[Mesa-dev] [RFC 3/4] nvc0: Add new launch descriptor format for GP100
Boyan Ding
boyan.j.ding at gmail.com
Sun Apr 23 15:19:22 UTC 2017
Signed-off-by: Boyan Ding <boyan.j.ding at gmail.com>
---
src/gallium/drivers/nouveau/nvc0/nve4_compute.c | 105 ++++++++++++++++++++++--
src/gallium/drivers/nouveau/nvc0/nve4_compute.h | 67 +++++++++++++++
2 files changed, 166 insertions(+), 6 deletions(-)
diff --git a/src/gallium/drivers/nouveau/nvc0/nve4_compute.c b/src/gallium/drivers/nouveau/nvc0/nve4_compute.c
index 798761debd..c536381233 100644
--- a/src/gallium/drivers/nouveau/nvc0/nve4_compute.c
+++ b/src/gallium/drivers/nouveau/nvc0/nve4_compute.c
@@ -29,6 +29,7 @@
#ifdef DEBUG
static void nve4_compute_dump_launch_desc(const struct nve4_cp_launch_desc *);
+static void gp100_compute_dump_launch_desc(const struct gp100_cp_launch_desc *);
#endif
@@ -585,7 +586,45 @@ nve4_compute_setup_launch_desc(struct nvc0_context *nvc0,
NVC0_CB_AUX_INFO(5), 1 << 11);
}
-static inline struct nve4_cp_launch_desc *
+static void
+gp100_compute_setup_launch_desc(struct nvc0_context *nvc0,
+ struct gp100_cp_launch_desc *desc,
+ const struct pipe_grid_info *info)
+{
+ const struct nvc0_screen *screen = nvc0->screen;
+ const struct nvc0_program *cp = nvc0->compprog;
+
+ gp100_cp_launch_desc_init_default(desc);
+
+ desc->entry = nvc0_program_symbol_offset(cp, info->pc);
+
+ desc->griddim_x = info->grid[0];
+ desc->griddim_y = info->grid[1];
+ desc->griddim_z = info->grid[2];
+ desc->blockdim_x = info->block[0];
+ desc->blockdim_y = info->block[1];
+ desc->blockdim_z = info->block[2];
+
+ desc->shared_size = align(cp->cp.smem_size, 0x100);
+ desc->local_size_p = (cp->hdr[1] & 0xfffff0) + align(cp->cp.lmem_size, 0x10);
+ desc->local_size_n = 0;
+ desc->cstack_size = 0x800;
+
+ desc->gpr_alloc = cp->num_gprs;
+ desc->bar_alloc = cp->num_barriers;
+
+ // Only bind user uniforms and the driver constant buffer through the
+ // launch descriptor because UBOs are sticked to the driver cb to avoid the
+ // limitation of 8 CBs.
+ if (nvc0->constbuf[5][0].user || cp->parm_size) {
+ gp100_cp_launch_desc_set_cb(desc, 0, screen->uniform_bo,
+ NVC0_CB_USR_INFO(5), 1 << 16);
+ }
+ gp100_cp_launch_desc_set_cb(desc, 7, screen->uniform_bo,
+ NVC0_CB_AUX_INFO(5), 1 << 11);
+}
+
+static inline void *
nve4_compute_alloc_launch_desc(struct nouveau_context *nv,
struct nouveau_bo **pbo, uint64_t *pgpuaddr)
{
@@ -597,7 +636,7 @@ nve4_compute_alloc_launch_desc(struct nouveau_context *nv,
ptr += adj;
*pgpuaddr += adj;
}
- return (struct nve4_cp_launch_desc *)ptr;
+ return ptr;
}
void
@@ -605,7 +644,7 @@ nve4_launch_grid(struct pipe_context *pipe, const struct pipe_grid_info *info)
{
struct nvc0_context *nvc0 = nvc0_context(pipe);
struct nouveau_pushbuf *push = nvc0->base.pushbuf;
- struct nve4_cp_launch_desc *desc;
+ void *desc;
uint64_t desc_gpuaddr;
struct nouveau_bo *desc_bo;
int ret;
@@ -622,13 +661,20 @@ nve4_launch_grid(struct pipe_context *pipe, const struct pipe_grid_info *info)
if (ret)
goto out;
- nve4_compute_setup_launch_desc(nvc0, desc, info);
+ if (nvc0->screen->base.device->chipset >= 0x130)
+ gp100_compute_setup_launch_desc(nvc0, desc, info);
+ else
+ nve4_compute_setup_launch_desc(nvc0, desc, info);
nve4_compute_upload_input(nvc0, info);
#ifdef DEBUG
- if (debug_get_num_option("NV50_PROG_DEBUG", 0))
- nve4_compute_dump_launch_desc(desc);
+ if (debug_get_num_option("NV50_PROG_DEBUG", 0)) {
+ if (nvc0->screen->base.device->chipset >= 0x130)
+ gp100_compute_dump_launch_desc(desc);
+ else
+ nve4_compute_dump_launch_desc(desc);
+ }
#endif
if (unlikely(info->indirect)) {
@@ -831,6 +877,53 @@ nve4_compute_dump_launch_desc(const struct nve4_cp_launch_desc *desc)
i, address, size, valid ? "" : " (invalid)");
}
}
+
+static void
+gp100_compute_dump_launch_desc(const struct gp100_cp_launch_desc *desc)
+{
+ const uint32_t *data = (const uint32_t *)desc;
+ unsigned i;
+ bool zero = false;
+
+ debug_printf("COMPUTE LAUNCH DESCRIPTOR:\n");
+
+ for (i = 0; i < sizeof(*desc); i += 4) {
+ if (data[i / 4]) {
+ debug_printf("[%x]: 0x%08x\n", i, data[i / 4]);
+ zero = false;
+ } else
+ if (!zero) {
+ debug_printf("...\n");
+ zero = true;
+ }
+ }
+
+ debug_printf("entry = 0x%x\n", desc->entry);
+ debug_printf("grid dimensions = %ux%ux%u\n",
+ desc->griddim_x, desc->griddim_y, desc->griddim_z);
+ debug_printf("block dimensions = %ux%ux%u\n",
+ desc->blockdim_x, desc->blockdim_y, desc->blockdim_z);
+ debug_printf("s[] size: 0x%x\n", desc->shared_size);
+ debug_printf("l[] size: -0x%x / +0x%x\n",
+ desc->local_size_n, desc->local_size_p);
+ debug_printf("stack size: 0x%x\n", desc->cstack_size);
+ debug_printf("barrier count: %u\n", desc->bar_alloc);
+ debug_printf("$r count: %u\n", desc->gpr_alloc);
+ debug_printf("linked tsc: %d\n", desc->linked_tsc);
+
+ for (i = 0; i < 8; ++i) {
+ uint64_t address;
+ uint32_t size = desc->cb[i].size_sh4 << 4;
+ bool valid = !!(desc->cb_mask & (1 << i));
+
+ address = ((uint64_t)desc->cb[i].address_h << 32) | desc->cb[i].address_l;
+
+ if (!valid && !address && !size)
+ continue;
+ debug_printf("CB[%u]: address = 0x%"PRIx64", size 0x%x%s\n",
+ i, address, size, valid ? "" : " (invalid)");
+ }
+}
#endif
#ifdef NOUVEAU_NVE4_MP_TRAP_HANDLER
diff --git a/src/gallium/drivers/nouveau/nvc0/nve4_compute.h b/src/gallium/drivers/nouveau/nvc0/nve4_compute.h
index 75b169323b..7ff6935cc3 100644
--- 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;
--
2.12.2
More information about the mesa-dev
mailing list