[Mesa-dev] [PATCH 08/15] radeonsi: add GDS support to CP DMA

Samuel Pitoiset samuel.pitoiset at gmail.com
Thu Oct 4 07:27:15 UTC 2018


Did you benchmark copies and clears with GDS?

On 10/3/18 12:35 AM, Marek Olšák wrote:
> From: Marek Olšák <marek.olsak at amd.com>
> 
> ---
>   src/gallium/drivers/radeonsi/si_cp_dma.c | 104 ++++++++++++++++++-----
>   src/gallium/drivers/radeonsi/si_pipe.c   |   4 +
>   src/gallium/drivers/radeonsi/si_pipe.h   |   2 +
>   3 files changed, 89 insertions(+), 21 deletions(-)
> 
> diff --git a/src/gallium/drivers/radeonsi/si_cp_dma.c b/src/gallium/drivers/radeonsi/si_cp_dma.c
> index e85bb9b1acf..c1ecd5fb3e8 100644
> --- a/src/gallium/drivers/radeonsi/si_cp_dma.c
> +++ b/src/gallium/drivers/radeonsi/si_cp_dma.c
> @@ -32,22 +32,24 @@
>   #define CP_DMA_CLEAR_PERF_THRESHOLD	(32 * 1024) /* guess (clear is much slower) */
>   
>   /* 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 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 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);
> @@ -83,27 +85,37 @@ static void si_emit_cp_dma(struct si_context *sctx, uint64_t dst_va,
>   			command |= S_414_DISABLE_WR_CONFIRM_GFX6(1);
>   	}
>   
>   	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 (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, header);
>   		radeon_emit(cs, src_va);	/* SRC_ADDR_LO [31:0] */
>   		radeon_emit(cs, src_va >> 32);	/* SRC_ADDR_HI [31:0] */
> @@ -179,33 +191,35 @@ static void si_cp_dma_prepare(struct si_context *sctx, struct pipe_resource *dst
>   			      unsigned *packet_flags)
>   {
>   	/* Fast exit for a CPDMA prefetch. */
>   	if ((user_flags & SI_CPDMA_SKIP_ALL) == SI_CPDMA_SKIP_ALL) {
>   		*is_first = false;
>   		return;
>   	}
>   
>   	if (!(user_flags & SI_CPDMA_SKIP_BO_LIST_UPDATE)) {
>   		/* Count memory usage in so that need_cs_space can take it into account. */
> -		si_context_add_resource_size(sctx, dst);
> +		if (dst)
> +			si_context_add_resource_size(sctx, dst);
>   		if (src)
>   			si_context_add_resource_size(sctx, src);
>   	}
>   
>   	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)) {
> -		radeon_add_to_buffer_list(sctx, sctx->gfx_cs,
> -					  r600_resource(dst),
> -					  RADEON_USAGE_WRITE, RADEON_PRIO_CP_DMA);
> +		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);
>   	}
>   
>   	/* 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)
> @@ -227,57 +241,59 @@ static void si_cp_dma_prepare(struct si_context *sctx, struct pipe_resource *dst
>   			*packet_flags |= CP_DMA_PFP_SYNC_ME;
>   	}
>   }
>   
>   void si_cp_dma_clear_buffer(struct si_context *sctx, struct pipe_resource *dst,
>   			    uint64_t offset, uint64_t size, unsigned value,
>   			    enum si_coherency coher,
>   			    enum si_cache_policy cache_policy)
>   {
>   	struct r600_resource *rdst = r600_resource(dst);
> -	uint64_t va = rdst->gpu_address + offset;
> +	uint64_t va = (rdst ? rdst->gpu_address : 0) + offset;
>   	bool is_first = true;
>   
>   	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 + size);
> +	if (rdst)
> +		util_range_add(&rdst->valid_buffer_range, offset, offset + size);
>   
>   	/* Flush the caches. */
>   	sctx->flags |= SI_CONTEXT_PS_PARTIAL_FLUSH |
>   		       SI_CONTEXT_CS_PARTIAL_FLUSH |
>   		       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;
> +		unsigned dma_flags = CP_DMA_CLEAR | (rdst ? 0 : CP_DMA_DST_IS_GDS);
>   
>   		si_cp_dma_prepare(sctx, dst, NULL, byte_count, size, 0, coher,
>   				  &is_first, &dma_flags);
>   
>   		/* Emit the clear packet. */
>   		si_emit_cp_dma(sctx, va, value, byte_count, dma_flags, cache_policy);
>   
>   		size -= byte_count;
>   		va += byte_count;
>   	}
>   
> -	if (cache_policy != L2_BYPASS)
> +	if (rdst && cache_policy != L2_BYPASS)
>   		rdst->TC_L2_dirty = true;
>   
>   	/* If it's not a framebuffer fast clear... */
>   	if (coher == SI_COHERENCY_SHADER)
>   		sctx->num_cp_dma_calls++;
>   }
>   
> +/* dst == NULL means GDS. */
>   void si_clear_buffer(struct si_context *sctx, struct pipe_resource *dst,
>   		     uint64_t offset, uint64_t size, unsigned value,
>   		     enum si_coherency coher)
>   {
>   	struct radeon_winsys *ws = sctx->ws;
>   	struct r600_resource *rdst = r600_resource(dst);
>   	enum si_cache_policy cache_policy = get_cache_policy(sctx, coher);
>   	uint64_t dma_clear_size;
>   
>   	if (!size)
> @@ -311,20 +327,21 @@ void si_clear_buffer(struct si_context *sctx, struct pipe_resource *dst,
>   
>   		offset += dma_clear_size;
>   		size -= dma_clear_size;
>   	}
>   
>   	if (size) {
>   		/* Handle non-dword alignment.
