radeondb: avoid duplicate state emission

Avoid reemitting same state allow nice cut in the cs stream
size from the classic driver (up to 2 time smaller than what
classic does for same rendering)

Signed-off-by: Jerome Glisse <jglisse@redhat.com>

9 files changed, 69 insertions, 11 deletions

diff --git a/Makefile.am b/Makefile.am
index 4196901..a39844f 100644
--- a/Makefile.am
+++ b/Makefile.am
@@ -20,4 +20,7 @@
 #  CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
 AUTOMAKE_OPTIONS = foreign
 
+debug:
+	$(MAKE) all "CFLAGS=-O0 -g3 -DDEBUG"
+
 SUBDIRS = src
diff --git a/src/r600_ctx.c b/src/r600_ctx.c
index 2181d51..5f1fdff 100644
--- a/src/r600_ctx.c
+++ b/src/r600_ctx.c
@@ -178,6 +178,27 @@ static int cs_draw_index(struct radeon_ctx *ctx)
 	return r600_bof_ctx_new_draw(ctx);
 }
 
+static int cs_draw_index_immd(struct radeon_ctx *ctx)
+{
+	unsigned i;
+	int r;
+
+	r = radeon_ctx_draw_set_reg(ctx, R_008970_VGT_NUM_INDICES, ctx->pm4[ctx->id + 1]);
+	if (r)
+		return r;
+	r = radeon_ctx_draw_set_reg(ctx, R_0287F0_VGT_DRAW_INITIATOR, ctx->pm4[ctx->id + 2]);
+	if (r)
+		return r;
+	ctx->cdraw->state[R600_DRAW]->nimmd = PKT_COUNT_G(ctx->pm4[ctx->id]) - 1;
+	ctx->cdraw->state[R600_DRAW]->immd = calloc(1, ctx->cdraw->state[R600_DRAW]->nimmd * 4);
+	if (ctx->cdraw->state[R600_DRAW]->immd == NULL)
+		return -ENOMEM;
+	for (i = 0; i < ctx->cdraw->state[R600_DRAW]->nimmd; i++) {
+		ctx->cdraw->state[R600_DRAW]->immd[i] = ctx->pm4[ctx->id + 3 + i];
+	}
+	return r600_bof_ctx_new_draw(ctx);
+}
+
 static int cs_set_alu_const(struct radeon_ctx *ctx)
 {
 	unsigned i, count, offset;
@@ -296,7 +317,7 @@ struct cs_packet3 r600_packet3[] = {
 	{0x2A, "INDEX_TYPE", &cs_index_type},
 	{0x2B, "DRAW_INDEX", &cs_draw_index},
 	{0x2D, "DRAW_INDEX_AUTO", &cs_draw_auto},
-	{0x2E, "DRAW_INDEX_IMMD", NULL},
+	{0x2E, "DRAW_INDEX_IMMD", &cs_draw_index_immd},
 	{0x2F, "NUM_INSTANCES", &cs_num_instances},
 	{0x32, "INDIRECT_BUFFER", NULL},
 	{0x34, "STRMOUT_BUFFER_UPDATE", NULL},
diff --git a/src/r600_state.c b/src/r600_state.c
index 739e260..397c5ad 100644
--- a/src/r600_state.c
+++ b/src/r600_state.c
@@ -286,6 +286,7 @@ static int r600_state_pm4_vgt(struct radeon_state *state)
 
 static int r600_state_pm4_draw(struct radeon_state *state)
 {
+	unsigned i;
 	int r;
 
