From: Eric Anholt Date: Mon, 13 Oct 2014 15:20:01 +0000 (+0100) Subject: vc4: Write the VPM read setup multiple times to queue all the inputs. X-Git-Url: https://git.libre-soc.org/?a=commitdiff_plain;h=615bbf0ca641d356d975f12a5491f2fd56549ed8;p=mesa.git vc4: Write the VPM read setup multiple times to queue all the inputs. There's a 4-element fifo, and the size (number of dwords per vertex) field is just 4 bits. Fixes glsl-routing on sim. --- diff --git a/src/gallium/drivers/vc4/vc4_qpu_emit.c b/src/gallium/drivers/vc4/vc4_qpu_emit.c index 397e6f26dd2..99e634eb493 100644 --- a/src/gallium/drivers/vc4/vc4_qpu_emit.c +++ b/src/gallium/drivers/vc4/vc4_qpu_emit.c @@ -238,15 +238,30 @@ vc4_generate_code(struct vc4_context *vc4, struct vc4_compile *c) { struct qpu_reg *temp_registers = vc4_register_allocate(vc4, c); bool discard = false; + uint32_t inputs_remaining = c->num_inputs; + uint32_t vpm_read_fifo_count = 0; + uint32_t vpm_read_offset = 0; make_empty_list(&c->qpu_inst_list); switch (c->stage) { case QSTAGE_VERT: case QSTAGE_COORD: - queue(c, qpu_load_imm_ui(qpu_vrsetup(), - (0x00001a00 + - 0x00100000 * c->num_inputs))); + /* There's a 4-entry FIFO for VPMVCD reads, each of which can + * load up to 16 dwords (4 vec4s) per vertex. + */ + while (inputs_remaining) { + uint32_t num_entries = MIN2(inputs_remaining, 16); + queue(c, qpu_load_imm_ui(qpu_vrsetup(), + vpm_read_offset | + 0x00001a00 | + ((num_entries & 0xf) << 20))); + inputs_remaining -= num_entries; + vpm_read_offset += num_entries; + vpm_read_fifo_count++; + } + assert(vpm_read_fifo_count <= 4); + queue(c, qpu_load_imm_ui(qpu_vwsetup(), 0x00001a00)); break; case QSTAGE_FRAG: