/*
* Copyright 2013 Advanced Micro Devices, Inc.
+ * All Rights Reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the "Software"),
* 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:
- * Marek Olšák <maraeo@gmail.com>
*/
#include "si_pipe.h"
#include "sid.h"
-#include "radeon/r600_cs.h"
-
-/* Set this if you want the 3D engine to wait until CP DMA is done.
+/* Set this if you want the ME to wait until CP DMA is done.
* It should be set on the last CP DMA packet. */
-#define R600_CP_DMA_SYNC (1 << 0) /* R600+ */
+#define CP_DMA_SYNC (1 << 0)
/* Set this if the source data was used as a destination in a previous CP DMA
* packet. It's for preventing a read-after-write (RAW) hazard between two
* CP DMA packets. */
-#define SI_CP_DMA_RAW_WAIT (1 << 1) /* SI+ */
-#define CIK_CP_DMA_USE_L2 (1 << 2)
+#define CP_DMA_RAW_WAIT (1 << 1)
+#define CP_DMA_DST_IS_GDS (1 << 2)
+#define CP_DMA_CLEAR (1 << 3)
+#define CP_DMA_PFP_SYNC_ME (1 << 4)
+#define CP_DMA_SRC_IS_GDS (1 << 5)
+
+/* The max number of bytes that can be copied per packet. */
+static inline unsigned cp_dma_max_byte_count(struct si_context *sctx)
+{
+ unsigned max = sctx->chip_class >= GFX9 ?
+ S_414_BYTE_COUNT_GFX9(~0u) :
+ S_414_BYTE_COUNT_GFX6(~0u);
+
+ /* make it aligned for optimal performance */
+ return max & ~(SI_CPDMA_ALIGNMENT - 1);
+}
+
-/* Emit a CP DMA packet to do a copy from one buffer to another.
- * The size must fit in bits [20:0].
+/* Emit a CP DMA packet to do a copy from one buffer to another, or to clear
+ * a buffer. The size must fit in bits [20:0]. If CP_DMA_CLEAR is set, src_va is a 32-bit
+ * clear value.
*/
-static void si_emit_cp_dma_copy_buffer(struct si_context *sctx,
- uint64_t dst_va, uint64_t src_va,
- unsigned size, unsigned flags)
+static void si_emit_cp_dma(struct si_context *sctx, struct radeon_cmdbuf *cs,
+ uint64_t dst_va, uint64_t src_va, unsigned size,
+ unsigned flags, enum si_cache_policy cache_policy)
{
- struct radeon_winsys_cs *cs = sctx->b.rings.gfx.cs;
- uint32_t sync_flag = flags & R600_CP_DMA_SYNC ? S_411_CP_SYNC(1) : 0;
- uint32_t raw_wait = flags & SI_CP_DMA_RAW_WAIT ? S_414_RAW_WAIT(1) : 0;
- uint32_t sel = flags & CIK_CP_DMA_USE_L2 ?
- S_411_SRC_SEL(V_411_SRC_ADDR_TC_L2) |
- S_411_DSL_SEL(V_411_DST_ADDR_TC_L2) : 0;
+ uint32_t header = 0, command = 0;
+
+ assert(size <= cp_dma_max_byte_count(sctx));
+ assert(sctx->chip_class != SI || cache_policy == L2_BYPASS);
+
+ if (sctx->chip_class >= GFX9)
+ command |= S_414_BYTE_COUNT_GFX9(size);
+ else
+ command |= S_414_BYTE_COUNT_GFX6(size);
+
+ /* Sync flags. */
+ if (flags & CP_DMA_SYNC)
+ header |= S_411_CP_SYNC(1);
+ else {
+ if (sctx->chip_class >= GFX9)
+ command |= S_414_DISABLE_WR_CONFIRM_GFX9(1);
+ else
+ command |= S_414_DISABLE_WR_CONFIRM_GFX6(1);
+ }
- assert(size);
- assert((size & ((1<<21)-1)) == size);
+ if (flags & CP_DMA_RAW_WAIT)
+ command |= S_414_RAW_WAIT(1);
+
+ /* Src and dst flags. */
+ if (sctx->chip_class >= GFX9 && !(flags & CP_DMA_CLEAR) &&
+ src_va == dst_va) {
+ header |= S_411_DST_SEL(V_411_NOWHERE); /* prefetch only */
+ } else if (flags & CP_DMA_DST_IS_GDS) {
+ header |= S_411_DST_SEL(V_411_GDS);
