summaryrefslogtreecommitdiff
path: root/src/gallium/drivers/nv50/nv50_state_validate.c
diff options
context:
space:
mode:
Diffstat (limited to 'src/gallium/drivers/nv50/nv50_state_validate.c')
-rw-r--r--src/gallium/drivers/nv50/nv50_state_validate.c168
1 files changed, 122 insertions, 46 deletions
diff --git a/src/gallium/drivers/nv50/nv50_state_validate.c b/src/gallium/drivers/nv50/nv50_state_validate.c
index fd27620371..f83232f43c 100644
--- a/src/gallium/drivers/nv50/nv50_state_validate.c
+++ b/src/gallium/drivers/nv50/nv50_state_validate.c
@@ -23,14 +23,29 @@
#include "nv50_context.h"
#include "nouveau/nouveau_stateobj.h"
+#define NV50_CBUF_FORMAT_CASE(n) \
+ case PIPE_FORMAT_##n: so_data(so, NV50TCL_RT_FORMAT_##n); break
+
+#define NV50_ZETA_FORMAT_CASE(n) \
+ case PIPE_FORMAT_##n: so_data(so, NV50TCL_ZETA_FORMAT_##n); break
+
static void
nv50_state_validate_fb(struct nv50_context *nv50)
{
struct nouveau_grobj *tesla = nv50->screen->tesla;
- struct nouveau_stateobj *so = so_new(128, 18);
+ struct nouveau_stateobj *so = so_new(32, 79, 18);
struct pipe_framebuffer_state *fb = &nv50->framebuffer;
unsigned i, w, h, gw = 0;
+ /* Set nr of active RTs and select RT for each colour output.
+ * FP result 0 always goes to RT[0], bits 4 - 6 are ignored.
+ * Ambiguous assignment results in no rendering (no DATA_ERROR).
+ */
+ so_method(so, tesla, NV50TCL_RT_CONTROL, 1);
+ so_data (so, fb->nr_cbufs |
+ (0 << 4) | (1 << 7) | (2 << 10) | (3 << 13) |
+ (4 << 16) | (5 << 19) | (6 << 22) | (7 << 25));
+
for (i = 0; i < fb->nr_cbufs; i++) {
struct pipe_texture *pt = fb->cbufs[i]->texture;
struct nouveau_bo *bo = nv50_miptree(pt)->base.bo;
@@ -54,15 +69,14 @@ nv50_state_validate_fb(struct nv50_context *nv50)
so_reloc (so, bo, fb->cbufs[i]->offset, NOUVEAU_BO_VRAM |
NOUVEAU_BO_LOW | NOUVEAU_BO_RDWR, 0, 0);
switch (fb->cbufs[i]->format) {
- case PIPE_FORMAT_A8R8G8B8_UNORM:
- so_data(so, NV50TCL_RT_FORMAT_A8R8G8B8_UNORM);
- break;
- case PIPE_FORMAT_X8R8G8B8_UNORM:
- so_data(so, NV50TCL_RT_FORMAT_X8R8G8B8_UNORM);
- break;
- case PIPE_FORMAT_R5G6B5_UNORM:
- so_data(so, NV50TCL_RT_FORMAT_R5G6B5_UNORM);
- break;
+ NV50_CBUF_FORMAT_CASE(A8R8G8B8_UNORM);
+ NV50_CBUF_FORMAT_CASE(X8R8G8B8_UNORM);
+ NV50_CBUF_FORMAT_CASE(R5G6B5_UNORM);
+ NV50_CBUF_FORMAT_CASE(R16G16B16A16_SNORM);
+ NV50_CBUF_FORMAT_CASE(R16G16B16A16_UNORM);
+ NV50_CBUF_FORMAT_CASE(R32G32B32A32_FLOAT);
+ NV50_CBUF_FORMAT_CASE(R16G16_SNORM);
+ NV50_CBUF_FORMAT_CASE(R16G16_UNORM);
default:
NOUVEAU_ERR("AIIII unknown format %s\n",
pf_name(fb->cbufs[i]->format));
@@ -73,7 +87,7 @@ nv50_state_validate_fb(struct nv50_context *nv50)
level[fb->cbufs[i]->level].tile_mode << 4);
so_data(so, 0x00000000);
- so_method(so, tesla, 0x1224, 1);
+ so_method(so, tesla, NV50TCL_RT_ARRAY_MODE, 1);
so_data (so, 1);
}
@@ -96,18 +110,10 @@ nv50_state_validate_fb(struct nv50_context *nv50)
so_reloc (so, bo, fb->zsbuf->offset, NOUVEAU_BO_VRAM |
NOUVEAU_BO_LOW | NOUVEAU_BO_RDWR, 0, 0);
switch (fb->zsbuf->format) {
- case PIPE_FORMAT_Z32_FLOAT:
- so_data(so, NV50TCL_ZETA_FORMAT_Z32_FLOAT);
- break;
- case PIPE_FORMAT_Z24S8_UNORM:
- so_data(so, NV50TCL_ZETA_FORMAT_Z24S8_UNORM);
- break;
- case PIPE_FORMAT_X8Z24_UNORM:
- so_data(so, NV50TCL_ZETA_FORMAT_X8Z24_UNORM);
- break;
- case PIPE_FORMAT_S8Z24_UNORM:
- so_data(so, NV50TCL_ZETA_FORMAT_S8Z24_UNORM);
- break;
+ NV50_ZETA_FORMAT_CASE(S8Z24_UNORM);
+ NV50_ZETA_FORMAT_CASE(X8Z24_UNORM);
+ NV50_ZETA_FORMAT_CASE(Z24S8_UNORM);
+ NV50_ZETA_FORMAT_CASE(Z32_FLOAT);
default:
NOUVEAU_ERR("AIIII unknown format %s\n",
pf_name(fb->zsbuf->format));
@@ -118,19 +124,22 @@ nv50_state_validate_fb(struct nv50_context *nv50)
level[fb->zsbuf->level].tile_mode << 4);
so_data(so, 0x00000000);
- so_method(so, tesla, 0x1538, 1);
+ so_method(so, tesla, NV50TCL_ZETA_ENABLE, 1);
so_data (so, 1);
so_method(so, tesla, NV50TCL_ZETA_HORIZ, 3);
so_data (so, fb->zsbuf->width);
so_data (so, fb->zsbuf->height);
so_data (so, 0x00010001);
+ } else {
+ so_method(so, tesla, NV50TCL_ZETA_ENABLE, 1);
+ so_data (so, 0);
}
- so_method(so, tesla, NV50TCL_VIEWPORT_HORIZ, 2);
+ so_method(so, tesla, NV50TCL_VIEWPORT_HORIZ(0), 2);
so_data (so, w << 16);
so_data (so, h << 16);
/* set window lower left corner */
- so_method(so, tesla, NV50TCL_WINDOW_LEFT, 2);
+ so_method(so, tesla, NV50TCL_WINDOW_OFFSET_X, 2);
so_data (so, 0);
so_data (so, 0);
/* set screen scissor rectangle */
@@ -147,11 +156,38 @@ nv50_state_validate_fb(struct nv50_context *nv50)
}
static void
+nv50_validate_samplers(struct nv50_context *nv50, struct nouveau_stateobj *so,
+ unsigned p)
+{
+ struct nouveau_grobj *eng2d = nv50->screen->eng2d;
+ unsigned i, j, dw = nv50->sampler_nr[p] * 8;
+
+ if (!dw)
+ return;
+ nv50_so_init_sifc(nv50, so, nv50->screen->tsc, NOUVEAU_BO_VRAM,
+ p * (32 * 8 * 4), dw * 4);
+
+ so_method(so, eng2d, NV50_2D_SIFC_DATA | (2 << 29), dw);
+
+ for (i = 0; i < nv50->sampler_nr[p]; ++i) {
+ if (nv50->sampler[p][i])
+ so_datap(so, nv50->sampler[p][i]->tsc, 8);
+ else {
+ for (j = 0; j < 8; ++j) /* you get punished */
+ so_data(so, 0); /* ... for leaving holes */
+ }
+ }
+}
+
+static void
