X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;ds=sidebyside;f=src%2Fgallium%2Fdrivers%2Fradeonsi%2Fsi_cp_dma.c;h=d1b44b894b14337d3355df2fdf80ef873a77e454;hb=50fde5b20801a82bd95f14301707c32cf5df38a2;hp=1381d8cfc9462e519b7bf8715db28e5a986ccee0;hpb=71ad666414f00791047302dc31a79dfd41608c02;p=mesa.git diff --git a/src/gallium/drivers/radeonsi/si_cp_dma.c b/src/gallium/drivers/radeonsi/si_cp_dma.c index 1381d8cfc94..d1b44b894b1 100644 --- a/src/gallium/drivers/radeonsi/si_cp_dma.c +++ b/src/gallium/drivers/radeonsi/si_cp_dma.c @@ -1,5 +1,6 @@ /* * 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"), @@ -19,14 +20,10 @@ * 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 */ #include "si_pipe.h" #include "sid.h" -#include "radeon/r600_cs.h" /* Set this if you want the ME to wait until CP DMA is done. * It should be set on the last CP DMA packet. */ @@ -36,13 +33,15 @@ * packet. It's for preventing a read-after-write (RAW) hazard between two * CP DMA packets. */ #define CP_DMA_RAW_WAIT (1 << 1) -#define CP_DMA_USE_L2 (1 << 2) /* CIK+ */ +#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->b.chip_class >= GFX9 ? + unsigned max = sctx->chip_class >= GFX9 ? S_414_BYTE_COUNT_GFX9(~0u) : S_414_BYTE_COUNT_GFX6(~0u); @@ -55,17 +54,16 @@ static inline unsigned cp_dma_max_byte_count(struct si_context *sctx) * 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(struct si_context *sctx, uint64_t dst_va, - uint64_t src_va, unsigned size, unsigned flags, - enum r600_coherency coher) +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.gfx.cs; uint32_t header = 0, command = 0; - assert(size); assert(size <= cp_dma_max_byte_count(sctx)); + assert(sctx->chip_class != GFX6 || cache_policy == L2_BYPASS); - if (sctx->b.chip_class >= GFX9) + if (sctx->chip_class >= GFX9) command |= S_414_BYTE_COUNT_GFX9(size); else command |= S_414_BYTE_COUNT_GFX6(size); @@ -74,7 +72,7 @@ static void si_emit_cp_dma(struct si_context *sctx, uint64_t dst_va, if (flags & CP_DMA_SYNC) header |= S_411_CP_SYNC(1); else { - if (sctx->b.chip_class >= GFX9) + if (sctx->chip_class >= GFX9) command |= S_414_DISABLE_WR_CONFIRM_GFX9(1); else command |= S_414_DISABLE_WR_CONFIRM_GFX6(1); @@ -84,17 +82,32 @@ static void si_emit_cp_dma(struct si_context *sctx, uint64_t dst_va, command |= S_414_RAW_WAIT(1); /* Src and dst flags. */ - if (sctx->b.chip_class >= GFX9 && src_va == dst_va) - header |= S_411_DSL_SEL(V_411_NOWHERE); /* prefetch only */ - else if (flags & CP_DMA_USE_L2) - header |= S_411_DSL_SEL(V_411_DST_ADDR_TC_L2); + 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 >= GFX7 && 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 (flags & CP_DMA_CLEAR) + if (flags & CP_DMA_CLEAR) { header |= S_411_SRC_SEL(V_411_DATA); - else if (flags & CP_DMA_USE_L2) - header |= S_411_SRC_SEL(V_411_SRC_ADDR_TC_L2); + } 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 >= GFX7 && 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->b.chip_class >= CIK) { + if (sctx->chip_class >= GFX7) { radeon_emit(cs, PKT3(PKT3_DMA_DATA, 5, 0)); radeon_emit(cs, header); radeon_emit(cs, src_va); /* SRC_ADDR_LO [31:0] */ @@ -118,37 +131,28 @@ static void si_emit_cp_dma(struct si_context *sctx, uint64_t dst_va, * indices. If we wanted to execute CP DMA in PFP, this packet * should precede it. */ - if (coher == R600_COHERENCY_SHADER && flags & CP_DMA_SYNC) { + if (sctx->has_graphics && flags & CP_DMA_PFP_SYNC_ME) { radeon_emit(cs, PKT3(PKT3_PFP_SYNC_ME, 0, 0)); radeon_emit(cs, 0); } } -static unsigned get_flush_flags(struct si_context *sctx, enum r600_coherency coher) -{ - switch (coher) { - default: - case R600_COHERENCY_NONE: - return 0; - case R600_COHERENCY_SHADER: - return SI_CONTEXT_INV_SMEM_L1 | - SI_CONTEXT_INV_VMEM_L1 | - (sctx->b.chip_class == SI ? SI_CONTEXT_INV_GLOBAL_L2 : 0); - case R600_COHERENCY_CB_META: - return SI_CONTEXT_FLUSH_AND_INV_CB; - } -} - -static unsigned get_tc_l2_flag(struct si_context *sctx, enum r600_coherency coher) +void si_cp_dma_wait_for_idle(struct si_context *sctx) { - return coher == R600_COHERENCY_SHADER && - sctx->b.chip_class >= CIK ? CP_DMA_USE_L2 : 0; + /* 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); } 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, - bool *is_first, unsigned *packet_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) { @@ -158,32 +162,35 @@ static void si_cp_dma_prepare(struct si_context *sctx, struct pipe_resource *dst if (!(user_flags & SI_CPDMA_SKIP_BO_LIST_UPDATE)) { /* Count memory usage in so that need_cs_space can take it into account. */ - r600_context_add_resource_size(&sctx->b.b, dst); + if (dst) + si_context_add_resource_size(sctx, dst); if (src) - r600_context_add_resource_size(&sctx->b.b, src); + si_context_add_resource_size(sctx, src); } if (!(user_flags & SI_CPDMA_SKIP_CHECK_CS_SPACE)) - si_need_cs_space(sctx); + si_need_gfx_cs_space(sctx); /* This must be done after need_cs_space. */ if (!(user_flags & SI_CPDMA_SKIP_BO_LIST_UPDATE)) { - radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx, - (struct r600_resource*)dst, - RADEON_USAGE_WRITE, RADEON_PRIO_CP_DMA); + if (dst) + radeon_add_to_buffer_list(sctx, sctx->gfx_cs, + si_resource(dst), + RADEON_USAGE_WRITE, RADEON_PRIO_CP_DMA); if (src) - radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx, - (struct r600_resource*)src, + radeon_add_to_buffer_list(sctx, sctx->gfx_cs, + si_resource(src), RADEON_USAGE_READ, RADEON_PRIO_CP_DMA); } /* 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->b.flags) - si_emit_cache_flush(sctx); + if (!(user_flags & SI_CPDMA_SKIP_GFX_SYNC) && sctx->flags) + sctx->emit_cache_flush(sctx); - if (!(user_flags & SI_CPDMA_SKIP_SYNC_BEFORE) && *is_first) + if (!(user_flags & SI_CPDMA_SKIP_SYNC_BEFORE) && *is_first && + !(*packet_flags & CP_DMA_CLEAR)) *packet_flags |= CP_DMA_RAW_WAIT; *is_first = false; @@ -192,93 +199,59 @@ static void si_cp_dma_prepare(struct si_context *sctx, struct pipe_resource *dst * is written to memory. */ if (!(user_flags & SI_CPDMA_SKIP_SYNC_AFTER) && - byte_count == remaining_size) + byte_count == remaining_size) { *packet_flags |= CP_DMA_SYNC; + + if (coher == SI_COHERENCY_SHADER) + *packet_flags |= CP_DMA_PFP_SYNC_ME; + } } -static void si_clear_buffer(struct pipe_context *ctx, struct pipe_resource *dst, - uint64_t offset, uint64_t size, unsigned value, - enum r600_coherency coher) +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 si_context *sctx = (struct si_context*)ctx; - struct radeon_winsys *ws = sctx->b.ws; - struct r600_resource *rdst = r600_resource(dst); - unsigned tc_l2_flag = get_tc_l2_flag(sctx, coher); - unsigned flush_flags = get_flush_flags(sctx, coher); - uint64_t dma_clear_size; + struct si_resource *sdst = si_resource(dst); + uint64_t va = (sdst ? sdst->gpu_address : 0) + offset; bool is_first = true; - if (!size) - return; - - dma_clear_size = size & ~3llu; + 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. */ - util_range_add(&rdst->valid_buffer_range, offset, - offset + dma_clear_size); - - /* dma_clear_buffer can use clear_buffer on failure. Make sure that - * doesn't happen. We don't want an infinite recursion: */ - if (sctx->b.dma.cs && - (offset % 4 == 0) && - /* CP DMA is very