radv: Set shared VGPR count in radv_postprocess_config.

This commit allows RADV to set the shared VGPR count according to
the shader config.

Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

diff --git a/src/amd/vulkan/radv_pipeline.c b/src/amd/vulkan/radv_pipeline.c
index c107d55..ddf9462 100644 (file)
--- a/src/amd/vulkan/radv_pipeline.c
+++ b/src/amd/vulkan/radv_pipeline.c
@@ -4831,8 +4831,8 @@ radv_compute_generate_pm4(struct radv_pipeline *pipeline)
        unsigned max_waves_per_sh = 0;
        uint64_t va;
 
-       pipeline->cs.buf = malloc(20 * 4);
-       pipeline->cs.max_dw = 20;
+       pipeline->cs.max_dw = device->physical_device->rad_info.chip_class >= GFX10 ? 22 : 20;
+       pipeline->cs.buf = malloc(pipeline->cs.max_dw * 4);
 
        compute_shader = pipeline->shaders[MESA_SHADER_COMPUTE];
        va = radv_buffer_get_va(compute_shader->bo) + compute_shader->bo_offset;
@@ -4844,6 +4844,9 @@ radv_compute_generate_pm4(struct radv_pipeline *pipeline)
        radeon_set_sh_reg_seq(&pipeline->cs, R_00B848_COMPUTE_PGM_RSRC1, 2);
        radeon_emit(&pipeline->cs, compute_shader->config.rsrc1);
        radeon_emit(&pipeline->cs, compute_shader->config.rsrc2);
+       if (device->physical_device->rad_info.chip_class >= GFX10) {
+               radeon_set_sh_reg(&pipeline->cs, R_00B8A0_COMPUTE_PGM_RSRC3, compute_shader->config.rsrc3);
+       }
 
        radeon_set_sh_reg(&pipeline->cs, R_00B860_COMPUTE_TMPRING_SIZE,
                          S_00B860_WAVES(pipeline->max_waves) |

diff --git a/src/amd/vulkan/radv_shader.c b/src/amd/vulkan/radv_shader.c
index 2f3d85d..0561b46 100644 (file)
--- a/src/amd/vulkan/radv_shader.c
+++ b/src/amd/vulkan/radv_shader.c
@@ -701,10 +701,16 @@ static void radv_postprocess_config(const struct radv_physical_device *pdevice,
        unsigned num_vgprs = MAX2(config_in->num_vgprs, num_input_vgprs);
        /* +3 for scratch wave offset and VCC */
        unsigned num_sgprs = MAX2(config_in->num_sgprs, info->num_input_sgprs + 3);
+       unsigned num_shared_vgprs = config_in->num_shared_vgprs;
+       /* shared VGPRs are introduced in Navi and are allocated in blocks of 8 (RDNA ref 3.6.5) */
+       assert((pdevice->rad_info.chip_class >= GFX10 && num_shared_vgprs % 8 == 0)
+              || (pdevice->rad_info.chip_class < GFX10 && num_shared_vgprs == 0));
+       unsigned num_shared_vgpr_blocks = num_shared_vgprs / 8;
 
        *config_out = *config_in;
        config_out->num_vgprs = num_vgprs;
        config_out->num_sgprs = num_sgprs;
+       config_out->num_shared_vgprs = num_shared_vgprs;
 
        /* Enable 64-bit and 16-bit denormals, because there is no performance
         * cost.
@@ -760,6 +766,7 @@ static void radv_postprocess_config(const struct radv_physical_device *pdevice,
                        config_out->rsrc1 |= S_00B128_MEM_ORDERED(pdevice->rad_info.chip_class >= GFX10);
                        config_out->rsrc2 |= S_00B12C_OC_LDS_EN(1);
                }
+               config_out->rsrc2 |= S_00B22C_SHARED_VGPR_CNT(num_shared_vgpr_blocks);
                break;
        case MESA_SHADER_TESS_CTRL:
                if (pdevice->rad_info.chip_class >= GFX9) {
@@ -777,6 +784,7 @@ static void radv_postprocess_config(const struct radv_physical_device *pdevice,
                }
                config_out->rsrc1 |= S_00B428_MEM_ORDERED(pdevice->rad_info.chip_class >= GFX10) |
                                     S_00B848_WGP_MODE(pdevice->rad_info.chip_class >= GFX10);
+               config_out->rsrc2 |= S_00B42C_SHARED_VGPR_CNT(num_shared_vgpr_blocks);
                break;
        case MESA_SHADER_VERTEX:
                if (info->is_ngg) {
@@ -808,14 +816,17 @@ static void radv_postprocess_config(const struct radv_physical_device *pdevice,
                        }
 
                        config_out->rsrc1 |= S_00B128_MEM_ORDERED(pdevice->rad_info.chip_class >= GFX10);
+                       config_out->rsrc2 |= S_00B12C_SHARED_VGPR_CNT(num_shared_vgpr_blocks);
                }
                break;
        case MESA_SHADER_FRAGMENT:
                config_out->rsrc1 |= S_00B028_MEM_ORDERED(pdevice->rad_info.chip_class >= GFX10);
+               config_out->rsrc2 |= S_00B02C_SHARED_VGPR_CNT(num_shared_vgpr_blocks);
                break;
        case MESA_SHADER_GEOMETRY:
                config_out->rsrc1 |= S_00B228_MEM_ORDERED(pdevice->rad_info.chip_class >= GFX10) |
                                     S_00B848_WGP_MODE(pdevice->rad_info.chip_class >= GFX10);
+               config_out->rsrc2 |= S_00B22C_SHARED_VGPR_CNT(num_shared_vgpr_blocks);
                break;
        case MESA_SHADER_COMPUTE:
                config_out->rsrc1 |= S_00B848_MEM_ORDERED(pdevice->rad_info.chip_class >= GFX10) |
@@ -828,6 +839,8 @@ static void radv_postprocess_config(const struct radv_physical_device *pdevice,
                                                info->cs.uses_thread_id[1] ? 1 : 0) |
                        S_00B84C_TG_SIZE_EN(info->cs.uses_local_invocation_idx) |
                        S_00B84C_LDS_SIZE(config_in->lds_size);
+               config_out->rsrc3 |= S_00B8A0_SHARED_VGPR_CNT(num_shared_vgpr_blocks);
+
                break;
        default:
                unreachable("unsupported shader type");