* OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
* USE OR OTHER DEALINGS IN THE SOFTWARE.
*/
-#include "r600_asm.h"
-#include "r600_context.h"
+#include <stdio.h>
+#include <errno.h>
#include "util/u_memory.h"
+#include "r600_pipe.h"
#include "r600_sq.h"
#include "r600_opcodes.h"
-#include <stdio.h>
-#include <errno.h>
+#include "r600_asm.h"
static inline unsigned int r600_bc_get_num_operands(struct r600_bc_alu *alu)
{
case V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SIN:
case V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_COS:
return 1;
-
default: R600_ERR(
"Need instruction operand number for 0x%x.\n", alu->inst);
};
case CHIP_RV740:
bc->chiprev = 1;
break;
+ case CHIP_CEDAR:
+ case CHIP_REDWOOD:
+ case CHIP_JUNIPER:
+ case CHIP_CYPRESS:
+ case CHIP_HEMLOCK:
+ bc->chiprev = 2;
+ break;
default:
R600_ERR("unknown family %d\n", bc->family);
return -EINVAL;
return 0;
}
+#if 0
static int reserve_gpr(struct r600_bc_alu *alu, unsigned sel, unsigned chan, unsigned cycle)
{
if (alu->hw_gpr[cycle][chan] < 0)
return ret;
}
-static int is_const(int sel)
-{
- if (sel > 255 && sel < 512)
- return 1;
- if (sel >= V_SQ_ALU_SRC_0 && sel <= V_SQ_ALU_SRC_LITERAL)
- return 1;
- return 0;
-}
+
static void update_chan_counter(struct r600_bc_alu *alu, int *chan_counter)
{
}
}
-#if 0
/* we need something like this I think - but this is bogus */
int check_read_slots(struct r600_bc *bc, struct r600_bc_alu *alu_first)
{
}
#endif
+static int is_const(int sel)
+{
+ if (sel > 255 && sel < 512)
+ return 1;
+ if (sel >= V_SQ_ALU_SRC_0 && sel <= V_SQ_ALU_SRC_LITERAL)
+ return 1;
+ return 0;
+}
+
static int check_scalar(struct r600_bc *bc, struct r600_bc_alu *alu)
{
unsigned swizzle_key;
+ if (alu->bank_swizzle_force) {
+ alu->bank_swizzle = alu->bank_swizzle_force;
+ return 0;
+ }
swizzle_key = (is_const(alu->src[0].sel) ? 4 : 0 ) +
(is_const(alu->src[1].sel) ? 2 : 0 ) +
(is_const(alu->src[2].sel) ? 1 : 0 );
{
unsigned swizzle_key;
+ if (alu->bank_swizzle_force) {
+ alu->bank_swizzle = alu->bank_swizzle_force;
+ return 0;
+ }
swizzle_key = (is_const(alu->src[0].sel) ? 4 : 0 ) +
(is_const(alu->src[1].sel) ? 2 : 0 ) +
(is_const(alu->src[2].sel) ? 1 : 0 );
static int check_and_set_bank_swizzle(struct r600_bc *bc, struct r600_bc_alu *alu_first)
{
- struct r600_bc_alu *alu;
+ struct r600_bc_alu *alu = NULL;
int num_instr = 1;
init_gpr(alu_first);
{
struct r600_bc_alu *nalu = r600_bc_alu();
struct r600_bc_alu *lalu;
- struct r600_bc_alu *curr_bs_head;
int i, r;
if (nalu == NULL)
/* cf can contains only alu or only vtx or only tex */
if (bc->cf_last == NULL || bc->cf_last->inst != (type << 3) ||
bc->force_add_cf) {
- /* at most 128 slots, one add alu can add 4 slots + 4 constant worst case */
r = r600_bc_add_cf(bc);
if (r) {
free(nalu);
} else {
LIST_ADDTAIL(&nalu->bs_list, &bc->cf_last->curr_bs_head->bs_list);
}
- if (alu->last && (bc->cf_last->ndw >> 1) >= 124) {
+ /* at most 128 slots, one add alu can add 4 slots + 4 constants(2 slots)
+ * worst case */
+ if (alu->last && (bc->cf_last->ndw >> 1) >= 120) {
bc->force_add_cf = 1;
}
/* number of gpr == the last gpr used in any alu */
bc->cf_last->ndw += 2;
bc->ndw += 2;
- if (bc->use_mem_constant)
- bc->cf_last->kcache0_mode = 2;
+ bc->cf_last->kcache0_mode = 2;
/* process cur ALU instructions for bank swizzle */
if (alu->last) {
int r600_bc_add_alu(struct r600_bc *bc, const struct r600_bc_alu *alu)
{
- return r600_bc_add_alu_type(bc, alu, V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU);
+ return r600_bc_add_alu_type(bc, alu, BC_INST(bc, V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU));
}
int r600_bc_add_literal(struct r600_bc *bc, const u32 *value)
if (bc->cf_last->inst == V_SQ_CF_WORD1_SQ_CF_INST_TEX) {
return 0;
}
+ /* all same on EG */
if (bc->cf_last->inst == V_SQ_CF_WORD1_SQ_CF_INST_JUMP ||
bc->cf_last->inst == V_SQ_CF_WORD1_SQ_CF_INST_ELSE ||
bc->cf_last->inst == V_SQ_CF_WORD1_SQ_CF_INST_LOOP_START_NO_AL ||
bc->cf_last->inst == V_SQ_CF_WORD1_SQ_CF_INST_POP) {
return 0;
}
+ /* same on EG */
if (((bc->cf_last->inst != (V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU << 3)) &&
(bc->cf_last->inst != (V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_PUSH_BEFORE << 3))) ||
LIST_IS_EMPTY(&bc->cf_last->alu)) {
/* cf can contains only alu or only vtx or only tex */
if (bc->cf_last == NULL ||
(bc->cf_last->inst != V_SQ_CF_WORD1_SQ_CF_INST_VTX &&
- bc->cf_last->inst != V_SQ_CF_WORD1_SQ_CF_INST_VTX_TC)) {
+ bc->cf_last->inst != V_SQ_CF_WORD1_SQ_CF_INST_VTX_TC) ||
+ bc->force_add_cf) {
r = r600_bc_add_cf(bc);
if (r) {
free(nvtx);
/* each fetch use 4 dwords */
bc->cf_last->ndw += 4;
bc->ndw += 4;
+ if ((bc->ndw / 4) > 7)
+ bc->force_add_cf = 1;
return 0;
}
/* cf can contains only alu or only vtx or only tex */
if (bc->cf_last == NULL ||
- bc->cf_last->inst != V_SQ_CF_WORD1_SQ_CF_INST_TEX) {
+ bc->cf_last->inst != V_SQ_CF_WORD1_SQ_CF_INST_TEX ||
+ bc->force_add_cf) {
r = r600_bc_add_cf(bc);
if (r) {
free(ntex);
/* each texture fetch use 4 dwords */
bc->cf_last->ndw += 4;
bc->ndw += 4;
+ if ((bc->ndw / 4) > 7)
+ bc->force_add_cf = 1;
return 0;
}
return 0;
}
+/* common to all 3 families */
static int r600_bc_vtx_build(struct r600_bc *bc, struct r600_bc_vtx *vtx, unsigned id)
{
bc->bytecode[id++] = S_SQ_VTX_WORD0_BUFFER_ID(vtx->buffer_id) |
S_SQ_VTX_WORD1_DST_SEL_Y(vtx->dst_sel_y) |
S_SQ_VTX_WORD1_DST_SEL_Z(vtx->dst_sel_z) |
S_SQ_VTX_WORD1_DST_SEL_W(vtx->dst_sel_w) |
- S_SQ_VTX_WORD1_USE_CONST_FIELDS(1) |
+ S_SQ_VTX_WORD1_USE_CONST_FIELDS(vtx->use_const_fields) |
+ S_SQ_VTX_WORD1_DATA_FORMAT(vtx->data_format) |
+ S_SQ_VTX_WORD1_NUM_FORMAT_ALL(vtx->num_format_all) |
+ S_SQ_VTX_WORD1_FORMAT_COMP_ALL(vtx->format_comp_all) |
+ S_SQ_VTX_WORD1_SRF_MODE_ALL(vtx->srf_mode_all) |
S_SQ_VTX_WORD1_GPR_DST_GPR(vtx->dst_gpr);
bc->bytecode[id++] = S_SQ_VTX_WORD2_MEGA_FETCH(1);
bc->bytecode[id++] = 0;
return 0;
}
+/* common to all 3 families */
static int r600_bc_tex_build(struct r600_bc *bc, struct r600_bc_tex *tex, unsigned id)
{
bc->bytecode[id++] = S_SQ_TEX_WORD0_TEX_INST(tex->inst) |
return 0;
}
+/* r600 only, r700/eg bits in r700_asm.c */
static int r600_bc_alu_build(struct r600_bc *bc, struct r600_bc_alu *alu, unsigned id)
{
unsigned i;
return 0;
}
+/* common for r600/r700 - eg in eg_asm.c */
static int r600_bc_cf_build(struct r600_bc *bc, struct r600_bc_cf *cf)
{
unsigned id = cf->id;
bc->bytecode[id++] = S_SQ_CF_ALU_WORD1_CF_INST(cf->inst >> 3) |
S_SQ_CF_ALU_WORD1_BARRIER(1) |
+ S_SQ_CF_ALU_WORD1_USES_WATERFALL(bc->chiprev == 0 ? cf->r6xx_uses_waterfall : 0) |
S_SQ_CF_ALU_WORD1_COUNT((cf->ndw / 2) - 1);
break;
case V_SQ_CF_WORD1_SQ_CF_INST_TEX:
break;
case V_SQ_CF_ALLOC_EXPORT_WORD1_SQ_CF_INST_EXPORT:
case V_SQ_CF_ALLOC_EXPORT_WORD1_SQ_CF_INST_EXPORT_DONE:
+ case EG_V_SQ_CF_ALLOC_EXPORT_WORD1_SQ_CF_INST_EXPORT:
+ case EG_V_SQ_CF_ALLOC_EXPORT_WORD1_SQ_CF_INST_EXPORT_DONE:
break;
case V_SQ_CF_WORD1_SQ_CF_INST_JUMP:
case V_SQ_CF_WORD1_SQ_CF_INST_ELSE:
return -ENOMEM;
LIST_FOR_EACH_ENTRY(cf, &bc->cf, list) {
addr = cf->addr;
- r = r600_bc_cf_build(bc, cf);
+ if (bc->chiprev == 2)
+ r = eg_bc_cf_build(bc, cf);
+ else
+ r = r600_bc_cf_build(bc, cf);
if (r)
return r;
switch (cf->inst) {
r = r600_bc_alu_build(bc, alu, addr);
break;
case 1:
+ case 2: /* eg alu is same encoding as r700 */
r = r700_bc_alu_build(bc, alu, addr);
break;
default:
break;
case V_SQ_CF_ALLOC_EXPORT_WORD1_SQ_CF_INST_EXPORT:
case V_SQ_CF_ALLOC_EXPORT_WORD1_SQ_CF_INST_EXPORT_DONE:
+ case EG_V_SQ_CF_ALLOC_EXPORT_WORD1_SQ_CF_INST_EXPORT:
+ case EG_V_SQ_CF_ALLOC_EXPORT_WORD1_SQ_CF_INST_EXPORT_DONE:
case V_SQ_CF_WORD1_SQ_CF_INST_LOOP_START_NO_AL:
case V_SQ_CF_WORD1_SQ_CF_INST_LOOP_END:
case V_SQ_CF_WORD1_SQ_CF_INST_LOOP_CONTINUE: