[Mesa-dev] [PATCH] radeonsi: Add CIK SDMA support
Michel Dänzer
michel at daenzer.net
Wed Oct 1 00:53:32 PDT 2014
From: Michel Dänzer <michel.daenzer at amd.com>
Based on the corresponding SI support.
Signed-off-by: Michel Dänzer <michel.daenzer at amd.com>
---
No piglit regressions[0], even with R600_DEBUG=forcedma.
[0] The arb_texture_multisample-sample-mask-execution tests are failing
randormly, but I can't get them to fail when run manually, and they
don't seem to actually hit any DMA copies, so probably not directly
related to this change.
src/gallium/drivers/radeonsi/Makefile.sources | 1 +
src/gallium/drivers/radeonsi/cik_sdma.c | 305 ++++++++++++++++++++++++++
src/gallium/drivers/radeonsi/si_dma.c | 20 --
src/gallium/drivers/radeonsi/si_pipe.h | 9 +
src/gallium/drivers/radeonsi/si_state.c | 22 +-
src/gallium/drivers/radeonsi/si_state.h | 1 +
src/gallium/drivers/radeonsi/sid.h | 24 ++
7 files changed, 361 insertions(+), 21 deletions(-)
create mode 100644 src/gallium/drivers/radeonsi/cik_sdma.c
diff --git a/src/gallium/drivers/radeonsi/Makefile.sources b/src/gallium/drivers/radeonsi/Makefile.sources
index 0c998f4..9cbd028 100644
--- a/src/gallium/drivers/radeonsi/Makefile.sources
+++ b/src/gallium/drivers/radeonsi/Makefile.sources
@@ -1,4 +1,5 @@
C_SOURCES := \
+ cik_sdma.c \
si_blit.c \
si_commands.c \
si_compute.c \
diff --git a/src/gallium/drivers/radeonsi/cik_sdma.c b/src/gallium/drivers/radeonsi/cik_sdma.c
new file mode 100644
index 0000000..afac2b1
--- /dev/null
+++ b/src/gallium/drivers/radeonsi/cik_sdma.c
@@ -0,0 +1,305 @@
+/*
+ * Copyright 2010 Jerome Glisse <glisse at freedesktop.org>
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * on the rights to use, copy, modify, merge, publish, distribute, sub
+ * license, and/or sell copies of the Software, and to permit persons to whom
+ * the Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
+ * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
+ * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
+ * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
+ * USE OR OTHER DEALINGS IN THE SOFTWARE.
+ *
+ * Authors:
+ * Jerome Glisse
+ */
+
+#include "sid.h"
+#include "si_pipe.h"
+#include "../radeon/r600_cs.h"
+
+#include "util/u_format.h"
+
+static uint32_t cik_micro_tile_mode(struct si_screen *sscreen, unsigned tile_mode)
+{
+ if (sscreen->b.info.si_tile_mode_array_valid) {
+ uint32_t gb_tile_mode = sscreen->b.info.si_tile_mode_array[tile_mode];
+
+ return (gb_tile_mode >> 22) & 0x7;
+ }
+
+ /* The kernel cannod return the tile mode array. Guess? */
+ return 0;
+}
+
+static void cik_sdma_do_copy_buffer(struct si_context *ctx,
+ struct pipe_resource *dst,
+ struct pipe_resource *src,
+ uint64_t dst_offset,
+ uint64_t src_offset,
+ uint64_t size)
+{
+ struct radeon_winsys_cs *cs = ctx->b.rings.dma.cs;
+ unsigned i, ncopy, csize;
