/**
* This function initializes all the compute specific registers that need to
* be initialized for each compute command stream. Registers that are common
* to both compute and 3D will be initialized at the beginning of each compute
* command stream by the start_cs_cmd atom. However, since the SET_CONTEXT_REG
* packet requires that the shader type bit be set, we must initialize all
* context registers needed for compute in this function. The registers
* intialized by the start_cs_cmd atom can be found in evereen_state.c in the
* functions evergreen_init_atom_start_cs or cayman_init_atom_start_cs depending
* on the GPU family.
*/
void evergreen_init_atom_start_compute_cs(struct r600_context *ctx)
{
struct r600_command_buffer *cb = &ctx->start_compute_cs_cmd;
int num_threads;
int num_stack_entries;
/* since all required registers are initialised in the
* start_compute_cs_cmd atom, we can EMIT_EARLY here.
*/
r600_init_command_buffer(ctx, cb, 1, 256);
cb->pkt_flags = RADEON_CP_PACKET3_COMPUTE_MODE;
switch (ctx->family) {
case CHIP_CEDAR:
default:
num_threads = 128;
num_stack_entries = 256;
break;
case CHIP_REDWOOD:
num_threads = 128;
num_stack_entries = 256;
break;
case CHIP_JUNIPER:
num_threads = 128;
num_stack_entries = 512;
break;
case CHIP_CYPRESS:
case CHIP_HEMLOCK:
num_threads = 128;
num_stack_entries = 512;
break;
case CHIP_PALM:
num_threads = 128;
num_stack_entries = 256;
break;
case CHIP_SUMO:
num_threads = 128;
num_stack_entries = 256;
break;
case CHIP_SUMO2:
num_threads = 128;
num_stack_entries = 512;
break;
case CHIP_BARTS:
num_threads = 128;
num_stack_entries = 512;
break;
case CHIP_TURKS:
num_threads = 128;
num_stack_entries = 256;
break;
case CHIP_CAICOS:
num_threads = 128;
num_stack_entries = 256;
break;
}
/* Config Registers */
evergreen_init_common_regs(cb, ctx->chip_class
, ctx->family, ctx->screen->info.drm_minor);
/* The primitive type always needs to be POINTLIST for compute. */
r600_store_config_reg(cb, R_008958_VGT_PRIMITIVE_TYPE,
V_008958_DI_PT_POINTLIST);
if (ctx->chip_class < CAYMAN) {
/* These registers control which simds can be used by each stage.
* The default for these registers is 0xffffffff, which means
* all simds are available for each stage. It's possible we may
* want to play around with these in the future, but for now
* the default value is fine.
*
* R_008E20_SQ_STATIC_THREAD_MGMT1
* R_008E24_SQ_STATIC_THREAD_MGMT2
* R_008E28_SQ_STATIC_THREAD_MGMT3
*/
/* XXX: We may need to adjust the thread and stack resouce
* values for 3D/compute interop */
r600_store_config_reg_seq(cb, R_008C18_SQ_THREAD_RESOURCE_MGMT_1, 5);
/* R_008C18_SQ_THREAD_RESOURCE_MGMT_1
* Set the number of threads used by the PS/VS/GS/ES stage to
* 0.
*/
r600_store_value(cb, 0);
/* R_008C1C_SQ_THREAD_RESOURCE_MGMT_2
* Set the number of threads used by the CS (aka LS) stage to
* the maximum number of threads and set the number of threads
* for the HS stage to 0. */
r600_store_value(cb, S_008C1C_NUM_LS_THREADS(num_threads));
/* R_008C20_SQ_STACK_RESOURCE_MGMT_1
* Set the Control Flow stack entries to 0 for PS/VS stages */
r600_store_value(cb, 0);
/* R_008C24_SQ_STACK_RESOURCE_MGMT_2
* Set the Control Flow stack entries to 0 for GS/ES stages */
r600_store_value(cb, 0);
/* R_008C28_SQ_STACK_RESOURCE_MGMT_3
* Set the Contol Flow stack entries to 0 for the HS stage, and
* set it to the maximum value for the CS (aka LS) stage. */
r600_store_value(cb,
S_008C28_NUM_LS_STACK_ENTRIES(num_stack_entries));
}
/* Context Registers */
if (ctx->chip_class < CAYMAN) {
/* workaround for hw issues with dyn gpr - must set all limits
* to 240 instead of 0, 0x1e == 240 / 8
*/
r600_store_context_reg(cb, R_028838_SQ_DYN_GPR_RESOURCE_LIMIT_1,
S_028838_PS_GPRS(0x1e) |
S_028838_VS_GPRS(0x1e) |
S_028838_GS_GPRS(0x1e) |
S_028838_ES_GPRS(0x1e) |
S_028838_HS_GPRS(0x1e) |
S_028838_LS_GPRS(0x1e));
}
/* XXX: Investigate setting bit 15, which is FAST_COMPUTE_MODE */
r600_store_context_reg(cb, R_028A40_VGT_GS_MODE,
S_028A40_COMPUTE_MODE(1) | S_028A40_PARTIAL_THD_AT_EOI(1));
r600_store_context_reg(cb, R_028B54_VGT_SHADER_STAGES_EN, 2/*CS_ON*/);
r600_store_context_reg(cb, R_0286E8_SPI_COMPUTE_INPUT_CNTL,
S_0286E8_TID_IN_GROUP_ENA
| S_0286E8_TGID_ENA
| S_0286E8_DISABLE_INDEX_PACK)
;
/* The LOOP_CONST registers are an optimizations for loops that allows
* you to store the initial counter, increment value, and maximum
* counter value in a register so that hardware can calculate the
* correct number of iterations for the loop, so that you don't need
* to have the loop counter in your shader code. We don't currently use
* this optimization, so we must keep track of the counter in the
* shader and use a break instruction to exit loops. However, the
* hardware will still uses this register to determine when to exit a
* loop, so we need to initialize the counter to 0, set the increment
* value to 1 and the maximum counter value to the 4095 (0xfff) which
* is the maximum value allowed. This gives us a maximum of 4096
* iterations for our loops, but hopefully our break instruction will
* execute before some time before the 4096th iteration.
*/
eg_store_loop_const(cb, R_03A200_SQ_LOOP_CONST_0 + (160 * 4), 0x1000FFF);
}
void evergreen_compute_init_config(struct r600_context *ctx)
{
struct evergreen_compute_resource* res =
get_empty_res(ctx->cs_shader, COMPUTE_RESOURCE_CONFIG, 0);
int num_threads;
int num_stack_entries;
int num_temp_gprs;
enum radeon_family family;
unsigned tmp;
family = ctx->family;
switch (family) {
case CHIP_CEDAR:
default:
num_temp_gprs = 4;
num_threads = 128;
num_stack_entries = 256;
break;
case CHIP_REDWOOD:
num_temp_gprs = 4;
num_threads = 128;
num_stack_entries = 256;
break;
case CHIP_JUNIPER:
num_temp_gprs = 4;
num_threads = 128;
num_stack_entries = 512;
break;
case CHIP_CYPRESS:
case CHIP_HEMLOCK:
num_temp_gprs = 4;
num_threads = 128;
num_stack_entries = 512;
break;
case CHIP_PALM:
num_temp_gprs = 4;
num_threads = 128;
num_stack_entries = 256;
break;
case CHIP_SUMO:
num_temp_gprs = 4;
num_threads = 128;
num_stack_entries = 256;
break;
case CHIP_SUMO2:
num_temp_gprs = 4;
num_threads = 128;
num_stack_entries = 512;
break;
case CHIP_BARTS:
num_temp_gprs = 4;
num_threads = 128;
num_stack_entries = 512;
break;
case CHIP_TURKS:
num_temp_gprs = 4;
num_threads = 128;
num_stack_entries = 256;
break;
case CHIP_CAICOS:
num_temp_gprs = 4;
num_threads = 128;
num_stack_entries = 256;
break;
}
tmp = 0x00000000;
