r300g: atomize texture and sampler states

src/gallium/drivers/r300/r300_blit.c

@@ -98,6 +98,8 @@ static void r300_hw_copy(struct pipe_context* pipe,
                          unsigned width, unsigned height)
 {
     struct r300_context* r300 = r300_context(pipe);
+    struct r300_textures_state* state =
+        (struct r300_textures_state*)r300->textures_state.state;
 
     /* Yeah we have to save all those states to ensure this blitter operation
      * is really transparent. The states will be restored by the blitter once
@@ -106,11 +108,11 @@ static void r300_hw_copy(struct pipe_context* pipe,
     util_blitter_save_framebuffer(r300->blitter, r300->fb_state.state);
 
     util_blitter_save_fragment_sampler_states(
-        r300->blitter, r300->sampler_count, (void**)r300->sampler_states);
+        r300->blitter, state->sampler_count, (void**)state->sampler_states);
 
     util_blitter_save_fragment_sampler_textures(
-        r300->blitter, r300->texture_count,
-        (struct pipe_texture**)r300->textures);
+        r300->blitter, state->texture_count,
+        (struct pipe_texture**)state->textures);
 
     /* Do a copy */
     util_blitter_copy(r300->blitter,

src/gallium/drivers/r300/r300_context.c

@@ -59,6 +59,7 @@ static void r300_destroy_context(struct pipe_context* context)
     FREE(r300->fb_state.state);
     FREE(r300->rs_block_state.state);
     FREE(r300->scissor_state.state);
+    FREE(r300->textures_state.state);
     FREE(r300->vertex_stream_state.state);
     FREE(r300->vap_output_state.state);
     FREE(r300->viewport_state.state);
@@ -133,6 +134,7 @@ static void r300_setup_atoms(struct r300_context* r300)
     R300_INIT_ATOM(pvs_flush, 2);
     R300_INIT_ATOM(vs_state, 0);
     R300_INIT_ATOM(texture_cache_inval, 2);
+    R300_INIT_ATOM(textures_state, 0);
 
     /* Some non-CSO atoms need explicit space to store the state locally. */
     r300->blend_color_state.state = CALLOC_STRUCT(r300_blend_color_state);
@@ -140,6 +142,7 @@ static void r300_setup_atoms(struct r300_context* r300)
     r300->fb_state.state = CALLOC_STRUCT(pipe_framebuffer_state);
     r300->rs_block_state.state = CALLOC_STRUCT(r300_rs_block);
     r300->scissor_state.state = CALLOC_STRUCT(pipe_scissor_state);
+    r300->textures_state.state = CALLOC_STRUCT(r300_textures_state);
     r300->vertex_stream_state.state = CALLOC_STRUCT(r300_vertex_stream_state);
     r300->vap_output_state.state = CALLOC_STRUCT(r300_vap_output_state);
     r300->viewport_state.state = CALLOC_STRUCT(r300_viewport_state);

src/gallium/drivers/r300/r300_context.h

@@ -118,7 +118,7 @@ struct r300_sampler_state {
     unsigned min_lod, max_lod;
 };
 
-struct r300_texture_state {
+struct r300_texture_format_state {
     uint32_t format0; /* R300_TX_FORMAT0: 0x4480 */
     uint32_t format1; /* R300_TX_FORMAT1: 0x44c0 */
     uint32_t format2; /* R300_TX_FORMAT2: 0x4500 */
@@ -134,6 +134,25 @@ struct r300_texture_fb_state {
     uint32_t zb_format; /* R300_ZB_FORMAT */
 };
 
+struct r300_textures_state {
+    /* Textures. */
+    struct r300_texture *textures[8];
+    int texture_count;
+    /* Sampler states. */
+    struct r300_sampler_state *sampler_states[8];
+    int sampler_count;
+
+    /* These is the merge of the texture and sampler states. */
+    unsigned count;
+    uint32_t tx_enable;         /* R300_TX_ENABLE: 0x4101 */
+    struct r300_texture_sampler_state {
+        uint32_t format[3];     /* R300_TX_FORMAT[0-2] */
+        uint32_t filter[2];     /* R300_TX_FILTER[0-1] */
+        uint32_t border_color;  /* R300_TX_BORDER_COLOR: 0x45c0 */
+        uint32_t tile_config;   /* R300_TX_OFFSET (subset thereof) */
+    } regs[8];
+};
+
 struct r300_vertex_stream_state {
     /* R300_VAP_PROG_STREAK_CNTL_[0-7] */
     uint32_t vap_prog_stream_cntl[8];
@@ -165,10 +184,6 @@ struct r300_ztop_state {
 
 #define R300_NEW_FRAGMENT_SHADER 0x00000020
 #define R300_NEW_FRAGMENT_SHADER_CONSTANTS    0x00000040
-#define R300_NEW_SAMPLER         0x00000200
-#define R300_ANY_NEW_SAMPLERS    0x0001fe00
-#define R300_NEW_TEXTURE         0x00040000
-#define R300_ANY_NEW_TEXTURES    0x03fc0000
 #define R300_NEW_VERTEX_SHADER_CONSTANTS    0x10000000
 #define R300_NEW_QUERY           0x40000000
 #define R300_NEW_KITCHEN_SINK    0x7fffffff
@@ -254,7 +269,7 @@ struct r300_texture {
     struct pipe_buffer* buffer;
 
     /* Registers carrying texture format data. */
-    struct r300_texture_state state;
+    struct r300_texture_format_state state;
     struct r300_texture_fb_state fb_state;
 
     /* Buffer tiling */
@@ -306,14 +321,10 @@ struct r300_context {
     struct r300_atom rs_state;
     /* RS block state. */
     struct r300_atom rs_block_state;
-    /* Sampler states. */
-    struct r300_sampler_state* sampler_states[8];
-    int sampler_count;
     /* Scissor state. */
     struct r300_atom scissor_state;
-    /* Texture states. */
-    struct r300_texture* textures[8];
-    int texture_count;
+    /* Textures state. */
+    struct r300_atom textures_state;
     /* Vertex stream formatting state. */
     struct r300_atom vertex_stream_state;
     /* VAP (vertex shader) output mapping state. */

src/gallium/drivers/r300/r300_emit.c

@@ -146,6 +146,8 @@ static const float * get_shader_constant(
 {
     struct r300_viewport_state* viewport =
         (struct r300_viewport_state*)r300->viewport_state.state;
+    struct r300_textures_state* texstate =
+        (struct r300_textures_state*)r300->textures_state.state;
     static float vec[4] = { 0.0, 0.0, 0.0, 1.0 };
     struct pipe_texture *tex;
 
@@ -161,7 +163,7 @@ static const float * get_shader_constant(
                 /* Factor for converting rectangle coords to
                  * normalized coords. Should only show up on non-r500. */
                 case RC_STATE_R300_TEXRECT_FACTOR:
-                    tex = &r300->textures[constant->u.State[1]]->tex;
+                    tex = &texstate->textures[constant->u.State[1]]->tex;
                     vec[0] = 1.0 / tex->width0;
                     vec[1] = 1.0 / tex->height0;
                     break;
@@ -728,49 +730,35 @@ void r300_emit_scissor_state(struct r300_context* r300,
     END_CS;
 }
 
-void r300_emit_texture(struct r300_context* r300,
-                       struct r300_sampler_state* sampler,
-                       struct r300_texture* tex,
-                       unsigned offset)
+void r300_emit_textures_state(struct r300_context *r300,
+                              unsigned size, void *state)
 {
-    uint32_t filter0 = sampler->filter0;
-    uint32_t format0 = tex->state.format0;
-    unsigned min_level, max_level;
+    struct r300_textures_state *allstate = (struct r300_textures_state*)state;
+    struct r300_texture_sampler_state *texstate;
+    unsigned i;
     CS_LOCALS(r300);
 
-    /* to emulate 1D textures through 2D ones correctly */
-    if (tex->tex.target == PIPE_TEXTURE_1D) {
-        filter0 &= ~R300_TX_WRAP_T_MASK;
-        filter0 |= R300_TX_WRAP_T(R300_TX_CLAMP_TO_EDGE);
-    }
+    BEGIN_CS(size);
+    OUT_CS_REG(R300_TX_ENABLE, allstate->tx_enable);
 