>   		 *
>   		 * This function is called for embedded texture metadata clears,
>   		 * but those should always be properly aligned. */
> +		assert(dst);
>   		assert(dst->target == PIPE_BUFFER);
>   		assert(size < 4);
>   
>   		pipe_buffer_write(&sctx->b, dst, offset, size, &value);
>   	}
>   }
>   
>   static void si_pipe_clear_buffer(struct pipe_context *ctx,
>   				 struct pipe_resource *dst,
>   				 unsigned offset, unsigned size,
> @@ -423,99 +440,108 @@ static void si_cp_dma_realign_engine(struct si_context *sctx, unsigned size,
>   			  &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, 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_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 gds_flags = (dst ? 0 : CP_DMA_DST_IS_GDS) |
> +			     (src ? 0 : CP_DMA_SRC_IS_GDS);
>   	bool is_first = true;
>   
>   	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(&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_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);
>   			size -= skipped_size;
>   		}
>   	}
>   
>   	/* Flush the caches. */
> -	if (!(user_flags & SI_CPDMA_SKIP_GFX_SYNC)) {
> +	if ((dst || src) && !(user_flags & SI_CPDMA_SKIP_GFX_SYNC)) {
>   		sctx->flags |= SI_CONTEXT_PS_PARTIAL_FLUSH |
>   			       SI_CONTEXT_CS_PARTIAL_FLUSH |
>   			       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 byte_count = MIN2(size, cp_dma_max_byte_count(sctx));
> -		unsigned dma_flags = 0;
> +		unsigned dma_flags = gds_flags;
>   
>   		si_cp_dma_prepare(sctx, dst, src, byte_count,
>   				  size + skipped_size + realign_size,
>   				  user_flags, coher, &is_first, &dma_flags);
>   
>   		si_emit_cp_dma(sctx, main_dst_offset, main_src_offset,
>   			       byte_count, dma_flags, cache_policy);
>   
>   		size -= byte_count;
>   		main_src_offset += byte_count;
>   		main_dst_offset += byte_count;
>   	}
>   
>   	/* Copy the part we skipped because src wasn't aligned. */
>   	if (skipped_size) {
> -		unsigned dma_flags = 0;
> +		unsigned dma_flags = gds_flags;
>   
>   		si_cp_dma_prepare(sctx, dst, src, skipped_size,
>   				  skipped_size + realign_size, user_flags,
>   				  coher, &is_first, &dma_flags);
>   
>   		si_emit_cp_dma(sctx, dst_offset, src_offset, skipped_size,
>   			       dma_flags, cache_policy);
>   	}
>   
>   	/* Finally, realign the engine if the size wasn't aligned. */
> @@ -676,14 +702,50 @@ void cik_emit_prefetch_L2(struct si_context *sctx, bool vertex_stage_only)
>   			}
>   		}
>   	}
>   
>   	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, src, 0, 4, 0xabcdef01, SI_COHERENCY_SHADER, L2_BYPASS);
> +	si_cp_dma_clear_buffer(sctx, src, 4, 4, 0x23456789, SI_COHERENCY_SHADER, L2_BYPASS);
> +	si_cp_dma_clear_buffer(sctx, src, 8, 4, 0x87654321, SI_COHERENCY_SHADER, L2_BYPASS);
> +	si_cp_dma_clear_buffer(sctx, src, 12, 4, 0xfedcba98, SI_COHERENCY_SHADER, L2_BYPASS);
