[Mesa-dev] [PATCH 45/71] st/nine: Rework constant buffer state handling
Axel Davy
axel.davy at ens.fr
Sun Aug 16 08:28:09 PDT 2015
We have two paths:
. One that uses a fixed constant buffer, and updates it when needed
. One that uses a user constant buffer, and upload it when needed.
This patch separates the preparation of the constant buffer
and the commit.
It also removes NineDevice9_RestoreNonCSOState, which was
used to restore all states. Instead the commit of the constant
buffer is moved to nine_state, and the other field settings
moved to other functions where more appropriate.
Signed-off-by: Axel Davy <axel.davy at ens.fr>
---
src/gallium/state_trackers/nine/device9.c | 69 +--
src/gallium/state_trackers/nine/device9.h | 4 -
src/gallium/state_trackers/nine/nine_state.c | 795 ++++++++++++++-------------
src/gallium/state_trackers/nine/nine_state.h | 5 +
src/gallium/state_trackers/nine/swapchain9.c | 2 +-
5 files changed, 435 insertions(+), 440 deletions(-)
diff --git a/src/gallium/state_trackers/nine/device9.c b/src/gallium/state_trackers/nine/device9.c
index 1416a38..c568761 100644
--- a/src/gallium/state_trackers/nine/device9.c
+++ b/src/gallium/state_trackers/nine/device9.c
@@ -119,68 +119,6 @@ NineDevice9_SetDefaultState( struct NineDevice9 *This, boolean is_reset )
This, (IDirect3DSurface9 *)This->swapchains[0]->zsbuf);
}
-void
-NineDevice9_RestoreNonCSOState( struct NineDevice9 *This, unsigned mask )
-{
- struct pipe_context *pipe = This->pipe;
-
- DBG("This=%p mask=%u\n", This, mask);
-
- if (mask & 0x1) {
- struct pipe_constant_buffer cb;
- cb.buffer_offset = 0;
- cb.buffer_size = This->vs_const_size;
-
- if (This->prefer_user_constbuf) {
- cb.buffer = NULL;
- cb.user_buffer = This->state.vs_const_f;
- if (!This->driver_caps.user_cbufs) {
- u_upload_data(This->constbuf_uploader,
- 0,
- cb.buffer_size,
- cb.user_buffer,
- &cb.buffer_offset,
- &cb.buffer);
- u_upload_unmap(This->constbuf_uploader);
- cb.user_buffer = NULL;
- }
- } else {
- cb.buffer = This->constbuf_vs;
- cb.user_buffer = NULL;
- }
- pipe->set_constant_buffer(pipe, PIPE_SHADER_VERTEX, 0, &cb);
-
- cb.buffer_size = This->ps_const_size;
- if (This->prefer_user_constbuf) {
- cb.user_buffer = This->state.ps_const_f;
- if (!This->driver_caps.user_cbufs) {
- u_upload_data(This->constbuf_uploader,
- 0,
- cb.buffer_size,
- cb.user_buffer,
- &cb.buffer_offset,
- &cb.buffer);
- u_upload_unmap(This->constbuf_uploader);
- cb.user_buffer = NULL;
- }
- } else {
- cb.buffer = This->constbuf_ps;
- }
- pipe->set_constant_buffer(pipe, PIPE_SHADER_FRAGMENT, 0, &cb);
- }
-
- if (mask & 0x2) {
- struct pipe_poly_stipple stipple;
- memset(&stipple, ~0, sizeof(stipple));
- pipe->set_polygon_stipple(pipe, &stipple);
- }
-
- This->state.changed.group = NINE_STATE_ALL;
- This->state.changed.vtxbuf = (1ULL << This->caps.MaxStreams) - 1;
- This->state.changed.ucp = (1 << PIPE_MAX_CLIP_PLANES) - 1;
- This->state.changed.texture = NINE_PS_SAMPLERS_MASK | NINE_VS_SAMPLERS_MASK;
-}
-
#define GET_PCAP(n) pScreen->get_param(pScreen, PIPE_CAP_##n)
HRESULT
NineDevice9_ctor( struct NineDevice9 *This,
@@ -455,7 +393,12 @@ NineDevice9_ctor( struct NineDevice9 *This,
nine_ff_init(This); /* initialize fixed function code */
NineDevice9_SetDefaultState(This, FALSE);
- NineDevice9_RestoreNonCSOState(This, ~0);
+
+ {
+ struct pipe_poly_stipple stipple;
+ memset(&stipple, ~0, sizeof(stipple));
+ This->pipe->set_polygon_stipple(This->pipe, &stipple);
+ }
This->update = &This->state;
nine_update_state(This);
diff --git a/src/gallium/state_trackers/nine/device9.h b/src/gallium/state_trackers/nine/device9.h
index 8b955a7..f109f3c 100644
--- a/src/gallium/state_trackers/nine/device9.h
+++ b/src/gallium/state_trackers/nine/device9.h
@@ -184,10 +184,6 @@ NineDevice9_GetCSO( struct NineDevice9 *This );
const D3DCAPS9 *
NineDevice9_GetCaps( struct NineDevice9 *This );
-/* Mask: 0x1 = constant buffers, 0x2 = stipple */
-void
-NineDevice9_RestoreNonCSOState( struct NineDevice9 *This, unsigned mask );
-
/*** Direct3D public ***/
HRESULT WINAPI
diff --git a/src/gallium/state_trackers/nine/nine_state.c b/src/gallium/state_trackers/nine/nine_state.c
index 8c2b6eb..1b9622b 100644
--- a/src/gallium/state_trackers/nine/nine_state.c
+++ b/src/gallium/state_trackers/nine/nine_state.c
@@ -61,6 +61,275 @@ prepare_rasterizer(struct NineDevice9 *device)
device->state.commit |= NINE_STATE_COMMIT_RASTERIZER;
}
+#define DO_UPLOAD_CONST_F(buf,p,c,d) \
