/* vertex shader driver params: */
IR3_DP_VTXID_BASE = 0,
IR3_DP_VTXCNT_MAX = 1,
+ IR3_DP_INSTID_BASE = 2,
/* user-clip-plane components, up to 8x vec4's: */
IR3_DP_UCP0_X = 4,
/* .... */
#define IR3_MAX_SHADER_BUFFERS 32
#define IR3_MAX_SHADER_IMAGES 32
#define IR3_MAX_SO_BUFFERS 4
+#define IR3_MAX_SO_STREAMS 4
#define IR3_MAX_SO_OUTPUTS 64
-#define IR3_MAX_CONSTANT_BUFFERS 32
+#define IR3_MAX_UBO_PUSH_RANGES 32
/**
* Describes the layout of shader consts. This includes:
- * + Driver lowered UBO ranges
+ * + User consts + driver lowered UBO ranges
* + SSBO sizes
* + Image sizes/dimensions
* + Driver params (ie. IR3_DP_*)
*/
struct ir3_const_state {
unsigned num_ubos;
+ unsigned num_reserved_user_consts;
unsigned num_driver_params; /* scalar */
struct {
* encode the return type (in 3 bits) but it hasn't been verified yet.
*/
#define IR3_SAMPLER_PREFETCH_CMD 0x4
+#define IR3_SAMPLER_BINDLESS_PREFETCH_CMD 0x6
/**
* Stream output for texture sampling pre-dispatches.
uint8_t src;
uint8_t samp_id;
uint8_t tex_id;
+ uint16_t samp_bindless_id;
+ uint16_t tex_bindless_id;
uint8_t dst;
uint8_t wrmask;
uint8_t half_precision;
/* Configuration key used to identify a shader variant.. different
* shader variants can be used to implement features not supported
* in hw (two sided color), binning-pass vertex shader, etc.
+ *
+ * When adding to this struct, please update ir3_shader_variant()'s debug
+ * output.
*/
struct ir3_shader_key {
union {
unsigned sample_shading : 1;
unsigned msaa : 1;
unsigned color_two_side : 1;
- unsigned half_precision : 1;
/* used when shader needs to handle flat varyings (a4xx)
* for front/back color inputs to frag shader:
*/
uint16_t vastc_srgb, fastc_srgb;
};
+static inline unsigned
+ir3_tess_mode(unsigned gl_tess_mode)
+{
+ switch (gl_tess_mode) {
+ case GL_ISOLINES:
+ return IR3_TESS_ISOLINES;
+ case GL_TRIANGLES:
+ return IR3_TESS_TRIANGLES;
+ case GL_QUADS:
+ return IR3_TESS_QUADS;
+ default:
+ unreachable("bad tessmode");
+ }
+}
+
static inline bool
ir3_shader_key_equal(struct ir3_shader_key *a, struct ir3_shader_key *b)
{
if (last_key->color_two_side != key->color_two_side)
return true;
- if (last_key->half_precision != key->half_precision)
- return true;
-
if (last_key->rasterflat != key->rasterflat)
return true;
return false;
}
-/* clears shader-key flags which don't apply to the given shader
- * stage
- */
-static inline void
-ir3_normalize_key(struct ir3_shader_key *key, gl_shader_stage type)
-{
- switch (type) {
- case MESA_SHADER_FRAGMENT:
- if (key->has_per_samp) {
- key->vsaturate_s = 0;
- key->vsaturate_t = 0;
- key->vsaturate_r = 0;
- key->vastc_srgb = 0;
- key->vsamples = 0;
- key->has_gs = false; /* FS doesn't care */
- key->tessellation = IR3_TESS_NONE;
- }
- break;
- case MESA_SHADER_VERTEX:
- case MESA_SHADER_GEOMETRY:
- key->color_two_side = false;
- key->half_precision = false;
- key->rasterflat = false;
- if (key->has_per_samp) {
- key->fsaturate_s = 0;
- key->fsaturate_t = 0;
- key->fsaturate_r = 0;
- key->fastc_srgb = 0;
- key->fsamples = 0;
- }
-
- /* VS and GS only care about whether or not we're tessellating. */
- key->tessellation = !!key->tessellation;
- break;
- case MESA_SHADER_TESS_CTRL:
- case MESA_SHADER_TESS_EVAL:
- key->color_two_side = false;
- key->half_precision = false;
- key->rasterflat = false;
- if (key->has_per_samp) {
- key->fsaturate_s = 0;
- key->fsaturate_t = 0;
- key->fsaturate_r = 0;
- key->fastc_srgb = 0;
- key->fsamples = 0;
- key->vsaturate_s = 0;
- key->vsaturate_t = 0;
- key->vsaturate_r = 0;
- key->vastc_srgb = 0;
- key->vsamples = 0;
- }
- break;
- default:
- /* TODO */
- break;
- }
-}
-
/**
* On a4xx+a5xx, Images share state with textures and SSBOs:
*
*/
struct ir3_ibo_mapping {
#define IBO_INVALID 0xff
- /* Maps logical SSBO state to hw state: */
- uint8_t ssbo_to_ibo[IR3_MAX_SHADER_BUFFERS];
+ /* Maps logical SSBO state to hw tex state: */
uint8_t ssbo_to_tex[IR3_MAX_SHADER_BUFFERS];
- /* Maps logical Image state to hw state: */
- uint8_t image_to_ibo[IR3_MAX_SHADER_IMAGES];
+ /* Maps logical Image state to hw tex state: */
uint8_t image_to_tex[IR3_MAX_SHADER_IMAGES];
/* Maps hw state back to logical SSBO or Image state:
* hw slot is used for SSBO state vs Image state.