+ /* GDS increments the address, not CP. */
+ command |= S_414_DAS(V_414_REGISTER) |
+ S_414_DAIC(V_414_NO_INCREMENT);
+ } else if (sctx->chip_class >= CIK && cache_policy != L2_BYPASS) {
+ header |= S_411_DST_SEL(V_411_DST_ADDR_TC_L2) |
+ S_500_DST_CACHE_POLICY(cache_policy == L2_STREAM);
+ }
- if (sctx->b.chip_class >= CIK) {
+ if (flags & CP_DMA_CLEAR) {
+ header |= S_411_SRC_SEL(V_411_DATA);
+ } else if (flags & CP_DMA_SRC_IS_GDS) {
+ header |= S_411_SRC_SEL(V_411_GDS);
+ /* Both of these are required for GDS. It does increment the address. */
+ command |= S_414_SAS(V_414_REGISTER) |
+ S_414_SAIC(V_414_NO_INCREMENT);
+ } else if (sctx->chip_class >= CIK && cache_policy != L2_BYPASS) {
+ header |= S_411_SRC_SEL(V_411_SRC_ADDR_TC_L2) |
+ S_500_SRC_CACHE_POLICY(cache_policy == L2_STREAM);
+ }
+
+ if (sctx->chip_class >= CIK) {
radeon_emit(cs, PKT3(PKT3_DMA_DATA, 5, 0));
- radeon_emit(cs, sync_flag | sel); /* CP_SYNC [31] */
- radeon_emit(cs, src_va); /* SRC_ADDR_LO [31:0] */
- radeon_emit(cs, src_va >> 32); /* SRC_ADDR_HI [31:0] */
- radeon_emit(cs, dst_va); /* DST_ADDR_LO [31:0] */
- radeon_emit(cs, dst_va >> 32); /* DST_ADDR_HI [31:0] */
- radeon_emit(cs, size | raw_wait); /* COMMAND [29:22] | BYTE_COUNT [20:0] */
+ radeon_emit(cs, header);
+ radeon_emit(cs, src_va); /* SRC_ADDR_LO [31:0] */
+ radeon_emit(cs, src_va >> 32); /* SRC_ADDR_HI [31:0] */
+ radeon_emit(cs, dst_va); /* DST_ADDR_LO [31:0] */
+ radeon_emit(cs, dst_va >> 32); /* DST_ADDR_HI [31:0] */
+ radeon_emit(cs, command);
} else {
+ header |= S_411_SRC_ADDR_HI(src_va >> 32);
+
radeon_emit(cs, PKT3(PKT3_CP_DMA, 4, 0));
- radeon_emit(cs, src_va); /* SRC_ADDR_LO [31:0] */
- radeon_emit(cs, sync_flag | ((src_va >> 32) & 0xffff)); /* CP_SYNC [31] | SRC_ADDR_HI [15:0] */
- radeon_emit(cs, dst_va); /* DST_ADDR_LO [31:0] */
- radeon_emit(cs, (dst_va >> 32) & 0xffff); /* DST_ADDR_HI [15:0] */
- radeon_emit(cs, size | raw_wait); /* COMMAND [29:22] | BYTE_COUNT [20:0] */
+ radeon_emit(cs, src_va); /* SRC_ADDR_LO [31:0] */
+ radeon_emit(cs, header); /* SRC_ADDR_HI [15:0] + flags. */
+ radeon_emit(cs, dst_va); /* DST_ADDR_LO [31:0] */
+ radeon_emit(cs, (dst_va >> 32) & 0xffff); /* DST_ADDR_HI [15:0] */
+ radeon_emit(cs, command);
}
-}
-
-/* Emit a CP DMA packet to clear a buffer. The size must fit in bits [20:0]. */
-static void si_emit_cp_dma_clear_buffer(struct si_context *sctx,
- uint64_t dst_va, unsigned size,
- uint32_t clear_value, unsigned flags)
-{
- struct radeon_winsys_cs *cs = sctx->b.rings.gfx.cs;
- uint32_t sync_flag = flags & R600_CP_DMA_SYNC ? S_411_CP_SYNC(1) : 0;
- uint32_t raw_wait = flags & SI_CP_DMA_RAW_WAIT ? S_414_RAW_WAIT(1) : 0;
- uint32_t dst_sel = flags & CIK_CP_DMA_USE_L2 ? S_411_DSL_SEL(V_411_DST_ADDR_TC_L2) : 0;
- assert(size);
- assert((size & ((1<<21)-1)) == size);
-
- if (sctx->b.chip_class >= CIK) {
- radeon_emit(cs, PKT3(PKT3_DMA_DATA, 5, 0));
- radeon_emit(cs, sync_flag | dst_sel | S_411_SRC_SEL(V_411_DATA)); /* CP_SYNC [31] | SRC_SEL[30:29] */
- radeon_emit(cs, clear_value); /* DATA [31:0] */
+ /* CP DMA is executed in ME, but index buffers are read by PFP.
