#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
case 0x120:
obj_class = GM200_COMPUTE_CLASS;
break;
+ case 0x130:
+ obj_class = (dev->chipset == 0x130 || dev->chipset == 0x13b) ?
+ GP100_COMPUTE_CLASS : GP104_COMPUTE_CLASS;
+ break;
default:
NOUVEAU_ERR("unsupported chipset: NV%02x\n", dev->chipset);
return -1;
return ret;
}
- ret = nouveau_bo_new(dev, NV_VRAM_DOMAIN(&screen->base), 0, 1 << 12, NULL,
- &screen->parm);
- if (ret)
- return ret;
-
BEGIN_NVC0(push, SUBC_CP(NV01_SUBCHAN_OBJECT), 1);
PUSH_DATA (push, screen->compute->oclass);
BEGIN_NVC0(push, NVE4_CP(TEX_CB_INDEX), 1);
PUSH_DATA (push, 7); /* does not interfere with 3D */
+ /* Disabling this UNK command avoid a read fault when using texelFetch()
+ * from a compute shader for weird reasons.
if (obj_class == NVF0_COMPUTE_CLASS)
IMMED_NVC0(push, SUBC_CP(0x02c4), 1);
+ */
address = screen->uniform_bo->offset + NVC0_CB_AUX_INFO(5);
return 0;
}
-
static void
-nve4_compute_validate_surfaces(struct nvc0_context *nvc0)
+gm107_compute_validate_surfaces(struct nvc0_context *nvc0,
+ struct pipe_image_view *view, int slot)
{
- struct nvc0_screen *screen = nvc0->screen;
+ struct nv04_resource *res = nv04_resource(view->resource);
struct nouveau_pushbuf *push = nvc0->base.pushbuf;
- struct nv50_surface *sf;
- struct nv04_resource *res;
- uint32_t mask;
- unsigned i;
- const unsigned t = 1;
+ struct nvc0_screen *screen = nvc0->screen;
+ struct nouveau_bo *txc = nvc0->screen->txc;
+ struct nv50_tic_entry *tic;
uint64_t address;
+ const int s = 5;
- address = screen->uniform_bo->offset + NVC0_CB_AUX_INFO(5);
+ tic = nv50_tic_entry(nvc0->images_tic[s][slot]);
+
+ res = nv04_resource(tic->pipe.texture);
+ nvc0_update_tic(nvc0, tic, res);
- mask = nvc0->surfaces_dirty[t];
- while (mask) {
- i = ffs(mask) - 1;
- mask &= ~(1 << i);
+ if (tic->id < 0) {
+ tic->id = nvc0_screen_tic_alloc(nvc0->screen, tic);
- /*
- * NVE4's surface load/store instructions receive all the information
- * directly instead of via binding points, so we have to supply them.
- */
+ /* upload the texture view */
+ PUSH_SPACE(push, 16);
BEGIN_NVC0(push, NVE4_CP(UPLOAD_DST_ADDRESS_HIGH), 2);
- PUSH_DATAh(push, address + NVC0_CB_AUX_BUF_INFO(i));
- PUSH_DATA (push, address + NVC0_CB_AUX_BUF_INFO(i));
+ PUSH_DATAh(push, txc->offset + (tic->id * 32));
+ PUSH_DATA (push, txc->offset + (tic->id * 32));
BEGIN_NVC0(push, NVE4_CP(UPLOAD_LINE_LENGTH_IN), 2);
- PUSH_DATA (push, 64);
+ PUSH_DATA (push, 32);
PUSH_DATA (push, 1);
- BEGIN_1IC0(push, NVE4_CP(UPLOAD_EXEC), 17);
+ BEGIN_1IC0(push, NVE4_CP(UPLOAD_EXEC), 9);
PUSH_DATA (push, NVE4_COMPUTE_UPLOAD_EXEC_LINEAR | (0x20 << 1));
+ PUSH_DATAp(push, &tic->tic[0], 8);
+
+ BEGIN_NIC0(push, NVE4_CP(TIC_FLUSH), 1);
+ PUSH_DATA (push, (tic->id << 4) | 1);
+ } else
+ if (res->status & NOUVEAU_BUFFER_STATUS_GPU_WRITING) {
+ BEGIN_NIC0(push, NVE4_CP(TEX_CACHE_CTL), 1);
+ PUSH_DATA (push, (tic->id << 4) | 1);
+ }
+ nvc0->screen->tic.lock[tic->id / 32] |= 1 << (tic->id % 32);
- nve4_set_surface_info(push, nvc0->surfaces[t][i], screen);
+ res->status &= ~NOUVEAU_BUFFER_STATUS_GPU_WRITING;
+ res->status |= NOUVEAU_BUFFER_STATUS_GPU_READING;
