#include "ir3_image.h"
-/* Images get mapped into SSBO/image state (for store/atomic) and texture
- * state block (for load). To simplify things, invert the image id and
- * map it from end of state block, ie. image 0 becomes num-1, image 1
- * becomes num-2, etc. This potentially avoids needing to re-emit texture
- * state when switching shaders.
- *
- * TODO is max # of samplers and SSBOs the same. This shouldn't be hard-
- * coded. Also, since all the gl shader stages (ie. everything but CS)
- * share the same SSBO/image state block, this might require some more
- * logic if we supported images in anything other than FS..
- */
-unsigned
-ir3_get_image_slot(struct ir3_context *ctx, nir_deref_instr *deref)
-{
- unsigned int loc = 0;
- unsigned inner_size = 1;
- while (deref->deref_type != nir_deref_type_var) {
- assert(deref->deref_type == nir_deref_type_array);
- nir_const_value *const_index = nir_src_as_const_value(deref->arr.index);
- assert(const_index);
+/*
+ * SSBO/Image to/from IBO/tex hw mapping table:
+ */
- /* Go to the next instruction */
- deref = nir_deref_instr_parent(deref);
+void
+ir3_ibo_mapping_init(struct ir3_ibo_mapping *mapping, unsigned num_textures)
+{
+ memset(mapping, IBO_INVALID, sizeof(*mapping));
+ mapping->num_tex = 0;
+ mapping->tex_base = num_textures;
+}
- assert(glsl_type_is_array(deref->type));
- const unsigned array_len = glsl_get_length(deref->type);
- loc += MIN2(const_index->u32[0], array_len - 1) * inner_size;
+struct ir3_instruction *
+ir3_ssbo_to_ibo(struct ir3_context *ctx, nir_src src)
+{
+ if (ir3_bindless_resource(src)) {
+ ctx->so->bindless_ibo = true;
+ return ir3_get_src(ctx, &src)[0];
+ } else {
+ /* can this be non-const buffer_index? how do we handle that? */
+ int ssbo_idx = nir_src_as_uint(src);
+ return create_immed(ctx->block, ssbo_idx);
+ }
+}
- /* Update the inner size */
- inner_size *= array_len;
+unsigned
+ir3_ssbo_to_tex(struct ir3_ibo_mapping *mapping, unsigned ssbo)
+{
+ if (mapping->ssbo_to_tex[ssbo] == IBO_INVALID) {
+ unsigned tex = mapping->num_tex++;
+ mapping->ssbo_to_tex[ssbo] = tex;
+ mapping->tex_to_image[tex] = IBO_SSBO | ssbo;
}
+ return mapping->ssbo_to_tex[ssbo] + mapping->tex_base;
+}
- loc += deref->var->data.driver_location;
+struct ir3_instruction *
+ir3_image_to_ibo(struct ir3_context *ctx, nir_src src)
+{
+ if (ir3_bindless_resource(src)) {
+ ctx->so->bindless_ibo = true;
+ return ir3_get_src(ctx, &src)[0];
+ } else {
+ /* can this be non-const buffer_index? how do we handle that? */
+ int image_idx = nir_src_as_uint(src);
+ return create_immed(ctx->block, ctx->s->info.num_ssbos + image_idx);
+ }
+}
- /* TODO figure out real limit per generation, and don't hardcode: */
- const unsigned max_samplers = 16;
- return max_samplers - loc - 1;
+unsigned
+ir3_image_to_tex(struct ir3_ibo_mapping *mapping, unsigned image)
+{
+ if (mapping->image_to_tex[image] == IBO_INVALID) {
+ unsigned tex = mapping->num_tex++;
+ mapping->image_to_tex[image] = tex;
+ mapping->tex_to_image[tex] = image;
+ }
+ return mapping->image_to_tex[image] + mapping->tex_base;
}
/* see tex_info() for equiv logic for texture instructions.. it would be
* nice if this could be better unified..