+ * This ensures that ME (CP DMA) is idle before PFP starts fetching
+ * indices. If we wanted to execute CP DMA in PFP, this packet
+ * should precede it.
+ */
+ if (flags & CP_DMA_PFP_SYNC_ME) {
+ radeon_emit(cs, PKT3(PKT3_PFP_SYNC_ME, 0, 0));
radeon_emit(cs, 0);
- radeon_emit(cs, dst_va); /* DST_ADDR_LO [31:0] */
- radeon_emit(cs, dst_va >> 32); /* DST_ADDR_HI [15:0] */
- radeon_emit(cs, size | raw_wait); /* COMMAND [29:22] | BYTE_COUNT [20:0] */
- } else {
- radeon_emit(cs, PKT3(PKT3_CP_DMA, 4, 0));
- radeon_emit(cs, clear_value); /* DATA [31:0] */
- radeon_emit(cs, sync_flag | S_411_SRC_SEL(V_411_DATA)); /* CP_SYNC [31] | SRC_SEL[30:29] */
- radeon_emit(cs, dst_va); /* DST_ADDR_LO [31:0] */
- radeon_emit(cs, (dst_va >> 32) & 0xffff); /* DST_ADDR_HI [15:0] */
- radeon_emit(cs, size | raw_wait); /* COMMAND [29:22] | BYTE_COUNT [20:0] */
}
}
-/* The max number of bytes to copy per packet. */
-#define CP_DMA_MAX_BYTE_COUNT ((1 << 21) - 8)
-
-static void si_clear_buffer(struct pipe_context *ctx, struct pipe_resource *dst,
- unsigned offset, unsigned size, unsigned value,
- bool is_framebuffer)
+void si_cp_dma_wait_for_idle(struct si_context *sctx)
{
- struct si_context *sctx = (struct si_context*)ctx;
- unsigned flush_flags, tc_l2_flag;
-
- if (!size)
- return;
+ /* Issue a dummy DMA that copies zero bytes.
+ *
+ * The DMA engine will see that there's no work to do and skip this
+ * DMA request, however, the CP will see the sync flag and still wait
+ * for all DMAs to complete.
+ */
+ si_emit_cp_dma(sctx, sctx->gfx_cs, 0, 0, 0, CP_DMA_SYNC, L2_BYPASS);
+}
- /* 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(&r600_resource(dst)->valid_buffer_range, offset,
- offset + size);
-
- /* Fallback for unaligned clears. */
- if (offset % 4 != 0 || size % 4 != 0) {
- uint32_t *map = sctx->b.ws->buffer_map(r600_resource(dst)->cs_buf,
- sctx->b.rings.gfx.cs,
- PIPE_TRANSFER_WRITE);
- size /= 4;
- for (unsigned i = 0; i < size; i++)
- *map++ = value;
+static void si_cp_dma_prepare(struct si_context *sctx, struct pipe_resource *dst,
+ struct pipe_resource *src, unsigned byte_count,
+ uint64_t remaining_size, unsigned user_flags,
+ enum si_coherency coher, bool *is_first,
+ unsigned *packet_flags)
+{
+ /* Fast exit for a CPDMA prefetch. */
+ if ((user_flags & SI_CPDMA_SKIP_ALL) == SI_CPDMA_SKIP_ALL) {
+ *is_first = false;
return;
}
- uint64_t va = r600_resource(dst)->gpu_address + offset;
+ if (!(user_flags & SI_CPDMA_SKIP_BO_LIST_UPDATE)) {
+ /* Count memory usage in so that need_cs_space can take it into account. */
+ if (dst)
+ si_context_add_resource_size(sctx, dst);
+ if (src)
+ si_context_add_resource_size(sctx, src);
+ }
- /* Flush the caches where the resource is bound. */
- if (is_framebuffer) {
- flush_flags = SI_CONTEXT_FLUSH_AND_INV_FRAMEBUFFER;
- tc_l2_flag = 0;
- } else {
- flush_flags = SI_CONTEXT_INV_TC_L1 |
- (sctx->b.chip_class == SI ? SI_CONTEXT_INV_TC_L2 : 0) |
- SI_CONTEXT_INV_KCACHE;
- tc_l2_flag = sctx->b.chip_class == SI ? 0 : CIK_CP_DMA_USE_L2;
+ if (!(user_flags & SI_CPDMA_SKIP_CHECK_CS_SPACE))
+ si_need_gfx_cs_space(sctx);
+
+ /* This must be done after need_cs_space. */
+ if (!(user_flags & SI_CPDMA_SKIP_BO_LIST_UPDATE)) {
+ if (dst)
+ radeon_add_to_buffer_list(sctx, sctx->gfx_cs,
+ r600_resource(dst),
+ RADEON_USAGE_WRITE, RADEON_PRIO_CP_DMA);
+ if (src)
+ radeon_add_to_buffer_list(sctx, sctx->gfx_cs,
+ r600_resource(src),
+ RADEON_USAGE_READ, RADEON_PRIO_CP_DMA);
}
- sctx->b.flags |= SI_CONTEXT_PS_PARTIAL_FLUSH |
- flush_flags;
+ /* Flush the caches for the first copy only.