+ do { \
+ DBG("upload ConstantF [%u .. %u]\n", x, (x) + (c) - 1); \
+ box.x = (p) * 4 * sizeof(float); \
+ box.width = (c) * 4 * sizeof(float); \
+ pipe->transfer_inline_write(pipe, buf, 0, usage, &box, &((d)[p * 4]), \
+ 0, 0); \
+ } while(0)
+
+/* OK, this is a bit ugly ... */
+static void
+upload_constants(struct NineDevice9 *device, unsigned shader_type)
+{
+ struct pipe_context *pipe = device->pipe;
+ struct pipe_resource *buf;
+ struct pipe_box box;
+ const void *data;
+ const float *const_f;
+ const int *const_i;
+ const BOOL *const_b;
+ uint32_t data_b[NINE_MAX_CONST_B];
+ uint16_t dirty_i;
+ uint16_t dirty_b;
+ const unsigned usage = PIPE_TRANSFER_WRITE | PIPE_TRANSFER_DISCARD_RANGE;
+ unsigned x = 0; /* silence warning */
+ unsigned i, c;
+ struct nine_range *r, *p, *lconstf_ranges;
+ float *lconstf_data;
+
+ box.y = 0;
+ box.z = 0;
+ box.height = 1;
+ box.depth = 1;
+
+ if (shader_type == PIPE_SHADER_VERTEX) {
+ DBG("VS\n");
+ buf = device->constbuf_vs;
+
+ const_f = device->state.vs_const_f;
+ for (p = r = device->state.changed.vs_const_f; r; p = r, r = r->next)
+ DO_UPLOAD_CONST_F(buf, r->bgn, r->end - r->bgn, const_f);
+ if (p) {
+ nine_range_pool_put_chain(&device->range_pool,
+ device->state.changed.vs_const_f, p);
+ device->state.changed.vs_const_f = NULL;
+ }
+
+ dirty_i = device->state.changed.vs_const_i;
+ device->state.changed.vs_const_i = 0;
+ const_i = &device->state.vs_const_i[0][0];
+
+ dirty_b = device->state.changed.vs_const_b;
+ device->state.changed.vs_const_b = 0;
+ const_b = device->state.vs_const_b;
+
+ lconstf_ranges = device->state.vs->lconstf.ranges;
+ lconstf_data = device->state.vs->lconstf.data;
+
+ device->state.ff.clobber.vs_const = TRUE;
+ device->state.changed.group &= ~NINE_STATE_VS_CONST;
+ } else {
+ DBG("PS\n");
+ buf = device->constbuf_ps;
+
+ const_f = device->state.ps_const_f;
+ for (p = r = device->state.changed.ps_const_f; r; p = r, r = r->next)
+ DO_UPLOAD_CONST_F(buf, r->bgn, r->end - r->bgn, const_f);
+ if (p) {
+ nine_range_pool_put_chain(&device->range_pool,
+ device->state.changed.ps_const_f, p);
+ device->state.changed.ps_const_f = NULL;
+ }
+
+ dirty_i = device->state.changed.ps_const_i;
+ device->state.changed.ps_const_i = 0;
+ const_i = &device->state.ps_const_i[0][0];
+
+ dirty_b = device->state.changed.ps_const_b;
+ device->state.changed.ps_const_b = 0;
+ const_b = device->state.ps_const_b;
+
+ lconstf_ranges = NULL;
+ lconstf_data = NULL;
+
+ device->state.ff.clobber.ps_const = TRUE;
+ device->state.changed.group &= ~NINE_STATE_PS_CONST;
+ }
+
+ /* write range from min to max changed, it's not much data */
+ /* bool1 */
+ if (dirty_b) {
+ c = util_last_bit(dirty_b);
+ i = ffs(dirty_b) - 1;
+ x = buf->width0 - (NINE_MAX_CONST_B - i) * 4;
+ c -= i;
+ memcpy(data_b, &(const_b[i]), c * sizeof(uint32_t));
+ box.x = x;
+ box.width = c * 4;
+ DBG("upload ConstantB [%u .. %u]\n", x, x + c - 1);
+ pipe->transfer_inline_write(pipe, buf, 0, usage, &box, data_b, 0, 0);
+ }
+
+ /* int4 */
+ for (c = 0, i = 0; dirty_i; i++, dirty_i >>= 1) {
+ if (dirty_i & 1) {
+ if (!c)
+ x = i;
+ ++c;
+ } else
+ if (c) {
+ DBG("upload ConstantI [%u .. %u]\n", x, x + c - 1);
+ data = &const_i[x * 4];
+ box.x = buf->width0 - (NINE_MAX_CONST_I * 4 + NINE_MAX_CONST_B) * 4;
+ box.x += x * 4 * sizeof(int);
+ box.width = c * 4 * sizeof(int);
+ c = 0;
+ pipe->transfer_inline_write(pipe, buf, 0, usage, &box, data, 0, 0);
+ }
+ }
+ if (c) {
+ DBG("upload ConstantI [%u .. %u]\n", x, x + c - 1);
+ data = &const_i[x * 4];
+ box.x = buf->width0 - (NINE_MAX_CONST_I * 4 + NINE_MAX_CONST_B) * 4;
+ box.x += x * 4 * sizeof(int);
+ box.width = c * 4 * sizeof(int);
+ pipe->transfer_inline_write(pipe, buf, 0, usage, &box, data, 0, 0);
+ }
+
+ /* TODO: only upload these when shader itself changes */
+ if (lconstf_ranges) {
+ unsigned n = 0;
+ struct nine_range *r = lconstf_ranges;
+ while (r) {
+ box.x = r->bgn * 4 * sizeof(float);
+ n += r->end - r->bgn;
+ box.width = (r->end - r->bgn) * 4 * sizeof(float);
+ data = &lconstf_data[4 * n];
+ pipe->transfer_inline_write(pipe, buf, 0, usage, &box, data, 0, 0);
+ r = r->next;
+ }
+ }
+}
+
+static void
+prepare_vs_constants_userbuf(struct NineDevice9 *device)
+{
+ struct nine_state *state = &device->state;
+ struct pipe_constant_buffer cb;
+ cb.buffer = NULL;
+ cb.buffer_offset = 0;
+ cb.buffer_size = device->state.vs->const_used_size;
+ cb.user_buffer = device->state.vs_const_f;
+
+ if (!cb.buffer_size)
+ return;
+
+ if (state->changed.vs_const_i) {
+ int *idst = (int *)&state->vs_const_f[4 * device->max_vs_const_f];
+ memcpy(idst, state->vs_const_i, sizeof(state->vs_const_i));
+ state->changed.vs_const_i = 0;
+ }
+ if (state->changed.vs_const_b) {
+ int *idst = (int *)&state->vs_const_f[4 * device->max_vs_const_f];
+ uint32_t *bdst = (uint32_t *)&idst[4 * NINE_MAX_CONST_I];
+ memcpy(bdst, state->vs_const_b, sizeof(state->vs_const_b));
+ state->changed.vs_const_b = 0;
+ }
+
+ if (device->state.vs->lconstf.ranges) {
+ /* TODO: Can we make it so that we don't have to copy everything ? */
+ const struct nine_lconstf *lconstf = &device->state.vs->lconstf;
+ const struct nine_range *r = lconstf->ranges;
+ unsigned n = 0;
+ float *dst = device->state.vs_lconstf_temp;
+ float *src = (float *)cb.user_buffer;
+ memcpy(dst, src, cb.buffer_size);
+ while (r) {
+ unsigned p = r->bgn;
+ unsigned c = r->end - r->bgn;
+ memcpy(&dst[p * 4], &lconstf->data[n * 4], c * 4 * sizeof(float));
+ n += c;
+ r = r->next;
+ }
+ cb.user_buffer = dst;
+ }
+
+ if (!device->driver_caps.user_cbufs) {
+ u_upload_data(device->constbuf_uploader,
+ 0,
+ cb.buffer_size,
+ cb.user_buffer,
+ &cb.buffer_offset,
+ &cb.buffer);
+ u_upload_unmap(device->constbuf_uploader);
+ cb.user_buffer = NULL;
+ }
+
+ state->pipe.cb_vs = cb;
+
+ if (device->state.changed.vs_const_f) {
+ struct nine_range *r = device->state.changed.vs_const_f;
+ struct nine_range *p = r;
+ while (p->next)
+ p = p->next;
+ nine_range_pool_put_chain(&device->range_pool, r, p);
+ device->state.changed.vs_const_f = NULL;
+ }
+ state->changed.group &= ~NINE_STATE_VS_CONST;
+ state->commit |= NINE_STATE_COMMIT_CONST_VS;
+}
+
+static void
+prepare_ps_constants_userbuf(struct NineDevice9 *device)
+{
+ struct nine_state *state = &device->state;
+ struct pipe_constant_buffer cb;
+ cb.buffer = NULL;
+ cb.buffer_offset = 0;
+ cb.buffer_size = device->state.ps->const_used_size;
+ cb.user_buffer = device->state.ps_const_f;
+
+ if (!cb.buffer_size)
+ return;
+
+ if (state->changed.ps_const_i) {
+ int *idst = (int *)&state->ps_const_f[4 * device->max_ps_const_f];
+ memcpy(idst, state->ps_const_i, sizeof(state->ps_const_i));
+ state->changed.ps_const_i = 0;
+ }
+ if (state->changed.ps_const_b) {
+ int *idst = (int *)&state->ps_const_f[4 * device->max_ps_const_f];
+ uint32_t *bdst = (uint32_t *)&idst[4 * NINE_MAX_CONST_I];
+ memcpy(bdst, state->ps_const_b, sizeof(state->ps_const_b));
+ state->changed.ps_const_b = 0;
+ }
+
+ /* Upload special constants needed to implement PS1.x instructions like TEXBEM,TEXBEML and BEM */
+ if (device->state.ps->bumpenvmat_needed) {
+ memcpy(device->state.ps_lconstf_temp, cb.user_buffer, cb.buffer_size);
+ memcpy(&device->state.ps_lconstf_temp[4 * 8], &device->state.bumpmap_vars, sizeof(device->state.bumpmap_vars));
+
+ cb.user_buffer = device->state.ps_lconstf_temp;
+ }
+
+ if (!device->driver_caps.user_cbufs) {
+ u_upload_data(device->constbuf_uploader,
+ 0,
+ cb.buffer_size,
+ cb.user_buffer,
+ &cb.buffer_offset,
+ &cb.buffer);
+ u_upload_unmap(device->constbuf_uploader);
+ cb.user_buffer = NULL;
+ }
+
+ state->pipe.cb_ps = cb;
+
+ if (device->state.changed.ps_const_f) {
+ struct nine_range *r = device->state.changed.ps_const_f;
+ struct nine_range *p = r;
+ while (p->next)
+ p = p->next;
+ nine_range_pool_put_chain(&device->range_pool, r, p);
+ device->state.changed.ps_const_f = NULL;
+ }
+ state->changed.group &= ~NINE_STATE_PS_CONST;
+ state->commit |= NINE_STATE_COMMIT_CONST_PS;
+}
+
/* State preparation incremental */
/* State preparation + State commit */
@@ -252,404 +521,134 @@ update_vertex_elements(struct NineDevice9 *device)
for (i = 0; i < device->caps.MaxStreams; i++ ) {
if (!used_streams[i]) {
dummy_vbo_stream = i;
- break;
- }
- }
- }
- /* there are less vertex shader inputs than stream slots,
- * so if we need a slot for the dummy vbo, we should have found one */
- assert (!need_dummy_vbo || dummy_vbo_stream != -1);
-
- for (n = 0; n < vs->num_inputs; ++n) {
- index = vdecl_index_map[n];
- if (index >= 0) {
- ve[n] = vdecl->elems[index];
- b = ve[n].vertex_buffer_index;
- state->stream_usage_mask |= 1 << b;
- /* XXX wine just uses 1 here: */
- if (state->stream_freq[b] & D3DSTREAMSOURCE_INSTANCEDATA)
- ve[n].instance_divisor = state->stream_freq[b] & 0x7FFFFF;
- } else {
- /* if the vertex declaration is incomplete compared to what the
- * vertex shader needs, we bind a dummy vbo with 0 0 0 0.
- * This is not precised by the spec, but is the behaviour
- * tested on win */
- ve[n].vertex_buffer_index = dummy_vbo_stream;
- ve[n].src_format = PIPE_FORMAT_R32G32B32A32_FLOAT;
- ve[n].src_offset = 0;
- ve[n].instance_divisor = 0;
- }
- }
-
- if (state->dummy_vbo_bound_at != dummy_vbo_stream) {
- if (state->dummy_vbo_bound_at >= 0)
- state->changed.vtxbuf |= 1 << state->dummy_vbo_bound_at;
- if (dummy_vbo_stream >= 0) {
- state->changed.vtxbuf |= 1 << dummy_vbo_stream;
- state->vbo_bound_done = FALSE;
- }
- state->dummy_vbo_bound_at = dummy_vbo_stream;
- }
-
- cso_set_vertex_elements(device->cso, vs->num_inputs, ve);
-
- state->changed.stream_freq = 0;
-}
-
-static inline uint32_t
-update_shader_variant_keys(struct NineDevice9 *device)
-{
- struct nine_state *state = &device->state;
- uint32_t mask = 0;
- uint32_t vs_key = state->samplers_shadow;
- uint32_t ps_key = state->samplers_shadow;
-
- vs_key = (vs_key & NINE_VS_SAMPLERS_MASK) >> NINE_SAMPLER_VS(0);
- ps_key = (ps_key & NINE_PS_SAMPLERS_MASK) >> NINE_SAMPLER_PS(0);
-
- if (state->vs) vs_key &= state->vs->sampler_mask;
- if (state->ps) {
- if (unlikely(state->ps->byte_code.version < 0x20)) {
- /* no depth textures, but variable targets */
- uint32_t m = state->ps->sampler_mask;
- ps_key = 0;
- while (m) {
- int s = ffs(m) - 1;
- m &= ~(1 << s);
- ps_key |= (state->texture[s] ? state->texture[s]->pstype : 1) << (s * 2);
- }
- } else {
- ps_key &= state->ps->sampler_mask;
- }
- }
-
- if (state->vs && state->vs_key != vs_key) {
- state->vs_key = vs_key;
- mask |= NINE_STATE_VS;
- }
- if (state->ps && state->ps_key != ps_key) {
- state->ps_key = ps_key;
- mask |= NINE_STATE_PS;
- }
- return mask;
-}
-
-static inline uint32_t
-update_vs(struct NineDevice9 *device)
-{
- struct nine_state *state = &device->state;
- struct NineVertexShader9 *vs = state->vs;
- uint32_t changed_group = 0;
-
- /* likely because we dislike FF */
- if (likely(vs)) {
- state->cso.vs = NineVertexShader9_GetVariant(vs, state->vs_key);
- } else {
- vs = device->ff.vs;
- state->cso.vs = vs->variant.cso;
- }
- device->pipe->bind_vs_state(device->pipe, state->cso.vs);
-
- if (state->rs[NINED3DRS_VSPOINTSIZE] != vs->point_size) {
- state->rs[NINED3DRS_VSPOINTSIZE] = vs->point_size;
- changed_group |= NINE_STATE_RASTERIZER;
- }
-
- if ((state->bound_samplers_mask_vs & vs->sampler_mask) != vs->sampler_mask)
- /* Bound dummy sampler. */
- changed_group |= NINE_STATE_SAMPLER;
- return changed_group;
-}
-
-static inline uint32_t
-update_ps(struct NineDevice9 *device)
-{
- struct nine_state *state = &device->state;
- struct NinePixelShader9 *ps = state->ps;
- uint32_t changed_group = 0;
-
- if (likely(ps)) {
- state->cso.ps = NinePixelShader9_GetVariant(ps, state->ps_key);
- } else {
- ps = device->ff.ps;
- state->cso.ps = ps->variant.cso;
- }
- device->pipe->bind_fs_state(device->pipe, state->cso.ps);
-
- if ((state->bound_samplers_mask_ps & ps->sampler_mask) != ps->sampler_mask)
- /* Bound dummy sampler. */
- changed_group |= NINE_STATE_SAMPLER;
- return changed_group;
-}
-
-#define DO_UPLOAD_CONST_F(buf,p,c,d) \
- do { \
- DBG("upload ConstantF [%u .. %u]\n", x, (x) + (c) - 1); \
- box.x = (p) * 4 * sizeof(float); \
- box.width = (c) * 4 * sizeof(float); \
- pipe->transfer_inline_write(pipe, buf, 0, usage, &box, &((d)[p * 4]), \
- 0, 0); \
- } while(0)
-
-/* OK, this is a bit ugly ... */
-static void
-update_constants(struct NineDevice9 *device, unsigned shader_type)
-{
- struct pipe_context *pipe = device->pipe;
- struct pipe_resource *buf;
- struct pipe_box box;
- const void *data;
- const float *const_f;
- const int *const_i;
- const BOOL *const_b;
- uint32_t data_b[NINE_MAX_CONST_B];
- uint16_t dirty_i;
- uint16_t dirty_b;
- const unsigned usage = PIPE_TRANSFER_WRITE | PIPE_TRANSFER_DISCARD_RANGE;
- unsigned x = 0; /* silence warning */
- unsigned i, c;
- struct nine_range *r, *p, *lconstf_ranges;
- float *lconstf_data;
-
- box.y = 0;
- box.z = 0;
- box.height = 1;
- box.depth = 1;
-
- if (shader_type == PIPE_SHADER_VERTEX) {
- DBG("VS\n");
- buf = device->constbuf_vs;
-
- const_f = device->state.vs_const_f;
- for (p = r = device->state.changed.vs_const_f; r; p = r, r = r->next)
- DO_UPLOAD_CONST_F(buf, r->bgn, r->end - r->bgn, const_f);
- if (p) {
- nine_range_pool_put_chain(&device->range_pool,
- device->state.changed.vs_const_f, p);
- device->state.changed.vs_const_f = NULL;
- }
-
- dirty_i = device->state.changed.vs_const_i;
- device->state.changed.vs_const_i = 0;
- const_i = &device->state.vs_const_i[0][0];
-
- dirty_b = device->state.changed.vs_const_b;
- device->state.changed.vs_const_b = 0;
- const_b = device->state.vs_const_b;
-
- lconstf_ranges = device->state.vs->lconstf.ranges;
- lconstf_data = device->state.vs->lconstf.data;
-
- device->state.ff.clobber.vs_const = TRUE;
- device->state.changed.group &= ~NINE_STATE_VS_CONST;
- } else {
- DBG("PS\n");
- buf = device->constbuf_ps;
-
- const_f = device->state.ps_const_f;
- for (p = r = device->state.changed.ps_const_f; r; p = r, r = r->next)
- DO_UPLOAD_CONST_F(buf, r->bgn, r->end - r->bgn, const_f);
- if (p) {
- nine_range_pool_put_chain(&device->range_pool,
- device->state.changed.ps_const_f, p);
- device->state.changed.ps_const_f = NULL;
- }
-
- dirty_i = device->state.changed.ps_const_i;
- device->state.changed.ps_const_i = 0;
- const_i = &device->state.ps_const_i[0][0];
-
- dirty_b = device->state.changed.ps_const_b;
- device->state.changed.ps_const_b = 0;
- const_b = device->state.ps_const_b;
-
- lconstf_ranges = NULL;
- lconstf_data = NULL;
-
- device->state.ff.clobber.ps_const = TRUE;
- device->state.changed.group &= ~NINE_STATE_PS_CONST;
- }
-
- /* write range from min to max changed, it's not much data */
- /* bool1 */
- if (dirty_b) {
- c = util_last_bit(dirty_b);
- i = ffs(dirty_b) - 1;
- x = buf->width0 - (NINE_MAX_CONST_B - i) * 4;
- c -= i;
- memcpy(data_b, &(const_b[i]), c * sizeof(uint32_t));
- box.x = x;
- box.width = c * 4;
- DBG("upload ConstantB [%u .. %u]\n", x, x + c - 1);
- pipe->transfer_inline_write(pipe, buf, 0, usage, &box, data_b, 0, 0);
+ break;
+ }
+ }
}
+ /* there are less vertex shader inputs than stream slots,
+ * so if we need a slot for the dummy vbo, we should have found one */
+ assert (!need_dummy_vbo || dummy_vbo_stream != -1);
- /* int4 */
- for (c = 0, i = 0; dirty_i; i++, dirty_i >>= 1) {
- if (dirty_i & 1) {
- if (!c)
- x = i;
- ++c;
- } else
- if (c) {
- DBG("upload ConstantI [%u .. %u]\n", x, x + c - 1);
- data = &const_i[x * 4];
- box.x = buf->width0 - (NINE_MAX_CONST_I * 4 + NINE_MAX_CONST_B) * 4;
- box.x += x * 4 * sizeof(int);
- box.width = c * 4 * sizeof(int);
- c = 0;
- pipe->transfer_inline_write(pipe, buf, 0, usage, &box, data, 0, 0);
+ for (n = 0; n < vs->num_inputs; ++n) {
+ index = vdecl_index_map[n];
+ if (index >= 0) {
+ ve[n] = vdecl->elems[index];
+ b = ve[n].vertex_buffer_index;
+ state->stream_usage_mask |= 1 << b;
+ /* XXX wine just uses 1 here: */
+ if (state->stream_freq[b] & D3DSTREAMSOURCE_INSTANCEDATA)
+ ve[n].instance_divisor = state->stream_freq[b] & 0x7FFFFF;
+ } else {
+ /* if the vertex declaration is incomplete compared to what the
+ * vertex shader needs, we bind a dummy vbo with 0 0 0 0.
+ * This is not precised by the spec, but is the behaviour
+ * tested on win */
+ ve[n].vertex_buffer_index = dummy_vbo_stream;
+ ve[n].src_format = PIPE_FORMAT_R32G32B32A32_FLOAT;
+ ve[n].src_offset = 0;
+ ve[n].instance_divisor = 0;
}
}
- if (c) {
- DBG("upload ConstantI [%u .. %u]\n", x, x + c - 1);
- data = &const_i[x * 4];
- box.x = buf->width0 - (NINE_MAX_CONST_I * 4 + NINE_MAX_CONST_B) * 4;
- box.x += x * 4 * sizeof(int);
- box.width = c * 4 * sizeof(int);
- pipe->transfer_inline_write(pipe, buf, 0, usage, &box, data, 0, 0);
- }
- /* TODO: only upload these when shader itself changes */
- if (lconstf_ranges) {
- unsigned n = 0;
- struct nine_range *r = lconstf_ranges;
- while (r) {
- box.x = r->bgn * 4 * sizeof(float);
- n += r->end - r->bgn;
- box.width = (r->end - r->bgn) * 4 * sizeof(float);
- data = &lconstf_data[4 * n];
- pipe->transfer_inline_write(pipe, buf, 0, usage, &box, data, 0, 0);
- r = r->next;
+ if (state->dummy_vbo_bound_at != dummy_vbo_stream) {
+ if (state->dummy_vbo_bound_at >= 0)
+ state->changed.vtxbuf |= 1 << state->dummy_vbo_bound_at;
+ if (dummy_vbo_stream >= 0) {
+ state->changed.vtxbuf |= 1 << dummy_vbo_stream;
+ state->vbo_bound_done = FALSE;
}
+ state->dummy_vbo_bound_at = dummy_vbo_stream;
}
+
+ cso_set_vertex_elements(device->cso, vs->num_inputs, ve);
+
+ state->changed.stream_freq = 0;
}
-static void
-update_vs_constants_userbuf(struct NineDevice9 *device)
+static inline uint32_t
+update_shader_variant_keys(struct NineDevice9 *device)
{
struct nine_state *state = &device->state;
- struct pipe_context *pipe = device->pipe;
- struct pipe_constant_buffer cb;
- cb.buffer = NULL;
- cb.buffer_offset = 0;
- cb.buffer_size = device->state.vs->const_used_size;
- cb.user_buffer = device->state.vs_const_f;
-
- if (!cb.buffer_size)
- return;
+ uint32_t mask = 0;
+ uint32_t vs_key = state->samplers_shadow;
+ uint32_t ps_key = state->samplers_shadow;
- if (state->changed.vs_const_i) {
- int *idst = (int *)&state->vs_const_f[4 * device->max_vs_const_f];
- memcpy(idst, state->vs_const_i, sizeof(state->vs_const_i));
- state->changed.vs_const_i = 0;
- }
- if (state->changed.vs_const_b) {
- int *idst = (int *)&state->vs_const_f[4 * device->max_vs_const_f];
- uint32_t *bdst = (uint32_t *)&idst[4 * NINE_MAX_CONST_I];
- memcpy(bdst, state->vs_const_b, sizeof(state->vs_const_b));
- state->changed.vs_const_b = 0;
- }
+ vs_key = (vs_key & NINE_VS_SAMPLERS_MASK) >> NINE_SAMPLER_VS(0);
+ ps_key = (ps_key & NINE_PS_SAMPLERS_MASK) >> NINE_SAMPLER_PS(0);
- if (device->state.vs->lconstf.ranges) {
- /* TODO: Can we make it so that we don't have to copy everything ? */
- const struct nine_lconstf *lconstf = &device->state.vs->lconstf;
- const struct nine_range *r = lconstf->ranges;
- unsigned n = 0;
- float *dst = device->state.vs_lconstf_temp;
- float *src = (float *)cb.user_buffer;
- memcpy(dst, src, cb.buffer_size);
- while (r) {
- unsigned p = r->bgn;
- unsigned c = r->end - r->bgn;
- memcpy(&dst[p * 4], &lconstf->data[n * 4], c * 4 * sizeof(float));
- n += c;
- r = r->next;
+ if (state->vs) vs_key &= state->vs->sampler_mask;
+ if (state->ps) {
+ if (unlikely(state->ps->byte_code.version < 0x20)) {
+ /* no depth textures, but variable targets */
+ uint32_t m = state->ps->sampler_mask;
+ ps_key = 0;
+ while (m) {
+ int s = ffs(m) - 1;
+ m &= ~(1 << s);
+ ps_key |= (state->texture[s] ? state->texture[s]->pstype : 1) << (s * 2);
+ }
+ } else {
+ ps_key &= state->ps->sampler_mask;
}
- cb.user_buffer = dst;
}
- if (!device->driver_caps.user_cbufs) {
- u_upload_data(device->constbuf_uploader,
- 0,
- cb.buffer_size,
- cb.user_buffer,
- &cb.buffer_offset,
- &cb.buffer);
- u_upload_unmap(device->constbuf_uploader);
- cb.user_buffer = NULL;
+ if (state->vs && state->vs_key != vs_key) {
+ state->vs_key = vs_key;
+ mask |= NINE_STATE_VS;
}
-
- pipe->set_constant_buffer(pipe, PIPE_SHADER_VERTEX, 0, &cb);
-
- if (device->state.changed.vs_const_f) {
- struct nine_range *r = device->state.changed.vs_const_f;
- struct nine_range *p = r;
- while (p->next)
- p = p->next;
- nine_range_pool_put_chain(&device->range_pool, r, p);
- device->state.changed.vs_const_f = NULL;
+ if (state->ps && state->ps_key != ps_key) {
+ state->ps_key = ps_key;
+ mask |= NINE_STATE_PS;
}
- state->changed.group &= ~NINE_STATE_VS_CONST;
+ return mask;
}
-static void
-update_ps_constants_userbuf(struct NineDevice9 *device)
+static inline uint32_t
+update_vs(struct NineDevice9 *device)
{
struct nine_state *state = &device->state;
- struct pipe_context *pipe = device->pipe;
- struct pipe_constant_buffer cb;
- int i;
- cb.buffer = NULL;
- cb.buffer_offset = 0;
- cb.buffer_size = device->state.ps->const_used_size;
- cb.user_buffer = device->state.ps_const_f;
-
- if (!cb.buffer_size)
- return;
+ struct NineVertexShader9 *vs = state->vs;
+ uint32_t changed_group = 0;
- if (state->changed.ps_const_i) {
- int *idst = (int *)&state->ps_const_f[4 * device->max_ps_const_f];
- memcpy(idst, state->ps_const_i, sizeof(state->ps_const_i));
- state->changed.ps_const_i = 0;
- }
- if (state->changed.ps_const_b) {
- int *idst = (int *)&state->ps_const_f[4 * device->max_ps_const_f];
- uint32_t *bdst = (uint32_t *)&idst[4 * NINE_MAX_CONST_I];
- memcpy(bdst, state->ps_const_b, sizeof(state->ps_const_b));
- state->changed.ps_const_b = 0;
+ /* likely because we dislike FF */
+ if (likely(vs)) {
+ state->cso.vs = NineVertexShader9_GetVariant(vs, state->vs_key);
+ } else {
+ vs = device->ff.vs;
+ state->cso.vs = vs->variant.cso;
}
+ device->pipe->bind_vs_state(device->pipe, state->cso.vs);
- /* Upload special constants needed to implement PS1.x instructions like TEXBEM,TEXBEML and BEM */
- if (device->state.ps->bumpenvmat_needed) {
- memcpy(device->state.ps_lconstf_temp, cb.user_buffer, cb.buffer_size);
- memcpy(&device->state.ps_lconstf_temp[4 * 8], &device->state.bumpmap_vars, sizeof(device->state.bumpmap_vars));
-
- cb.user_buffer = device->state.ps_lconstf_temp;
+ if (state->rs[NINED3DRS_VSPOINTSIZE] != vs->point_size) {
+ state->rs[NINED3DRS_VSPOINTSIZE] = vs->point_size;
+ changed_group |= NINE_STATE_RASTERIZER;
}
- if (!device->driver_caps.user_cbufs) {
- u_upload_data(device->constbuf_uploader,
- 0,
- cb.buffer_size,
- cb.user_buffer,
- &cb.buffer_offset,
- &cb.buffer);
- u_upload_unmap(device->constbuf_uploader);
- cb.user_buffer = NULL;
- }
+ if ((state->bound_samplers_mask_vs & vs->sampler_mask) != vs->sampler_mask)
+ /* Bound dummy sampler. */
+ changed_group |= NINE_STATE_SAMPLER;
+ return changed_group;
+}
- pipe->set_constant_buffer(pipe, PIPE_SHADER_FRAGMENT, 0, &cb);
+static inline uint32_t
+update_ps(struct NineDevice9 *device)
+{
+ struct nine_state *state = &device->state;
+ struct NinePixelShader9 *ps = state->ps;
+ uint32_t changed_group = 0;
- if (device->state.changed.ps_const_f) {
- struct nine_range *r = device->state.changed.ps_const_f;
- struct nine_range *p = r;
- while (p->next)
- p = p->next;
- nine_range_pool_put_chain(&device->range_pool, r, p);
- device->state.changed.ps_const_f = NULL;
+ if (likely(ps)) {
+ state->cso.ps = NinePixelShader9_GetVariant(ps, state->ps_key);
+ } else {
+ ps = device->ff.ps;
+ state->cso.ps = ps->variant.cso;
}
- state->changed.group &= ~NINE_STATE_PS_CONST;
+ device->pipe->bind_fs_state(device->pipe, state->cso.ps);
+
+ if ((state->bound_samplers_mask_ps & ps->sampler_mask) != ps->sampler_mask)
+ /* Bound dummy sampler. */
+ changed_group |= NINE_STATE_SAMPLER;
+ return changed_group;
}
static void
@@ -905,6 +904,22 @@ commit_index_buffer(struct NineDevice9 *device)
pipe->set_index_buffer(pipe, NULL);
}
+static inline void
+commit_vs_constants(struct NineDevice9 *device)
+{
+ struct pipe_context *pipe = device->pipe;
+
+ pipe->set_constant_buffer(pipe, PIPE_SHADER_VERTEX, 0, &device->state.pipe.cb_vs);
+}
+
+static inline void
+commit_ps_constants(struct NineDevice9 *device)
+{
+ struct pipe_context *pipe = device->pipe;
+
+ pipe->set_constant_buffer(pipe, PIPE_SHADER_FRAGMENT, 0, &device->state.pipe.cb_ps);
+}
+
/* State Update */
#define NINE_STATE_FREQ_GROUP_0 \
@@ -1034,14 +1049,14 @@ nine_update_state(struct NineDevice9 *device)
if (device->prefer_user_constbuf) {
if ((group & (NINE_STATE_VS_CONST | NINE_STATE_VS)) && state->vs)
- update_vs_constants_userbuf(device);
+ prepare_vs_constants_userbuf(device);
if ((group & (NINE_STATE_PS_CONST | NINE_STATE_PS)) && state->ps)
- update_ps_constants_userbuf(device);
+ prepare_ps_constants_userbuf(device);
