planb: multi-ib command submissioncs-start

3 files changed, 69 insertions, 61 deletions

diff --git a/include/drm/radeon_drm.h b/include/drm/radeon_drm.h
index 42409daa..99316b22 100644
--- a/include/drm/radeon_drm.h
+++ b/include/drm/radeon_drm.h
@@ -874,7 +874,7 @@ struct drm_radeon_gem_pwrite {
 
 #define RADEON_CHUNK_ID_RELOCS	0x01
 #define RADEON_CHUNK_ID_IB	0x02
-#define RADEON_CHUNK_ID_IB_SETUP  0x03
+#define RADEON_CHUNK_ID_IB_CONTINUED 0x03
 
 struct drm_radeon_cs_chunk {
 	uint32_t		chunk_id;
diff --git a/radeon/radeon_cs_gem.c b/radeon/radeon_cs_gem.c
index e0cdbaa2..9f6ac4b5 100644
--- a/radeon/radeon_cs_gem.c
+++ b/radeon/radeon_cs_gem.c
@@ -49,6 +49,8 @@
 
 #define CS_BOF_DUMP 0
 
+#define CS_MAX_IBS 4
+
 struct radeon_cs_manager_gem {
     struct radeon_cs_manager    base;
     uint32_t                    device_id;
@@ -68,12 +70,13 @@ struct cs_reloc_gem {
 
 struct cs_gem {
     struct radeon_cs_int        base;
-    struct drm_radeon_cs        cs;
-    struct drm_radeon_cs_chunk  chunks[3];
     unsigned                    nrelocs;
+    unsigned                    reloc_length_dw;
     uint32_t                    *relocs;
     struct radeon_bo_int        **relocs_bo;
-    struct radeon_cs        alt;
+    struct radeon_cs            ibs[CS_MAX_IBS];
+    unsigned num_ibs_alloced; 
+    unsigned num_ibs_used;
 };
 
 static pthread_mutex_t id_mutex = PTHREAD_MUTEX_INITIALIZER;
@@ -120,21 +123,11 @@ static void free_id(uint32_t id)
     pthread_mutex_unlock( &id_mutex );
 }
 
-void radeon_cs_gem_switch_alt(struct radeon_cs *cs)
-{
-    struct cs_gem *csg = (struct cs_gem*)cs;
-    struct radeon_cs tmp;
-
-    /* swap alt and tmp */
-    tmp = *cs;
-    *cs = csg->alt;
-    csg->alt = tmp;
-}
-
 struct radeon_cs *radeon_cs_gem_create_setup(struct radeon_cs_manager *csm,
-					     uint32_t ndw, uint32_t setup_ndw)
+					     uint32_t ndw, uint32_t num_ibs)
 {
     struct cs_gem *csg;
+    uint32_t i;
 
     /* max cmd buffer size is 64Kb */
     if (ndw > (64 * 1024 / 4)) {
@@ -145,12 +138,18 @@ struct radeon_cs *radeon_cs_gem_create_setup(struct radeon_cs_manager *csm,
         return NULL;
     }
     csg->base.csm = csm;
-    csg->base.ndw = 64 * 1024 / 4;
-    csg->base.packets = (uint32_t*)calloc(1, 64 * 1024);
-    if (csg->base.packets == NULL) {
-        free(csg);
-        return NULL;
+
+    for (i = 0; i < num_ibs; i++) {
+	csg->ibs[i].packets = (uint32_t *)calloc(1, 64 * 1024);
+	if (!csg->ibs[i].packets) {
+	    free(csg);
+	    return NULL;
+	}
     }
+    csg->num_ibs_alloced = num_ibs;
+    csg->num_ibs_used = 1;
+    csg->base.ndw = 64 * 1024 / 4;
+    csg->base.packets = csg->ibs[0].packets;
     csg->base.relocs_total_size = 0;
     csg->base.crelocs = 0;
     csg->base.id = generate_id();
@@ -158,39 +157,37 @@ struct radeon_cs *radeon_cs_gem_create_setup(struct radeon_cs_manager *csm,
     csg->relocs_bo = (struct radeon_bo_int**)calloc(1,
                                                 csg->nrelocs*sizeof(void*));
     if (csg->relocs_bo == NULL) {
-        free(csg->base.packets);
         free(csg);
         return NULL;
     }
     csg->base.relocs = csg->relocs = (uint32_t*)calloc(1, 4096);
     if (csg->relocs == NULL) {
         free(csg->relocs_bo);
-        free(csg->base.packets);
         free(csg);
         return NULL;
     }
-    csg->alt.ndw = setup_ndw;
-    csg->alt.packets = (uint32_t*)calloc(1, 4*1024);
-    if (csg->alt.packets == NULL) {
-	free(csg);
-	return NULL;
-    }
-    csg->chunks[0].chunk_id = RADEON_CHUNK_ID_IB;
-    csg->chunks[0].length_dw = 0;
-    csg->chunks[0].chunk_data = (uint64_t)(uintptr_t)csg->base.packets;
-    csg->chunks[1].chunk_id = RADEON_CHUNK_ID_RELOCS;
-    csg->chunks[1].length_dw = 0;
-    csg->chunks[1].chunk_data = (uint64_t)(uintptr_t)csg->relocs;
-    csg->chunks[2].chunk_id = RADEON_CHUNK_ID_IB_SETUP;
-    csg->chunks[2].length_dw = 0;
-    csg->chunks[2].chunk_data = (uint64_t)(uintptr_t)csg->alt.packets;
     return (struct radeon_cs*)csg;
 }
 