- sf = nv50_surface(nvc0->surfaces[t][i]);
- if (sf) {
- res = nv04_resource(sf->base.texture);
+ BCTX_REFN(nvc0->bufctx_cp, CP_SUF, res, RD);
- if (sf->base.writable)
- BCTX_REFN(nvc0->bufctx_cp, CP_SUF, res, RDWR);
- else
- BCTX_REFN(nvc0->bufctx_cp, CP_SUF, res, RD);
- }
- }
- if (nvc0->surfaces_dirty[t]) {
- BEGIN_NVC0(push, NVE4_CP(FLUSH), 1);
- PUSH_DATA (push, NVE4_COMPUTE_FLUSH_CB);
- }
+ address = screen->uniform_bo->offset + NVC0_CB_AUX_INFO(s);
+
+ /* upload the texture handle */
+ BEGIN_NVC0(push, NVE4_CP(UPLOAD_DST_ADDRESS_HIGH), 2);
+ PUSH_DATAh(push, address + NVC0_CB_AUX_TEX_INFO(slot + 32));
+ PUSH_DATA (push, address + NVC0_CB_AUX_TEX_INFO(slot + 32));
+ BEGIN_NVC0(push, NVE4_CP(UPLOAD_LINE_LENGTH_IN), 2);
+ PUSH_DATA (push, 4);
+ PUSH_DATA (push, 0x1);
+ BEGIN_1IC0(push, NVE4_CP(UPLOAD_EXEC), 2);
+ PUSH_DATA (push, NVE4_COMPUTE_UPLOAD_EXEC_LINEAR | (0x20 << 1));
+ PUSH_DATA (push, tic->id);
+
+ BEGIN_NVC0(push, NVE4_CP(FLUSH), 1);
+ PUSH_DATA (push, NVE4_COMPUTE_FLUSH_CB);
+}
+
+static void
+nve4_compute_validate_surfaces(struct nvc0_context *nvc0)
+{
+ struct nouveau_pushbuf *push = nvc0->base.pushbuf;
+ uint64_t address;
+ const int s = 5;
+ int i, j;
+
+ if (!nvc0->images_dirty[s])
+ return;
+
+ address = nvc0->screen->uniform_bo->offset + NVC0_CB_AUX_INFO(s);
+
+ for (i = 0; i < NVC0_MAX_IMAGES; ++i) {
+ struct pipe_image_view *view = &nvc0->images[s][i];
+
+ BEGIN_NVC0(push, NVE4_CP(UPLOAD_DST_ADDRESS_HIGH), 2);
+ PUSH_DATAh(push, address + NVC0_CB_AUX_SU_INFO(i));
+ PUSH_DATA (push, address + NVC0_CB_AUX_SU_INFO(i));
+ BEGIN_NVC0(push, NVE4_CP(UPLOAD_LINE_LENGTH_IN), 2);
+ PUSH_DATA (push, 16 * 4);
+ PUSH_DATA (push, 0x1);
+ BEGIN_1IC0(push, NVE4_CP(UPLOAD_EXEC), 1 + 16);
+ PUSH_DATA (push, NVE4_COMPUTE_UPLOAD_EXEC_LINEAR | (0x20 << 1));
- /* re-reference non-dirty surfaces */
- mask = nvc0->surfaces_valid[t] & ~nvc0->surfaces_dirty[t];
- while (mask) {
- i = ffs(mask) - 1;
- mask &= ~(1 << i);
+ if (view->resource) {
+ struct nv04_resource *res = nv04_resource(view->resource);
- sf = nv50_surface(nvc0->surfaces[t][i]);
- res = nv04_resource(sf->base.texture);
+ if (res->base.target == PIPE_BUFFER) {
+ if (view->access & PIPE_IMAGE_ACCESS_WRITE)
+ nvc0_mark_image_range_valid(view);
+ }
- if (sf->base.writable)
+ nve4_set_surface_info(push, view, nvc0);
BCTX_REFN(nvc0->bufctx_cp, CP_SUF, res, RDWR);
- else
- BCTX_REFN(nvc0->bufctx_cp, CP_SUF, res, RD);
- }
- nvc0->surfaces_dirty[t] = 0;
+ if (nvc0->screen->base.class_3d >= GM107_3D_CLASS)
+ gm107_compute_validate_surfaces(nvc0, view, i);
+ } else {
+ for (j = 0; j < 16; j++)
+ PUSH_DATA(push, 0);
+ }
+ }
}
-
/* Thankfully, textures with samplers follow the normal rules. */
static void
nve4_compute_validate_samplers(struct nvc0_context *nvc0)
BEGIN_NVC0(nvc0->base.pushbuf, NVE4_CP(TSC_FLUSH), 1);
PUSH_DATA (nvc0->base.pushbuf, 0);
}
+
+ /* Invalidate all 3D samplers because they are aliased. */
+ for (int s = 0; s < 5; s++)
+ nvc0->samplers_dirty[s] = ~0;
+ nvc0->dirty_3d |= NVC0_NEW_3D_SAMPLERS;
}
+
/* (Code duplicated at bottom for various non-convincing reasons.