slow. Always use SDMA for big clears. This - * alone improves DeusEx:MD performance by 70%. */ - (size > 128 * 1024 || - /* Buffers not used by the GFX IB yet will be cleared by SDMA. - * This happens to move most buffer clears to SDMA, including - * DCC and CMASK clears, because pipe->clear clears them before - * si_emit_framebuffer_state (in a draw call) adds them. - * For example, DeusEx:MD has 21 buffer clears per frame and all - * of them are moved to SDMA thanks to this. */ - !ws->cs_is_buffer_referenced(sctx->b.gfx.cs, rdst->buf, - RADEON_USAGE_READWRITE))) { - sctx->b.dma_clear_buffer(ctx, dst, offset, dma_clear_size, value); - - offset += dma_clear_size; - size -= dma_clear_size; - } else if (dma_clear_size >= 4) { - uint64_t va = rdst->gpu_address + offset; - - offset += dma_clear_size; - size -= dma_clear_size; - - /* Flush the caches. */ - sctx->b.flags |= SI_CONTEXT_PS_PARTIAL_FLUSH | - SI_CONTEXT_CS_PARTIAL_FLUSH | flush_flags; - - while (dma_clear_size) { - unsigned byte_count = MIN2(dma_clear_size, cp_dma_max_byte_count(sctx)); - unsigned dma_flags = tc_l2_flag | CP_DMA_CLEAR; - - si_cp_dma_prepare(sctx, dst, NULL, byte_count, dma_clear_size, 0, - &is_first, &dma_flags); - - /* Emit the clear packet. */ - si_emit_cp_dma(sctx, va, value, byte_count, dma_flags, coher); - - dma_clear_size -= byte_count; - va += byte_count; - } + if (sdst) + util_range_add(&sdst->valid_buffer_range, offset, offset + size); - if (tc_l2_flag) - rdst->TC_L2_dirty = true; + /* Flush the caches. */ + if (sdst && !(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 | (sdst ? 0 : CP_DMA_DST_IS_GDS); - /* If it's not a framebuffer fast clear... */ - if (coher == R600_COHERENCY_SHADER) - sctx->b.num_cp_dma_calls++; + 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(sctx, cs, va, value, byte_count, dma_flags, cache_policy); + + size -= byte_count; + va += byte_count; } - if (size) { - /* Handle non-dword alignment. - * - * This function is called for embedded texture metadata clears, - * but those should always be properly aligned. */ - assert(dst->target == PIPE_BUFFER); - assert(size < 4); + if (sdst && cache_policy != L2_BYPASS) + sdst->TC_L2_dirty = true; - pipe_buffer_write(ctx, dst, offset, size, &value); + /* If it's not a framebuffer fast clear... */ + if (coher == SI_COHERENCY_SHADER) { + sctx->num_cp_dma_calls++; + si_prim_discard_signal_next_compute_ib_start(sctx); } } @@ -289,7 +262,9 @@ static void si_clear_buffer(struct pipe_context *ctx, struct pipe_resource *dst, * \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, bool *is_first) + unsigned user_flags, enum si_coherency coher, + enum si_cache_policy cache_policy, + bool *is_first) { uint64_t va; unsigned dma_flags = 0; @@ -302,61 +277,66 @@ static void si_cp_dma_realign_engine(struct si_context *sctx, unsigned size, */ if (!sctx->scratch_buffer || sctx->scratch_buffer->b.b.width0 < scratch_size) { - r600_resource_reference(&sctx->scratch_buffer, NULL); - sctx->scratch_buffer = (struct r600_resource*) - r600_aligned_buffer_create(&sctx->screen->b.b, - R600_RESOURCE_FLAG_UNMAPPABLE, + si_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->scratch_state); + 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, - is_first, &dma_flags); + coher, is_first, &dma_flags); va = sctx->scratch_buffer->gpu_address; - si_emit_cp_dma(sctx, va, va + SI_CPDMA_ALIGNMENT, size, dma_flags, - R600_COHERENCY_SHADER); + si_emit_cp_dma(sctx, sctx->gfx_cs, va, va + SI_CPDMA_ALIGNMENT, size, dma_flags, + cache_policy); } /** * 