*/
#include "nvc0/nvc0_context.h"
-#include "nvc0/nvc0_compute.h"
+
+#include "nvc0/nvc0_compute.xml.h"
int
nvc0_screen_compute_setup(struct nvc0_screen *screen,
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);
PUSH_DATA (push, screen->txc->offset + 65536);
PUSH_DATA (push, NVC0_TSC_MAX_ENTRIES - 1);
- return 0;
-}
-
-bool
-nvc0_compute_validate_program(struct nvc0_context *nvc0)
-{
- struct nvc0_program *prog = nvc0->compprog;
-
- if (prog->mem)
- return true;
+ /* MS sample coordinate offsets */
+ BEGIN_NVC0(push, NVC0_CP(CB_SIZE), 3);
+ PUSH_DATA (push, NVC0_CB_AUX_SIZE);
+ PUSH_DATAh(push, screen->uniform_bo->offset + NVC0_CB_AUX_INFO(5));
+ PUSH_DATA (push, screen->uniform_bo->offset + NVC0_CB_AUX_INFO(5));
+ BEGIN_1IC0(push, NVC0_CP(CB_POS), 1 + 2 * 8);
+ PUSH_DATA (push, NVC0_CB_AUX_MS_INFO);
+ PUSH_DATA (push, 0); /* 0 */
+ PUSH_DATA (push, 0);
+ PUSH_DATA (push, 1); /* 1 */
+ PUSH_DATA (push, 0);
+ PUSH_DATA (push, 0); /* 2 */
+ PUSH_DATA (push, 1);
+ PUSH_DATA (push, 1); /* 3 */
+ PUSH_DATA (push, 1);
+ PUSH_DATA (push, 2); /* 4 */
+ PUSH_DATA (push, 0);
+ PUSH_DATA (push, 3); /* 5 */
+ PUSH_DATA (push, 0);
+ PUSH_DATA (push, 2); /* 6 */
+ PUSH_DATA (push, 1);
+ PUSH_DATA (push, 3); /* 7 */
+ PUSH_DATA (push, 1);
- if (!prog->translated) {
- prog->translated = nvc0_program_translate(
- prog, nvc0->screen->base.device->chipset, &nvc0->base.debug);
- if (!prog->translated)
- return false;
- }
- if (unlikely(!prog->code_size))
- return false;
-
- if (likely(prog->code_size)) {
- if (nvc0_program_upload_code(nvc0, prog)) {
- struct nouveau_pushbuf *push = nvc0->base.pushbuf;
- BEGIN_NVC0(push, NVC0_CP(FLUSH), 1);
- PUSH_DATA (push, NVC0_COMPUTE_FLUSH_CODE);
- return true;
- }
- }
- return false;
+ return 0;
}
static void
BEGIN_NVC0(nvc0->base.pushbuf, NVC0_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;
}
static void
BEGIN_NVC0(nvc0->base.pushbuf, NVC0_CP(TIC_FLUSH), 1);
PUSH_DATA (nvc0->base.pushbuf, 0);
}
+
+ /* 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;
+}
+
+static inline void
+nvc0_compute_invalidate_constbufs(struct nvc0_context *nvc0)
+{
+ int s;
+
+ /* Invalidate all 3D constbufs because they are aliased with COMPUTE. */
+ for (s = 0; s < 5; s++) {
+ nvc0->constbuf_dirty[s] |= nvc0->constbuf_valid[s];
+ nvc0->state.uniform_buffer_bound[s] = 0;
+ }
+ nvc0->dirty_3d |= NVC0_NEW_3D_CONSTBUF;
}
static void
if (nvc0->constbuf[s][i].user) {
struct nouveau_bo *bo = nvc0->screen->uniform_bo;
- const unsigned base = s << 16;
+ const unsigned base = NVC0_CB_USR_INFO(s);
const unsigned size = nvc0->constbuf[s][0].size;
assert(i == 0); /* we really only want OpenGL uniforms here */
