[Mesa-dev] [RFC 11/12] anv/cmd_buffer: Emit instanced draws for multiple views
Iago Toral
itoral at igalia.com
Mon Mar 27 12:33:07 UTC 2017
On Fri, 2017-03-24 at 16:57 -0700, Jason Ekstrand wrote:
> On Fri, Mar 24, 2017 at 5:53 AM, Iago Toral <itoral at igalia.com>
> wrote:
> > On Wed, 2017-03-22 at 21:01 -0700, Jason Ekstrand wrote:
> > ---
> > src/intel/vulkan/anv_private.h | 6 ++
> > src/intel/vulkan/genX_cmd_buffer.c | 141
> > ++++++++++++++++++++++++++++++++++++-
> > src/intel/vulkan/genX_pipeline.c | 10 ++-
> > 3 files changed, 152 insertions(+), 5 deletions(-)
> >
> >
> > diff --git a/src/intel/vulkan/anv_private.h
> > b/src/intel/vulkan/anv_private.h
> > index b402bc3..253dce2 100644
> > --- a/src/intel/vulkan/anv_private.h
> > +++ b/src/intel/vulkan/anv_private.h
> > @@ -1980,6 +1980,12 @@ struct anv_subpass {
> > bool has_resolve;
> > };
> >
> > +static inline unsigned
> > +anv_subpass_view_count(const struct anv_subpass *subpass)
> > +{
> > + return MAX2(1, _mesa_bitcount(subpass->view_mask));
> > +}
> > +
> > enum anv_subpass_usage {
> > ANV_SUBPASS_USAGE_DRAW = (1 << 0),
> > ANV_SUBPASS_USAGE_INPUT = (1 << 1),
> > diff --git a/src/intel/vulkan/genX_cmd_buffer.c
> > b/src/intel/vulkan/genX_cmd_buffer.c
> > index aafe7fd..8c21c33 100644
> > --- a/src/intel/vulkan/genX_cmd_buffer.c
> > +++ b/src/intel/vulkan/genX_cmd_buffer.c
> > @@ -26,6 +26,7 @@
> >
> > #include "anv_private.h"
> > #include "vk_format_info.h"
> > +#include "util/vk_util.h"
> >
> > #include "common/gen_l3_config.h"
> > #include "genxml/gen_macros.h"
> > @@ -50,6 +51,17 @@ emit_lri(struct anv_batch *batch, uint32_t reg,
> > uint32_t imm)
> > }
> > }
> >
> > +#if GEN_IS_HASWELL || GEN_GEN >= 8
> > +static void
> > +emit_lrr(struct anv_batch *batch, uint32_t dst, uint32_t src)
> > +{
> > + anv_batch_emit(batch, GENX(MI_LOAD_REGISTER_REG), lrr) {
> > + lrr.SourceRegisterAddress = src;
> > + lrr.DestinationRegisterAddress = dst;
> > + }
> > +}
> > +#endif
> > +
> > void
> > genX(cmd_buffer_emit_state_base_address)(struct anv_cmd_buffer
> > *cmd_buffer)
> > {
> > @@ -1525,7 +1537,13 @@ genX(cmd_buffer_flush_state)(struct
> > anv_cmd_buffer *cmd_buffer)
> > .MemoryObjectControlState = GENX(MOCS),
> > #else
> > .BufferAccessType = pipeline->instancing_enable[vb] ?
> > INSTANCEDATA : VERTEXDATA,
> > - .InstanceDataStepRate = 1,
> > + /* Our implementation of VK_KHR_multiview uses
> > instancing to draw
> > + * the different views. If the client asks for
> > instancing, we
> > + * need to use the Instance Data Step Rate to ensure
> > that we
> > + * repeat the client's per-instance data once for each
> > view.
> > + */
> > + .InstanceDataStepRate =
> > + _mesa_bitcount(pipeline->subpass->view_mask),
> >
> > mmm... shouldn't this be:
> >
> > .InstanceDataStepRate = anv_subpass_view_count(pipeline->subpass);
> >
> > so that we set this to 1 when multiview is not in use? (view_mask
> > == 0)
> Good call! You're absolutely right. This line is older than
> anv_subpass_view_count, I think.