+ struct r600_resource *rdst = (struct r600_resource*)dst;
+ struct r600_resource *rsrc = (struct r600_resource*)src;
+
+ dst_offset += r600_resource(dst)->gpu_address;
+ src_offset += r600_resource(src)->gpu_address;
+
+ ncopy = (size / CIK_SDMA_COPY_MAX_SIZE) + !!(size % CIK_SDMA_COPY_MAX_SIZE);
+
+ r600_context_bo_reloc(&ctx->b, &ctx->b.rings.dma, rsrc, RADEON_USAGE_READ,
+ RADEON_PRIO_MIN);
+ r600_context_bo_reloc(&ctx->b, &ctx->b.rings.dma, rdst, RADEON_USAGE_WRITE,
+ RADEON_PRIO_MIN);
+
+ r600_need_dma_space(&ctx->b, ncopy * 7);
+ for (i = 0; i < ncopy; i++) {
+ csize = size < CIK_SDMA_COPY_MAX_SIZE ? size : CIK_SDMA_COPY_MAX_SIZE;
+ cs->buf[cs->cdw++] = CIK_SDMA_PACKET(CIK_SDMA_OPCODE_COPY,
+ CIK_SDMA_COPY_SUB_OPCODE_LINEAR,
+ 0);
+ cs->buf[cs->cdw++] = csize;
+ cs->buf[cs->cdw++] = 0; /* src/dst endian swap */
+ cs->buf[cs->cdw++] = src_offset;
+ cs->buf[cs->cdw++] = src_offset >> 32;
+ cs->buf[cs->cdw++] = dst_offset;
+ cs->buf[cs->cdw++] = dst_offset >> 32;
+ dst_offset += csize;
+ src_offset += csize;
+ size -= csize;
+ }
+}
+
+static void cik_sdma_copy_buffer(struct si_context *ctx,
+ struct pipe_resource *dst,
+ struct pipe_resource *src,
+ uint64_t dst_offset,
+ uint64_t src_offset,
+ uint64_t size)
+{
+ struct r600_resource *rdst = (struct r600_resource*)dst;
+
+ /* Mark the buffer range of destination as valid (initialized),
+ * so that transfer_map knows it should wait for the GPU when mapping
+ * that range. */
+ util_range_add(&rdst->valid_buffer_range, dst_offset,
+ dst_offset + size);
+
+ cik_sdma_do_copy_buffer(ctx, dst, src, dst_offset, src_offset, size);
+}
+
+static void cik_sdma_copy_tile(struct si_context *ctx,
+ struct pipe_resource *dst,
+ unsigned dst_level,
+ unsigned x,
+ unsigned y,
+ unsigned z,
+ struct pipe_resource *src,
+ unsigned src_level,
+ unsigned copy_height,
+ unsigned pitch,
+ unsigned bpp)
+{
+ struct radeon_winsys_cs *cs = ctx->b.rings.dma.cs;
+ struct si_screen *sscreen = ctx->screen;
+ struct r600_texture *rsrc = (struct r600_texture*)src;
+ struct r600_texture *rdst = (struct r600_texture*)dst;
+ struct r600_texture *rlinear, *rtiled;
+ unsigned linear_lvl, tiled_lvl;
+ unsigned array_mode, lbpp, pitch_tile_max, slice_tile_max, size;
+ unsigned ncopy, height, cheight, detile, i, src_mode, dst_mode;
+ unsigned sub_op, bank_h, bank_w, mt_aspect, nbanks, tile_split, mt;
+ uint64_t base, addr;
+ unsigned pipe_config, tile_mode_index;
+
+ dst_mode = rdst->surface.level[dst_level].mode;
+ src_mode = rsrc->surface.level[src_level].mode;
+ /* downcast linear aligned to linear to simplify test */
+ src_mode = src_mode == RADEON_SURF_MODE_LINEAR_ALIGNED ? RADEON_SURF_MODE_LINEAR : src_mode;
+ dst_mode = dst_mode == RADEON_SURF_MODE_LINEAR_ALIGNED ? RADEON_SURF_MODE_LINEAR : dst_mode;
+ assert(dst_mode != src_mode);
+
+ sub_op = CIK_SDMA_COPY_SUB_OPCODE_TILED;
