r300g: implement TRUNC correctly
[mesa.git] / src / gallium / drivers / radeonsi / si_state.c
index 5df22dd5f3ade89295cbc6ec95aebcb90526b76e..fced24cc4ec86eba2e121ac5008054615cec7ae6 100644 (file)
@@ -996,6 +996,53 @@ static uint32_t si_colorformat_endian_swap(uint32_t colorformat)
        }
 }
 
+/* Returns the size in bits of the widest component of a CB format */
+static unsigned si_colorformat_max_comp_size(uint32_t colorformat)
+{
+       switch(colorformat) {
+       case V_028C70_COLOR_4_4_4_4:
+               return 4;
+
+       case V_028C70_COLOR_1_5_5_5:
+       case V_028C70_COLOR_5_5_5_1:
+               return 5;
+
+       case V_028C70_COLOR_5_6_5:
+               return 6;
+
+       case V_028C70_COLOR_8:
+       case V_028C70_COLOR_8_8:
+       case V_028C70_COLOR_8_8_8_8:
+               return 8;
+
+       case V_028C70_COLOR_10_10_10_2:
+       case V_028C70_COLOR_2_10_10_10:
+               return 10;
+
+       case V_028C70_COLOR_10_11_11:
+       case V_028C70_COLOR_11_11_10:
+               return 11;
+
+       case V_028C70_COLOR_16:
+       case V_028C70_COLOR_16_16:
+       case V_028C70_COLOR_16_16_16_16:
+               return 16;
+
+       case V_028C70_COLOR_8_24:
+       case V_028C70_COLOR_24_8:
+               return 24;
+
+       case V_028C70_COLOR_32:
+       case V_028C70_COLOR_32_32:
+       case V_028C70_COLOR_32_32_32_32:
+       case V_028C70_COLOR_X24_8_32_FLOAT:
+               return 32;
+       }
+
+       assert(!"Unknown maximum component size");
+       return 0;
+}
+
 static uint32_t si_translate_dbformat(enum pipe_format format)
 {
        switch (format) {
@@ -1409,6 +1456,7 @@ static void si_cb(struct r600_context *rctx, struct si_pm4_state *pm4,
        const struct util_format_description *desc;
        int i;
        unsigned blend_clamp = 0, blend_bypass = 0;
+       unsigned max_comp_size;
 
        surf = (struct r600_surface *)state->cbufs[cb];
        rtex = (struct r600_resource_texture*)state->cbufs[cb]->texture;
@@ -1549,6 +1597,17 @@ static void si_cb(struct r600_context *rctx, struct si_pm4_state *pm4,
        }
        si_pm4_set_reg(pm4, R_028C70_CB_COLOR0_INFO + cb * 0x3C, color_info);
        si_pm4_set_reg(pm4, R_028C74_CB_COLOR0_ATTRIB + cb * 0x3C, color_attrib);
+
+       /* Determine pixel shader export format */
+       max_comp_size = si_colorformat_max_comp_size(format);
+       if (ntype == V_028C70_NUMBER_SRGB ||
+           ((ntype == V_028C70_NUMBER_UNORM || ntype == V_028C70_NUMBER_SNORM) &&
+            max_comp_size <= 10) ||
+           (ntype == V_028C70_NUMBER_FLOAT && max_comp_size <= 16)) {
+               rctx->export_16bpc |= 1 << cb;
+               rctx->spi_shader_col_format |= V_028714_SPI_SHADER_FP16_ABGR << (4 * cb);
+       } else
+               rctx->spi_shader_col_format |= V_028714_SPI_SHADER_32_ABGR << (4 * cb);
 }
 
 static void si_db(struct r600_context *rctx, struct si_pm4_state *pm4,
@@ -1667,9 +1726,12 @@ static void si_set_framebuffer_state(struct pipe_context *ctx,
 
        /* build states */
        rctx->have_depth_fb = 0;
+       rctx->export_16bpc = 0;
+       rctx->spi_shader_col_format = 0;
        for (int i = 0; i < state->nr_cbufs; i++) {
                si_cb(rctx, pm4, state, i);
        }
+       assert(!(rctx->export_16bpc & ~0xff));
        si_db(rctx, pm4, state);
 
        shader_mask = 0;
@@ -1706,6 +1768,8 @@ static void si_set_framebuffer_state(struct pipe_context *ctx,
        si_pm4_set_reg(pm4, R_028200_PA_SC_WINDOW_OFFSET, 0x00000000);
        si_pm4_set_reg(pm4, R_028230_PA_SC_EDGERULE, 0xAAAAAAAA);
        si_pm4_set_reg(pm4, R_02823C_CB_SHADER_MASK, shader_mask);
+       si_pm4_set_reg(pm4, R_028714_SPI_SHADER_COL_FORMAT,
+                      rctx->spi_shader_col_format);
        si_pm4_set_reg(pm4, R_028BE0_PA_SC_AA_CONFIG, 0x00000000);
 
        si_pm4_set_state(rctx, framebuffer, pm4);
@@ -1717,77 +1781,201 @@ static void si_set_framebuffer_state(struct pipe_context *ctx,
  * shaders
  */
 
