From d0bdd4ce983ddd52f9f4b70dced4e471c60a130c Mon Sep 17 00:00:00 2001 From: Iago Toral Quiroga Date: Wed, 9 Jul 2014 08:46:17 +0200 Subject: [PATCH] i965/gen6/gs: Implement GS_OPCODE_FF_SYNC. This implements the FF_SYNC message required in gen6 geometry shaders to get the initial URB handle. Acked-by: Kenneth Graunke Reviewed-by: Jordan Justen --- src/mesa/drivers/dri/i965/brw_defines.h | 15 ++++++++ src/mesa/drivers/dri/i965/brw_shader.cpp | 2 ++ src/mesa/drivers/dri/i965/brw_vec4.cpp | 2 ++ src/mesa/drivers/dri/i965/brw_vec4.h | 3 ++ src/mesa/drivers/dri/i965/brw_vec4_generator.cpp | 45 ++++++++++++++++++++++++ 5 files changed, 67 insertions(+) diff --git a/src/mesa/drivers/dri/i965/brw_defines.h b/src/mesa/drivers/dri/i965/brw_defines.h index 0ef43e9..27a513a 100644 --- a/src/mesa/drivers/dri/i965/brw_defines.h +++ b/src/mesa/drivers/dri/i965/brw_defines.h @@ -1014,6 +1014,21 @@ enum opcode { * - dst is the GRF for gl_InvocationID. */ GS_OPCODE_GET_INSTANCE_ID, + + /** + * Send a FF_SYNC message to allocate initial URB handles (gen6). + * + * - dst will be used as the writeback register for the FF_SYNC operation. + * + * - src0 is the number of primitives written. + * + * Note: This opcode uses an implicit MRF register for the ff_sync message + * header, so the caller is expected to set inst->base_mrf and initialize + * that MRF register to r0. This opcode will also write to this MRF register + * to include the allocated URB handle so it can then be reused directly as + * the header in the URB write operation we are allocating the handle for. + */ + GS_OPCODE_FF_SYNC, }; enum brw_derivative_quality { diff --git a/src/mesa/drivers/dri/i965/brw_shader.cpp b/src/mesa/drivers/dri/i965/brw_shader.cpp index 1a18169..4eea74e 100644 --- a/src/mesa/drivers/dri/i965/brw_shader.cpp +++ b/src/mesa/drivers/dri/i965/brw_shader.cpp @@ -522,6 +522,8 @@ brw_instruction_name(enum opcode op) return "set_channel_masks"; case GS_OPCODE_GET_INSTANCE_ID: return "get_instance_id"; + case GS_OPCODE_FF_SYNC: + return "ff_sync"; default: /* Yes, this leaks. It's in debug code, it should never occur, and if diff --git a/src/mesa/drivers/dri/i965/brw_vec4.cpp b/src/mesa/drivers/dri/i965/brw_vec4.cpp index 9102be2..fa99276 100644 --- a/src/mesa/drivers/dri/i965/brw_vec4.cpp +++ b/src/mesa/drivers/dri/i965/brw_vec4.cpp @@ -276,6 +276,8 @@ vec4_visitor::implied_mrf_writes(vec4_instruction *inst) case GS_OPCODE_URB_WRITE: case GS_OPCODE_THREAD_END: return 0; + case GS_OPCODE_FF_SYNC: + return 1; case SHADER_OPCODE_SHADER_TIME_ADD: return 0; case SHADER_OPCODE_TEX: diff --git a/src/mesa/drivers/dri/i965/brw_vec4.h b/src/mesa/drivers/dri/i965/brw_vec4.h index 186667c..d2d2a8f 100644 --- a/src/mesa/drivers/dri/i965/brw_vec4.h +++ b/src/mesa/drivers/dri/i965/brw_vec4.h @@ -656,6 +656,9 @@ private: void generate_gs_prepare_channel_masks(struct brw_reg dst); void generate_gs_set_channel_masks(struct brw_reg dst, struct brw_reg src); void generate_gs_get_instance_id(struct brw_reg dst); + void generate_gs_ff_sync(vec4_instruction *inst, + struct brw_reg dst, + struct brw_reg src0); void generate_oword_dual_block_offsets(struct brw_reg m1, struct brw_reg index); void generate_scratch_write(vec4_instruction *inst, diff --git a/src/mesa/drivers/dri/i965/brw_vec4_generator.cpp b/src/mesa/drivers/dri/i965/brw_vec4_generator.cpp index 954ce64..c1eda12 100644 --- a/src/mesa/drivers/dri/i965/brw_vec4_generator.cpp +++ b/src/mesa/drivers/dri/i965/brw_vec4_generator.cpp @@ -659,6 +659,47 @@ vec4_generator::generate_gs_get_instance_id(struct brw_reg dst) } void +vec4_generator::generate_gs_ff_sync(vec4_instruction *inst, + struct brw_reg dst, + struct brw_reg src0) +{ + /* This opcode uses an implied MRF register for: + * - the header of the ff_sync message. And as such it is expected to be + * initialized to r0 before calling here. + * - the destination where we will write the allocated URB handle. + */ + struct brw_reg header = + retype(brw_message_reg(inst->base_mrf), BRW_REGISTER_TYPE_UD); + + /* Overwrite dword 0 of the header (cleared for now since we are not doing + * transform feedback) and dword 1 (to hold the number of primitives + * written). + */ + brw_push_insn_state(p); + brw_set_default_mask_control(p, BRW_MASK_DISABLE); + brw_set_default_access_mode(p, BRW_ALIGN_1); + brw_MOV(p, get_element_ud(header, 0), brw_imm_ud(0)); + brw_MOV(p, get_element_ud(header, 1), get_element_ud(src0, 0)); + brw_pop_insn_state(p); + + /* Allocate URB handle in dst */ + brw_ff_sync(p, + dst, + 0, + header, + 1, /* allocate */ + 1, /* response length */ + 0 /* eot */); + + /* Now put allocated urb handle in header.0 */ + brw_push_insn_state(p); + brw_set_default_access_mode(p, BRW_ALIGN_1); + brw_set_default_mask_control(p, BRW_MASK_DISABLE); + brw_MOV(p, get_element_ud(header, 0), get_element_ud(dst, 0)); + brw_pop_insn_state(p); +} + +void vec4_generator::generate_oword_dual_block_offsets(struct brw_reg m1, struct brw_reg index) { @@ -1280,6 +1321,10 @@ vec4_generator::generate_code(const cfg_t *cfg) generate_gs_set_vertex_count(dst, src[0]); break; + case GS_OPCODE_FF_SYNC: + generate_gs_ff_sync(inst, dst, src[0]); + break; + case GS_OPCODE_SET_DWORD_2_IMMED: generate_gs_set_dword_2_immed(dst, src[0]); break; -- 2.7.4