nvc0: Add new launch descriptor format for GP100

[mesa.git] / src / gallium / drivers / nouveau / nvc0 / nve4_compute.h

diff --git a/src/gallium/drivers/nouveau/nvc0/nve4_compute.h b/src/gallium/drivers/nouveau/nvc0/nve4_compute.h
index 75b169323bc9fa88865e4434dd32ded80d29e61b..7ff6935cc3d00d3bd773a40487e4b2264439d86e 100644 (file)
--- a/src/gallium/drivers/nouveau/nvc0/nve4_compute.h
+++ b/src/gallium/drivers/nouveau/nvc0/nve4_compute.h
@@ -45,6 +45,46 @@ struct nve4_cp_launch_desc
    u32 unk48[16];
 };
 
+struct gp100_cp_launch_desc
+{
+   u32 unk0[8];
+   u32 entry;
+   u32 unk9[2];
+   u32 unk11_0      : 30;
+   u32 linked_tsc   : 1;
+   u32 unk11_31     : 1;
+   u32 griddim_x    : 31;
+   u32 unk12        : 1;
+   u16 griddim_y;
+   u16 unk13;
+   u16 griddim_z;
+   u16 unk14;
+   u32 unk15[2];
+   u32 shared_size  : 18;
+   u32 unk17        : 14;
+   u16 unk18;
+   u16 blockdim_x;
+   u16 blockdim_y;
+   u16 blockdim_z;
+   u32 cb_mask      : 8;
+   u32 unk20        : 24;
+   u32 unk21[8];
+   u32 local_size_p : 24;
+   u32 unk29        : 3;
+   u32 bar_alloc    : 5;
+   u32 local_size_n : 24;
+   u32 gpr_alloc    : 8;
+   u32 cstack_size  : 24;
+   u32 unk31        : 8;
+   struct {
+      u32 address_l;
+      u32 address_h : 17;
+      u32 reserved  : 2;
+      u32 size_sh4  : 13;
+   } cb[8];
+   u32 unk48[16];
+};
+
 static inline void
 nve4_cp_launch_desc_init_default(struct nve4_cp_launch_desc *desc)
 {
@@ -73,6 +113,33 @@ nve4_cp_launch_desc_set_cb(struct nve4_cp_launch_desc *desc,
    desc->cb_mask |= 1 << index;
 }
 
+static inline void
+gp100_cp_launch_desc_init_default(struct gp100_cp_launch_desc *desc)
+{
+   memset(desc, 0, sizeof(*desc));
+
+   desc->unk0[4]  = 0x40;
+   desc->unk11_0  = 0x04014000;
+}
+
+static inline void
+gp100_cp_launch_desc_set_cb(struct gp100_cp_launch_desc *desc,
+                            unsigned index,
+                            struct nouveau_bo *bo,
+                            uint32_t base, uint32_t size)
+{
+   uint64_t address = bo->offset + base;
+
+   assert(index < 8);
+   assert(!(base & 0xff));
+
+   desc->cb[index].address_l = address;
+   desc->cb[index].address_h = address >> 32;
+   desc->cb[index].size_sh4 = DIV_ROUND_UP(size, 16);
+
+   desc->cb_mask |= 1 << index;
+}
+
 struct nve4_mp_trap_info {
    u32 lock;
    u32 pc;