assert(nvc0->constbuf[s][0].u.data);
}
}
+ nvc0_compute_invalidate_constbufs(nvc0);
+
BEGIN_NVC0(push, NVC0_CP(FLUSH), 1);
PUSH_DATA (push, NVC0_COMPUTE_FLUSH_CB);
}
struct nvc0_screen *screen = nvc0->screen;
BEGIN_NVC0(push, NVC0_CP(CB_SIZE), 3);
- PUSH_DATA (push, 1024);
- PUSH_DATAh(push, screen->uniform_bo->offset + (6 << 16) + (5 << 10));
- PUSH_DATA (push, screen->uniform_bo->offset + (6 << 16) + (5 << 10));
+ PUSH_DATA (push, NVC0_CB_AUX_SIZE);
+ PUSH_DATAh(push, screen->uniform_bo->offset + NVC0_CB_AUX_INFO(5));
+ PUSH_DATA (push, screen->uniform_bo->offset + NVC0_CB_AUX_INFO(5));
BEGIN_NVC0(push, NVC0_CP(CB_BIND), 1);
PUSH_DATA (push, (15 << 8) | 1);
nvc0_compute_validate_buffers(struct nvc0_context *nvc0)
{
struct nouveau_pushbuf *push = nvc0->base.pushbuf;
+ struct nvc0_screen *screen = nvc0->screen;
const int s = 5;
int i;
BEGIN_NVC0(push, NVC0_CP(CB_SIZE), 3);
- PUSH_DATA (push, 1024);
- PUSH_DATAh(push, nvc0->screen->uniform_bo->offset + (6 << 16) + (s << 10));
- PUSH_DATA (push, nvc0->screen->uniform_bo->offset + (6 << 16) + (s << 10));
+ PUSH_DATA (push, NVC0_CB_AUX_SIZE);
+ PUSH_DATAh(push, screen->uniform_bo->offset + NVC0_CB_AUX_INFO(s));
+ PUSH_DATA (push, screen->uniform_bo->offset + NVC0_CB_AUX_INFO(s));
BEGIN_1IC0(push, NVC0_CP(CB_POS), 1 + 4 * NVC0_MAX_BUFFERS);
- PUSH_DATA (push, 512);
+ PUSH_DATA (push, NVC0_CB_AUX_BUF_INFO(0));
for (i = 0; i < NVC0_MAX_BUFFERS; i++) {
if (nvc0->buffers[s][i].buffer) {
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);
}
}
+void
+nvc0_compute_validate_globals(struct nvc0_context *nvc0)
+{
+ unsigned i;
+
+ for (i = 0; i < nvc0->global_residents.size / sizeof(struct pipe_resource *);
+ ++i) {
+ struct pipe_resource *res = *util_dynarray_element(
+ &nvc0->global_residents, struct pipe_resource *, i);
+ if (res)
+ nvc0_add_resident(nvc0->bufctx_cp, NVC0_BIND_CP_GLOBAL,
+ nv04_resource(res), NOUVEAU_BO_RDWR);
+ }
+}
+
+static inline void
+nvc0_compute_invalidate_surfaces(struct nvc0_context *nvc0, const int s)
+{
+ struct nouveau_pushbuf *push = nvc0->base.pushbuf;
+ int i;
+
+ for (i = 0; i < NVC0_MAX_IMAGES; ++i) {
+ if (s == 5)
+ BEGIN_NVC0(push, NVC0_CP(IMAGE(i)), 6);
+ else
+ BEGIN_NVC0(push, NVC0_3D(IMAGE(i)), 6);
+ PUSH_DATA(push, 0);
+ PUSH_DATA(push, 0);
+ PUSH_DATA(push, 0);
+ PUSH_DATA(push, 0);
+ PUSH_DATA(push, 0x14000);
+ PUSH_DATA(push, 0);
+ }
+}
+
+static void
+nvc0_compute_validate_surfaces(struct nvc0_context *nvc0)
+{
+ /* TODO: Invalidating both 3D and CP surfaces before validating surfaces for
+ * compute is probably not really necessary, but we didn't find any better