-    if (tex->is_npot) {
-        /* NPOT textures don't support mip filter, unfortunately.
-         * This prevents incorrect rendering. */
-        filter0 &= ~R300_TX_MIN_FILTER_MIP_MASK;
-    } else {
-        /* determine min/max levels */
-        /* the MAX_MIP level is the largest (finest) one */
-        max_level = MIN2(sampler->max_lod, tex->tex.last_level);
-        min_level = MIN2(sampler->min_lod, max_level);
-        format0 |= R300_TX_NUM_LEVELS(max_level);
-        filter0 |= R300_TX_MAX_MIP_LEVEL(min_level);
-    }
+    for (i = 0; i < allstate->count; i++) {
+        if ((1 << i) & allstate->tx_enable) {
+            texstate = &allstate->regs[i];
+
+            OUT_CS_REG(R300_TX_FILTER0_0 + (i * 4), texstate->filter[0]);
+            OUT_CS_REG(R300_TX_FILTER1_0 + (i * 4), texstate->filter[1]);
+            OUT_CS_REG(R300_TX_BORDER_COLOR_0 + (i * 4),
+                       texstate->border_color);
+
+            OUT_CS_REG(R300_TX_FORMAT0_0 + (i * 4), texstate->format[0]);
+            OUT_CS_REG(R300_TX_FORMAT1_0 + (i * 4), texstate->format[1]);
+            OUT_CS_REG(R300_TX_FORMAT2_0 + (i * 4), texstate->format[2]);
 
-    BEGIN_CS(16);
-    OUT_CS_REG(R300_TX_FILTER0_0 + (offset * 4), filter0 |
-        (offset << 28));
-    OUT_CS_REG(R300_TX_FILTER1_0 + (offset * 4), sampler->filter1);
-    OUT_CS_REG(R300_TX_BORDER_COLOR_0 + (offset * 4), sampler->border_color);
-
-    OUT_CS_REG(R300_TX_FORMAT0_0 + (offset * 4), format0);
-    OUT_CS_REG(R300_TX_FORMAT1_0 + (offset * 4), tex->state.format1);
-    OUT_CS_REG(R300_TX_FORMAT2_0 + (offset * 4), tex->state.format2);
-    OUT_CS_REG_SEQ(R300_TX_OFFSET_0 + (offset * 4), 1);
-    OUT_CS_RELOC(tex->buffer,
-                 R300_TXO_MACRO_TILE(tex->macrotile) |
-                 R300_TXO_MICRO_TILE(tex->microtile),
-                 RADEON_GEM_DOMAIN_GTT | RADEON_GEM_DOMAIN_VRAM, 0, 0);
+            OUT_CS_REG_SEQ(R300_TX_OFFSET_0 + (i * 4), 1);
+            OUT_CS_RELOC(allstate->textures[i]->buffer, texstate->tile_config,
+                         RADEON_GEM_DOMAIN_GTT | RADEON_GEM_DOMAIN_VRAM, 0, 0);
+        }
+    }
     END_CS;
 }
 
@@ -976,27 +964,6 @@ void r300_emit_viewport_state(struct r300_context* r300,
     }
 }
 
-void r300_emit_texture_count(struct r300_context* r300)
-{
-    uint32_t tx_enable = 0;
-    int i;
-    CS_LOCALS(r300);
-
-    /* Notice that texture_count and sampler_count are just sizes
-     * of the respective arrays. We still have to check for the individual
-     * elements. */
-    for (i = 0; i < MIN2(r300->sampler_count, r300->texture_count); i++) {
-        if (r300->textures[i]) {
-            tx_enable |= 1 << i;
-        }
-    }
-
-    BEGIN_CS(2);
-    OUT_CS_REG(R300_TX_ENABLE, tx_enable);
-    END_CS;
-
-}
-
 void r300_emit_ztop_state(struct r300_context* r300,
                           unsigned size, void* state)
 {
@@ -1023,6 +990,8 @@ void r300_emit_buffer_validate(struct r300_context *r300,
 {
     struct pipe_framebuffer_state* fb =
         (struct pipe_framebuffer_state*)r300->fb_state.state;
+    struct r300_textures_state *texstate =
+        (struct r300_textures_state*)r300->textures_state.state;
     struct r300_texture* tex;
     struct pipe_vertex_buffer *vbuf = r300->vertex_buffer;
     struct pipe_vertex_element *velem = r300->vertex_element;
@@ -1055,9 +1024,9 @@ validate:
         }
     }
     /* ...textures... */
-    for (i = 0; i < r300->texture_count; i++) {
-        tex = r300->textures[i];
-        if (!tex)
+    for (i = 0; i < texstate->count; i++) {
+        tex = texstate->textures[i];
+        if (!tex || !texstate->sampler_states[i])
             continue;
         if (!r300->winsys->add_buffer(r300->winsys, tex->buffer,
                     RADEON_GEM_DOMAIN_GTT | RADEON_GEM_DOMAIN_VRAM, 0)) {
@@ -1136,7 +1105,6 @@ void r300_emit_dirty_state(struct r300_context* r300)
 {
     struct r300_screen* r300screen = r300_screen(r300->context.screen);
     struct r300_atom* atom;
-    unsigned i;
 
     if (r300->dirty_state & R300_NEW_QUERY) {
         r300_emit_query_start(r300);
@@ -1171,27 +1139,6 @@ void r300_emit_dirty_state(struct r300_context* r300)
         r300->dirty_state &= ~R300_NEW_FRAGMENT_SHADER_CONSTANTS;
     }
 
-    /* Samplers and textures are tracked separately but emitted together. */
-    if (r300->dirty_state &
-            (R300_ANY_NEW_SAMPLERS | R300_ANY_NEW_TEXTURES)) {
-        r300_emit_texture_count(r300);
-
-        for (i = 0; i < MIN2(r300->sampler_count, r300->texture_count); i++) {
-  	    if (r300->dirty_state &
-		((R300_NEW_SAMPLER << i) | (R300_NEW_TEXTURE << i))) {
-		if (r300->textures[i]) {
-		    r300_emit_texture(r300,
-				      r300->sampler_states[i],
-				      r300->textures[i],
-				      i);
-                }
-                r300->dirty_state &=
-                    ~((R300_NEW_SAMPLER << i) | (R300_NEW_TEXTURE << i));
-            }
-        }
-        r300->dirty_state &= ~(R300_ANY_NEW_SAMPLERS | R300_ANY_NEW_TEXTURES);
-    }
-
     if (r300->dirty_state & R300_NEW_VERTEX_SHADER_CONSTANTS) {
         struct r300_vertex_shader* vs = r300->vs_state.state;
         r300_emit_vs_constant_buffer(r300, &vs->code.constants);

src/gallium/drivers/r300/r300_emit.h

@@ -70,10 +70,8 @@ void r300_emit_rs_block_state(struct r300_context* r300,
 void r300_emit_scissor_state(struct r300_context* r300,
                              unsigned size, void* state);
 
-void r300_emit_texture(struct r300_context* r300,
-                       struct r300_sampler_state* sampler,
-                       struct r300_texture* tex,
-                       unsigned offset);
+void r300_emit_textures_state(struct r300_context *r300,
+                              unsigned size, void *state);
 
 void r300_emit_vertex_buffer(struct r300_context* r300);
 
@@ -83,9 +81,6 @@ void r300_emit_vertex_stream_state(struct r300_context* r300,
 void r300_emit_vap_output_state(struct r300_context* r300,
                                unsigned size, void* state);
 
-void r300_emit_vertex_program_code(struct r300_context* r300,
-                                   struct r300_vertex_program_code* code);
-
 void r300_emit_vs_constant_buffer(struct r300_context* r300,
                                   struct rc_constant_list* constants);
 
@@ -94,8 +89,6 @@ void r300_emit_vs_state(struct r300_context* r300, unsigned size, void* state);
 void r300_emit_viewport_state(struct r300_context* r300,
                               unsigned size, void* state);
 
-void r300_emit_texture_count(struct r300_context* r300);
-
 void r300_emit_ztop_state(struct r300_context* r300,
                           unsigned size, void* state);
 

src/gallium/drivers/r300/r300_fs.c

@@ -133,10 +133,13 @@ static void get_compare_state(
     struct r300_fragment_program_external_state* state,
     unsigned shadow_samplers)
 {
+    struct r300_textures_state *texstate =
+        (struct r300_textures_state*)r300->textures_state.state;
+
     memset(state, 0, sizeof(*state));
 