* E.g. we might want to use the COMPUTE subchannel to upload TIC/TSC
* entries to avoid a subchannel switch.
uint64_t address
= nvc0->screen->uniform_bo->offset + NVC0_CB_AUX_INFO(s);
- assert(i > 0); /* we really only want uniform buffer objects */
-
- BEGIN_NVC0(push, NVE4_CP(UPLOAD_DST_ADDRESS_HIGH), 2);
- PUSH_DATAh(push, address + NVC0_CB_AUX_UBO_INFO(i - 1));
- PUSH_DATA (push, address + NVC0_CB_AUX_UBO_INFO(i - 1));
- BEGIN_NVC0(push, NVE4_CP(UPLOAD_LINE_LENGTH_IN), 2);
- PUSH_DATA (push, 4 * 4);
- PUSH_DATA (push, 0x1);
- BEGIN_1IC0(push, NVE4_CP(UPLOAD_EXEC), 1 + 4);
- PUSH_DATA (push, NVE4_COMPUTE_UPLOAD_EXEC_LINEAR | (0x20 << 1));
-
- PUSH_DATA (push, res->address + nvc0->constbuf[s][i].offset);
- PUSH_DATAh(push, res->address + nvc0->constbuf[s][i].offset);
- PUSH_DATA (push, nvc0->constbuf[5][i].size);
- PUSH_DATA (push, 0);
- BCTX_REFN(nvc0->bufctx_cp, CP_CB(i), res, RD);
+ /* constbufs above 0 will are fetched via ubo info in the shader */
+ if (i > 0) {
+ BEGIN_NVC0(push, NVE4_CP(UPLOAD_DST_ADDRESS_HIGH), 2);
+ PUSH_DATAh(push, address + NVC0_CB_AUX_UBO_INFO(i - 1));
+ PUSH_DATA (push, address + NVC0_CB_AUX_UBO_INFO(i - 1));
+ BEGIN_NVC0(push, NVE4_CP(UPLOAD_LINE_LENGTH_IN), 2);
+ PUSH_DATA (push, 4 * 4);
+ PUSH_DATA (push, 0x1);
+ BEGIN_1IC0(push, NVE4_CP(UPLOAD_EXEC), 1 + 4);
+ PUSH_DATA (push, NVE4_COMPUTE_UPLOAD_EXEC_LINEAR | (0x20 << 1));
+
+ PUSH_DATA (push, res->address + nvc0->constbuf[s][i].offset);
+ PUSH_DATAh(push, res->address + nvc0->constbuf[s][i].offset);
+ PUSH_DATA (push, nvc0->constbuf[s][i].size);
+ PUSH_DATA (push, 0);
+ }
+ BCTX_REFN(nvc0->bufctx_cp, CP_CB(i), res, RD);
res->cb_bindings[s] |= 1 << i;
}
}
PUSH_DATA (push, nvc0->buffers[s][i].buffer_size);
PUSH_DATA (push, 0);
BCTX_REFN(nvc0->bufctx_cp, CP_BUF, res, RDWR);
+ util_range_add(&res->valid_buffer_range,
+ nvc0->buffers[s][i].buffer_offset,
+ nvc0->buffers[s][i].buffer_offset +
+ nvc0->buffers[s][i].buffer_size);
} else {
PUSH_DATA (push, 0);
PUSH_DATA (push, 0);
static void
nve4_compute_upload_input(struct nvc0_context *nvc0,
- struct nve4_cp_launch_desc *desc,
const struct pipe_grid_info *info)
{
struct nvc0_screen *screen = nvc0->screen;
if (cp->parm_size) {
BEGIN_NVC0(push, NVE4_CP(UPLOAD_DST_ADDRESS_HIGH), 2);
- PUSH_DATAh(push, screen->parm->offset);
- PUSH_DATA (push, screen->parm->offset);
+ PUSH_DATAh(push, screen->uniform_bo->offset + NVC0_CB_USR_INFO(5));
+ PUSH_DATA (push, screen->uniform_bo->offset + NVC0_CB_USR_INFO(5));
BEGIN_NVC0(push, NVE4_CP(UPLOAD_LINE_LENGTH_IN), 2);
PUSH_DATA (push, cp->parm_size);
PUSH_DATA (push, 0x1);
BEGIN_1IC0(push, NVE4_CP(UPLOAD_EXEC), 1 + (cp->parm_size / 4));
PUSH_DATA (push, NVE4_COMPUTE_UPLOAD_EXEC_LINEAR | (0x20 << 1));