>
> >
> > .VertexBufferMemoryObjectControlState = GENX(MOCS),
> > #endif
> >
> > @@ -1715,6 +1733,11 @@ void genX(CmdDraw)(
> > if (vs_prog_data->uses_drawid)
> > emit_draw_index(cmd_buffer, 0);
> >
> > + /* Our implementation of VK_KHR_multiview uses instancing to
> > draw
> > the
> > + * different views. We need to multiply instanceCount by the
> > view count.
> > + */
> > + instanceCount *= anv_subpass_view_count(cmd_buffer-
> > >state.subpass);
> > +
> > anv_batch_emit(&cmd_buffer->batch, GENX(3DPRIMITIVE), prim) {
> > prim.VertexAccessType = SEQUENTIAL;
> > prim.PrimitiveTopologyType = pipeline->topology;
> > @@ -1748,6 +1771,11 @@ void genX(CmdDrawIndexed)(
> > if (vs_prog_data->uses_drawid)
> > emit_draw_index(cmd_buffer, 0);
> >
> > + /* Our implementation of VK_KHR_multiview uses instancing to
> > draw
> > the
> > + * different views. We need to multiply instanceCount by the
> > view count.
> > + */
> > + instanceCount *= anv_subpass_view_count(cmd_buffer-
> > >state.subpass);
> > +
> > anv_batch_emit(&cmd_buffer->batch, GENX(3DPRIMITIVE), prim) {
> > prim.VertexAccessType = RANDOM;
> > prim.PrimitiveTopologyType = pipeline->topology;
> > @@ -1767,6 +1795,90 @@ void genX(CmdDrawIndexed)(
> > #define GEN7_3DPRIM_START_INSTANCE 0x243C
> > #define GEN7_3DPRIM_BASE_VERTEX 0x2440
> >
> > +/* MI_MATH only exists on Haswell+ */
> > +#if GEN_IS_HASWELL || GEN_GEN >= 8
> > +
> > +#define alu_opcode(v) __gen_uint((v), 20, 31)
> > +#define alu_operand1(v) __gen_uint((v), 10, 19)
> > +#define alu_operand2(v) __gen_uint((v), 0, 9)
> > +#define alu(opcode, operand1, operand2) \
> > + alu_opcode(opcode) | alu_operand1(operand1) |
> > alu_operand2(operand2)
> > +
> > +#define OPCODE_NOOP 0x000
> > +#define OPCODE_LOAD 0x080
> > +#define OPCODE_LOADINV 0x480
> > +#define OPCODE_LOAD0 0x081
> > +#define OPCODE_LOAD1 0x481
> > +#define OPCODE_ADD 0x100
> > +#define OPCODE_SUB 0x101
> > +#define OPCODE_AND 0x102
> > +#define OPCODE_OR 0x103
> > +#define OPCODE_XOR 0x104
> > +#define OPCODE_STORE 0x180
> > +#define OPCODE_STOREINV 0x580
> > +
> > +#define OPERAND_R0 0x00
> > +#define OPERAND_R1 0x01
> > +#define OPERAND_R2 0x02
> > +#define OPERAND_R3 0x03
> > +#define OPERAND_R4 0x04
> > +#define OPERAND_SRCA 0x20
> > +#define OPERAND_SRCB 0x21
> > +#define OPERAND_ACCU 0x31
> > +#define OPERAND_ZF 0x32
> > +#define OPERAND_CF 0x33
> > +
> > +#define CS_GPR(n) (0x2600 + (n) * 8)
> > +
> > +/* Emit dwords to multiply GPR0 by N */
> > +static void
> > +build_alu_multiply_gpr0(uint32_t *dw, unsigned *dw_count, uint32_t
> > N)
> > +{
> > + VK_OUTARRAY_MAKE(out, dw, dw_count);
> > +
> > +#define append_alu(opcode, operand1, operand2) \
> > + vk_outarray_append(&out, alu_dw) *alu_dw = alu(opcode,
> > operand1,
> > operand2)
> > +
> > + assert(N > 0);
> > + unsigned top_bit = 31 - __builtin_clz(N);
> > + for (int i = top_bit - 1; i >= 0; i--) {
> > + /* We get our initial data in GPR0 and we write the final
> > data
> > out to
> > + * GPR0 but we use GPR1 as our scratch register.
> > + */
> > + unsigned src_reg = i == top_bit - 1 ? OPERAND_R0 :
> > OPERAND_R1;
> > + unsigned dst_reg = i == 0 ? OPERAND_R0 : OPERAND_R1;
> > +
> > + /* Shift the current value left by 1 */
> > + append_alu(OPCODE_LOAD, OPERAND_SRCA, src_reg);
> > + append_alu(OPCODE_LOAD, OPERAND_SRCB, src_reg);
> > + append_alu(OPCODE_AND, 0, 0);
> > +
> > + if (N & (1 << i)) {
> > + /* Store ACCU to R1 and add R0 to R1 */
> > + append_alu(OPCODE_STORE, OPERAND_R1, OPERAND_ACCU);
> > + append_alu(OPCODE_LOAD, OPERAND_SRCA, OPERAND_R0);
> > + append_alu(OPCODE_LOAD, OPERAND_SRCB, OPERAND_R1);
> > + append_alu(OPCODE_AND, 0, 0);
> > + }
> > +
> > + append_alu(OPCODE_STORE, dst_reg, OPERAND_ACCU);
> > + }
> > +
> > +#undef append_alu
> > +}
> > +
> > +static void
> > +emit_mul_gpr0(struct anv_batch *batch, uint32_t N)
> > +{
> > + uint32_t num_dwords;
> > > + build_alu_multiply_gpr0(NULL, &num_dwords, N);
> >
> > I don't get this, it seems that the expectation of this code is
> > that build_alu_multiply_gpr0() fills in num_dwords when we call it
> > with NULL in the first argument, but I don't see where it does
> > that,
> > what am I missing?
> build_alu_multiply_gpr0 has the same semantics as the Vulkan query
> functions that return arrays. If you call it with NULL, it returns
> the number of dwords required. Otherwise, it assumes num_dwords is
> the number of dwords that it has available to write. I made use of
> vk_outarray to get this behavior. Yeah, it's a bit deceptive, but it
> worked out really well. :-) Feel free to tell me that it's a bit too
> clever and that I should make it simpler.
No, it is okay, I understood what the intended behavior was but somehow
I didn't see clearly how the macros in vk-util achieved that and got
confused, I looked at it more carefully now and I think it is clear and
make the implementation easier too.
> >
> > +
> >
> > + uint32_t *dw = anv_batch_emitn(batch, 1 + num_dwords,
> >
> > GENX(MI_MATH));
> >
> > + build_alu_multiply_gpr0(> > dw, &num_dwords, N);
> >
> > +}
> >
> > +
> >
> > +#endif /* GEN_IS_HASWELL || GEN_GEN >= 8 */
> >
> > +
> >
> > static void
> >
> > load_indirect_parameters(> > struct anv_cmd_buffer *cmd_buffer,
> >
> > > > struct anv_buffer *buffer, uint64_t offset,
> >
> > @@ -1777,7 +1889,22 @@ load_indirect_parameters(> > struct anv_cmd_buffer
> >
> > *cmd_buffer,
> >
> > uint32_t bo_offset = buffer->offset + offset;
> >
> >
> >
> > emit_lrm(batch, GEN7_3DPRIM_VERTEX_COUNT, bo, bo_offset);
> >
> > - emit_lrm(batch, GEN7_3DPRIM_INSTANCE_COUNT, bo, bo_offset + 4);
> >
> > +
> >
> > + unsigned view_count = anv_subpass_view_count(cmd_> > buffer-
> >
> > >state.subpass);
> >
> > + if (view_count > 1) {
> >
> > +#if GEN_IS_HASWELL || GEN_GEN >= 8
> >
> > + emit_lrm(batch, CS_GPR(0), bo, bo_offset + 4);
> >
> > + emit_mul_gpr0(batch, view_count);
> >
> > + emit_lrr(batch, GEN7_3DPRIM_INSTANCE_COUNT, CS_GPR(0));