*/
#define IBO_SSBO 0x80
- uint8_t ibo_to_image[32];
uint8_t tex_to_image[32];
- uint8_t num_ibo;
uint8_t num_tex; /* including real textures */
uint8_t tex_base; /* the number of real textures, ie. image/ssbo start here */
};
struct ir3_info info;
struct ir3 *ir;
+ /* The actual binary shader instructions, size given by info.sizedwords: */
+ uint32_t *bin;
+
/* Levels of nesting of flow control:
*/
unsigned branchstack;
* + From the vert shader, we only need the output regid
*/
- bool frag_coord, frag_face, color0_mrt;
+ bool frag_face, color0_mrt;
+ uint8_t fragcoord_compmask;
/* NOTE: for input/outputs, slot is:
* gl_vert_attrib - for VS inputs
/* do we have one or more SSBO instructions: */
bool has_ssbo;
+ /* Which bindless resources are used, for filling out sp_xs_config */
+ bool bindless_tex;
+ bool bindless_samp;
+ bool bindless_ibo;
+ bool bindless_ubo;
+
/* do we need derivatives: */
bool need_pixlod;
- /* do we have kill, image write, etc (which prevents early-z): */
+ bool need_fine_derivatives;
+
+ /* do we have image write, etc (which prevents early-z): */
bool no_earlyz;
+ /* do we have kill, which also prevents early-z, but not necessarily
+ * early-lrz (as long as lrz-write is disabled, which must be handled
+ * outside of ir3. Unlike other no_earlyz cases, kill doesn't have
+ * side effects that prevent early-lrz discard.
+ */
+ bool has_kill;
+
bool per_samp;
/* for astc srgb workaround, the number/base of additional
}
struct ir3_ubo_range {
- uint32_t offset; /* start offset of this block in const register file */
+ uint32_t offset; /* start offset to push in the const register file */
+ uint32_t block; /* Which constant block */
uint32_t start, end; /* range of block that's actually used */
+ uint16_t bindless_base; /* For bindless, which base register is used */
+ bool bindless;
};
struct ir3_ubo_analysis_state {
- struct ir3_ubo_range range[IR3_MAX_CONSTANT_BUFFERS];
- uint32_t enabled;
+ struct ir3_ubo_range range[IR3_MAX_UBO_PUSH_RANGES];
+ uint32_t num_enabled;
uint32_t size;
uint32_t lower_count;
uint32_t cmdstream_size; /* for per-gen backend to stash required cmdstream size */
uint32_t id;
uint32_t variant_count;
- /* so we know when we can disable TGSI related hacks: */
- bool from_tgsi;
+ /* Set by freedreno after shader_state_create, so we can emit debug info
+ * when recompiling a shader at draw time.
+ */
+ bool initial_variants_done;
struct ir3_compiler *compiler;
struct ir3_ubo_analysis_state ubo_state;
+
+ /* Number of UBOs loaded by LDC, as opposed to LDG through pointers in
+ * ubo_state.
+ */
+ unsigned num_ubos;
+
struct ir3_const_state const_state;
struct nir_shader *nir;
/* Map from driver_location to byte offset in per-primitive storage */
unsigned output_loc[32];
+
+ /* Bitmask of bits of the shader key used by this shader. Used to avoid
+ * recompiles for GL NOS that doesn't actually apply to the shader.
+ */
+ struct ir3_shader_key key_mask;
};
void * ir3_shader_assemble(struct ir3_shader_variant *v, uint32_t gpu_id);
struct ir3_shader_variant * ir3_shader_get_variant(struct ir3_shader *shader,
struct ir3_shader_key *key, bool binning_pass, bool *created);
-struct ir3_shader * ir3_shader_from_nir(struct ir3_compiler *compiler, nir_shader *nir);
+struct ir3_shader * ir3_shader_from_nir(struct ir3_compiler *compiler, nir_shader *nir,
+ unsigned reserved_user_consts, struct ir3_stream_output_info *stream_output);
void ir3_shader_destroy(struct ir3_shader *shader);
void ir3_shader_disasm(struct ir3_shader_variant *so, uint32_t *bin, FILE *out);
uint64_t ir3_shader_outputs(const struct ir3_shader *so);
* Helper/util:
*/
+/* clears shader-key flags which don't apply to the given shader.