 	if (state->nbo) {
@@ -299,6 +300,13 @@ static int r600_state_pm4_draw(struct radeon_state *state)
 		if (r)
 			return r;
 		state->pm4[state->cpm4++] = state->bo[0]->handle;
+	} else if  (state->nimmd) {
+		state->pm4[state->cpm4++] = PKT3(PKT3_DRAW_INDEX_IMMD, state->nimmd + 1);
+		state->pm4[state->cpm4++] = state->states[R600_DRAW__VGT_NUM_INDICES];
+		state->pm4[state->cpm4++] = state->states[R600_DRAW__VGT_DRAW_INITIATOR];
+		for (i = 0; i < state->nimmd; i++) {
+			state->pm4[state->cpm4++] = state->immd[i];
+		}
 	} else {
 		state->pm4[state->cpm4++] = PKT3(PKT3_DRAW_INDEX_AUTO, 1);
 		state->pm4[state->cpm4++] = state->states[R600_DRAW__VGT_NUM_INDICES];
diff --git a/src/radeon.c b/src/radeon.c
index 19ed4c8..66dbfc3 100644
--- a/src/radeon.c
+++ b/src/radeon.c
@@ -251,8 +251,8 @@ int radeon_flush(struct radeon *radeon)
 	ctx->pm4 = malloc(ctx->cpm4 * 4);
 	if (ctx->pm4 == NULL)
 		return -EINVAL;
-	for (i = 0, ctx->id = 0; i < ctx->ndraw; i++) {
-		r = radeon_ctx_draw_schedule(ctx, i);
+	for (i = 0, ctx->id = 0; i < ctx->nstate; i++) {
+		r = radeon_ctx_state_schedule(ctx, ctx->state[i]);
 		if (r)
 			return r;
 	}
@@ -261,7 +261,10 @@ int radeon_flush(struct radeon *radeon)
 			__func__, ctx->cpm4, ctx->id);
 		return -EINVAL;
 	}
-	return radeon_ctx_submit(ctx);
+	r = radeon_ctx_submit(ctx);
+	radeon_ctx_destroy(ctx);
+	radeon->ctx = radeon_ctx(radeon, 0);
+	return r;
 }
 
 int radeon_reg_id(struct radeon *radeon, unsigned offset, unsigned *typeid, unsigned *stateid, unsigned *id)
@@ -313,7 +316,8 @@ int radeon_schedule_draw(struct radeon *radeon, struct radeon_draw *draw)
 	struct radeon_ctx *ctx = radeon->ctx;
 	struct radeon_draw *pdraw = NULL;
 	struct radeon_state *nstate, *ostate;
-	unsigned cpm4, i;
+	unsigned cpm4, i, cstate;
+	void *tmp;
 	int r = 0;
 
 	if (ctx == NULL) {
@@ -331,22 +335,36 @@ int radeon_schedule_draw(struct radeon *radeon, struct radeon_draw *draw)
 		return -EINVAL;
 	}
 reprocess:
+	tmp = realloc(ctx->state, (ctx->nstate + draw->nstate) * sizeof(void*));
+	if (tmp == NULL)
+		return -ENOMEM;
+	ctx->state = tmp;
 	pdraw = ctx->cdraw;
-	for (i = 0, cpm4 = 0; i < draw->nstate; i++) {
+	for (i = 0, cpm4 = 0, cstate = ctx->nstate; i < draw->nstate - 1; i++) {
 		nstate = draw->state[i];
 		if (nstate) {
 			if (pdraw && pdraw->state[i]) {
 				ostate = pdraw->state[i];
 				if (ostate->pm4_crc == nstate->pm4_crc) {
-					cpm4 += nstate->cpm4;
+//					ctx->state[cstate++] = nstate;
+//					cpm4 += nstate->cpm4;
 				} else {
+					ctx->state[cstate++] = nstate;
 					cpm4 += nstate->cpm4;
 				}
 			} else {
+				ctx->state[cstate++] = nstate;
 				cpm4 += nstate->cpm4;
 			}
 		}
 	}
+	/* The last state is the draw state always add it */
+	if (draw->state[i] == NULL) {
+		fprintf(stderr, "%s no draw command\n", __func__);
+		return -EINVAL;
+	}
+	ctx->state[cstate++] = draw->state[i];
+	cpm4 += draw->state[i]->cpm4;
 	if ((ctx->cpm4 + cpm4) > RADEON_CTX_MAX_PM4) {
 		r = radeon_flush(radeon);
 		if (r)
@@ -356,7 +374,9 @@ reprocess:
 	r = radeon_ctx_set_draw_new(ctx, draw);
 	if (r)
 		goto out_err;
+	radeon_draw_incref(draw);
 	ctx->cpm4 += cpm4;
+	ctx->nstate = cstate;
 out_err:
 	return r;
 }
diff --git a/src/radeon.h b/src/radeon.h
index a787550..806d85c 100644
--- a/src/radeon.h
+++ b/src/radeon.h
@@ -105,6 +105,8 @@ struct radeon_state {
 	unsigned			cpm4;
 	u32				pm4_crc;
 	u32				*pm4;
+	u32				nimmd;
+	u32				*immd;
 	unsigned			nbo;
 	struct radeon_object		*bo[4];
 	unsigned			nreloc;
diff --git a/src/radeon_ctx.c b/src/radeon_ctx.c
index c625091..4e0ba56 100644
--- a/src/radeon_ctx.c
+++ b/src/radeon_ctx.c
@@ -341,7 +341,7 @@ int radeon_ctx_state_schedule(struct radeon_ctx *ctx, struct radeon_state *state
 	ctx->id += state->cpm4;
 	return 0;
 }
-
+#if 0
 int radeon_ctx_draw_schedule(struct radeon_ctx *ctx, unsigned id)
 {
 	struct radeon_draw *draw;
@@ -356,7 +356,7 @@ int radeon_ctx_draw_schedule(struct radeon_ctx *ctx, unsigned id)
 	}
 	return 0;
 }
-
+#endif
 int radeon_ctx_draw_set_reg(struct radeon_ctx *ctx, unsigned offset, u32 value)
 {
 	struct radeon_draw *draw;
diff --git a/src/radeon_priv.h b/src/radeon_priv.h
index 758c6c6..a30ba73 100644
--- a/src/radeon_priv.h
+++ b/src/radeon_priv.h
@@ -151,6 +151,8 @@ struct radeon_ctx {
 	unsigned			ndraw;
 	struct radeon_draw		*cdraw;
 	struct radeon_draw		**draw;
+	unsigned			nstate;
+	struct radeon_state		**state;
 };
 