+ lbpp = util_logbase2(bpp);
+ pitch_tile_max = ((pitch / bpp) / 8) - 1;
+
+ detile = dst_mode == RADEON_SURF_MODE_LINEAR;
+ rlinear = detile ? rdst : rsrc;
+ rtiled = detile ? rsrc : rdst;
+ linear_lvl = detile ? dst_level : src_level;
+ tiled_lvl = detile ? src_level : dst_level;
+
+ assert(!util_format_is_depth_and_stencil(rtiled->resource.b.b.format));
+
+ array_mode = si_array_mode(rtiled->surface.level[tiled_lvl].mode);
+ slice_tile_max = (rtiled->surface.level[tiled_lvl].nblk_x *
+ rtiled->surface.level[tiled_lvl].nblk_y) / (8*8) - 1;
+ /* linear height must be the same as the slice tile max height, it's ok even
+ * if the linear destination/source have smaller heigh as the size of the
+ * dma packet will be using the copy_height which is always smaller or equal
+ * to the linear height
+ */
+ height = rtiled->surface.level[tiled_lvl].nblk_y;
+ base = rtiled->surface.level[tiled_lvl].offset;
+ addr = rlinear->surface.level[linear_lvl].offset;
+ bank_h = cik_bank_wh(rtiled->surface.bankh);
+ bank_w = cik_bank_wh(rtiled->surface.bankw);
+ mt_aspect = cik_macro_tile_aspect(rtiled->surface.mtilea);
+ tile_split = cik_tile_split(rtiled->surface.tile_split);
+ tile_mode_index = si_tile_mode_index(rtiled, tiled_lvl, false);
+ nbanks = si_num_banks(sscreen, rtiled);
+ base += rtiled->resource.gpu_address;
+ addr += rlinear->resource.gpu_address;
+
+ pipe_config = cik_db_pipe_config(sscreen, tile_mode_index);
+ mt = cik_micro_tile_mode(sscreen, tile_mode_index);
+ size = (copy_height * pitch) / 4;
+ ncopy = (size / CIK_SDMA_COPY_MAX_SIZE) + !!(size % CIK_SDMA_COPY_MAX_SIZE);
+ r600_need_dma_space(&ctx->b, ncopy * 12);
+
+ r600_context_bo_reloc(&ctx->b, &ctx->b.rings.dma, &rsrc->resource,
+ RADEON_USAGE_READ, RADEON_PRIO_MIN);
+ r600_context_bo_reloc(&ctx->b, &ctx->b.rings.dma, &rdst->resource,
+ RADEON_USAGE_WRITE, RADEON_PRIO_MIN);
+
+ for (i = 0; i < ncopy; i++) {
+ cheight = copy_height;
+ if (((cheight * pitch) / 4) > CIK_SDMA_COPY_MAX_SIZE) {
+ cheight = (CIK_SDMA_COPY_MAX_SIZE * 4) / pitch;
+ }
+ size = (cheight * pitch) / 4;
+
+ cs->buf[cs->cdw++] = CIK_SDMA_PACKET(CIK_SDMA_OPCODE_COPY,
+ sub_op, detile << 15);
+ cs->buf[cs->cdw++] = base;
+ cs->buf[cs->cdw++] = base >> 32;
+ cs->buf[cs->cdw++] = ((height - 1) << 16) | pitch_tile_max;
+ cs->buf[cs->cdw++] = slice_tile_max;
+ cs->buf[cs->cdw++] = (pipe_config << 26) | (mt_aspect << 24) |
+ (nbanks << 21) | (bank_h << 18) | (bank_w << 15) |
+ (tile_split << 11) | (mt << 8) | (array_mode << 3) |
+ lbpp;
+ cs->buf[cs->cdw++] = (y << 16) | x;
+ cs->buf[cs->cdw++] = z;
+ cs->buf[cs->cdw++] = addr & 0xfffffffc;
+ cs->buf[cs->cdw++] = addr >> 32;
+ cs->buf[cs->cdw++] = (pitch / bpp) - 1;
+ cs->buf[cs->cdw++] = size;
+
+ copy_height -= cheight;
+ y += cheight;
+ }
+}
+
+void cik_sdma_copy(struct pipe_context *ctx,
+ struct pipe_resource *dst,
+ unsigned dst_level,
+ unsigned dstx, unsigned dsty, unsigned dstz,
+ struct pipe_resource *src,
+ unsigned src_level,
+ const struct pipe_box *src_box)
+{
+ struct si_context *sctx = (struct si_context *)ctx;
+ struct r600_texture *rsrc = (struct r600_texture*)src;
+ struct r600_texture *rdst = (struct r600_texture*)dst;
+ unsigned dst_pitch, src_pitch, bpp, dst_mode, src_mode;
+ unsigned src_w, dst_w;
+ unsigned src_x, src_y;
+ unsigned dst_x = dstx, dst_y = dsty, dst_z = dstz;
+
+ if (sctx->b.rings.dma.cs == NULL) {
+ goto fallback;
+ }
+
+ if (dst->target == PIPE_BUFFER && src->target == PIPE_BUFFER) {
+ cik_sdma_copy_buffer(sctx, dst, src, dst_x, src_box->x, src_box->width);
+ return;
+ }
+
+ if (src->format != dst->format || src_box->depth > 1 ||
+ src_box->z != dst_z ||
+ rdst->dirty_level_mask != 0 ||
+ rdst->cmask.size || rdst->fmask.size ||
+ rsrc->cmask.size || rsrc->fmask.size) {
+ goto fallback;
+ }
+
+ if (rsrc->dirty_level_mask) {
+ ctx->flush_resource(ctx, src);
+ }
+
+ src_x = util_format_get_nblocksx(src->format, src_box->x);
+ dst_x = util_format_get_nblocksx(src->format, dst_x);
+ src_y = util_format_get_nblocksy(src->format, src_box->y);
+ dst_y = util_format_get_nblocksy(src->format, dst_y);
+
+ bpp = rdst->surface.bpe;
+ dst_pitch = rdst->surface.level[dst_level].pitch_bytes;
+ src_pitch = rsrc->surface.level[src_level].pitch_bytes;
+ src_w = rsrc->surface.level[src_level].npix_x;
+ dst_w = rdst->surface.level[dst_level].npix_x;
+
+ dst_mode = rdst->surface.level[dst_level].mode;
+ src_mode = rsrc->surface.level[src_level].mode;
+ /* downcast linear aligned to linear to simplify test */
+ src_mode = src_mode == RADEON_SURF_MODE_LINEAR_ALIGNED ? RADEON_SURF_MODE_LINEAR : src_mode;
+ dst_mode = dst_mode == RADEON_SURF_MODE_LINEAR_ALIGNED ? RADEON_SURF_MODE_LINEAR : dst_mode;
+
+ if (src_pitch != dst_pitch || src_box->x || dst_x || src_w != dst_w ||
+ src_box->width != src_w ||
+ src_box->height != rsrc->surface.level[src_level].npix_y ||
+ src_box->height != rdst->surface.level[dst_level].npix_y ||
+ rsrc->surface.level[src_level].nblk_y !=
+ rdst->surface.level[dst_level].nblk_y) {
+ /* FIXME CIK can do partial blit */
+ goto fallback;
+ }
+ /* the x test here are currently useless (because we don't support partial blit)
+ * but keep them around so we don't forget about those
+ */
+ if ((src_pitch % 8) || (src_box->x % 8) || (dst_x % 8) ||
+ (src_box->y % 8) || (dst_y % 8) || (src_box->height % 8)) {
+ goto fallback;
+ }
+
+ if (src_mode == dst_mode) {
+ uint64_t dst_offset, src_offset;
+ /* simple dma blit would do NOTE code here assume :
+ * src_box.x/y == 0
+ * dst_x/y == 0
+ * dst_pitch == src_pitch
+ */
+ src_offset= rsrc->surface.level[src_level].offset;
+ src_offset += rsrc->surface.level[src_level].slice_size * src_box->z;
+ src_offset += src_y * src_pitch + src_x * bpp;