switch (family) {
case CHIP_CEDAR:
case CHIP_PALM:
case CHIP_SUMO:
case CHIP_SUMO2:
case CHIP_CAICOS:
break;
default:
tmp |= S_008C00_VC_ENABLE(1);
break;
}
tmp |= S_008C00_EXPORT_SRC_C(1);
tmp |= S_008C00_CS_PRIO(0);
tmp |= S_008C00_LS_PRIO(0);
tmp |= S_008C00_HS_PRIO(0);
tmp |= S_008C00_PS_PRIO(0);
tmp |= S_008C00_VS_PRIO(0);
tmp |= S_008C00_GS_PRIO(0);
tmp |= S_008C00_ES_PRIO(0);
evergreen_reg_set(res, R_008C00_SQ_CONFIG, tmp);
evergreen_reg_set(res, R_008C04_SQ_GPR_RESOURCE_MGMT_1,
S_008C04_NUM_CLAUSE_TEMP_GPRS(num_temp_gprs));
if (ctx->chip_class < CAYMAN) {
evergreen_reg_set(res, R_008C08_SQ_GPR_RESOURCE_MGMT_2, 0);
}
evergreen_reg_set(res, R_008C10_SQ_GLOBAL_GPR_RESOURCE_MGMT_1, 0);
evergreen_reg_set(res, R_008C14_SQ_GLOBAL_GPR_RESOURCE_MGMT_2, 0);
evergreen_reg_set(res, R_008D8C_SQ_DYN_GPR_CNTL_PS_FLUSH_REQ, (1 << 8));
/* workaround for hw issues with dyn gpr - must set all limits to 240
* instead of 0, 0x1e == 240/8 */
if (ctx->chip_class < CAYMAN) {
evergreen_reg_set(res, R_028838_SQ_DYN_GPR_RESOURCE_LIMIT_1,
S_028838_PS_GPRS(0x1e) |
S_028838_VS_GPRS(0x1e) |
S_028838_GS_GPRS(0x1e) |
S_028838_ES_GPRS(0x1e) |
S_028838_HS_GPRS(0x1e) |
S_028838_LS_GPRS(0x1e));
} else {
evergreen_reg_set(res, 0x286f8,
S_028838_PS_GPRS(0x1e) |
S_028838_VS_GPRS(0x1e) |
S_028838_GS_GPRS(0x1e) |
S_028838_ES_GPRS(0x1e) |
S_028838_HS_GPRS(0x1e) |
S_028838_LS_GPRS(0x1e));
}
if (ctx->chip_class < CAYMAN) {
evergreen_reg_set(res, R_008E20_SQ_STATIC_THREAD_MGMT1, 0xFFFFFFFF);
evergreen_reg_set(res, R_008E24_SQ_STATIC_THREAD_MGMT2, 0xFFFFFFFF);
evergreen_reg_set(res, R_008E20_SQ_STATIC_THREAD_MGMT1, 0xFFFFFFFF);
evergreen_reg_set(res, R_008E24_SQ_STATIC_THREAD_MGMT2, 0xFFFFFFFF);
evergreen_reg_set(res, R_008E28_SQ_STATIC_THREAD_MGMT3, 0xFFFFFFFF);
evergreen_reg_set(res, R_008C18_SQ_THREAD_RESOURCE_MGMT_1, 0);
tmp = S_008C1C_NUM_LS_THREADS(num_threads);
evergreen_reg_set(res, R_008C1C_SQ_THREAD_RESOURCE_MGMT_2, tmp);
evergreen_reg_set(res, R_008C20_SQ_STACK_RESOURCE_MGMT_1, 0);
evergreen_reg_set(res, R_008C24_SQ_STACK_RESOURCE_MGMT_2, 0);
tmp = S_008C28_NUM_LS_STACK_ENTRIES(num_stack_entries);
evergreen_reg_set(res, R_008C28_SQ_STACK_RESOURCE_MGMT_3, tmp);
}
evergreen_reg_set(res, R_0286CC_SPI_PS_IN_CONTROL_0, S_0286CC_LINEAR_GRADIENT_ENA(1));
evergreen_reg_set(res, R_0286D0_SPI_PS_IN_CONTROL_1, 0);
evergreen_reg_set(res, R_0286E4_SPI_PS_IN_CONTROL_2, 0);
evergreen_reg_set(res, R_0286D8_SPI_INPUT_Z, 0);
evergreen_reg_set(res, R_0286E0_SPI_BARYC_CNTL, 1 << 20);
tmp = S_0286E8_TID_IN_GROUP_ENA | S_0286E8_TGID_ENA | S_0286E8_DISABLE_INDEX_PACK;
evergreen_reg_set(res, R_0286E8_SPI_COMPUTE_INPUT_CNTL, tmp);
tmp = S_028A40_COMPUTE_MODE(1) | S_028A40_PARTIAL_THD_AT_EOI(1);
evergreen_reg_set(res, R_028A40_VGT_GS_MODE, tmp);
evergreen_reg_set(res, R_028B54_VGT_SHADER_STAGES_EN, 2/*CS_ON*/);
evergreen_reg_set(res, R_028800_DB_DEPTH_CONTROL, 0);
evergreen_reg_set(res, R_02880C_DB_SHADER_CONTROL, 0);
evergreen_reg_set(res, R_028000_DB_RENDER_CONTROL, S_028000_COLOR_DISABLE(1));
evergreen_reg_set(res, R_02800C_DB_RENDER_OVERRIDE, 0);
evergreen_reg_set(res, R_0286E8_SPI_COMPUTE_INPUT_CNTL,
S_0286E8_TID_IN_GROUP_ENA
| S_0286E8_TGID_ENA
| S_0286E8_DISABLE_INDEX_PACK)
;
}