+int radeon_cs_gem_switch_next_ib(struct radeon_cs *cs)
+{
+    struct cs_gem *csg = (struct cs_gem *)cs;
+
+    if (csg->num_ibs_used + 1 > csg->num_ibs_alloced)
+	return -1;
+
+    csg->ibs[csg->num_ibs_used - 1].cdw = csg->base.cdw;
+
+    csg->base.packets = csg->ibs[csg->num_ibs_used].packets;
+    csg->base.cdw = 0;
+    csg->num_ibs_used++;
+    return 0;
+}
+
 static struct radeon_cs_int *cs_gem_create(struct radeon_cs_manager *csm,
 					   uint32_t ndw)
 {
-    return (struct radeon_cs_int *)radeon_cs_gem_create_setup(csm, ndw, 0);
+    return (struct radeon_cs_int *)radeon_cs_gem_create_setup(csm, ndw, 1);
 }
 
 static int cs_gem_write_reloc(struct radeon_cs_int *cs,
@@ -278,7 +275,6 @@ static int cs_gem_write_reloc(struct radeon_cs_int *cs,
         }
         cs->relocs = csg->relocs = tmp;
         csg->nrelocs += 1;
-        csg->chunks[1].chunk_data = (uint64_t)(uintptr_t)csg->relocs;
     }
     csg->relocs_bo[csg->base.crelocs] = boi;
     idx = (csg->base.crelocs++) * RELOC_SIZE;
@@ -287,7 +283,7 @@ static int cs_gem_write_reloc(struct radeon_cs_int *cs,
     reloc->read_domain = read_domain;
     reloc->write_domain = write_domain;
     reloc->flags = flags;
-    csg->chunks[1].length_dw += RELOC_SIZE;
+    csg->reloc_length_dw += RELOC_SIZE;
     radeon_bo_ref(bo);
     /* bo might be referenced from another context so have to use atomic opertions */
     atomic_add((atomic_t *)radeon_gem_get_reloc_in_cs(bo), cs->id);
@@ -446,30 +442,42 @@ out_err:
 static int cs_gem_emit(struct radeon_cs_int *cs)
 {
     struct cs_gem *csg = (struct cs_gem*)cs;
-    uint64_t chunk_array[3];
+    uint64_t chunk_array[8];
+    struct drm_radeon_cs_chunk chunks[8];
+    struct drm_radeon_cs cs_pkt;
     unsigned i;
     int r;
 
 #if CS_BOF_DUMP
     cs_gem_dump_bof(cs);
 #endif
-    csg->chunks[0].length_dw = cs->cdw;
 