+ * Also wait for the previous CP DMA operations.
+ */
+ if (!(user_flags & SI_CPDMA_SKIP_GFX_SYNC) && sctx->flags)
+ si_emit_cache_flush(sctx);
- while (size) {
- unsigned byte_count = MIN2(size, CP_DMA_MAX_BYTE_COUNT);
- unsigned dma_flags = tc_l2_flag;
+ if (!(user_flags & SI_CPDMA_SKIP_SYNC_BEFORE) && *is_first &&
+ !(*packet_flags & CP_DMA_CLEAR))
+ *packet_flags |= CP_DMA_RAW_WAIT;
- si_need_cs_space(sctx);
+ *is_first = false;
- /* This must be done after need_cs_space. */
- radeon_add_to_buffer_list(&sctx->b, &sctx->b.rings.gfx,
- (struct r600_resource*)dst, RADEON_USAGE_WRITE,
- RADEON_PRIO_CP_DMA);
+ /* Do the synchronization after the last dma, so that all data
+ * is written to memory.
+ */
+ if (!(user_flags & SI_CPDMA_SKIP_SYNC_AFTER) &&
+ byte_count == remaining_size) {
+ *packet_flags |= CP_DMA_SYNC;
- /* Flush the caches for the first copy only.
- * Also wait for the previous CP DMA operations. */
- if (sctx->b.flags) {
- si_emit_cache_flush(sctx, NULL);
- dma_flags |= SI_CP_DMA_RAW_WAIT; /* same as WAIT_UNTIL=CP_DMA_IDLE */
- }
+ if (coher == SI_COHERENCY_SHADER)
+ *packet_flags |= CP_DMA_PFP_SYNC_ME;
+ }
+}
+
+void si_cp_dma_clear_buffer(struct si_context *sctx, struct radeon_cmdbuf *cs,
+ struct pipe_resource *dst, uint64_t offset,
+ uint64_t size, unsigned value, unsigned user_flags,
+ enum si_coherency coher, enum si_cache_policy cache_policy)
+{
+ struct r600_resource *rdst = r600_resource(dst);
+ uint64_t va = (rdst ? rdst->gpu_address : 0) + offset;
+ bool is_first = true;
- /* Do the synchronization after the last copy, so that all data is written to memory. */
- if (size == byte_count)
- dma_flags |= R600_CP_DMA_SYNC;
+ assert(size && size % 4 == 0);
+
+ /* Mark the buffer range of destination as valid (initialized),
+ * so that transfer_map knows it should wait for the GPU when mapping
+ * that range. */
+ if (rdst)
+ util_range_add(&rdst->valid_buffer_range, offset, offset + size);
+
+ /* Flush the caches. */
+ if (rdst && !(user_flags & SI_CPDMA_SKIP_GFX_SYNC)) {
+ sctx->flags |= SI_CONTEXT_PS_PARTIAL_FLUSH |
+ SI_CONTEXT_CS_PARTIAL_FLUSH |
+ si_get_flush_flags(sctx, coher, cache_policy);
+ }
+
+ while (size) {
+ unsigned byte_count = MIN2(size, cp_dma_max_byte_count(sctx));
+ unsigned dma_flags = CP_DMA_CLEAR | (rdst ? 0 : CP_DMA_DST_IS_GDS);
+
+ si_cp_dma_prepare(sctx, dst, NULL, byte_count, size, user_flags,
+ coher, &is_first, &dma_flags);
/* Emit the clear packet. */
- si_emit_cp_dma_clear_buffer(sctx, va, byte_count, value, dma_flags);
+ si_emit_cp_dma(sctx, cs, va, value, byte_count, dma_flags, cache_policy);
size -= byte_count;
va += byte_count;
}
- /* Flush the caches again in case the 3D engine has been prefetching
- * the resource. */
- sctx->b.flags |= flush_flags;
+ if (rdst && cache_policy != L2_BYPASS)
+ rdst->TC_L2_dirty = true;
- if (tc_l2_flag)
- r600_resource(dst)->TC_L2_dirty = true;
+ /* If it's not a framebuffer fast clear... */
+ if (coher == SI_COHERENCY_SHADER)
+ sctx->num_cp_dma_calls++;
+}
+
+/**
+ * Realign the CP DMA engine. This must be done after a copy with an unaligned
+ * size.