 extern struct radeon_ctx *radeon_ctx(struct radeon *radeon, unsigned device);
@@ -165,6 +167,7 @@ extern int radeon_ctx_allocate_bo(struct radeon_ctx *ctx);
 extern int radeon_ctx_submit(struct radeon_ctx *ctx);
 extern int radeon_ctx_draw(struct radeon_ctx *ctx);
 extern int radeon_ctx_set_draw_new(struct radeon_ctx *ctx, struct radeon_draw *draw);
+extern int radeon_ctx_state_schedule(struct radeon_ctx *ctx, struct radeon_state *state);
 extern int radeon_ctx_draw_schedule(struct radeon_ctx *ctx, unsigned id);
 extern int radeon_ctx_draw_set_reg(struct radeon_ctx *ctx, unsigned offset, u32 value);
 
diff --git a/src/radeon_state.c b/src/radeon_state.c
index 2a568a9..6cba23d 100644
--- a/src/radeon_state.c
+++ b/src/radeon_state.c
@@ -101,6 +101,7 @@ struct radeon_state *radeon_state_decref(struct radeon_state *state)
 	for (i = 0; i < state->nbo; i++) {
 		state->bo[i] = radeon_object_decref(state->bo[i]);
 	}
+	free(state->immd);
 	free(state->states);
 	free(state->pm4);
 	memset(state, 0, sizeof(*state));
@@ -152,7 +153,7 @@ int radeon_state_pm4(struct radeon_state *state)
 	if (r) {
 		return r;
 	}
-	state->pm4_crc = crc32(state->pm4, state->cpm4);
+	state->pm4_crc = crc32(state->pm4, state->cpm4 * 4);
 	return 0;
 }
 
diff --git a/src/radeondb.c b/src/radeondb.c
index c29e89e..b81fb5f 100644
--- a/src/radeondb.c
+++ b/src/radeondb.c
@@ -216,7 +216,7 @@ static int json(const char *file)
 		goto out;
 	}
 	radeon_blit(&dst, &src, 0, 0, 0, 0, src.width, src.height);
-	for (unsigned i = 0; i < ctx->ndraw - 0; i++) {
+	for (unsigned i = 0; i < ctx->ndraw; i++) {
 		if (radeon_draw_pm4(ctx->draw[i])) {
 			fprintf(stderr, "failed to pm4 %d\n", i);
 			goto out;