> >
> > +#else
> >
> > + anv_finishme("Multiview + indirect draw requires MI_MATH\n"
> >
> > + "MI_MATH is not supported on Ivy Bridge");
> >
> > + emit_lrm(batch, GEN7_3DPRIM_INSTANCE_COUNT, bo, bo_offset +
> >
> > 4);
> >
> > +#endif
> >
> > + } else {
> >
> > + emit_lrm(batch, GEN7_3DPRIM_INSTANCE_COUNT, bo, bo_offset +
> >
> > 4);
> >
> > + }
> >
> > +
> >
> > emit_lrm(batch, GEN7_3DPRIM_START_VERTEX, bo, bo_offset + 8);
> >
> >
> >
> > if (indexed) {
> >
> > @@ -2531,6 +2658,16 @@ genX(cmd_buffer_set_subpass)(> > struct
> >
> > anv_cmd_buffer *cmd_buffer,
> >
> >
> >
> > cmd_buffer->state.dirty |= ANV_CMD_DIRTY_RENDER_TARGETS;
> >
> >
> >
> > + /* Our implementation of VK_KHR_multiview uses instancing to draw
> >
> > the
> >
> > + * different views. If the client asks for instancing, we need
> >
> > to use the
> >
> > + * Instance Data Step Rate to ensure that we repeat the client's
> >
> > + * per-instance data once for each view. Since this bit is in
> >
> > + * VERTEX_BUFFER_STATE on gen7, we need to dirty vertex buffers
> >
> > at the top
> >
> > + * of each subpass.
> >
> > + */
> >
> > + if (GEN_GEN == 7)
> >
> > + cmd_buffer->state.vb_> > dirty |= ~0;
> >
> > > +
> >
> >
> > Shouldn't we do:
> >
> >
> > if (GEN_GEN == 7 && subpass->view_mask != 0)
> >
> > cmd_buffer->state.vb_dirty |= ~0;
> >
> >
> > So this doesn't affect cases that don't use multiview? > > Yes, we could do that because we're guaranteed that, for each render pass, either all subpass view masks are 0 or all are non-zero. This seems safer though and I doubt flushing vertex buffers at the subpass boundary will hurt anyone.
Right, makes sense.
Reviewed-by: Iago Toral Quiroga <itoral at igalia.com>
> >
> > > /* Perform transitions to the subpass layout before any writes
> >
> > > have
> >
> > * occurred.
> >
> > */
> >
> > diff --git a/src/intel/vulkan/genX_> > pipeline.c
> >
> > b/src/intel/vulkan/genX_> > pipeline.c
> >
> > index 85a9e4f..9aaa8a0 100644
> >
> > --- a/src/intel/vulkan/genX_> > pipeline.c
> >
> > +++ b/src/intel/vulkan/genX_> > pipeline.c
> >
> > @@ -155,9 +155,13 @@ emit_vertex_input(struct anv_pipeline *pipeline,
> >
> > anv_batch_emit(&> > pipeline->batch, GENX(3DSTATE_VF_INSTANCING),
> >
> > vfi) {
> >
> > vfi.InstancingEnable = pipeline->instancing_enable[> > desc-
> >
> > >binding];
> >
> > vfi.> > VertexElementIndex = slot;
> >
> > - /* Vulkan so far doesn't have an instance divisor, so
> >
> > - * this is always 1 (ignored if not instancing). */
> >
> > - vfi.> > InstanceDataStepRate = 1;
> >
> > + /* Our implementation of VK_KHX_multiview uses instancing
> >
> > to draw
> >
> > + * the different views. If the client asks for instancing,
> >
> > we
> >
> > + * need to use the Instance Data Step Rate to ensure that
> >
> > we
> >
> > + * repeat the client's per-instance data once for each
> >
> > view.
