* USE OR OTHER DEALINGS IN THE SOFTWARE.
*/
+#include "zink_context.h"
#include "zink_compiler.h"
+#include "zink_program.h"
#include "zink_screen.h"
#include "nir_to_spirv/nir_to_spirv.h"
nir_instr_remove(&instr->instr);
return true;
}
- assert(instr->intrinsic != nir_intrinsic_discard ||
- nir_block_last_instr(instr->instr.block) == &instr->instr);
+ /* a shader like this (shaders@glsl-fs-discard-04):
+
+ uniform int j, k;
+
+ void main()
+ {
+ for (int i = 0; i < j; i++) {
+ if (i > k)
+ continue;
+ discard;
+ }
+ gl_FragColor = vec4(0.0, 1.0, 0.0, 0.0);
+ }
+
+
+
+ will generate nir like:
+
+ loop {
+ //snip
+ if ssa_11 {
+ block block_5:
+ / preds: block_4 /
+ vec1 32 ssa_17 = iadd ssa_50, ssa_31
+ / succs: block_7 /
+ } else {
+ block block_6:
+ / preds: block_4 /
+ intrinsic discard () () <-- not last instruction
+ vec1 32 ssa_23 = iadd ssa_50, ssa_31 <-- dead code loop itr increment
+ / succs: block_7 /
+ }
+ //snip
+ }
+
+ which means that we can't assert like this:
+
+ assert(instr->intrinsic != nir_intrinsic_discard ||
+ nir_block_last_instr(instr->instr.block) == &instr->instr);
+
+
+ and it's unnecessary anyway since post-vtn optimizing will dce the instructions following the discard
+ */
return false;
}
.lower_flrp32 = true,
.lower_fpow = true,
.lower_fsat = true,
+ .lower_extract_byte = true,
+ .lower_extract_word = true,
+ .lower_mul_high = true,
+ .lower_rotate = true,
+ .lower_uadd_carry = true,
};
const void *
fprintf(stderr, "---8<---\n\n");
}
- return tgsi_to_nir(tokens, screen);
+ return tgsi_to_nir(tokens, screen, false);
}
static void
NIR_PASS(progress, s, nir_opt_algebraic);
NIR_PASS(progress, s, nir_opt_constant_folding);
NIR_PASS(progress, s, nir_opt_undef);
+ NIR_PASS(progress, s, zink_nir_lower_b2b);
} while (progress);
}
+/* check for a genuine gl_PointSize output vs one from nir_lower_point_size_mov */
+static bool
+check_psiz(struct nir_shader *s)
+{
+ nir_foreach_shader_out_variable(var, s) {
+ if (var->data.location == VARYING_SLOT_PSIZ) {
+ /* genuine PSIZ outputs will have this set */
+ return !!var->data.explicit_location;
+ }
+ }
+ return false;
+}
+
+/* semi-copied from iris */
+static void
+update_so_info(struct zink_shader *sh,
+ uint64_t outputs_written, bool have_psiz)
+{
+ uint8_t reverse_map[64] = {};
+ unsigned slot = 0;
+ while (outputs_written) {
+ int bit = u_bit_scan64(&outputs_written);
+ /* PSIZ from nir_lower_point_size_mov breaks stream output, so always skip it */
+ if (bit == VARYING_SLOT_PSIZ && !have_psiz)
+ continue;
+ reverse_map[slot++] = bit;
+ }
+
+ for (unsigned i = 0; i < sh->streamout.so_info.num_outputs; i++) {
+ struct pipe_stream_output *output = &sh->streamout.so_info.output[i];
+ /* Map Gallium's condensed "slots" back to real VARYING_SLOT_* enums */
+ sh->streamout.so_info_slots[i] = reverse_map[output->register_index];
+ }
+}
+
+VkShaderModule
+zink_shader_compile(struct zink_screen *screen, struct zink_shader *zs)
+{
+ VkShaderModule mod = VK_NULL_HANDLE;
+ void *streamout = zs->streamout.so_info_slots ? &zs->streamout : NULL;
+ struct spirv_shader *spirv = nir_to_spirv(zs->nir, streamout);
+ assert(spirv);
+
+ if (zink_debug & ZINK_DEBUG_SPIRV) {
+ char buf[256];
+ static int i;
+ snprintf(buf, sizeof(buf), "dump%02d.spv", i++);
+ FILE *fp = fopen(buf, "wb");
+ if (fp) {
+ fwrite(spirv->words, sizeof(uint32_t), spirv->num_words, fp);
+ fclose(fp);
+ fprintf(stderr, "wrote '%s'...\n", buf);
+ }
+ }
+
+ VkShaderModuleCreateInfo smci = {};
+ smci.sType = VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO;
+ smci.codeSize = spirv->num_words * sizeof(uint32_t);
+ smci.pCode = spirv->words;
+
+ if (vkCreateShaderModule(screen->dev, &smci, NULL, &mod) != VK_SUCCESS)
+ mod = VK_NULL_HANDLE;
+
+ /* TODO: determine if there's any reason to cache spirv output? */
+ free(spirv->words);
+ free(spirv);
+ return mod;
+}
+
struct zink_shader *
-zink_compile_nir(struct zink_screen *screen, struct nir_shader *nir)
+zink_shader_create(struct zink_screen *screen, struct nir_shader *nir,
+ const struct pipe_stream_output_info *so_info)
{
struct zink_shader *ret = CALLOC_STRUCT(zink_shader);
+ bool have_psiz = false;
+
+ ret->programs = _mesa_pointer_set_create(NULL);
NIR_PASS_V(nir, nir_lower_uniforms_to_ubo, 1);
NIR_PASS_V(nir, nir_lower_clip_halfz);
+ if (nir->info.stage == MESA_SHADER_VERTEX)
+ have_psiz = check_psiz(nir);
NIR_PASS_V(nir, nir_lower_regs_to_ssa);
optimize_nir(nir);
- NIR_PASS_V(nir, nir_remove_dead_variables, nir_var_function_temp);
+ NIR_PASS_V(nir, nir_remove_dead_variables, nir_var_function_temp, NULL);
NIR_PASS_V(nir, lower_discard_if);
+ NIR_PASS_V(nir, nir_lower_fragcolor);
NIR_PASS_V(nir, nir_convert_from_ssa, true);
if (zink_debug & ZINK_DEBUG_NIR) {
}
ret->num_bindings = 0;
- nir_foreach_variable(var, &nir->uniforms) {
+ nir_foreach_variable_with_modes(var, nir, nir_var_uniform |
+ nir_var_mem_ubo) {
if (var->data.mode == nir_var_mem_ubo) {
int binding = zink_binding(nir->info.stage,
VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER,
}
}
- ret->info = nir->info;
-
- struct spirv_shader *spirv = nir_to_spirv(nir);
- assert(spirv);
-
- if (zink_debug & ZINK_DEBUG_SPIRV) {
- char buf[256];
- static int i;
- snprintf(buf, sizeof(buf), "dump%02d.spv", i++);
- FILE *fp = fopen(buf, "wb");
- fwrite(spirv->words, sizeof(uint32_t), spirv->num_words, fp);
- fclose(fp);
- fprintf(stderr, "wrote '%s'...\n", buf);
+ ret->nir = nir;
+ if (so_info) {
+ memcpy(&ret->streamout.so_info, so_info, sizeof(struct pipe_stream_output_info));
+ ret->streamout.so_info_slots = malloc(so_info->num_outputs * sizeof(unsigned int));
+ assert(ret->streamout.so_info_slots);
+ update_so_info(ret, nir->info.outputs_written, have_psiz);
}
- VkShaderModuleCreateInfo smci = {};
- smci.sType = VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO;
- smci.codeSize = spirv->num_words * sizeof(uint32_t);
- smci.pCode = spirv->words;
-
- if (vkCreateShaderModule(screen->dev, &smci, NULL, &ret->shader_module) != VK_SUCCESS)
- return NULL;
-
return ret;
}
void
-zink_shader_free(struct zink_screen *screen, struct zink_shader *shader)
+zink_shader_free(struct zink_context *ctx, struct zink_shader *shader)
{
- vkDestroyShaderModule(screen->dev, shader->shader_module, NULL);
+ struct zink_screen *screen = zink_screen(ctx->base.screen);
+ set_foreach(shader->programs, entry) {
+ struct zink_gfx_program *prog = (void*)entry->key;
+ _mesa_hash_table_remove_key(ctx->program_cache, prog->shaders);
+ prog->shaders[pipe_shader_type_from_mesa(shader->nir->info.stage)] = NULL;
+ zink_gfx_program_reference(screen, &prog, NULL);
+ }
+ _mesa_set_destroy(shader->programs, NULL);
+ free(shader->streamout.so_info_slots);
+ ralloc_free(shader->nir);
FREE(shader);
}