} else {
if ((group & NINE_STATE_VS_CONST) && state->vs)
- update_constants(device, PIPE_SHADER_VERTEX);
+ upload_constants(device, PIPE_SHADER_VERTEX);
if ((group & NINE_STATE_PS_CONST) && state->ps)
- update_constants(device, PIPE_SHADER_FRAGMENT);
+ upload_constants(device, PIPE_SHADER_FRAGMENT);
}
}
if (state->changed.vtxbuf)
@@ -1053,6 +1068,10 @@ nine_update_state(struct NineDevice9 *device)
commit_dsa(device);
if (state->commit & NINE_STATE_COMMIT_RASTERIZER)
commit_rasterizer(device);
+ if (state->commit & NINE_STATE_COMMIT_CONST_VS)
+ commit_vs_constants(device);
+ if (state->commit & NINE_STATE_COMMIT_CONST_PS)
+ commit_ps_constants(device);
state->commit = 0;
@@ -1219,6 +1238,18 @@ static const DWORD nine_samp_state_defaults[NINED3DSAMP_LAST + 1] =
[NINED3DSAMP_MINLOD] = 0,
[NINED3DSAMP_SHADOW] = 0
};
+
+void nine_state_restore_non_cso(struct NineDevice9 *device)
+{
+ struct nine_state *state = &device->state;
+
+ state->changed.group = NINE_STATE_ALL;
+ state->changed.vtxbuf = (1ULL << device->caps.MaxStreams) - 1;
+ state->changed.ucp = (1 << PIPE_MAX_CLIP_PLANES) - 1;
+ state->changed.texture = NINE_PS_SAMPLERS_MASK | NINE_VS_SAMPLERS_MASK;
+ state->commit |= NINE_STATE_COMMIT_CONST_VS | NINE_STATE_COMMIT_CONST_PS;
+}
+
void
nine_state_set_defaults(struct NineDevice9 *device, const D3DCAPS9 *caps,
boolean is_reset)
@@ -1256,6 +1287,9 @@ nine_state_set_defaults(struct NineDevice9 *device, const D3DCAPS9 *caps,
/* Set changed flags to initialize driver.
*/
state->changed.group = NINE_STATE_ALL;
+ state->changed.vtxbuf = (1ULL << device->caps.MaxStreams) - 1;
+ state->changed.ucp = (1 << PIPE_MAX_CLIP_PLANES) - 1;
+ state->changed.texture = NINE_PS_SAMPLERS_MASK | NINE_VS_SAMPLERS_MASK;
state->ff.changed.transform[0] = ~0;
state->ff.changed.transform[D3DTS_WORLD / 32] |= 1 << (D3DTS_WORLD % 32);
@@ -1272,6 +1306,23 @@ nine_state_set_defaults(struct NineDevice9 *device, const D3DCAPS9 *caps,
state->dummy_vbo_bound_at = -1;
state->vbo_bound_done = FALSE;
}
+
+ if (!device->prefer_user_constbuf) {
+ /* fill cb_vs and cb_ps for the non user constbuf path */
+ struct pipe_constant_buffer cb;
+
+ cb.buffer_offset = 0;
+ cb.buffer_size = device->vs_const_size;
+ cb.buffer = device->constbuf_vs;
+ cb.user_buffer = NULL;
+ state->pipe.cb_vs = cb;
+
+ cb.buffer_size = device->ps_const_size;
+ cb.buffer = device->constbuf_ps;
+ state->pipe.cb_ps = cb;
+
+ state->commit |= NINE_STATE_COMMIT_CONST_VS | NINE_STATE_COMMIT_CONST_PS;
+ }
}
void
diff --git a/src/gallium/state_trackers/nine/nine_state.h b/src/gallium/state_trackers/nine/nine_state.h
index 60e5d8f..ff6c180 100644
--- a/src/gallium/state_trackers/nine/nine_state.h
+++ b/src/gallium/state_trackers/nine/nine_state.h
@@ -81,6 +81,8 @@
#define NINE_STATE_COMMIT_DSA (1 << 0)
#define NINE_STATE_COMMIT_RASTERIZER (1 << 1)
#define NINE_STATE_COMMIT_BLEND (1 << 2)
+#define NINE_STATE_COMMIT_CONST_VS (1 << 3)
+#define NINE_STATE_COMMIT_CONST_PS (1 << 4)
#define NINE_MAX_SIMULTANEOUS_RENDERTARGETS 4
@@ -218,6 +220,8 @@ struct nine_state
struct pipe_depth_stencil_alpha_state dsa;
struct pipe_rasterizer_state rast;
struct pipe_blend_state blend;
+ struct pipe_constant_buffer cb_vs;
+ struct pipe_constant_buffer cb_ps;
} pipe;
};
@@ -235,6 +239,7 @@ struct NineDevice9;
void nine_update_state_framebuffer(struct NineDevice9 *);
boolean nine_update_state(struct NineDevice9 *);
+void nine_state_restore_non_cso(struct NineDevice9 *device);
void nine_state_set_defaults(struct NineDevice9 *, const D3DCAPS9 *,
boolean is_reset);
void nine_state_clear(struct nine_state *, const boolean device);
diff --git a/src/gallium/state_trackers/nine/swapchain9.c b/src/gallium/state_trackers/nine/swapchain9.c
index 6f8e066..3f5be26 100644
--- a/src/gallium/state_trackers/nine/swapchain9.c
+++ b/src/gallium/state_trackers/nine/swapchain9.c
@@ -597,7 +597,7 @@ handle_draw_cursor_and_hud( struct NineSwapChain9 *This, struct pipe_resource *r
if (device->hud && resource) {
hud_draw(device->hud, resource); /* XXX: no offset */
/* HUD doesn't clobber stipple */
- NineDevice9_RestoreNonCSOState(device, ~0x2);
+ nine_state_restore_non_cso(device);
}
}
--
2.1.0
More information about the mesa-dev
mailing list