> >
> > + */
> >
> > + vfi.> > InstanceDataStepRate =
> >
> > > + _mesa_bitcount(> > pipeline->subpass->view_mask);
> >
> >
> > Same comment as above, shouldn't this be:
> >
> >
> > vfi.InstanceDataStepRate = anv_subpass_view_count(> > pipeline->subpass);> > >
> > +
> >
> > + uint32_t *dw = anv_batch_emitn(batch, 1 + num_dwords,
> >
> > GENX(MI_MATH));
> >
> > + build_alu_multiply_gpr0(> > dw, &num_dwords, N);
> >
> > +}
> >
> > +
> >
> > +#endif /* GEN_IS_HASWELL || GEN_GEN >= 8 */
> >
> > +
> >
> > static void
> >
> > load_indirect_parameters(> > struct anv_cmd_buffer *cmd_buffer,
> >
> > > > struct anv_buffer *buffer, uint64_t offset,
> >
> > @@ -1777,7 +1889,22 @@ load_indirect_parameters(> > struct anv_cmd_buffer
> >
> > *cmd_buffer,
> >
> > uint32_t bo_offset = buffer->offset + offset;
> >
> >
> >
> > emit_lrm(batch, GEN7_3DPRIM_VERTEX_COUNT, bo, bo_offset);
> >
> > - emit_lrm(batch, GEN7_3DPRIM_INSTANCE_COUNT, bo, bo_offset + 4);
> >
> > +
> >
> > + unsigned view_count = anv_subpass_view_count(cmd_> > buffer-
> >
> > >state.subpass);
> >
> > + if (view_count > 1) {
> >
> > +#if GEN_IS_HASWELL || GEN_GEN >= 8
> >
> > + emit_lrm(batch, CS_GPR(0), bo, bo_offset + 4);
> >
> > + emit_mul_gpr0(batch, view_count);
> >
> > + emit_lrr(batch, GEN7_3DPRIM_INSTANCE_COUNT, CS_GPR(0));
> >
> > +#else
> >
> > + anv_finishme("Multiview + indirect draw requires MI_MATH\n"
> >
> > + "MI_MATH is not supported on Ivy Bridge");
> >
> > + emit_lrm(batch, GEN7_3DPRIM_INSTANCE_COUNT, bo, bo_offset +
> >
> > 4);
> >
> > +#endif
> >
> > + } else {
> >
> > + emit_lrm(batch, GEN7_3DPRIM_INSTANCE_COUNT, bo, bo_offset +
> >
> > 4);
> >
> > + }
> >
> > +
> >
> > emit_lrm(batch, GEN7_3DPRIM_START_VERTEX, bo, bo_offset + 8);
> >
> >
> >
> > if (indexed) {
> >
> > @@ -2531,6 +2658,16 @@ genX(cmd_buffer_set_subpass)(> > struct
> >
> > anv_cmd_buffer *cmd_buffer,
> >
> >
> >
> > cmd_buffer->state.dirty |= ANV_CMD_DIRTY_RENDER_TARGETS;
> >
> >
> >
> > + /* Our implementation of VK_KHR_multiview uses instancing to draw
> >
> > the
> >
> > + * different views. If the client asks for instancing, we need
> >
> > to use the
> >
> > + * Instance Data Step Rate to ensure that we repeat the client's
> >
> > + * per-instance data once for each view. Since this bit is in
> >
> > + * VERTEX_BUFFER_STATE on gen7, we need to dirty vertex buffers
> >
> > at the top
> >
> > + * of each subpass.
> >
> > + */
> >
> > + if (GEN_GEN == 7)
> >
> > + cmd_buffer->state.vb_> > dirty |= ~0;
> >
> > > +
> >
> >
> > Shouldn't we do:
> >
> >
> > if (GEN_GEN == 7 && subpass->view_mask != 0)
> >
> > cmd_buffer->state.vb_dirty |= ~0;
> >
> >
> > So this doesn't affect cases that don't use multiview? > > Yes, we could do that because we're guaranteed that, for each render pass, either all subpass view masks are 0 or all are non-zero. This seems safer though and I doubt flushing vertex buffers at the subpass boundary will hurt anyone.>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/mesa-dev/attachments/20170327/c015c2ad/attachment-0001.html>
More information about the mesa-dev
mailing list