+ *
+ * \param size Remaining size to the CP DMA alignment.
+ */
+static void si_cp_dma_realign_engine(struct si_context *sctx, unsigned size,
+ unsigned user_flags, enum si_coherency coher,
+ enum si_cache_policy cache_policy,
+ bool *is_first)
+{
+ uint64_t va;
+ unsigned dma_flags = 0;
+ unsigned scratch_size = SI_CPDMA_ALIGNMENT * 2;
+
+ assert(size < SI_CPDMA_ALIGNMENT);
+
+ /* Use the scratch buffer as the dummy buffer. The 3D engine should be
+ * idle at this point.
+ */
+ if (!sctx->scratch_buffer ||
+ sctx->scratch_buffer->b.b.width0 < scratch_size) {
+ r600_resource_reference(&sctx->scratch_buffer, NULL);
+ sctx->scratch_buffer =
+ si_aligned_buffer_create(&sctx->screen->b,
+ SI_RESOURCE_FLAG_UNMAPPABLE,
+ PIPE_USAGE_DEFAULT,
+ scratch_size, 256);
+ if (!sctx->scratch_buffer)
+ return;
+
+ si_mark_atom_dirty(sctx, &sctx->atoms.s.scratch_state);
+ }
+
+ si_cp_dma_prepare(sctx, &sctx->scratch_buffer->b.b,
+ &sctx->scratch_buffer->b.b, size, size, user_flags,
+ coher, is_first, &dma_flags);
+
+ va = sctx->scratch_buffer->gpu_address;
+ si_emit_cp_dma(sctx, sctx->gfx_cs, va, va + SI_CPDMA_ALIGNMENT, size, dma_flags,
+ cache_policy);
}
-void si_copy_buffer(struct si_context *sctx,
- struct pipe_resource *dst, struct pipe_resource *src,
- uint64_t dst_offset, uint64_t src_offset, unsigned size,
- bool is_framebuffer)
+/**
+ * Do memcpy between buffers using CP DMA.
+ * If src or dst is NULL, it means read or write GDS, respectively.
+ *
+ * \param user_flags bitmask of SI_CPDMA_*
+ */
+void si_cp_dma_copy_buffer(struct si_context *sctx,
+ struct pipe_resource *dst, struct pipe_resource *src,
+ uint64_t dst_offset, uint64_t src_offset, unsigned size,
+ unsigned user_flags, enum si_coherency coher,
+ enum si_cache_policy cache_policy)
{
- unsigned flush_flags, tc_l2_flag;
+ uint64_t main_dst_offset, main_src_offset;
+ unsigned skipped_size = 0;
+ unsigned realign_size = 0;
+ unsigned gds_flags = (dst ? 0 : CP_DMA_DST_IS_GDS) |
+ (src ? 0 : CP_DMA_SRC_IS_GDS);
+ bool is_first = true;
- if (!size)
- return;
+ assert(size);
- /* 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(&r600_resource(dst)->valid_buffer_range, dst_offset,
- dst_offset + size);
+ if (dst) {
+ /* Skip this for the L2 prefetch. */
+ if (dst != src || dst_offset != src_offset) {
+ /* 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(&r600_resource(dst)->valid_buffer_range, dst_offset,
+ dst_offset + size);
+ }
- dst_offset += r600_resource(dst)->gpu_address;
- src_offset += r600_resource(src)->gpu_address;
+ dst_offset += r600_resource(dst)->gpu_address;
+ }
+ if (src)
+ src_offset += r600_resource(src)->gpu_address;
+
+ /* The workarounds aren't needed on Fiji and beyond. */
+ if (sctx->family <= CHIP_CARRIZO ||
+ sctx->family == CHIP_STONEY) {
+ /* If the size is not aligned, we must add a dummy copy at the end
+ * just to align the internal counter. Otherwise, the DMA engine
+ * would slow down by an order of magnitude for following copies.
+ */
+ if (size % SI_CPDMA_ALIGNMENT)
+ realign_size = SI_CPDMA_ALIGNMENT - (size % SI_CPDMA_ALIGNMENT);
+
+ /* If the copy begins unaligned, we must start copying from the next
+ * aligned block and the skipped part should be copied after everything
+ * else has been copied. Only the src alignment matters, not dst.
+ *
+ * GDS doesn't need the source address to be aligned.