+/* Compute the key for the hw shader variant */
+static INLINE unsigned si_shader_selector_key(struct pipe_context *ctx,
+                                             struct si_pipe_shader_selector *sel)
+{
+       struct r600_context *rctx = (struct r600_context *)ctx;
+       unsigned key = 0;
+
+       if (sel->type == PIPE_SHADER_FRAGMENT) {
+               if (sel->fs_write_all)
+                       key |= rctx->framebuffer.nr_cbufs;
+               key |= rctx->export_16bpc << 4;
+               /*if (rctx->queued.named.rasterizer)
+                         key |= rctx->queued.named.rasterizer->flatshade << 12;*/
+               /*key |== rctx->two_side << 13;*/
+       }
+
+       return key;
+}
+
+/* Select the hw shader variant depending on the current state.
+ * (*dirty) is set to 1 if current variant was changed */
+int si_shader_select(struct pipe_context *ctx,
+                    struct si_pipe_shader_selector *sel,
+                    unsigned *dirty)
+{
+       unsigned key;
+       struct si_pipe_shader * shader = NULL;
+       int r;
+
+       key = si_shader_selector_key(ctx, sel);
+
+       /* Check if we don't need to change anything.
+        * This path is also used for most shaders that don't need multiple
+        * variants, it will cost just a computation of the key and this
+        * test. */
+       if (likely(sel->current && sel->current->key == key)) {
+               return 0;
+       }
+
+       /* lookup if we have other variants in the list */
+       if (sel->num_shaders > 1) {
+               struct si_pipe_shader *p = sel->current, *c = p->next_variant;
+
+               while (c && c->key != key) {
+                       p = c;
+                       c = c->next_variant;
+               }
+
+               if (c) {
+                       p->next_variant = c->next_variant;
+                       shader = c;
+               }
+       }
+
+       if (unlikely(!shader)) {
+               shader = CALLOC(1, sizeof(struct si_pipe_shader));
+               shader->selector = sel;
+
+               r = si_pipe_shader_create(ctx, shader);
+               if (unlikely(r)) {
+                       R600_ERR("Failed to build shader variant (type=%u, key=%u) %d\n",
+                                sel->type, key, r);
+                       sel->current = NULL;
+                       return r;
+               }
+
+               /* We don't know the value of fs_write_all property until we built
+                * at least one variant, so we may need to recompute the key (include
+                * rctx->framebuffer.nr_cbufs) after building first variant. */
+               if (sel->type == PIPE_SHADER_FRAGMENT &&
+                   sel->num_shaders == 0 &&
+                   shader->shader.fs_write_all) {
+                       sel->fs_write_all = 1;
+                       key = si_shader_selector_key(ctx, sel);
+               }
+
+               shader->key = key;
+               sel->num_shaders++;
+       }
+
+       if (dirty)
+               *dirty = 1;
+
+       shader->next_variant = sel->current;
+       sel->current = shader;
+
+       return 0;
+}
+
 static void *si_create_shader_state(struct pipe_context *ctx,
-                             const struct pipe_shader_state *state)
+                                   const struct pipe_shader_state *state,
+                                   unsigned pipe_shader_type)
 {
-       struct si_pipe_shader *shader = CALLOC_STRUCT(si_pipe_shader);
+       struct si_pipe_shader_selector *sel = CALLOC_STRUCT(si_pipe_shader_selector);
+       int r;
+
+       sel->type = pipe_shader_type;
+       sel->tokens = tgsi_dup_tokens(state->tokens);
+       sel->so = state->stream_output;
 
-       shader->tokens = tgsi_dup_tokens(state->tokens);
-       shader->so = state->stream_output;
+       r = si_shader_select(ctx, sel, NULL);
+       if (r) {
+           free(sel);
+           return NULL;
+       }
 