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_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) +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) { uint64_t main_dst_offset, main_src_offset; unsigned skipped_size = 0; unsigned realign_size = 0; - unsigned tc_l2_flag = get_tc_l2_flag(sctx, R600_COHERENCY_SHADER); - unsigned flush_flags = get_flush_flags(sctx, R600_COHERENCY_SHADER); + 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); - 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); - } + 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(&si_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 += si_resource(dst)->gpu_address; + } + if (src) + src_offset += si_resource(src)->gpu_address; /* The workarounds aren't needed on Fiji and beyond. */ - if (sctx->b.family <= CHIP_CARRIZO || - sctx->b.family == CHIP_STONEY) { + 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. @@ -367,8 +347,10 @@ void si_copy_buffer(struct si_context *sctx, /* 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_offset % SI_CPDMA_ALIGNMENT) { + 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); @@ -377,24 +359,26 @@ void si_copy_buffer(struct si_context *sctx, } /* Flush the caches. */ - if (!(user_flags & SI_CPDMA_SKIP_GFX_SYNC)) - sctx->b.flags |= SI_CONTEXT_PS_PARTIAL_FLUSH | - SI_CONTEXT_CS_PARTIAL_FLUSH | flush_flags; + 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); + } /* 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 dma_flags = tc_l2_flag; unsigned byte_count = MIN2(size, cp_dma_max_byte_count(sctx)); + unsigned dma_flags = gds_flags; si_cp_dma_prepare(sctx, dst, src, byte_count, size + skipped_size + realign_size, - user_flags, &is_first, &dma_flags); + user_flags, coher, &is_first, &dma_flags); - si_emit_cp_dma(sctx, main_dst_offset, main_src_offset, - byte_count, dma_flags, R600_COHERENCY_SHADER); + si_emit_cp_dma(sctx, sctx->gfx_cs, main_dst_offset, main_src_offset, + byte_count, dma_flags, cache_policy); size -= byte_count; main_src_offset += byte_count; @@ -403,77 +387,253 @@ void si_copy_buffer(struct si_context *sctx, /* Copy the part we skipped because src wasn't aligned. */ if (skipped_size) { - unsigned dma_flags = tc_l2_flag; + unsigned dma_flags = gds_flags; si_cp_dma_prepare(sctx, dst, src, skipped_size, skipped_size + realign_size, user_flags, - &is_first, &dma_flags); + coher, &is_first, &dma_flags); - si_emit_cp_dma(sctx, dst_offset, src_offset, skipped_size, - dma_flags, R600_COHERENCY_SHADER); + si_emit_cp_dma(sctx, sctx->gfx_cs, dst_offset, src_offset, skipped_size, + dma_flags, cache_policy); } /* Finally, realign the engine if the size wasn't aligned. */ - if (realign_size) - si_cp_dma_realign_engine(sctx, realign_size, user_flags, - &is_first); + if (realign_size) { + si_cp_dma_realign_engine(sctx, realign_size, user_flags, coher, + cache_policy, &is_first); + } - if (tc_l2_flag) - r600_resource(dst)->TC_L2_dirty = true; + if (dst && cache_policy != L2_BYPASS) + si_resource(dst)->TC_L2_dirty = true; - /* If it's not a prefetch... */ - if (dst_offset != src_offset) - sctx->b.num_cp_dma_calls++; + /* If it's not a prefetch or GDS copy... */ + if (dst && src && (dst != src || dst_offset != src_offset)) { + sctx->num_cp_dma_calls++; + si_prim_discard_signal_next_compute_ib_start(sctx); + } } void cik_prefetch_TC_L2_async(struct si_context *sctx, struct pipe_resource *buf, uint64_t offset, unsigned size) { - assert(sctx->b.chip_class >= CIK); + assert(sctx->chip_class >= GFX7); - si_copy_buffer(sctx, buf, buf, offset, offset, size, SI_CPDMA_SKIP_ALL); + 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) { - if (state) { - struct pipe_resource *bo = &state->bo[0]->b.b; - assert(state->nbo == 1); + struct pipe_resource *bo = &state->bo[0]->b.b; + assert(state->nbo == 1); - cik_prefetch_TC_L2_async(sctx, bo, 0, bo->width0); - } + cik_prefetch_TC_L2_async(sctx, bo, 0, bo->width0); } -static void cik_emit_prefetch_L2(struct si_context *sctx, struct r600_atom *atom) +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 (si_pm4_state_changed(sctx, ls)) - cik_prefetch_shader_async(sctx, sctx->queued.named.ls); - if (si_pm4_state_changed(sctx, hs)) - cik_prefetch_shader_async(sctx, sctx->queued.named.hs); - if (si_pm4_state_changed(sctx, es)) - cik_prefetch_shader_async(sctx, sctx->queued.named.es); - if (si_pm4_state_changed(sctx, gs)) - cik_prefetch_shader_async(sctx, sctx->queued.named.gs); - if (si_pm4_state_changed(sctx, vs)) - cik_prefetch_shader_async(sctx, sctx->queued.named.vs); - - /* Vertex buffer descriptors are uploaded uncached, so prefetch - * them right after the VS binary. */ - if (sctx->vertex_buffer_pointer_dirty) { - cik_prefetch_TC_L2_async(sctx, &sctx->vertex_buffers.buffer->b.b, - sctx->vertex_buffers.buffer_offset, - sctx->vertex_elements->desc_list_byte_size); + if (sctx->chip_class >= GFX9) { + /* Choose the right spot for the VBO prefetch. */ + if (sctx->queued.named.hs) { + 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->queued.named.gs) { + 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 { + /* GFX6-GFX8 */ + /* 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 (si_pm4_state_changed(sctx, ps)) + + if (mask & SI_PREFETCH_PS) cik_prefetch_shader_async(sctx, sctx->queued.named.ps); + + sctx->prefetch_L2_mask = 0; +} + +void si_test_gds(struct si_context *sctx) +{ + 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); } -void si_init_cp_dma_functions(struct si_context *sctx) +void si_cp_write_data(struct si_context *sctx, struct si_resource *buf, + unsigned offset, unsigned size, unsigned dst_sel, + unsigned engine, const void *data) { - sctx->b.clear_buffer = si_clear_buffer; + struct radeon_cmdbuf *cs = sctx->gfx_cs; + + assert(offset % 4 == 0); + assert(size % 4 == 0); + + if (sctx->chip_class == GFX6 && dst_sel == V_370_MEM) + dst_sel = V_370_MEM_GRBM; + + radeon_add_to_buffer_list(sctx, cs, buf, + RADEON_USAGE_WRITE, RADEON_PRIO_CP_DMA); + uint64_t va = buf->gpu_address + offset; + + radeon_emit(cs, PKT3(PKT3_WRITE_DATA, 2 + size/4, 0)); + radeon_emit(cs, S_370_DST_SEL(dst_sel) | + S_370_WR_CONFIRM(1) | + S_370_ENGINE_SEL(engine)); + radeon_emit(cs, va); + radeon_emit(cs, va >> 32); + radeon_emit_array(cs, (const uint32_t*)data, size/4); +} + +void si_cp_copy_data(struct si_context *sctx, struct radeon_cmdbuf *cs, + unsigned dst_sel, struct si_resource *dst, unsigned dst_offset, + unsigned src_sel, struct si_resource *src, unsigned src_offset) +{ + /* cs can point to the compute IB, which has the buffer list in gfx_cs. */ + if (dst) { + radeon_add_to_buffer_list(sctx, sctx->gfx_cs, dst, + RADEON_USAGE_WRITE, RADEON_PRIO_CP_DMA); + } + if (src) { + radeon_add_to_buffer_list(sctx, sctx->gfx_cs, src, + RADEON_USAGE_READ, RADEON_PRIO_CP_DMA); + } - si_init_atom(sctx, &sctx->prefetch_L2, &sctx->atoms.s.prefetch_L2, - cik_emit_prefetch_L2); + uint64_t dst_va = (dst ? dst->gpu_address : 0ull) + dst_offset; + uint64_t src_va = (src ? src->gpu_address : 0ull) + src_offset; + + radeon_emit(cs, PKT3(PKT3_COPY_DATA, 4, 0)); + radeon_emit(cs, COPY_DATA_SRC_SEL(src_sel) | + COPY_DATA_DST_SEL(dst_sel) | + COPY_DATA_WR_CONFIRM); + radeon_emit(cs, src_va); + radeon_emit(cs, src_va >> 32); + radeon_emit(cs, dst_va); + radeon_emit(cs, dst_va >> 32); }