+ * solutions for now. This fixes some invalidation issues when compute and
+ * fragment shaders are used inside the same context. Anyway, we definitely
+ * have invalidation issues between 3D and CP for other resources like SSBO
+ * and atomic counters. */
+ nvc0_compute_invalidate_surfaces(nvc0, 4);
+ nvc0_compute_invalidate_surfaces(nvc0, 5);
+
+ nvc0_validate_suf(nvc0, 5);
+
+ /* Invalidate all FRAGMENT images because they are aliased with COMPUTE. */
+ nouveau_bufctx_reset(nvc0->bufctx_3d, NVC0_BIND_3D_SUF);
+ nvc0->dirty_3d |= NVC0_NEW_3D_SURFACES;
+ nvc0->images_dirty[4] |= nvc0->images_valid[4];
+}
+
+static struct nvc0_state_validate
+validate_list_cp[] = {
+ { nvc0_compprog_validate, NVC0_NEW_CP_PROGRAM },
+ { nvc0_compute_validate_constbufs, NVC0_NEW_CP_CONSTBUF },
+ { nvc0_compute_validate_driverconst, NVC0_NEW_CP_DRIVERCONST },
+ { nvc0_compute_validate_buffers, NVC0_NEW_CP_BUFFERS },
+ { nvc0_compute_validate_textures, NVC0_NEW_CP_TEXTURES },
+ { nvc0_compute_validate_samplers, NVC0_NEW_CP_SAMPLERS },
+ { nvc0_compute_validate_globals, NVC0_NEW_CP_GLOBALS },
+ { nvc0_compute_validate_surfaces, NVC0_NEW_CP_SURFACES },
+};
+
static bool
-nvc0_compute_state_validate(struct nvc0_context *nvc0)
+nvc0_state_validate_cp(struct nvc0_context *nvc0, uint32_t mask)
{
- if (!nvc0_compute_validate_program(nvc0))
- return false;
- if (nvc0->dirty_cp & NVC0_NEW_CP_CONSTBUF)
- nvc0_compute_validate_constbufs(nvc0);
- if (nvc0->dirty_cp & NVC0_NEW_CP_DRIVERCONST)
- nvc0_compute_validate_driverconst(nvc0);
- if (nvc0->dirty_cp & NVC0_NEW_CP_BUFFERS)
- nvc0_compute_validate_buffers(nvc0);
- if (nvc0->dirty_cp & NVC0_NEW_CP_TEXTURES)
- nvc0_compute_validate_textures(nvc0);
- if (nvc0->dirty_cp & NVC0_NEW_CP_SAMPLERS)
- nvc0_compute_validate_samplers(nvc0);
-
- /* TODO: surfaces, global memory buffers */
-
- nvc0_bufctx_fence(nvc0, nvc0->bufctx_cp, false);
-
- nouveau_pushbuf_bufctx(nvc0->base.pushbuf, nvc0->bufctx_cp);
- if (unlikely(nouveau_pushbuf_validate(nvc0->base.pushbuf)))
- return false;
- if (unlikely(nvc0->state.flushed))
- nvc0_bufctx_fence(nvc0, nvc0->bufctx_cp, true);
+ bool ret;
- return true;
+ ret = nvc0_state_validate(nvc0, mask, validate_list_cp,
+ ARRAY_SIZE(validate_list_cp), &nvc0->dirty_cp,
+ nvc0->bufctx_cp);
+ if (unlikely(nvc0->state.flushed))
+ nvc0_bufctx_fence(nvc0, nvc0->bufctx_cp, true);
+ return ret;
}
static void
-nvc0_compute_upload_input(struct nvc0_context *nvc0, const void *input)
+nvc0_compute_upload_input(struct nvc0_context *nvc0,
+ const struct pipe_grid_info *info)
{
struct nouveau_pushbuf *push = nvc0->base.pushbuf;
struct nvc0_screen *screen = nvc0->screen;
struct nvc0_program *cp = nvc0->compprog;
