR_TCS_OFFCHIP_LAYOUT = GFX9_MERGED_NUM_USER_SGPR,
        GFX9_SGPR_TCS_OUT_OFFSETS,
        GFX9_SGPR_TCS_OUT_LAYOUT,
        GFX9_SGPR_TCS_OFFCHIP_ADDR_BASE64K,
        GFX9_SGPR_TCS_FACTOR_ADDR_BASE64K,
-#if !HAVE_32BIT_POINTERS
-       GFX9_SGPR_unused_to_align_the_next_pointer,
-#endif
-       GFX9_SGPR_TCS_CONST_AND_SHADER_BUFFERS,
-#if !HAVE_32BIT_POINTERS
-       GFX9_SGPR_TCS_CONST_AND_SHADER_BUFFERS_HI,
-#endif
-       GFX9_SGPR_TCS_SAMPLERS_AND_IMAGES,
-#if !HAVE_32BIT_POINTERS
-       GFX9_SGPR_TCS_SAMPLERS_AND_IMAGES_HI,
-#endif
        GFX9_TCS_NUM_USER_SGPR,
 
-       /* GFX9: Merged ES-GS (VS-GS or TES-GS). */
-       GFX9_SGPR_GS_CONST_AND_SHADER_BUFFERS = SI_VS_NUM_USER_SGPR,
-#if !HAVE_32BIT_POINTERS
-       GFX9_SGPR_GS_CONST_AND_SHADER_BUFFERS_HI,
-#endif
-       GFX9_SGPR_GS_SAMPLERS_AND_IMAGES,
-#if !HAVE_32BIT_POINTERS
-       GFX9_SGPR_GS_SAMPLERS_AND_IMAGES_HI,
-#endif
-       GFX9_GS_NUM_USER_SGPR,
-
        /* GS limits */
        GFX6_GS_NUM_USER_SGPR = SI_NUM_RESOURCE_SGPRS,
+       GFX9_GS_NUM_USER_SGPR = GFX9_MERGED_NUM_USER_SGPR,
        SI_GSCOPY_NUM_USER_SGPR = SI_SGPR_RW_BUFFERS + (HAVE_32BIT_POINTERS ? 1 : 2),
 
        /* PS only */