+ dst_offset = rdst->surface.level[dst_level].offset;
+ dst_offset += rdst->surface.level[dst_level].slice_size * dst_z;
+ dst_offset += dst_y * dst_pitch + dst_x * bpp;
+ cik_sdma_copy_buffer(sctx, dst, src, dst_offset, src_offset,
+ rsrc->surface.level[src_level].slice_size);
+ } else {
+ if (src_mode == RADEON_SURF_MODE_LINEAR)
+ goto fallback;
+ cik_sdma_copy_tile(sctx, dst, dst_level, dst_x, dst_y, dst_z,
+ src, src_level,
+ src_box->height / rsrc->surface.blk_h,
+ dst_pitch, bpp);
+ }
+ return;
+
+fallback:
+ si_resource_copy_region(ctx, dst, dst_level, dstx, dsty, dstz,
+ src, src_level, src_box);
+}
diff --git a/src/gallium/drivers/radeonsi/si_dma.c b/src/gallium/drivers/radeonsi/si_dma.c
index b1bd5e7..b354c3d 100644
--- a/src/gallium/drivers/radeonsi/si_dma.c
+++ b/src/gallium/drivers/radeonsi/si_dma.c
@@ -30,21 +30,6 @@
#include "util/u_format.h"
-static unsigned si_array_mode(unsigned mode)
-{
- switch (mode) {
- case RADEON_SURF_MODE_LINEAR_ALIGNED:
- return V_009910_ARRAY_LINEAR_ALIGNED;
- case RADEON_SURF_MODE_1D:
- return V_009910_ARRAY_1D_TILED_THIN1;
- case RADEON_SURF_MODE_2D:
- return V_009910_ARRAY_2D_TILED_THIN1;
- default:
- case RADEON_SURF_MODE_LINEAR:
- return V_009910_ARRAY_LINEAR_GENERAL;
- }
-}
-
static uint32_t si_micro_tile_mode(struct si_screen *sscreen, unsigned tile_mode)
{
if (sscreen->b.info.si_tile_mode_array_valid) {
@@ -240,11 +225,6 @@ void si_dma_copy(struct pipe_context *ctx,
goto fallback;
}
- /* TODO: Implement DMA copy for CIK */
- if (sctx->b.chip_class >= CIK) {
- goto fallback;
- }
-
if (dst->target == PIPE_BUFFER && src->target == PIPE_BUFFER) {
si_dma_copy_buffer(sctx, dst, src, dst_x, src_box->x, src_box->width);
return;
diff --git a/src/gallium/drivers/radeonsi/si_pipe.h b/src/gallium/drivers/radeonsi/si_pipe.h
index 5f5404d..91c7daa 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.h
+++ b/src/gallium/drivers/radeonsi/si_pipe.h
@@ -172,6 +172,15 @@ struct si_context {
unsigned ps_db_shader_control;
};
+/* cik_sdma.c */
+void cik_sdma_copy(struct pipe_context *ctx,
+ struct pipe_resource *dst,
+ unsigned dst_level,
+ unsigned dstx, unsigned dsty, unsigned dstz,
+ struct pipe_resource *src,
+ unsigned src_level,
+ const struct pipe_box *src_box);
+
/* si_blit.c */
void si_init_blit_functions(struct si_context *sctx);
void si_flush_depth_textures(struct si_context *sctx,
diff --git a/src/gallium/drivers/radeonsi/si_state.c b/src/gallium/drivers/radeonsi/si_state.c
index ed90f13..f951bf4 100644
--- a/src/gallium/drivers/radeonsi/si_state.c
+++ b/src/gallium/drivers/radeonsi/si_state.c
@@ -48,6 +48,21 @@ static void si_init_atom(struct r600_atom *atom, struct r600_atom **list_elem,
*list_elem = atom;
}
+unsigned si_array_mode(unsigned mode)
+{
+ switch (mode) {
+ case RADEON_SURF_MODE_LINEAR_ALIGNED:
+ return V_009910_ARRAY_LINEAR_ALIGNED;