if (cp->parm_size) {
+ struct nouveau_bo *bo = screen->uniform_bo;
+ const unsigned base = NVC0_CB_USR_INFO(5);
+
BEGIN_NVC0(push, NVC0_CP(CB_SIZE), 3);
PUSH_DATA (push, align(cp->parm_size, 0x100));
- PUSH_DATAh(push, screen->parm->offset);
- PUSH_DATA (push, screen->parm->offset);
+ PUSH_DATAh(push, bo->offset + base);
+ PUSH_DATA (push, bo->offset + base);
BEGIN_NVC0(push, NVC0_CP(CB_BIND), 1);
PUSH_DATA (push, (0 << 8) | 1);
/* NOTE: size is limited to 4 KiB, which is < NV04_PFIFO_MAX_PACKET_LEN */
BEGIN_1IC0(push, NVC0_CP(CB_POS), 1 + cp->parm_size / 4);
PUSH_DATA (push, 0);
- PUSH_DATAp(push, input, cp->parm_size / 4);
+ PUSH_DATAp(push, info->input, cp->parm_size / 4);
- BEGIN_NVC0(push, NVC0_CP(FLUSH), 1);
- PUSH_DATA (push, NVC0_COMPUTE_FLUSH_CB);
+ nvc0_compute_invalidate_constbufs(nvc0);
}
+
+ BEGIN_NVC0(push, NVC0_CP(CB_SIZE), 3);
+ PUSH_DATA (push, NVC0_CB_AUX_SIZE);
+ PUSH_DATAh(push, screen->uniform_bo->offset + NVC0_CB_AUX_INFO(5));
+ PUSH_DATA (push, screen->uniform_bo->offset + NVC0_CB_AUX_INFO(5));
+
+ BEGIN_1IC0(push, NVC0_CP(CB_POS), 1 + 1);
+ /* (7) as we only upload work_dim on nvc0, the rest uses special regs */
+ PUSH_DATA (push, NVC0_CB_AUX_GRID_INFO(7));
+ PUSH_DATA (push, info->work_dim);
+
+ BEGIN_NVC0(push, NVC0_CP(FLUSH), 1);
+ PUSH_DATA (push, NVC0_COMPUTE_FLUSH_CB);
}
void
struct nvc0_context *nvc0 = nvc0_context(pipe);
struct nouveau_pushbuf *push = nvc0->base.pushbuf;
struct nvc0_program *cp = nvc0->compprog;
- unsigned s;
int ret;
- ret = !nvc0_compute_state_validate(nvc0);
+ ret = !nvc0_state_validate_cp(nvc0, ~0);
if (ret) {
NOUVEAU_ERR("Failed to launch grid !\n");
return;
}
- nvc0_compute_upload_input(nvc0, info->input);
+ nvc0_compute_upload_input(nvc0, info);
BEGIN_NVC0(push, NVC0_CP(CP_START_ID), 1);
PUSH_DATA (push, nvc0_program_symbol_offset(cp, info->pc));
BEGIN_NVC0(push, NVC0_CP(LOCAL_POS_ALLOC), 3);
- PUSH_DATA (push, align(cp->cp.lmem_size, 0x10));
+ PUSH_DATA (push, (cp->hdr[1] & 0xfffff0) + align(cp->cp.lmem_size, 0x10));
PUSH_DATA (push, 0);
PUSH_DATA (push, 0x800); /* WARP_CSTACK_SIZE */
PUSH_DATA (push, 0x1);
}
- /* Invalidate all 3D constbufs because they are aliased with COMPUTE. */
- nvc0->dirty_3d |= NVC0_NEW_3D_CONSTBUF;
- for (s = 0; s < 5; s++) {
- nvc0->constbuf_dirty[s] |= nvc0->constbuf_valid[s];
- nvc0->state.uniform_buffer_bound[s] = 0;
- }
+ /* TODO: Not sure if this is really necessary. */
+ nvc0_compute_invalidate_surfaces(nvc0, 5);
+ nouveau_bufctx_reset(nvc0->bufctx_cp, NVC0_BIND_CP_SUF);
+ nvc0->dirty_cp |= NVC0_NEW_CP_SURFACES;
+ nvc0->images_dirty[5] |= nvc0->images_valid[5];
}
projects / mesa.git / blobdiff