+ */
+ if (src && src_offset % SI_CPDMA_ALIGNMENT) {
+ skipped_size = SI_CPDMA_ALIGNMENT - (src_offset % SI_CPDMA_ALIGNMENT);
+ /* The main part will be skipped if the size is too small. */
+ skipped_size = MIN2(skipped_size, size);
+ size -= skipped_size;
+ }
+ }
- /* Flush the caches where the resource is bound. */
- if (is_framebuffer) {
- flush_flags = SI_CONTEXT_FLUSH_AND_INV_FRAMEBUFFER;
- tc_l2_flag = 0;
- } else {
- flush_flags = SI_CONTEXT_INV_TC_L1 |
- (sctx->b.chip_class == SI ? SI_CONTEXT_INV_TC_L2 : 0) |
- SI_CONTEXT_INV_KCACHE;
- tc_l2_flag = sctx->b.chip_class == SI ? 0 : CIK_CP_DMA_USE_L2;
+ /* Flush the caches. */
+ if ((dst || src) && !(user_flags & SI_CPDMA_SKIP_GFX_SYNC)) {
+ sctx->flags |= SI_CONTEXT_PS_PARTIAL_FLUSH |
+ SI_CONTEXT_CS_PARTIAL_FLUSH |
+ si_get_flush_flags(sctx, coher, cache_policy);
}
- sctx->b.flags |= SI_CONTEXT_PS_PARTIAL_FLUSH |
- flush_flags;
+ /* This is the main part doing the copying. Src is always aligned. */
+ main_dst_offset = dst_offset + skipped_size;
+ main_src_offset = src_offset + skipped_size;
while (size) {
- unsigned sync_flags = tc_l2_flag;
- unsigned byte_count = MIN2(size, CP_DMA_MAX_BYTE_COUNT);
+ unsigned byte_count = MIN2(size, cp_dma_max_byte_count(sctx));
+ unsigned dma_flags = gds_flags;
- si_need_cs_space(sctx);
+ si_cp_dma_prepare(sctx, dst, src, byte_count,
+ size + skipped_size + realign_size,
+ user_flags, coher, &is_first, &dma_flags);
- /* Flush the caches for the first copy only. Also wait for old CP DMA packets to complete. */
- if (sctx->b.flags) {
- si_emit_cache_flush(sctx, NULL);
- sync_flags |= SI_CP_DMA_RAW_WAIT;
- }
+ si_emit_cp_dma(sctx, sctx->gfx_cs, main_dst_offset, main_src_offset,
+ byte_count, dma_flags, cache_policy);
- /* Do the synchronization after the last copy, so that all data is written to memory. */
- if (size == byte_count) {
- sync_flags |= R600_CP_DMA_SYNC;
- }
+ size -= byte_count;
+ main_src_offset += byte_count;
+ main_dst_offset += byte_count;
+ }
- /* This must be done after r600_need_cs_space. */
- radeon_add_to_buffer_list(&sctx->b, &sctx->b.rings.gfx, (struct r600_resource*)src,
- RADEON_USAGE_READ, RADEON_PRIO_CP_DMA);
- radeon_add_to_buffer_list(&sctx->b, &sctx->b.rings.gfx, (struct r600_resource*)dst,
- RADEON_USAGE_WRITE, RADEON_PRIO_CP_DMA);
+ /* Copy the part we skipped because src wasn't aligned. */
+ if (skipped_size) {
+ unsigned dma_flags = gds_flags;
- si_emit_cp_dma_copy_buffer(sctx, dst_offset, src_offset, byte_count, sync_flags);
+ si_cp_dma_prepare(sctx, dst, src, skipped_size,
+ skipped_size + realign_size, user_flags,
+ coher, &is_first, &dma_flags);
- size -= byte_count;
- src_offset += byte_count;
- dst_offset += byte_count;
+ si_emit_cp_dma(sctx, sctx->gfx_cs, dst_offset, src_offset, skipped_size,
+ dma_flags, cache_policy);
}
- /* Flush the caches again in case the 3D engine has been prefetching
- * the resource. */
- sctx->b.flags |= flush_flags;
+ /* Finally, realign the engine if the size wasn't aligned. */
+ if (realign_size) {
+ si_cp_dma_realign_engine(sctx, realign_size, user_flags, coher,
+ cache_policy, &is_first);
+ }
- if (tc_l2_flag)
+ if (dst && cache_policy != L2_BYPASS)
r600_resource(dst)->TC_L2_dirty = true;
+
+ /* If it's not a prefetch or GDS copy... */
+ if (dst && src && (dst != src || dst_offset != src_offset))
+ sctx->num_cp_dma_calls++;
+}
+
+void cik_prefetch_TC_L2_async(struct si_context *sctx, struct pipe_resource *buf,
+ uint64_t offset, unsigned size)
+{
+ assert(sctx->chip_class >= CIK);
+
+ si_cp_dma_copy_buffer(sctx, buf, buf, offset, offset, size,
+ SI_CPDMA_SKIP_ALL, SI_COHERENCY_SHADER, L2_LRU);
+}
+
+static void cik_prefetch_shader_async(struct si_context *sctx,
+ struct si_pm4_state *state)
+{
+ struct pipe_resource *bo = &state->bo[0]->b.b;
+ assert(state->nbo == 1);
+
+ cik_prefetch_TC_L2_async(sctx, bo, 0, bo->width0);
+}
+
+static void cik_prefetch_VBO_descriptors(struct si_context *sctx)
+{
+ if (!sctx->vertex_elements || !sctx->vertex_elements->desc_list_byte_size)
+ return;
+
+ cik_prefetch_TC_L2_async(sctx, &sctx->vb_descriptors_buffer->b.b,
+ sctx->vb_descriptors_offset,
+ sctx->vertex_elements->desc_list_byte_size);
+}
+
+/**
+ * Prefetch shaders and VBO descriptors.