PUSH_DATAp(push, info->input, cp->parm_size / 4);
-
- /* Bind user parameters coming from clover. */
- /* TODO: This should be harmonized with uniform_bo. */
- assert(!(desc->cb_mask & (1 << 0)));
- nve4_cp_launch_desc_set_cb(desc, 0, screen->parm, 0, 1 << 12);
}
BEGIN_NVC0(push, NVE4_CP(UPLOAD_DST_ADDRESS_HIGH), 2);
- PUSH_DATAh(push, address + NVC0_CB_AUX_GRID_INFO);
- PUSH_DATA (push, address + NVC0_CB_AUX_GRID_INFO);
+ PUSH_DATAh(push, address + NVC0_CB_AUX_GRID_INFO(0));
+ PUSH_DATA (push, address + NVC0_CB_AUX_GRID_INFO(0));
BEGIN_NVC0(push, NVE4_CP(UPLOAD_LINE_LENGTH_IN), 2);
- PUSH_DATA (push, 7 * 4);
+ PUSH_DATA (push, 8 * 4);
PUSH_DATA (push, 0x1);
if (unlikely(info->indirect)) {
struct nv04_resource *res = nv04_resource(info->indirect);
uint32_t offset = res->offset + info->indirect_offset;
- nouveau_pushbuf_space(push, 16, 0, 1);
+ nouveau_pushbuf_space(push, 32, 0, 1);
PUSH_REFN(push, res->bo, NOUVEAU_BO_RD | res->domain);
- BEGIN_1IC0(push, NVE4_CP(UPLOAD_EXEC), 1 + 7);
+ BEGIN_1IC0(push, NVE4_CP(UPLOAD_EXEC), 1 + 8);
PUSH_DATA (push, NVE4_COMPUTE_UPLOAD_EXEC_LINEAR | (0x20 << 1));
PUSH_DATAp(push, info->block, 3);
nouveau_pushbuf_data(push, res->bo, offset,
NVC0_IB_ENTRY_1_NO_PREFETCH | 3 * 4);
} else {
- BEGIN_1IC0(push, NVE4_CP(UPLOAD_EXEC), 1 + 7);
+ BEGIN_1IC0(push, NVE4_CP(UPLOAD_EXEC), 1 + 8);
PUSH_DATA (push, NVE4_COMPUTE_UPLOAD_EXEC_LINEAR | (0x20 << 1));
PUSH_DATAp(push, info->block, 3);
PUSH_DATAp(push, info->grid, 3);
}
PUSH_DATA (push, 0);
+ PUSH_DATA (push, info->work_dim);
BEGIN_NVC0(push, NVE4_CP(FLUSH), 1);
PUSH_DATA (push, NVE4_COMPUTE_FLUSH_CB);
return NVC1_3D_CACHE_SPLIT_16K_SHARED_48K_L1;
}
+static void
+nve4_compute_setup_buf_cb(struct nvc0_context *nvc0, bool gp100, void *desc)
+{
+ // only user constant buffers 0-6 can be put in the descriptor, the rest are
+ // loaded through global memory
+ for (int i = 0; i <= 6; i++) {
+ if (nvc0->constbuf[5][i].user || !nvc0->constbuf[5][i].u.buf)
+ continue;
+
+ struct nv04_resource *res =
+ nv04_resource(nvc0->constbuf[5][i].u.buf);
+
+ uint32_t base = res->offset + nvc0->constbuf[5][i].offset;
+ uint32_t size = nvc0->constbuf[5][i].size;
+ if (gp100)
+ gp100_cp_launch_desc_set_cb(desc, i, res->bo, base, size);
+ else
+ nve4_cp_launch_desc_set_cb(desc, i, res->bo, base, size);
+ }
+
+ // there is no need to do FLUSH(NVE4_COMPUTE_FLUSH_CB) because
+ // nve4_compute_upload_input() does it later
+}
+
static void
nve4_compute_setup_launch_desc(struct nvc0_context *nvc0,
struct nve4_cp_launch_desc *desc,
desc->blockdim_z = info->block[2];
desc->shared_size = align(cp->cp.smem_size, 0x100);
- desc->local_size_p = align(cp->cp.lmem_size, 0x10);