-    for (int i = 0; i < r300->sampler_count; i++) {
-        struct r300_sampler_state* s = r300->sampler_states[i];
+    for (int i = 0; i < texstate->sampler_count; i++) {
+        struct r300_sampler_state* s = texstate->sampler_states[i];
 
         if (s && s->state.compare_mode == PIPE_TEX_COMPARE_R_TO_TEXTURE) {
             /* XXX Gallium doesn't provide us with any information regarding

src/gallium/drivers/r300/r300_state.c

@@ -859,7 +859,7 @@ static void*
                                                    state->max_anisotropy > 0);
 
     /* Unfortunately, r300-r500 don't support floating-point mipmap lods. */
-    /* We must pass these to the emit function to clamp them properly. */
+    /* We must pass these to the merge function to clamp them properly. */
     sampler->min_lod = MAX2((unsigned)state->min_lod, 0);
     sampler->max_lod = MAX2((unsigned)ceilf(state->max_lod), 0);
 
@@ -885,23 +885,20 @@ static void r300_bind_sampler_states(struct pipe_context* pipe,
                                      void** states)
 {
     struct r300_context* r300 = r300_context(pipe);
-    int i;
+    struct r300_textures_state* state =
+        (struct r300_textures_state*)r300->textures_state.state;
 
     if (count > 8) {
         return;
     }
 
-    for (i = 0; i < count; i++) {
-        if (r300->sampler_states[i] != states[i]) {
-            r300->sampler_states[i] = (struct r300_sampler_state*)states[i];
-            r300->dirty_state |= (R300_NEW_SAMPLER << i);
-        }
-    }
+    memcpy(state->sampler_states, states, sizeof(void*) * count);
+    state->sampler_count = count;
 
-    r300->sampler_count = count;
+    r300->textures_state.dirty = TRUE;
 
     /* Pick a fragment shader based on the texture compare state. */
-    if (r300->fs && (r300->dirty_state & R300_ANY_NEW_SAMPLERS)) {
+    if (r300->fs && count) {
         if (r300_pick_fragment_shader(r300)) {
             r300->dirty_state |= R300_NEW_FRAGMENT_SHADER |
                                  R300_NEW_FRAGMENT_SHADER_CONSTANTS;
@@ -925,24 +922,25 @@ static void r300_set_sampler_textures(struct pipe_context* pipe,
                                       struct pipe_texture** texture)
 {
     struct r300_context* r300 = r300_context(pipe);
+    struct r300_textures_state* state =
+        (struct r300_textures_state*)r300->textures_state.state;
+    unsigned i;
     boolean is_r500 = r300_screen(r300->context.screen)->caps->is_r500;
     boolean dirty_tex = FALSE;
-    int i;
 
     /* XXX magic num */
     if (count > 8) {
         return;
     }
-    
+
     for (i = 0; i < count; i++) {
-        if (r300->textures[i] != (struct r300_texture*)texture[i]) {
-            pipe_texture_reference((struct pipe_texture**)&r300->textures[i],
-                texture[i]);
-            r300->dirty_state |= (R300_NEW_TEXTURE << i);
+        if (state->textures[i] != (struct r300_texture*)texture[i]) {
+            pipe_texture_reference((struct pipe_texture**)&state->textures[i],
+                                   texture[i]);
             dirty_tex = TRUE;
 
-            /* R300-specific - set the texrect factor in a fragment shader */
-            if (!is_r500 && r300->textures[i]->is_npot) {
+            /* R300-specific - set the texrect factor in the fragment shader */
+            if (!is_r500 && state->textures[i]->is_npot) {
                 /* XXX It would be nice to re-emit just 1 constant,
                  * XXX not all of them */
                 r300->dirty_state |= R300_NEW_FRAGMENT_SHADER_CONSTANTS;
@@ -951,14 +949,15 @@ static void r300_set_sampler_textures(struct pipe_context* pipe,
     }
 
     for (i = count; i < 8; i++) {
-        if (r300->textures[i]) {
-            pipe_texture_reference((struct pipe_texture**)&r300->textures[i],
+        if (state->textures[i]) {
+            pipe_texture_reference((struct pipe_texture**)&state->textures[i],
                 NULL);
-            r300->dirty_state |= (R300_NEW_TEXTURE << i);
         }
     }
 
-    r300->texture_count = count;
+    state->texture_count = count;
+
+    r300->textures_state.dirty = TRUE;
 
     if (dirty_tex) {
         r300->texture_cache_inval.dirty = TRUE;

src/gallium/drivers/r300/r300_state_derived.c

@@ -431,7 +431,6 @@ static void r300_update_rs_block(struct r300_context* r300,
     if (memcmp(r300->rs_block_state.state, &rs, sizeof(struct r300_rs_block))) {
         memcpy(r300->rs_block_state.state, &rs, sizeof(struct r300_rs_block));
         r300->rs_block_state.size = 5 + count*2;
-        r300->rs_block_state.dirty = TRUE;
     }
 }
 
@@ -529,6 +528,63 @@ static void r300_update_ztop(struct r300_context* r300)
     r300->ztop_state.dirty = TRUE;
 }
 
+static void r300_merge_textures_and_samplers(struct r300_context* r300)
+{
+    struct r300_textures_state *state =
+        (struct r300_textures_state*)r300->textures_state.state;
+    struct r300_texture_sampler_state *texstate;
+    struct r300_sampler_state *sampler;
+    struct r300_texture *tex;
+    unsigned min_level, max_level, i, size;
+    unsigned count = MIN2(state->texture_count, state->sampler_count);
+
+    state->tx_enable = 0;
+    size = 2;
+
+    for (i = 0; i < count; i++) {
+        if (state->textures[i] && state->sampler_states[i]) {
+            state->tx_enable |= 1 << i;
+
+            tex = state->textures[i];
+            sampler = state->sampler_states[i];
+
+            texstate = &state->regs[i];
+            memcpy(texstate->format, &tex->state, sizeof(uint32_t)*3);
+            texstate->filter[0] = sampler->filter0;
+            texstate->filter[1] = sampler->filter1;
+            texstate->border_color = sampler->border_color;
+            texstate->tile_config = R300_TXO_MACRO_TILE(tex->macrotile) |
+                                    R300_TXO_MICRO_TILE(tex->microtile);
+
+            /* to emulate 1D textures through 2D ones correctly */
+            if (tex->tex.target == PIPE_TEXTURE_1D) {
+                texstate->filter[0] &= ~R300_TX_WRAP_T_MASK;
+                texstate->filter[0] |= R300_TX_WRAP_T(R300_TX_CLAMP_TO_EDGE);
+            }
+
+            if (tex->is_npot) {
+                /* NPOT textures don't support mip filter, unfortunately.
+                 * This prevents incorrect rendering. */
+                texstate->filter[0] &= ~R300_TX_MIN_FILTER_MIP_MASK;
+            } else {
+                /* determine min/max levels */
+                /* the MAX_MIP level is the largest (finest) one */
+                max_level = MIN2(sampler->max_lod, tex->tex.last_level);
+                min_level = MIN2(sampler->min_lod, max_level);
+                texstate->format[0] |= R300_TX_NUM_LEVELS(max_level);
+                texstate->filter[0] |= R300_TX_MAX_MIP_LEVEL(min_level);
+            }
+
+            texstate->filter[0] |= i << 28;
+
+            size += 16;
+            state->count = i+1;
+        }
+    }
+
+    r300->textures_state.size = size;
+}
+
 void r300_update_derived_state(struct r300_context* r300)
 {
     if (r300->rs_block_state.dirty ||
@@ -537,5 +593,9 @@ void r300_update_derived_state(struct r300_context* r300)
         r300_update_derived_shader_state(r300);
     }
 
+    if (r300->textures_state.dirty) {
+        r300_merge_textures_and_samplers(r300);
+    }
+
     r300_update_ztop(r300);
 }

src/gallium/drivers/r300/r300_texture.c

@@ -505,7 +505,7 @@ boolean r300_is_sampler_format_supported(enum pipe_format format)
 
 static void r300_setup_texture_state(struct r300_screen* screen, struct r300_texture* tex)
 {
-    struct r300_texture_state* state = &tex->state;
+    struct r300_texture_format_state* state = &tex->state;
     struct pipe_texture *pt = &tex->tex;
     unsigned i;
     boolean is_r500 = screen->caps->is_r500;