*/
unsigned
-ir3_get_image_coords(const nir_variable *var, unsigned *flagsp)
+ir3_get_image_coords(const nir_intrinsic_instr *instr, unsigned *flagsp)
{
- const struct glsl_type *type = glsl_without_array(var->type);
- unsigned coords, flags = 0;
-
- switch (glsl_get_sampler_dim(type)) {
- case GLSL_SAMPLER_DIM_1D:
- case GLSL_SAMPLER_DIM_BUF:
- coords = 1;
- break;
- case GLSL_SAMPLER_DIM_2D:
- case GLSL_SAMPLER_DIM_RECT:
- case GLSL_SAMPLER_DIM_EXTERNAL:
- case GLSL_SAMPLER_DIM_MS:
- coords = 2;
- break;
- case GLSL_SAMPLER_DIM_3D:
- case GLSL_SAMPLER_DIM_CUBE:
+ unsigned coords = nir_image_intrinsic_coord_components(instr);
+ unsigned flags = 0;
+
+ if (coords == 3)
flags |= IR3_INSTR_3D;
- coords = 3;
- break;
- default:
- unreachable("bad sampler dim");
- return 0;
- }
- if (glsl_sampler_type_is_array(type)) {
- /* note: unlike tex_info(), adjust # of coords to include array idx: */
- coords++;
+ if (nir_intrinsic_image_array(instr))
flags |= IR3_INSTR_A;
- }
if (flagsp)
*flagsp = flags;
}
type_t
-ir3_get_image_type(const nir_variable *var)
+ir3_get_type_for_image_intrinsic(const nir_intrinsic_instr *instr)
{
- switch (glsl_get_sampler_result_type(glsl_without_array(var->type))) {
- case GLSL_TYPE_UINT:
- return TYPE_U32;
- case GLSL_TYPE_INT:
- return TYPE_S32;
- case GLSL_TYPE_FLOAT:
- return TYPE_F32;
- default:
- unreachable("bad sampler type.");
- return 0;
- }
+ const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic];
+ int bit_size = info->has_dest ? nir_dest_bit_size(instr->dest) : 32;
+ enum pipe_format format = nir_intrinsic_format(instr);
+
+ if (util_format_is_pure_uint(format))
+ return bit_size == 16 ? TYPE_U16 : TYPE_U32;
+ else if (util_format_is_pure_sint(format))
+ return bit_size == 16 ? TYPE_S16 : TYPE_S32;
+ else
+ return bit_size == 16 ? TYPE_F16 : TYPE_F32;
}
/* Returns the number of components for the different image formats
* GL_NV_image_formats extension.
*/
unsigned
-ir3_get_num_components_for_glformat(GLuint format)
+ir3_get_num_components_for_image_format(enum pipe_format format)
{
- switch (format) {
- case GL_R32F:
- case GL_R32I:
- case GL_R32UI:
- case GL_R16F:
- case GL_R16I:
- case GL_R16UI:
- case GL_R16:
- case GL_R16_SNORM:
- case GL_R8I:
- case GL_R8UI:
- case GL_R8:
- case GL_R8_SNORM:
- return 1;
-
- case GL_RG32F:
- case GL_RG32I:
- case GL_RG32UI:
- case GL_RG16F:
- case GL_RG16I:
- case GL_RG16UI:
- case GL_RG16:
- case GL_RG16_SNORM:
- case GL_RG8I:
- case GL_RG8UI:
- case GL_RG8:
- case GL_RG8_SNORM:
- return 2;
-
- case GL_R11F_G11F_B10F:
- return 3;
-
- case GL_RGBA32F:
- case GL_RGBA32I:
- case GL_RGBA32UI:
- case GL_RGBA16F:
- case GL_RGBA16I:
- case GL_RGBA16UI:
- case GL_RGBA16:
- case GL_RGBA16_SNORM:
- case GL_RGBA8I:
- case GL_RGBA8UI:
- case GL_RGBA8:
- case GL_RGBA8_SNORM:
- case GL_RGB10_A2UI:
- case GL_RGB10_A2:
- return 4;
-
- case GL_NONE:
- /* Omitting the image format qualifier is allowed on desktop GL
- * profiles. Assuming 4 components is always safe.
- */
+ if (format == PIPE_FORMAT_NONE)
return 4;
-
- default:
- /* Return 4 components also for all other formats we don't know
- * about. The format should have been validated already by
- * the higher level API, but drop a debug message just in case.
- */
- debug_printf("Unhandled GL format %u while emitting imageStore()\n",
- format);
- return 4;
- }
+ else
+ return util_format_get_nr_components(format);
}