-       return shader;
+       return sel;
+}
+
+static void *si_create_fs_state(struct pipe_context *ctx,
+                               const struct pipe_shader_state *state)
+{
+       return si_create_shader_state(ctx, state, PIPE_SHADER_FRAGMENT);
+}
+
+static void *si_create_vs_state(struct pipe_context *ctx,
+                               const struct pipe_shader_state *state)
+{
+       return si_create_shader_state(ctx, state, PIPE_SHADER_VERTEX);
 }
 
 static void si_bind_vs_shader(struct pipe_context *ctx, void *state)
 {
        struct r600_context *rctx = (struct r600_context *)ctx;
-       struct si_pipe_shader *shader = state;
+       struct si_pipe_shader_selector *sel = state;
 
-       if (rctx->vs_shader == state)
+       if (rctx->vs_shader == sel)
                return;
 
        rctx->shader_dirty = true;
-       rctx->vs_shader = shader;
+       rctx->vs_shader = sel;
 
-       if (shader) {
-               si_pm4_bind_state(rctx, vs, shader->pm4);
-       }
+       if (sel && sel->current)
+               si_pm4_bind_state(rctx, vs, sel->current->pm4);
+       else
+               si_pm4_bind_state(rctx, vs, rctx->dummy_pixel_shader->pm4);
 }
 
 static void si_bind_ps_shader(struct pipe_context *ctx, void *state)
 {
        struct r600_context *rctx = (struct r600_context *)ctx;
-       struct si_pipe_shader *shader = state;
+       struct si_pipe_shader_selector *sel = state;
 
-       if (rctx->ps_shader == state)
+       if (rctx->ps_shader == sel)
                return;
 
        rctx->shader_dirty = true;
-       rctx->ps_shader = shader;
+       rctx->ps_shader = sel;
 
-       if (shader) {
-               si_pm4_bind_state(rctx, ps, shader->pm4);
-       }
+       if (sel && sel->current)
+               si_pm4_bind_state(rctx, ps, sel->current->pm4);
+       else
+               si_pm4_bind_state(rctx, ps, rctx->dummy_pixel_shader->pm4);
 }
 
+static void si_delete_shader_selector(struct pipe_context *ctx,
+                                     struct si_pipe_shader_selector *sel)
+{
+       struct r600_context *rctx = (struct r600_context *)ctx;
+       struct si_pipe_shader *p = sel->current, *c;
+
+       while (p) {
+               c = p->next_variant;
+               si_pm4_delete_state(rctx, vs, p->pm4);
+               si_pipe_shader_destroy(ctx, p);
+               free(p);
+               p = c;
+       }
+
+       free(sel->tokens);
+       free(sel);
+ }
+
 static void si_delete_vs_shader(struct pipe_context *ctx, void *state)
 {
        struct r600_context *rctx = (struct r600_context *)ctx;
-       struct si_pipe_shader *shader = (struct si_pipe_shader *)state;
+       struct si_pipe_shader_selector *sel = (struct si_pipe_shader_selector *)state;
 
-       if (rctx->vs_shader == shader) {
+       if (rctx->vs_shader == sel) {
                rctx->vs_shader = NULL;
        }
 
-       si_pm4_delete_state(rctx, vs, shader->pm4);
-       free(shader->tokens);
-       si_pipe_shader_destroy(ctx, shader);
-       free(shader);
+       si_delete_shader_selector(ctx, sel);
 }
 
 static void si_delete_ps_shader(struct pipe_context *ctx, void *state)
 {
        struct r600_context *rctx = (struct r600_context *)ctx;
-       struct si_pipe_shader *shader = (struct si_pipe_shader *)state;
+       struct si_pipe_shader_selector *sel = (struct si_pipe_shader_selector *)state;
 
-       if (rctx->ps_shader == shader) {
+       if (rctx->ps_shader == sel) {
                rctx->ps_shader = NULL;
        }
 
-       si_pm4_delete_state(rctx, ps, shader->pm4);
-       free(shader->tokens);
-       si_pipe_shader_destroy(ctx, shader);
-       free(shader);
+       si_delete_shader_selector(ctx, sel);
 }
 
 /*
@@ -2269,8 +2457,8 @@ void si_init_state_functions(struct r600_context *rctx)
 
        rctx->context.set_framebuffer_state = si_set_framebuffer_state;
 
-       rctx->context.create_vs_state = si_create_shader_state;
-       rctx->context.create_fs_state = si_create_shader_state;
+       rctx->context.create_vs_state = si_create_vs_state;
+       rctx->context.create_fs_state = si_create_fs_state;
        rctx->context.bind_vs_state = si_bind_vs_shader;
        rctx->context.bind_fs_state = si_bind_ps_shader;
        rctx->context.delete_vs_state = si_delete_vs_shader;