+ case RADEON_SURF_MODE_1D:
+ return V_009910_ARRAY_1D_TILED_THIN1;
+ case RADEON_SURF_MODE_2D:
+ return V_009910_ARRAY_2D_TILED_THIN1;
+ default:
+ case RADEON_SURF_MODE_LINEAR:
+ return V_009910_ARRAY_LINEAR_GENERAL;
+ }
+}
+
uint32_t si_num_banks(struct si_screen *sscreen, struct r600_texture *tex)
{
if (sscreen->b.chip_class == CIK &&
@@ -3074,11 +3089,16 @@ void si_init_state_functions(struct si_context *sctx)
sctx->b.b.set_polygon_stipple = si_set_polygon_stipple;
sctx->b.b.set_min_samples = si_set_min_samples;
- sctx->b.dma_copy = si_dma_copy;
sctx->b.set_occlusion_query_state = si_set_occlusion_query_state;
sctx->b.need_gfx_cs_space = si_need_gfx_cs_space;
sctx->b.b.draw_vbo = si_draw_vbo;
+
+ if (sctx->b.chip_class >= CIK) {
+ sctx->b.dma_copy = cik_sdma_copy;
+ } else {
+ sctx->b.dma_copy = si_dma_copy;
+ }
}
void si_init_config(struct si_context *sctx)
diff --git a/src/gallium/drivers/radeonsi/si_state.h b/src/gallium/drivers/radeonsi/si_state.h
index 7b6de6b..3099583 100644
--- a/src/gallium/drivers/radeonsi/si_state.h
+++ b/src/gallium/drivers/radeonsi/si_state.h
@@ -263,6 +263,7 @@ unsigned cik_bank_wh(unsigned bankwh);
unsigned cik_db_pipe_config(struct si_screen *sscreen, unsigned tile_mode);
unsigned cik_macro_tile_aspect(unsigned macro_tile_aspect);
unsigned cik_tile_split(unsigned tile_split);
+unsigned si_array_mode(unsigned mode);
uint32_t si_num_banks(struct si_screen *sscreen, struct r600_texture *tex);
unsigned si_tile_mode_index(struct r600_texture *rtex, unsigned level, bool stencil);
diff --git a/src/gallium/drivers/radeonsi/sid.h b/src/gallium/drivers/radeonsi/sid.h
index 3241725..1653b7d 100644
--- a/src/gallium/drivers/radeonsi/sid.h
+++ b/src/gallium/drivers/radeonsi/sid.h
@@ -8665,5 +8665,29 @@
#define SI_DMA_PACKET_CONSTANT_FILL 0xd
#define SI_DMA_PACKET_NOP 0xf
+/* CIK async DMA packets */
+#define CIK_SDMA_PACKET(op, sub_op, n) ((((n) & 0xFFFF) << 16) | \
+ (((sub_op) & 0xFF) << 8) | \
+ (((op) & 0xFF) << 0))
+/* CIK async DMA packet types */
+#define CIK_SDMA_OPCODE_NOP 0x0
+#define CIK_SDMA_OPCODE_COPY 0x1
+#define CIK_SDMA_COPY_SUB_OPCODE_LINEAR 0x0
+#define CIK_SDMA_COPY_SUB_OPCODE_TILED 0x1
+#define CIK_SDMA_COPY_SUB_OPCODE_SOA 0x3
+#define CIK_SDMA_COPY_SUB_OPCODE_LINEAR_SUB_WINDOW 0x4
+#define CIK_SDMA_COPY_SUB_OPCODE_TILED_SUB_WINDOW 0x5
+#define CIK_SDMA_COPY_SUB_OPCODE_T2T_SUB_WINDOW 0x6
+#define CIK_SDMA_OPCODE_WRITE 0x2
+#define SDMA_WRITE_SUB_OPCODE_LINEAR 0x0
+#define SDMA_WRTIE_SUB_OPCODE_TILED 0x1
+#define CIK_SDMA_OPCODE_INDIRECT_BUFFER 0x4
+#define CIK_SDMA_PACKET_FENCE 0x5
+#define CIK_SDMA_PACKET_TRAP 0x6
+#define CIK_SDMA_PACKET_SEMAPHORE 0x7
+#define CIK_SDMA_PACKET_CONSTANT_FILL 0xb
+#define CIK_SDMA_PACKET_SRBM_WRITE 0xe
+#define CIK_SDMA_COPY_MAX_SIZE 0x1fffff
+
#endif /* _SID_H */
--
2.1.1
More information about the mesa-dev
mailing list