nv50_state_emit(struct nv50_context *nv50)
{
struct nv50_screen *screen = nv50->screen;
struct nouveau_channel *chan = screen->base.channel;
+ /* I don't want to copy headers from the winsys. */
+ screen->cur_ctx = nv50;
+
if (nv50->pctx_id != screen->cur_pctx) {
if (nv50->state.fb)
nv50->state.dirty |= NV50_NEW_FRAMEBUFFER;
@@ -192,7 +228,8 @@ nv50_state_emit(struct nv50_context *nv50)
so_emit(chan, nv50->state.vertprog);
if (nv50->state.dirty & NV50_NEW_FRAGPROG)
so_emit(chan, nv50->state.fragprog);
- if (nv50->state.dirty & (NV50_NEW_FRAGPROG | NV50_NEW_VERTPROG))
+ if (nv50->state.dirty & (NV50_NEW_FRAGPROG | NV50_NEW_VERTPROG |
+ NV50_NEW_RASTERIZER))
so_emit(chan, nv50->state.programs);
if (nv50->state.dirty & NV50_NEW_RASTERIZER)
so_emit(chan, nv50->state.rast);
@@ -222,6 +259,9 @@ nv50_state_flush_notify(struct nouveau_channel *chan)
{
struct nv50_context *nv50 = chan->user_private;
+ if (nv50->state.tic_upload && !(nv50->dirty & NV50_NEW_TEXTURE))
+ so_emit(chan, nv50->state.tic_upload);
+
so_emit_reloc_markers(chan, nv50->state.fb);
so_emit_reloc_markers(chan, nv50->state.vertprog);
so_emit_reloc_markers(chan, nv50->state.fragprog);
@@ -251,14 +291,15 @@ nv50_state_validate(struct nv50_context *nv50)
if (nv50->dirty & (NV50_NEW_FRAGPROG | NV50_NEW_FRAGPROG_CB))
nv50_fragprog_validate(nv50);
- if (nv50->dirty & (NV50_NEW_FRAGPROG | NV50_NEW_VERTPROG))
+ if (nv50->dirty & (NV50_NEW_FRAGPROG | NV50_NEW_VERTPROG |
+ NV50_NEW_RASTERIZER))
nv50_linkage_validate(nv50);
if (nv50->dirty & NV50_NEW_RASTERIZER)
so_ref(nv50->rasterizer->so, &nv50->state.rast);
if (nv50->dirty & NV50_NEW_BLEND_COLOUR) {
- so = so_new(5, 0);
+ so = so_new(1, 4, 0);
so_method(so, tesla, NV50TCL_BLEND_COLOR(0), 4);
so_data (so, fui(nv50->blend_colour.color[0]));
so_data (so, fui(nv50->blend_colour.color[1]));
@@ -269,10 +310,10 @@ nv50_state_validate(struct nv50_context *nv50)
}
if (nv50->dirty & NV50_NEW_STIPPLE) {
- so = so_new(33, 0);
+ so = so_new(1, 32, 0);
so_method(so, tesla, NV50TCL_POLYGON_STIPPLE_PATTERN(0), 32);
for (i = 0; i < 32; i++)
- so_data(so, nv50->stipple.stipple[i]);
+ so_data(so, util_bswap32(nv50->stipple.stipple[i]));
so_ref(so, &nv50->state.stipple);
so_ref(NULL, &so);
}
@@ -286,8 +327,8 @@ nv50_state_validate(struct nv50_context *nv50)
goto scissor_uptodate;
nv50->state.scissor_enabled = rast->scissor;
- so = so_new(3, 0);
- so_method(so, tesla, NV50TCL_SCISSOR_HORIZ, 2);
+ so = so_new(1, 2, 0);
+ so_method(so, tesla, NV50TCL_SCISSOR_HORIZ(0), 2);
if (nv50->state.scissor_enabled) {
so_data(so, (s->maxx << 16) | s->minx);