+ */
+static inline void
+ir3_key_clear_unused(struct ir3_shader_key *key, struct ir3_shader *shader)
+{
+ uint32_t *key_bits = (uint32_t *)key;
+ uint32_t *key_mask = (uint32_t *)&shader->key_mask;
+ STATIC_ASSERT(sizeof(*key) % 4 == 0);
+ for (int i = 0; i < sizeof(*key) >> 2; i++)
+ key_bits[i] &= key_mask[i];
+}
+
static inline int
ir3_find_output(const struct ir3_shader_variant *so, gl_varying_slot slot)
{
} else if (slot == VARYING_SLOT_COL1) {
slot = VARYING_SLOT_BFC1;
} else {
- return 0;
+ return -1;
}
for (j = 0; j < so->outputs_count; j++)
debug_assert(0);
- return 0;
+ return -1;
}
static inline int
}
struct ir3_shader_linkage {
+ /* Maximum location either consumed by the fragment shader or produced by
+ * the last geometry stage, i.e. the size required for each vertex in the
+ * VPC in DWORD's.
+ */
uint8_t max_loc;
+
+ /* Number of entries in var. */
uint8_t cnt;
+
+ /* Bitset of locations used, including ones which are only used by the FS.
+ */
+ uint32_t varmask[4];
+
+ /* Map from VS output to location. */
struct {
uint8_t regid;
uint8_t compmask;
uint8_t loc;
} var[32];
+
+ /* location for fixed-function gl_PrimitiveID passthrough */
+ uint8_t primid_loc;
};
static inline void
-ir3_link_add(struct ir3_shader_linkage *l, uint8_t regid, uint8_t compmask, uint8_t loc)
+ir3_link_add(struct ir3_shader_linkage *l, uint8_t regid_, uint8_t compmask, uint8_t loc)
{
- int i = l->cnt++;
- debug_assert(i < ARRAY_SIZE(l->var));
- l->var[i].regid = regid;
- l->var[i].compmask = compmask;
- l->var[i].loc = loc;
+ for (int j = 0; j < util_last_bit(compmask); j++) {
+ uint8_t comploc = loc + j;
+ l->varmask[comploc / 32] |= 1 << (comploc % 32);
+ }
+
l->max_loc = MAX2(l->max_loc, loc + util_last_bit(compmask));
+
+ if (regid_ != regid(63, 0)) {
+ int i = l->cnt++;
+ debug_assert(i < ARRAY_SIZE(l->var));
+
+ l->var[i].regid = regid_;
+ l->var[i].compmask = compmask;
+ l->var[i].loc = loc;
+ }
}
static inline void
ir3_link_shaders(struct ir3_shader_linkage *l,
const struct ir3_shader_variant *vs,
- const struct ir3_shader_variant *fs)
+ const struct ir3_shader_variant *fs,
+ bool pack_vs_out)
{
+ /* On older platforms, varmask isn't programmed at all, and it appears
+ * that the hardware generates a mask of used VPC locations using the VS
+ * output map, and hangs if a FS bary instruction references a location
+ * not in the list. This means that we need to have a dummy entry in the
+ * VS out map for things like gl_PointCoord which aren't written by the
+ * VS. Furthermore we can't use r63.x, so just pick a random register to
+ * use if there is no VS output.
+ */
+ const unsigned default_regid = pack_vs_out ? regid(63, 0) : regid(0, 0);
int j = -1, k;
+ l->primid_loc = 0xff;
+
while (l->cnt < ARRAY_SIZE(l->var)) {
j = ir3_next_varying(fs, j);
k = ir3_find_output(vs, fs->inputs[j].slot);
- ir3_link_add(l, vs->outputs[k].regid,
+ if (k < 0 && fs->inputs[j].slot == VARYING_SLOT_PRIMITIVE_ID) {
+ l->primid_loc = fs->inputs[j].inloc;
+ }
+
+ ir3_link_add(l, k >= 0 ? vs->outputs[k].regid : default_regid,
fs->inputs[j].compmask, fs->inputs[j].inloc);
}
}
return (2 * (v->info.max_reg + 1)) + (v->info.max_half_reg + 1);
}
+static inline uint32_t
+ir3_shader_nibo(const struct ir3_shader_variant *v)
+{
+ /* The dummy variant used in binning mode won't have an actual shader. */
+ if (!v->shader)
+ return 0;
+
+ return v->shader->nir->info.num_ssbos + v->shader->nir->info.num_images;
+}
+
#endif /* IR3_SHADER_H_ */
projects / mesa.git / blobdiff