+ desc->local_size_p = (cp->hdr[1] & 0xfffff0) + align(cp->cp.lmem_size, 0x10);
desc->local_size_n = 0;
desc->cstack_size = 0x800;
desc->cache_split = nve4_compute_derive_cache_split(nvc0, cp->cp.smem_size);
desc->gpr_alloc = cp->num_gprs;
desc->bar_alloc = cp->num_barriers;
- // Only bind OpenGL uniforms and the driver constant buffer through the
+ // 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) {
+ if (nvc0->constbuf[5][0].user || cp->parm_size) {
nve4_cp_launch_desc_set_cb(desc, 0, screen->uniform_bo,
NVC0_CB_USR_INFO(5), 1 << 16);
+
+ // Later logic will attempt to bind a real buffer at position 0. That
+ // should not happen if we've bound a user buffer.
+ assert(nvc0->constbuf[5][0].user || !nvc0->constbuf[5][0].u.buf);
}
nve4_cp_launch_desc_set_cb(desc, 7, screen->uniform_bo,
- NVC0_CB_AUX_INFO(5), 1 << 10);
+ NVC0_CB_AUX_INFO(5), 1 << 11);
+
+ nve4_compute_setup_buf_cb(nvc0, false, desc);
}
-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);
+
+ // Later logic will attempt to bind a real buffer at position 0. That
+ // should not happen if we've bound a user buffer.
+ assert(nvc0->constbuf[5][0].user || !nvc0->constbuf[5][0].u.buf);
+ }
+ gp100_cp_launch_desc_set_cb(desc, 7, screen->uniform_bo,
+ NVC0_CB_AUX_INFO(5), 1 << 11);
+
+ nve4_compute_setup_buf_cb(nvc0, true, desc);
+}
+
+static inline void *
nve4_compute_alloc_launch_desc(struct nouveau_context *nv,
struct nouveau_bo **pbo, uint64_t *pgpuaddr)
{
ptr += adj;
*pgpuaddr += adj;
}
- return (struct nve4_cp_launch_desc *)ptr;
+ return ptr;
+}
+
+static void
+nve4_upload_indirect_desc(struct nouveau_pushbuf *push,
+ struct nv04_resource *res, uint64_t gpuaddr,
+ uint32_t length, uint32_t bo_offset)
+{
+ BEGIN_NVC0(push, NVE4_CP(UPLOAD_DST_ADDRESS_HIGH), 2);
+ PUSH_DATAh(push, gpuaddr);
+ PUSH_DATA (push, gpuaddr);
+ BEGIN_NVC0(push, NVE4_CP(UPLOAD_LINE_LENGTH_IN), 2);
+ PUSH_DATA (push, length);
+ PUSH_DATA (push, 1);
+
+ nouveau_pushbuf_space(push, 32, 0, 1);
+ PUSH_REFN(push, res->bo, NOUVEAU_BO_RD | res->domain);
+
+ BEGIN_1IC0(push, NVE4_CP(UPLOAD_EXEC), 1 + (length / 4));
+ PUSH_DATA (push, NVE4_COMPUTE_UPLOAD_EXEC_LINEAR | (0x08 << 1));
+ nouveau_pushbuf_data(push, res->bo, bo_offset,
+ NVC0_IB_ENTRY_1_NO_PREFETCH | length);
}
void
nve4_launch_grid(struct pipe_context *pipe, const struct pipe_grid_info *info)
{
struct nvc0_context *nvc0 = nvc0_context(pipe);
+ struct nvc0_screen *screen = nvc0->screen;
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;
BCTX_REFN_bo(nvc0->bufctx_cp, CP_DESC, NOUVEAU_BO_GART | NOUVEAU_BO_RD,
desc_bo);
+ list_for_each_entry(struct nvc0_resident, resident, &nvc0->tex_head, list) {
+ nvc0_add_resident(nvc0->bufctx_cp, NVC0_BIND_CP_BINDLESS, resident->buf,