+ *
+ * \param vertex_stage_only Whether only the the API VS and VBO descriptors
+ * should be prefetched.
+ */
+void cik_emit_prefetch_L2(struct si_context *sctx, bool vertex_stage_only)
+{
+ unsigned mask = sctx->prefetch_L2_mask;
+ assert(mask);
+
+ /* Prefetch shaders and VBO descriptors to TC L2. */
+ if (sctx->chip_class >= GFX9) {
+ /* Choose the right spot for the VBO prefetch. */
+ if (sctx->tes_shader.cso) {
+ if (mask & SI_PREFETCH_HS)
+ cik_prefetch_shader_async(sctx, sctx->queued.named.hs);
+ if (mask & SI_PREFETCH_VBO_DESCRIPTORS)
+ cik_prefetch_VBO_descriptors(sctx);
+ if (vertex_stage_only) {
+ sctx->prefetch_L2_mask &= ~(SI_PREFETCH_HS |
+ SI_PREFETCH_VBO_DESCRIPTORS);
+ return;
+ }
+
+ if (mask & SI_PREFETCH_GS)
+ cik_prefetch_shader_async(sctx, sctx->queued.named.gs);
+ if (mask & SI_PREFETCH_VS)
+ cik_prefetch_shader_async(sctx, sctx->queued.named.vs);
+ } else if (sctx->gs_shader.cso) {
+ if (mask & SI_PREFETCH_GS)
+ cik_prefetch_shader_async(sctx, sctx->queued.named.gs);
+ if (mask & SI_PREFETCH_VBO_DESCRIPTORS)
+ cik_prefetch_VBO_descriptors(sctx);
+ if (vertex_stage_only) {
+ sctx->prefetch_L2_mask &= ~(SI_PREFETCH_GS |
+ SI_PREFETCH_VBO_DESCRIPTORS);
+ return;
+ }
+
+ if (mask & SI_PREFETCH_VS)
+ cik_prefetch_shader_async(sctx, sctx->queued.named.vs);
+ } else {
+ if (mask & SI_PREFETCH_VS)
+ cik_prefetch_shader_async(sctx, sctx->queued.named.vs);
+ if (mask & SI_PREFETCH_VBO_DESCRIPTORS)
+ cik_prefetch_VBO_descriptors(sctx);
+ if (vertex_stage_only) {
+ sctx->prefetch_L2_mask &= ~(SI_PREFETCH_VS |
+ SI_PREFETCH_VBO_DESCRIPTORS);
+ return;
+ }
+ }
+ } else {
+ /* SI-CI-VI */
+ /* Choose the right spot for the VBO prefetch. */
+ if (sctx->tes_shader.cso) {
+ if (mask & SI_PREFETCH_LS)
+ cik_prefetch_shader_async(sctx, sctx->queued.named.ls);
+ if (mask & SI_PREFETCH_VBO_DESCRIPTORS)
+ cik_prefetch_VBO_descriptors(sctx);
+ if (vertex_stage_only) {
+ sctx->prefetch_L2_mask &= ~(SI_PREFETCH_LS |
+ SI_PREFETCH_VBO_DESCRIPTORS);
+ return;
+ }
+
+ if (mask & SI_PREFETCH_HS)
+ cik_prefetch_shader_async(sctx, sctx->queued.named.hs);
+ if (mask & SI_PREFETCH_ES)
+ cik_prefetch_shader_async(sctx, sctx->queued.named.es);
+ if (mask & SI_PREFETCH_GS)
+ cik_prefetch_shader_async(sctx, sctx->queued.named.gs);
+ if (mask & SI_PREFETCH_VS)
+ cik_prefetch_shader_async(sctx, sctx->queued.named.vs);
+ } else if (sctx->gs_shader.cso) {
+ if (mask & SI_PREFETCH_ES)
+ cik_prefetch_shader_async(sctx, sctx->queued.named.es);
+ if (mask & SI_PREFETCH_VBO_DESCRIPTORS)
+ cik_prefetch_VBO_descriptors(sctx);
+ if (vertex_stage_only) {
+ sctx->prefetch_L2_mask &= ~(SI_PREFETCH_ES |
+ SI_PREFETCH_VBO_DESCRIPTORS);
+ return;
+ }
+
+ if (mask & SI_PREFETCH_GS)
+ cik_prefetch_shader_async(sctx, sctx->queued.named.gs);
+ if (mask & SI_PREFETCH_VS)
+ cik_prefetch_shader_async(sctx, sctx->queued.named.vs);
+ } else {
+ if (mask & SI_PREFETCH_VS)
+ cik_prefetch_shader_async(sctx, sctx->queued.named.vs);
+ if (mask & SI_PREFETCH_VBO_DESCRIPTORS)
+ cik_prefetch_VBO_descriptors(sctx);
+ if (vertex_stage_only) {
+ sctx->prefetch_L2_mask &= ~(SI_PREFETCH_VS |
+ SI_PREFETCH_VBO_DESCRIPTORS);
+ return;
+ }
+ }
+ }
+
+ if (mask & SI_PREFETCH_PS)
+ cik_prefetch_shader_async(sctx, sctx->queued.named.ps);
+
+ sctx->prefetch_L2_mask = 0;
}
-void si_init_cp_dma_functions(struct si_context *sctx)
+void si_test_gds(struct si_context *sctx)
{
- sctx->b.clear_buffer = si_clear_buffer;
+ struct pipe_context *ctx = &sctx->b;
+ struct pipe_resource *src, *dst;
+ unsigned r[4] = {};
+ unsigned offset = debug_get_num_option("OFFSET", 16);
+
+ src = pipe_buffer_create(ctx->screen, 0, PIPE_USAGE_DEFAULT, 16);
+ dst = pipe_buffer_create(ctx->screen, 0, PIPE_USAGE_DEFAULT, 16);
+ si_cp_dma_clear_buffer(sctx, sctx->gfx_cs, src, 0, 4, 0xabcdef01, 0, SI_COHERENCY_SHADER, L2_BYPASS);
+ si_cp_dma_clear_buffer(sctx, sctx->gfx_cs, src, 4, 4, 0x23456789, 0, SI_COHERENCY_SHADER, L2_BYPASS);
+ si_cp_dma_clear_buffer(sctx, sctx->gfx_cs, src, 8, 4, 0x87654321, 0, SI_COHERENCY_SHADER, L2_BYPASS);
+ si_cp_dma_clear_buffer(sctx, sctx->gfx_cs, src, 12, 4, 0xfedcba98, 0, SI_COHERENCY_SHADER, L2_BYPASS);
+ si_cp_dma_clear_buffer(sctx, sctx->gfx_cs, dst, 0, 16, 0xdeadbeef, 0, SI_COHERENCY_SHADER, L2_BYPASS);
+
+ si_cp_dma_copy_buffer(sctx, NULL, src, offset, 0, 16, 0, SI_COHERENCY_NONE, L2_BYPASS);
+ si_cp_dma_copy_buffer(sctx, dst, NULL, 0, offset, 16, 0, SI_COHERENCY_NONE, L2_BYPASS);
+
+ pipe_buffer_read(ctx, dst, 0, sizeof(r), r);
+ printf("GDS copy = %08x %08x %08x %08x -> %s\n", r[0], r[1], r[2], r[3],
+ r[0] == 0xabcdef01 && r[1] == 0x23456789 &&
+ r[2] == 0x87654321 && r[3] == 0xfedcba98 ? "pass" : "fail");
+
+ si_cp_dma_clear_buffer(sctx, sctx->gfx_cs, NULL, offset, 16, 0xc1ea4146, 0, SI_COHERENCY_NONE, L2_BYPASS);
+ si_cp_dma_copy_buffer(sctx, dst, NULL, 0, offset, 16, 0, SI_COHERENCY_NONE, L2_BYPASS);
+
+ pipe_buffer_read(ctx, dst, 0, sizeof(r), r);
+ printf("GDS clear = %08x %08x %08x %08x -> %s\n", r[0], r[1], r[2], r[3],
+ r[0] == 0xc1ea4146 && r[1] == 0xc1ea4146 &&
+ r[2] == 0xc1ea4146 && r[3] == 0xc1ea4146 ? "pass" : "fail");
+
+ pipe_resource_reference(&src, NULL);
+ pipe_resource_reference(&dst, NULL);
+ exit(0);
}
projects / mesa.git / blobdiff