#include <stdio.h>
#include <stdlib.h>
#include <string.h>
+#include "pan_bo.h"
#include "pan_context.h"
+#include "pan_util.h"
+#include "panfrost-quirks.h"
#include "compiler/nir/nir.h"
#include "nir/tgsi_to_nir.h"
#include "midgard/midgard_compile.h"
+#include "bifrost/bifrost_compile.h"
#include "util/u_dynarray.h"
#include "tgsi/tgsi_dump.h"
-void
-panfrost_shader_compile(struct panfrost_context *ctx, struct mali_shader_meta *meta, const char *src, int type, struct panfrost_shader_state *state)
+static unsigned
+pan_format_from_nir_base(nir_alu_type base)
{
- uint8_t *dst;
+ switch (base) {
+ case nir_type_int:
+ return MALI_FORMAT_SINT;
+ case nir_type_uint:
+ case nir_type_bool:
+ return MALI_FORMAT_UINT;
+ case nir_type_float:
+ return MALI_CHANNEL_FLOAT;
+ default:
+ unreachable("Invalid base");
+ }
+}
- nir_shader *s;
+static unsigned
+pan_format_from_nir_size(nir_alu_type base, unsigned size)
+{
+ if (base == nir_type_float) {
+ switch (size) {
+ case 16: return MALI_FORMAT_SINT;
+ case 32: return MALI_FORMAT_UNORM;
+ default:
+ unreachable("Invalid float size for format");
+ }
+ } else {
+ switch (size) {
+ case 1:
+ case 8: return MALI_CHANNEL_8;
+ case 16: return MALI_CHANNEL_16;
+ case 32: return MALI_CHANNEL_32;
+ default:
+ unreachable("Invalid int size for format");
+ }
+ }
+}
- struct pipe_shader_state *cso = state->base;
+static enum mali_format
+pan_format_from_glsl(const struct glsl_type *type, unsigned precision, unsigned frac)
+{
+ const struct glsl_type *column = glsl_without_array_or_matrix(type);
+ enum glsl_base_type glsl_base = glsl_get_base_type(column);
+ nir_alu_type t = nir_get_nir_type_for_glsl_base_type(glsl_base);
+ unsigned chan = glsl_get_components(column);
+
+ /* If we have a fractional location added, we need to increase the size
+ * so it will fit, i.e. a vec3 in YZW requires us to allocate a vec4.
+ * We could do better but this is an edge case as it is, normally
+ * packed varyings will be aligned. */
+ chan += frac;
+
+ assert(chan >= 1 && chan <= 4);
+
+ unsigned base = nir_alu_type_get_base_type(t);
+ unsigned size = nir_alu_type_get_type_size(t);
+
+ /* Demote to fp16 where possible. int16 varyings are TODO as the hw
+ * will saturate instead of wrap which is not conformant, so we need to
+ * insert i2i16/u2u16 instructions before the st_vary_32i/32u to get
+ * the intended behaviour */
+
+ bool is_16 = (precision == GLSL_PRECISION_MEDIUM)
+ || (precision == GLSL_PRECISION_LOW);
+
+ if (is_16 && base == nir_type_float)
+ size = 16;
+ else
+ size = 32;
+
+ return pan_format_from_nir_base(base) |
+ pan_format_from_nir_size(base, size) |
+ MALI_NR_CHANNELS(chan);
+}
- if (cso->type == PIPE_SHADER_IR_NIR) {
- s = nir_shader_clone(NULL, cso->ir.nir);
- } else {
- assert (cso->type == PIPE_SHADER_IR_TGSI);
- //tgsi_dump(cso->tokens, 0);
- s = tgsi_to_nir(cso->tokens, ctx->base.screen);
+static enum bifrost_shader_type
+bifrost_blend_type_from_nir(nir_alu_type nir_type)
+{
+ switch(nir_type) {
+ case 0: /* Render target not in use */
+ return 0;
+ case nir_type_float16:
+ return BIFROST_BLEND_F16;
+ case nir_type_float32:
+ return BIFROST_BLEND_F32;
+ case nir_type_int32:
+ return BIFROST_BLEND_I32;
+ case nir_type_uint32:
+ return BIFROST_BLEND_U32;
+ case nir_type_int16:
+ return BIFROST_BLEND_I16;
+ case nir_type_uint16:
+ return BIFROST_BLEND_U16;
+ default:
+ unreachable("Unsupported blend shader type for NIR alu type");
+ return 0;
}
+}
- s->info.stage = type == JOB_TYPE_VERTEX ? MESA_SHADER_VERTEX : MESA_SHADER_FRAGMENT;
+void
+panfrost_shader_compile(struct panfrost_context *ctx,
+ enum pipe_shader_ir ir_type,
+ const void *ir,
+ gl_shader_stage stage,
+ struct panfrost_shader_state *state,
+ uint64_t *outputs_written)
+{
+ struct panfrost_device *dev = pan_device(ctx->base.screen);
+ uint8_t *dst;
- if (s->info.stage == MESA_SHADER_FRAGMENT) {
- /* Inject the alpha test now if we need to */
+ nir_shader *s;
- if (state->alpha_state.enabled) {
- NIR_PASS_V(s, nir_lower_alpha_test, state->alpha_state.func, false);
- }
+ if (ir_type == PIPE_SHADER_IR_NIR) {
+ s = nir_shader_clone(NULL, ir);
+ } else {
+ assert (ir_type == PIPE_SHADER_IR_TGSI);
+ s = tgsi_to_nir(ir, ctx->base.screen, false);
}
+ s->info.stage = stage;
+
/* Call out to Midgard compiler given the above NIR */
- midgard_program program = {
+ panfrost_program program = {
.alpha_ref = state->alpha_state.ref_value
};
- midgard_compile_shader_nir(s, &program, false);
+ memcpy(program.rt_formats, state->rt_formats, sizeof(program.rt_formats));
+
+ if (dev->quirks & IS_BIFROST) {
+ bifrost_compile_shader_nir(s, &program, dev->gpu_id);
+ } else {
+ midgard_compile_shader_nir(s, &program, false, 0, dev->gpu_id,
+ dev->debug & PAN_DBG_PRECOMPILE, false);
+ }
/* Prepare the compiled binary for upload */
int size = program.compiled.size;
* I bet someone just thought that would be a cute pun. At least,
* that's how I'd do it. */
- meta->shader = panfrost_upload(&ctx->shaders, dst, size, true) | program.first_tag;
-
- util_dynarray_fini(&program.compiled);
-
- meta->midgard1.uniform_count = MIN2(program.uniform_count, program.uniform_cutoff);
- meta->attribute_count = program.attribute_count;
- meta->varying_count = program.varying_count;
- meta->midgard1.work_count = program.work_register_count;
-
- state->can_discard = program.can_discard;
- state->writes_point_size = program.writes_point_size;
-
- /* Separate as primary uniform count is truncated */
- state->uniform_count = program.uniform_count;
+ if (size) {
+ state->bo = panfrost_bo_create(dev, size, PAN_BO_EXECUTE);
+ memcpy(state->bo->cpu, dst, size);
+ }
- /* gl_Position eats up an extra spot */
- if (type == JOB_TYPE_VERTEX)
- meta->varying_count += 1;
+ if (!(dev->quirks & IS_BIFROST)) {
+ /* If size = 0, no shader. Use dummy tag to avoid
+ * INSTR_INVALID_ENC */
+ state->first_tag = size ? program.first_tag : 1;
+ }
- /* Note: gl_FragCoord does -not- eat an extra spot; it will be included
- * in our count if we need it */
+ util_dynarray_fini(&program.compiled);
- meta->midgard1.unknown2 = 8; /* XXX */
+ state->sysval_count = program.sysval_count;
+ memcpy(state->sysval, program.sysvals, sizeof(state->sysval[0]) * state->sysval_count);
+
+ bool vertex_id = s->info.system_values_read & (1 << SYSTEM_VALUE_VERTEX_ID);
+ bool instance_id = s->info.system_values_read & (1 << SYSTEM_VALUE_INSTANCE_ID);
+
+ /* On Bifrost it's a sysval, on Midgard it's a varying */
+ state->reads_frag_coord = s->info.system_values_read & (1 << SYSTEM_VALUE_FRAG_COORD);
+
+ state->writes_global = s->info.writes_memory;
+
+ switch (stage) {
+ case MESA_SHADER_VERTEX:
+ state->attribute_count = util_bitcount64(s->info.inputs_read);
+ state->varying_count = util_bitcount64(s->info.outputs_written);
+
+ if (vertex_id)
+ state->attribute_count = MAX2(state->attribute_count, PAN_VERTEX_ID + 1);
+
+ if (instance_id)
+ state->attribute_count = MAX2(state->attribute_count, PAN_INSTANCE_ID + 1);
+
+ break;
+ case MESA_SHADER_FRAGMENT:
+ state->attribute_count = 0;
+ state->varying_count = util_bitcount64(s->info.inputs_read);
+ if (s->info.outputs_written & BITFIELD64_BIT(FRAG_RESULT_DEPTH))
+ state->writes_depth = true;
+ if (s->info.outputs_written & BITFIELD64_BIT(FRAG_RESULT_STENCIL))
+ state->writes_stencil = true;
+
+ uint64_t outputs_read = s->info.outputs_read;
+ if (outputs_read & BITFIELD64_BIT(FRAG_RESULT_COLOR))
+ outputs_read |= BITFIELD64_BIT(FRAG_RESULT_DATA0);
+
+ state->outputs_read = outputs_read >> FRAG_RESULT_DATA0;
+
+ /* List of reasons we need to execute frag shaders when things
+ * are masked off */
+
+ state->fs_sidefx =
+ s->info.writes_memory ||
+ s->info.fs.uses_discard ||
+ s->info.fs.uses_demote;
+ break;
+ case MESA_SHADER_COMPUTE:
+ /* TODO: images */
+ state->attribute_count = 0;
+ state->varying_count = 0;
+ state->shared_size = s->info.cs.shared_size;
+ break;
+ default:
+ unreachable("Unknown shader state");
+ }
- /* Varyings are known only through the shader. We choose to upload this
- * information with the vertex shader, though the choice is perhaps
- * arbitrary */
+ state->can_discard = s->info.fs.uses_discard;
+ state->helper_invocations = s->info.fs.needs_helper_invocations;
+ state->stack_size = program.tls_size;
- if (type == JOB_TYPE_VERTEX) {
- struct panfrost_varyings *varyings = &state->varyings;
+ state->reads_frag_coord = s->info.inputs_read & (1 << VARYING_SLOT_POS);
+ state->reads_point_coord = s->info.inputs_read & (1 << VARYING_SLOT_PNTC);
+ state->reads_face = s->info.inputs_read & (1 << VARYING_SLOT_FACE);
+ state->writes_point_size = s->info.outputs_written & (1 << VARYING_SLOT_PSIZ);
- /* Measured in vec4 words. Don't include gl_Position */
- int varying_count = program.varying_count;
+ if (outputs_written)
+ *outputs_written = s->info.outputs_written;
- /* Setup two buffers, one for position, the other for normal
- * varyings, as seen in traces. TODO: Are there other
- * configurations we might use? */
+ /* Separate as primary uniform count is truncated. Sysvals are prefix
+ * uniforms */
+ state->uniform_count = s->num_uniforms + program.sysval_count;
+ state->uniform_cutoff = program.uniform_cutoff;
+ state->work_reg_count = program.work_register_count;
- varyings->varying_buffer_count = 2;
+ if (dev->quirks & IS_BIFROST)
+ for (unsigned i = 0; i < BIFROST_MAX_RENDER_TARGET_COUNT; i++)
+ state->blend_types[i] = bifrost_blend_type_from_nir(program.blend_types[i]);
- /* mediump vec4s sequentially */
- varyings->varyings_stride[0] = (2 * sizeof(float)) * varying_count;
+ /* Record the varying mapping for the command stream's bookkeeping */
- /* highp gl_Position */
- varyings->varyings_stride[1] = 4 * sizeof(float);
+ nir_variable_mode varying_mode =
+ stage == MESA_SHADER_VERTEX ? nir_var_shader_out : nir_var_shader_in;
- /* mediump gl_PointSize */
- if (program.writes_point_size) {
- ++varyings->varying_buffer_count;
- varyings->varyings_stride[2] = 2; /* sizeof(fp16) */
- }
+ nir_foreach_variable_with_modes(var, s, varying_mode) {
+ unsigned loc = var->data.driver_location;
+ unsigned sz = glsl_count_attribute_slots(var->type, FALSE);
- /* Setup gl_Position, its weirdo analogue, and gl_PointSize (optionally) */
- unsigned default_vec1_swizzle = panfrost_get_default_swizzle(1);
- unsigned default_vec4_swizzle = panfrost_get_default_swizzle(4);
-
- struct mali_attr_meta vertex_special_varyings[] = {
- {
- .index = 1,
- .format = MALI_VARYING_POS,
-
- .swizzle = default_vec4_swizzle,
- .unknown1 = 0x2,
- },
- {
- .index = 1,
- .format = MALI_RGBA16F,
-
- /* TODO: Wat? yyyy swizzle? */
- .swizzle = 0x249,
- .unknown1 = 0x0,
- },
- {
- .index = 2,
- .format = MALI_R16F,
- .swizzle = default_vec1_swizzle,
- .unknown1 = 0x2
- }
- };
-
- /* How many special vertex varyings are actually required? */
- int vertex_special_count = 2 + (program.writes_point_size ? 1 : 0);
-
- /* Setup actual varyings. XXX: Don't assume vec4 */
-
- struct mali_attr_meta mali_varyings[PIPE_MAX_ATTRIBS];
-
- for (int i = 0; i < varying_count; ++i) {
- struct mali_attr_meta vec4_varying_meta = {
- .index = 0,
- .format = MALI_RGBA16F,
- .swizzle = default_vec4_swizzle,
- .unknown1 = 0x2,
-
- /* Set offset to keep everything back-to-back in
- * the same buffer */
- .src_offset = 8 * i,
- };
-
- mali_varyings[i] = vec4_varying_meta;
+ for (int c = 0; c < sz; ++c) {
+ state->varyings_loc[loc + c] = var->data.location + c;
+ state->varyings[loc + c] = pan_format_from_glsl(var->type,
+ var->data.precision, var->data.location_frac);
}
-
- /* We don't count the weirdo gl_Position in our varying count */
- varyings->varying_count = varying_count - 1;
-
- /* In this context, position_meta represents the implicit
- * gl_FragCoord varying. So, upload all the varyings */
-
- unsigned varyings_size = sizeof(struct mali_attr_meta) * varyings->varying_count;
- unsigned vertex_special_size = sizeof(struct mali_attr_meta) * vertex_special_count;
- unsigned vertex_size = vertex_special_size + varyings_size;
- unsigned fragment_size = varyings_size + sizeof(struct mali_attr_meta);
-
- struct panfrost_transfer transfer = panfrost_allocate_chunk(ctx, vertex_size + fragment_size, HEAP_DESCRIPTOR);
-
- /* Copy varyings in the follow order:
- * - Position 1, 2
- * - Varyings 1, 2, ..., n
- * - Varyings 1, 2, ..., n (duplicate)
- * - Position 1
- */
-
- memcpy(transfer.cpu, vertex_special_varyings, vertex_special_size);
- memcpy(transfer.cpu + vertex_special_size, mali_varyings, varyings_size);
- memcpy(transfer.cpu + vertex_size, mali_varyings, varyings_size);
- memcpy(transfer.cpu + vertex_size + varyings_size, &vertex_special_varyings[0], sizeof(struct mali_attr_meta));
-
- /* Point to the descriptor */
- varyings->varyings_buffer_cpu = transfer.cpu;
- varyings->varyings_descriptor = transfer.gpu;
- varyings->varyings_descriptor_fragment = transfer.gpu + vertex_size;
}
}
projects / mesa.git / blobdiff