+ resident->flags);
+ }
+
+ list_for_each_entry(struct nvc0_resident, resident, &nvc0->img_head, list) {
+ nvc0_add_resident(nvc0->bufctx_cp, NVC0_BIND_CP_BINDLESS, resident->buf,
+ resident->flags);
+ }
+
ret = !nve4_state_validate_cp(nvc0, ~0);
if (ret)
goto out;
- nve4_compute_setup_launch_desc(nvc0, desc, info);
+ if (nvc0->screen->compute->oclass >= GP100_COMPUTE_CLASS)
+ gp100_compute_setup_launch_desc(nvc0, desc, info);
+ else
+ nve4_compute_setup_launch_desc(nvc0, desc, info);
- nve4_compute_upload_input(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->compute->oclass >= GP100_COMPUTE_CLASS)
+ gp100_compute_dump_launch_desc(desc);
+ else
+ nve4_compute_dump_launch_desc(desc);
+ }
#endif
if (unlikely(info->indirect)) {
PUSH_DATA (push, NVE4_COMPUTE_UPLOAD_EXEC_LINEAR | (0x08 << 1));
PUSH_DATAp(push, (const uint32_t *)desc, 256 / 4);
- /* overwrite griddim_x and griddim_y as two 32-bits integers even
- * if griddim_y must be a 16-bits integer */
- BEGIN_NVC0(push, NVE4_CP(UPLOAD_DST_ADDRESS_HIGH), 2);
- PUSH_DATAh(push, desc_gpuaddr + 48);
- PUSH_DATA (push, desc_gpuaddr + 48);
- BEGIN_NVC0(push, NVE4_CP(UPLOAD_LINE_LENGTH_IN), 2);
- PUSH_DATA (push, 8);
- PUSH_DATA (push, 1);
-
- nouveau_pushbuf_space(push, 16, 0, 1);
- PUSH_REFN(push, res->bo, NOUVEAU_BO_RD | res->domain);
-
- BEGIN_1IC0(push, NVE4_CP(UPLOAD_EXEC), 1 + (8 / 4));
- PUSH_DATA (push, NVE4_COMPUTE_UPLOAD_EXEC_LINEAR | (0x08 << 1));
- nouveau_pushbuf_data(push, res->bo, offset,
- NVC0_IB_ENTRY_1_NO_PREFETCH | 2 * 4);
+ if (nvc0->screen->compute->oclass >= GP100_COMPUTE_CLASS) {
+ nve4_upload_indirect_desc(push, res, desc_gpuaddr + 48, 12, offset);
+ } else {
+ /* overwrite griddim_x and griddim_y as two 32-bits integers even
+ * if griddim_y must be a 16-bits integer */
+ nve4_upload_indirect_desc(push, res, desc_gpuaddr + 48, 8, offset);
- /* overwrite the 16 high bits of griddim_y with griddim_z because
- * we need (z << 16) | x */
- BEGIN_NVC0(push, NVE4_CP(UPLOAD_DST_ADDRESS_HIGH), 2);
- PUSH_DATAh(push, desc_gpuaddr + 54);
- PUSH_DATA (push, desc_gpuaddr + 54);
- BEGIN_NVC0(push, NVE4_CP(UPLOAD_LINE_LENGTH_IN), 2);
- PUSH_DATA (push, 4);
- PUSH_DATA (push, 1);
- BEGIN_1IC0(push, NVE4_CP(UPLOAD_EXEC), 1 + (4 / 4));
- PUSH_DATA (push, NVE4_COMPUTE_UPLOAD_EXEC_LINEAR | (0x08 << 1));
- nouveau_pushbuf_data(push, res->bo, offset + 8,
- NVC0_IB_ENTRY_1_NO_PREFETCH | 1 * 4);
+ /* overwrite the 16 high bits of griddim_y with griddim_z because
+ * we need (z << 16) | x */
+ nve4_upload_indirect_desc(push, res, desc_gpuaddr + 54, 4, offset + 8);
+ }
}
/* upload descriptor and flush */
+ nouveau_pushbuf_space(push, 32, 1, 0);
+ PUSH_REFN(push, screen->text, NV_VRAM_DOMAIN(&screen->base) | NOUVEAU_BO_RD);
BEGIN_NVC0(push, NVE4_CP(LAUNCH_DESC_ADDRESS), 1);
PUSH_DATA (push, desc_gpuaddr >> 8);
BEGIN_NVC0(push, NVE4_CP(LAUNCH), 1);
BEGIN_NVC0(push, SUBC_CP(NV50_GRAPH_SERIALIZE), 1);
PUSH_DATA (push, 0);
+ nvc0_update_compute_invocations_counter(nvc0, info);
+
out:
if (ret)
NOUVEAU_ERR("Failed to launch grid !\n");
nouveau_scratch_done(&nvc0->base);
nouveau_bufctx_reset(nvc0->bufctx_cp, NVC0_BIND_CP_DESC);
+ nouveau_bufctx_reset(nvc0->bufctx_cp, NVC0_BIND_CP_BINDLESS);
}
continue;
}
res = nv04_resource(tic->pipe.texture);
+ nvc0_update_tic(nvc0, tic, res);
if (tic->id < 0) {
tic->id = nvc0_screen_tic_alloc(nvc0->screen, tic);
if (dirty)
BCTX_REFN(nvc0->bufctx_cp, CP_TEX(i), res, RD);
}
- for (; i < nvc0->state.num_textures[s]; ++i)
+ for (; i < nvc0->state.num_textures[s]; ++i) {
nvc0->tex_handles[s][i] |= NVE4_TIC_ENTRY_INVALID;
+ nvc0->textures_dirty[s] |= 1 << i;
+ }
if (n[0]) {
BEGIN_NIC0(push, NVE4_CP(TIC_FLUSH), n[0]);
}
nvc0->state.num_textures[s] = nvc0->num_textures[s];
+
+ /* Invalidate all 3D textures because they are aliased. */
+ for (int s = 0; s < 5; s++) {
+ for (int i = 0; i < nvc0->num_textures[s]; i++)
+ nouveau_bufctx_reset(nvc0->bufctx_3d, NVC0_BIND_3D_TEX(s, i));
+ nvc0->textures_dirty[s] = ~0;
+ }
+ nvc0->dirty_3d |= NVC0_NEW_3D_TEXTURES;
}
debug_printf("barrier count: %u\n", desc->bar_alloc);
debug_printf("$r count: %u\n", desc->gpr_alloc);
debug_printf("cache split: %s\n", nve4_cache_split_name(desc->cache_split));
+ debug_printf("linked tsc: %d\n", desc->linked_tsc);
for (i = 0; i < 8; ++i) {
uint64_t address;
i, address, size, valid ? "" : " (invalid)");
}
}
+
+static void
+gp100_compute_dump_launch_desc(const struct gp100_cp_launch_desc *desc)
+{
+ const uint32_t *data = (const uint32_t *)desc;
+ unsigned i;
+ bool zero = false;
+
+ debug_printf("COMPUTE LAUNCH DESCRIPTOR:\n");
+
+ for (i = 0; i < sizeof(*desc); i += 4) {
+ if (data[i / 4]) {
+ debug_printf("[%x]: 0x%08x\n", i, data[i / 4]);
+ zero = false;
+ } else
+ if (!zero) {
+ debug_printf("...\n");
+ zero = true;
+ }
+ }
+
+ debug_printf("entry = 0x%x\n", desc->entry);
+ debug_printf("grid dimensions = %ux%ux%u\n",
+ desc->griddim_x, desc->griddim_y, desc->griddim_z);
+ debug_printf("block dimensions = %ux%ux%u\n",
+ desc->blockdim_x, desc->blockdim_y, desc->blockdim_z);
+ debug_printf("s[] size: 0x%x\n", desc->shared_size);
+ debug_printf("l[] size: -0x%x / +0x%x\n",
+ desc->local_size_n, desc->local_size_p);
+ debug_printf("stack size: 0x%x\n", desc->cstack_size);
+ debug_printf("barrier count: %u\n", desc->bar_alloc);
+ debug_printf("$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
projects / mesa.git / blobdiff