#include "ir3/ir3_nir.h"
static void
-dump_shader_info(struct ir3_shader_variant *v, struct pipe_debug_callback *debug)
+dump_shader_info(struct ir3_shader_variant *v, bool binning_pass,
+ struct pipe_debug_callback *debug)
{
if (!unlikely(fd_mesa_debug & FD_DBG_SHADERDB))
return;
pipe_debug_message(debug, SHADER_INFO,
- "%s shader: %u inst, %u dwords, "
- "%u half, %u full, %u const, %u constlen, "
+ "%s%s shader: %u inst, %u dwords, "
+ "%u half, %u full, %u constlen, "
"%u (ss), %u (sy), %d max_sun, %d loops\n",
+ binning_pass ? "B" : "",
ir3_shader_stage(v->shader),
v->info.instrs_count,
v->info.sizedwords,
v->info.max_half_reg + 1,
v->info.max_reg + 1,
- v->info.max_const + 1,
v->constlen,
v->info.ss, v->info.sy,
v->max_sun, v->loops);
v = ir3_shader_get_variant(shader, &key, binning_pass, &created);
if (created) {
- dump_shader_info(v, debug);
+ dump_shader_info(v, binning_pass, debug);
}
return v;
static struct ir3_shader_key key;
memset(&key, 0, sizeof(key));
ir3_shader_variant(shader, key, false, debug);
+
+ if (nir->info.stage != MESA_SHADER_FRAGMENT)
+ ir3_shader_variant(shader, key, true, debug);
}
return shader;
}
}
static void
-emit_user_consts(struct fd_context *ctx, const struct ir3_shader_variant *v,
- struct fd_ringbuffer *ring, struct fd_constbuf_stateobj *constbuf)
+emit_const(struct fd_context *ctx, struct fd_ringbuffer *ring,
+ const struct ir3_shader_variant *v, uint32_t dst_offset,
+ uint32_t offset, uint32_t size,
+ const void *user_buffer, struct pipe_resource *buffer)
{
- const unsigned index = 0; /* user consts are index 0 */
+ assert(dst_offset + size <= v->constlen * 4);
- if (constbuf->enabled_mask & (1 << index)) {
- struct pipe_constant_buffer *cb = &constbuf->cb[index];
- /* size in dwords, aligned to vec4. (This works at least
- * with mesa/st, which seems to align constant buffer to
- * 16 bytes)
- */
- unsigned size = align(cb->buffer_size, 16) / 4;
-
- /* in particular, with binning shader we may end up with
- * unused consts, ie. we could end up w/ constlen that is
- * smaller than first_driver_param. In that case truncate
- * the user consts early to avoid HLSQ lockup caused by
- * writing too many consts
- */
- const struct ir3_const_state *const_state = &v->shader->const_state;
- uint32_t max_const = MIN2(const_state->num_uniforms, v->constlen);
-
- /* and even if the start of the const buffer is before
- * first_immediate, the end may not be:
- */
- size = MIN2(size, 4 * max_const);
-
- if (size > 0) {
- ring_wfi(ctx->batch, ring);
- ctx->emit_const(ring, v->type, 0,
- cb->buffer_offset, size,
- cb->user_buffer, cb->buffer);
- }
- }
+ ctx->emit_const(ring, v->type, dst_offset,
+ offset, size, user_buffer, buffer);
+}
+static void
+emit_user_consts(struct fd_context *ctx, const struct ir3_shader_variant *v,
+ struct fd_ringbuffer *ring, struct fd_constbuf_stateobj *constbuf)
+{
struct ir3_ubo_analysis_state *state;
state = &v->shader->ubo_state;
- for (uint32_t i = 1; i < ARRAY_SIZE(state->range); i++) {
+ for (uint32_t i = 0; i < ARRAY_SIZE(state->range); i++) {
struct pipe_constant_buffer *cb = &constbuf->cb[i];
if (state->range[i].start < state->range[i].end &&
uint32_t size = state->range[i].end - state->range[i].start;
uint32_t offset = cb->buffer_offset + state->range[i].start;
+
+ /* and even if the start of the const buffer is before
+ * first_immediate, the end may not be:
+ */
+ size = MIN2(size, (16 * v->constlen) - state->range[i].offset);
+
+ if (size == 0)
+ continue;
+
+ /* things should be aligned to vec4: */
debug_assert((state->range[i].offset % 16) == 0);
debug_assert((size % 16) == 0);
debug_assert((offset % 16) == 0);
- ctx->emit_const(ring, v->type, state->range[i].offset / 4,
+
+ emit_const(ctx, ring, v, state->range[i].offset / 4,
offset, size / 4, cb->user_buffer, cb->buffer);
}
}
}
}
+ assert(offset * 4 + params < v->constlen * 4);
+
ring_wfi(ctx->batch, ring);
ctx->emit_const_bo(ring, v->type, false, offset * 4, params, prscs, offsets);
}
}
ring_wfi(ctx->batch, ring);
- ctx->emit_const(ring, v->type, offset * 4,
+ emit_const(ctx, ring, v, offset * 4,
0, ARRAY_SIZE(sizes), sizes, NULL);
}
}
dims[off + 1] = ffs(dims[off + 0]) - 1;
}
}
+ uint32_t size = MIN2(ARRAY_SIZE(dims), v->constlen * 4 - offset * 4);
ring_wfi(ctx->batch, ring);
- ctx->emit_const(ring, v->type, offset * 4,
- 0, ARRAY_SIZE(dims), dims, NULL);
+ emit_const(ctx, ring, v, offset * 4, 0, size, dims, NULL);
}
}
if (size > 0) {
ring_wfi(ctx->batch, ring);
- ctx->emit_const(ring, v->type, base,
+ emit_const(ctx, ring, v, base,
0, size, const_state->immediates[0].val, NULL);
}
}
}
}
+ assert(offset * 4 + params < v->constlen * 4);
+
ring_wfi(ctx->batch, ring);
ctx->emit_const_bo(ring, v->type, true, offset * 4, params, prscs, offsets);
}
emit_common_consts(v, ring, ctx, PIPE_SHADER_VERTEX);
/* emit driver params every time: */
- /* TODO skip emit if shader doesn't use driver params to avoid WFI.. */
if (info) {
const struct ir3_const_state *const_state = &v->shader->const_state;
uint32_t offset = const_state->offsets.driver_param;
ctx->mem_to_mem(ring, vertex_params_rsc, 0,
indirect->buffer, src_off, 1);
- ctx->emit_const(ring, MESA_SHADER_VERTEX, offset * 4, 0,
+ emit_const(ctx, ring, v, offset * 4, 0,
vertex_params_size, NULL, vertex_params_rsc);
pipe_resource_reference(&vertex_params_rsc, NULL);
} else {
- ctx->emit_const(ring, MESA_SHADER_VERTEX, offset * 4, 0,
+ emit_const(ctx, ring, v, offset * 4, 0,
vertex_params_size, vertex_params, NULL);
}
indirect_offset = info->indirect_offset;
}
- ctx->emit_const(ring, MESA_SHADER_COMPUTE, offset * 4,
+ emit_const(ctx, ring, v, offset * 4,
indirect_offset, 4, NULL, indirect);
pipe_resource_reference(&indirect, NULL);
[IR3_DP_LOCAL_GROUP_SIZE_Y] = info->block[1],
[IR3_DP_LOCAL_GROUP_SIZE_Z] = info->block[2],
};
+ uint32_t size = MIN2(ARRAY_SIZE(compute_params),
+ v->constlen * 4 - offset * 4);
- ctx->emit_const(ring, MESA_SHADER_COMPUTE, offset * 4, 0,
- ARRAY_SIZE(compute_params), compute_params, NULL);
+ emit_const(ctx, ring, v, offset * 4, 0, size, compute_params, NULL);
}
}
}