so_data(so, (s->maxy << 16) | s->miny);
@@ -315,13 +356,13 @@ scissor_uptodate:
goto viewport_uptodate;
nv50->state.viewport_bypass = bypass;
- so = so_new(14, 0);
+ so = so_new(5, 9, 0);
if (!bypass) {
- so_method(so, tesla, NV50TCL_VIEWPORT_TRANSLATE(0), 3);
+ so_method(so, tesla, NV50TCL_VIEWPORT_TRANSLATE_X(0), 3);
so_data (so, fui(nv50->viewport.translate[0]));
so_data (so, fui(nv50->viewport.translate[1]));
so_data (so, fui(nv50->viewport.translate[2]));
- so_method(so, tesla, NV50TCL_VIEWPORT_SCALE(0), 3);
+ so_method(so, tesla, NV50TCL_VIEWPORT_SCALE_X(0), 3);
so_data (so, fui(nv50->viewport.scale[0]));
so_data (so, fui(nv50->viewport.scale[1]));
so_data (so, fui(nv50->viewport.scale[2]));
@@ -354,18 +395,20 @@ scissor_uptodate:
viewport_uptodate:
if (nv50->dirty & NV50_NEW_SAMPLER) {
- int i;
+ unsigned nr = 0;
- so = so_new(nv50->sampler_nr * 9 + 2, 0);
- so_method(so, tesla, NV50TCL_CB_ADDR, 1);
- so_data (so, NV50_CB_TSC);
- for (i = 0; i < nv50->sampler_nr; i++) {
- if (!nv50->sampler[i])
- continue;
+ for (i = 0; i < PIPE_SHADER_TYPES; ++i)
+ nr += nv50->sampler_nr[i];
+
+ so = so_new(1+ 5 * PIPE_SHADER_TYPES, 1+ 19 * PIPE_SHADER_TYPES
+ + nr * 8, PIPE_SHADER_TYPES * 2);
+
+ nv50_validate_samplers(nv50, so, PIPE_SHADER_VERTEX);
+ nv50_validate_samplers(nv50, so, PIPE_SHADER_FRAGMENT);
+
+ so_method(so, tesla, 0x1334, 1); /* flush TSC */
+ so_data (so, 0);
- so_method(so, tesla, NV50TCL_CB_DATA(0) | (2<<29), 8);
- so_datap (so, nv50->sampler[i]->tsc, 8);
- }
so_ref(so, &nv50->state.tsc_upload);
so_ref(NULL, &so);
}
@@ -383,3 +426,36 @@ viewport_uptodate:
return TRUE;
}
+void nv50_so_init_sifc(struct nv50_context *nv50,
+ struct nouveau_stateobj *so,
+ struct nouveau_bo *bo, unsigned reloc,
+ unsigned offset, unsigned size)
+{
+ struct nouveau_grobj *eng2d = nv50->screen->eng2d;
+
+ reloc |= NOUVEAU_BO_WR;
+
+ so_method(so, eng2d, NV50_2D_DST_FORMAT, 2);
+ so_data (so, NV50_2D_DST_FORMAT_R8_UNORM);
+ so_data (so, 1);
+ so_method(so, eng2d, NV50_2D_DST_PITCH, 5);
+ so_data (so, 262144);
+ so_data (so, 65536);
+ so_data (so, 1);
+ so_reloc (so, bo, offset, reloc | NOUVEAU_BO_HIGH, 0, 0);
+ so_reloc (so, bo, offset, reloc | NOUVEAU_BO_LOW, 0, 0);
+ so_method(so, eng2d, NV50_2D_SIFC_BITMAP_ENABLE, 2);
+ so_data (so, 0);
+ so_data (so, NV50_2D_SIFC_FORMAT_R8_UNORM);
+ so_method(so, eng2d, NV50_2D_SIFC_WIDTH, 10);
+ so_data (so, size);
+ so_data (so, 1);
+ so_data (so, 0);
+ so_data (so, 1);
+ so_data (so, 0);
+ so_data (so, 1);
+ so_data (so, 0);
+ so_data (so, 0);
+ so_data (so, 0);
+ so_data (so, 0);
+}