-    chunk_array[0] = (uint64_t)(uintptr_t)&csg->chunks[0];
-    chunk_array[1] = (uint64_t)(uintptr_t)&csg->chunks[1];
+    memset(&cs_pkt, 0, sizeof(struct drm_radeon_cs));
+    memset(chunks, 0, sizeof(chunks));
+
+    if (csg->num_ibs_used > 1)
+        fprintf(stderr,"reloc length is %d\n", csg->reloc_length_dw);
+
+    chunks[0].chunk_id = RADEON_CHUNK_ID_RELOCS;
+    chunks[0].length_dw = csg->reloc_length_dw;
+    chunks[0].chunk_data = (uint64_t)(uintptr_t)csg->relocs;
 
-    csg->cs.num_chunks = 2;
+    chunk_array[0] = (uint64_t)(uintptr_t)&chunks[0];
+    cs_pkt.num_chunks = 1;
 
-    if (csg->alt.cdw > 0) {
-        csg->chunks[2].length_dw = csg->alt.cdw;
-    	chunk_array[2] = (uint64_t)(uintptr_t)&csg->chunks[2];
-	csg->cs.num_chunks++;
+    csg->ibs[csg->num_ibs_used - 1].cdw = csg->base.cdw;
+    for (i = 0; i < csg->num_ibs_used; i++) {
+	chunks[i + 1].chunk_id = RADEON_CHUNK_ID_IB;
+	chunks[i + 1].chunk_data = (uint64_t)(uintptr_t)csg->ibs[i].packets;
+	chunks[i + 1].length_dw = csg->ibs[i].cdw;
+	chunk_array[i + 1] = (uint64_t)(uintptr_t)&chunks[i + 1];
+	cs_pkt.num_chunks++;
     }
 
-    csg->cs.chunks = (uint64_t)(uintptr_t)chunk_array;
+    cs_pkt.chunks = (uint64_t)(uintptr_t)chunk_array;
 
     r = drmCommandWriteRead(cs->csm->fd, DRM_RADEON_CS,
-                            &csg->cs, sizeof(struct drm_radeon_cs));
+                            &cs_pkt, sizeof(struct drm_radeon_cs));
     for (i = 0; i < csg->base.crelocs; i++) {
         csg->relocs_bo[i]->space_accounted = 0;
         /* bo might be referenced from another context so have to use atomic opertions */
@@ -487,12 +495,13 @@ static int cs_gem_emit(struct radeon_cs_int *cs)
 static int cs_gem_destroy(struct radeon_cs_int *cs)
 {
     struct cs_gem *csg = (struct cs_gem*)cs;
+    int i;
 
     free_id(cs->id);
     free(csg->relocs_bo);
     free(cs->relocs);
-    free(cs->packets);
-    free(csg->alt.packets);
+    for (i = 0; i < CS_MAX_IBS; i++)
+	free(csg->ibs[i].packets);
     free(cs);
     return 0;
 }
@@ -514,13 +523,11 @@ static int cs_gem_erase(struct radeon_cs_int *cs)
     }
     cs->relocs_total_size = 0;
     cs->cdw = 0;
-    csg->alt.cdw = 0;
     cs->section_ndw = 0;
-    csg->alt.section_ndw = 0;
     cs->crelocs = 0;
-    csg->chunks[0].length_dw = 0;
-    csg->chunks[1].length_dw = 0;
-    csg->chunks[2].length_dw = 0;
+    csg->num_ibs_used = 1;
+    csg->reloc_length_dw = 0;
+    csg->base.packets = csg->ibs[0].packets;
     return 0;
 }
 
diff --git a/radeon/radeon_cs_gem.h b/radeon/radeon_cs_gem.h
index fb51cd5d..340ed500 100644
--- a/radeon/radeon_cs_gem.h
+++ b/radeon/radeon_cs_gem.h
@@ -1,3 +1,4 @@
+
 /*
  * Copyright © 2008 Nicolai Haehnle
  * Copyright © 2008 Jérôme Glisse
@@ -38,7 +39,7 @@
 struct radeon_cs_manager *radeon_cs_manager_gem_ctor(int fd);
 void radeon_cs_manager_gem_dtor(struct radeon_cs_manager *csm);
 
-void radeon_cs_gem_switch_alt(struct radeon_cs *cs);
 struct radeon_cs *radeon_cs_gem_create_setup(struct radeon_cs_manager *csm,
-					     uint32_t ndw, uint32_t setup_ndw);
+					     uint32_t ndw, uint32_t num_ibs);
+int radeon_cs_gem_switch_next_ib(struct radeon_cs *cs);
 #endif