> +	si_cp_dma_clear_buffer(sctx, dst, 0, 16, 0xdeadbeef, 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, NULL, offset, 16, 0xc1ea4146, 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)
>   {
>   	sctx->b.clear_buffer = si_pipe_clear_buffer;
>   }
> diff --git a/src/gallium/drivers/radeonsi/si_pipe.c b/src/gallium/drivers/radeonsi/si_pipe.c
> index aa5340668f9..c0b23ee5f6e 100644
> --- a/src/gallium/drivers/radeonsi/si_pipe.c
> +++ b/src/gallium/drivers/radeonsi/si_pipe.c
> @@ -95,20 +95,21 @@ static const struct debug_named_value debug_options[] = {
>   	{ "nodccfb", DBG(NO_DCC_FB), "Disable separate DCC on the main framebuffer" },
>   	{ "nodccmsaa", DBG(NO_DCC_MSAA), "Disable DCC for MSAA" },
>   	{ "nofmask", DBG(NO_FMASK), "Disable MSAA compression" },
>   
>   	/* Tests: */
>   	{ "testdma", DBG(TEST_DMA), "Invoke SDMA tests and exit." },
>   	{ "testvmfaultcp", DBG(TEST_VMFAULT_CP), "Invoke a CP VM fault test and exit." },
>   	{ "testvmfaultsdma", DBG(TEST_VMFAULT_SDMA), "Invoke a SDMA VM fault test and exit." },
>   	{ "testvmfaultshader", DBG(TEST_VMFAULT_SHADER), "Invoke a shader VM fault test and exit." },
>   	{ "testdmaperf", DBG(TEST_DMA_PERF), "Test DMA performance" },
> +	{ "testgds", DBG(TEST_GDS), "Test GDS." },
>   
>   	DEBUG_NAMED_VALUE_END /* must be last */
>   };
>   
>   static void si_init_compiler(struct si_screen *sscreen,
>   			     struct ac_llvm_compiler *compiler)
>   {
>   	/* Only create the less-optimizing version of the compiler on APUs
>   	 * predating Ryzen (Raven). */
>   	bool create_low_opt_compiler = !sscreen->info.has_dedicated_vram &&
> @@ -1089,12 +1090,15 @@ struct pipe_screen *radeonsi_screen_create(struct radeon_winsys *ws,
>   
>   	if (sscreen->debug_flags & DBG(TEST_DMA_PERF)) {
>   		si_test_dma_perf(sscreen);
>   	}
>   
>   	if (sscreen->debug_flags & (DBG(TEST_VMFAULT_CP) |
>   				      DBG(TEST_VMFAULT_SDMA) |
>   				      DBG(TEST_VMFAULT_SHADER)))
>   		si_test_vmfault(sscreen);
>   
> +	if (sscreen->debug_flags & DBG(TEST_GDS))
> +		si_test_gds((struct si_context*)sscreen->aux_context);
> +
>   	return &sscreen->b;
>   }
> diff --git a/src/gallium/drivers/radeonsi/si_pipe.h b/src/gallium/drivers/radeonsi/si_pipe.h
> index bb851374c54..dad3029bc31 100644
> --- a/src/gallium/drivers/radeonsi/si_pipe.h
> +++ b/src/gallium/drivers/radeonsi/si_pipe.h
> @@ -159,20 +159,21 @@ enum {
>   	DBG_NO_DCC_FB,
>   	DBG_NO_DCC_MSAA,
>   	DBG_NO_FMASK,
>   
>   	/* Tests: */
>   	DBG_TEST_DMA,
>   	DBG_TEST_VMFAULT_CP,
>   	DBG_TEST_VMFAULT_SDMA,
>   	DBG_TEST_VMFAULT_SHADER,
>   	DBG_TEST_DMA_PERF,
> +	DBG_TEST_GDS,
>   };
>   
>   #define DBG_ALL_SHADERS		(((1 << (DBG_CS + 1)) - 1))
>   #define DBG(name)		(1ull << DBG_##name)
>   
>   struct si_compute;
>   struct hash_table;
>   struct u_suballocator;
>   
>   /* Only 32-bit buffer allocations are supported, gallium doesn't support more
> @@ -1131,20 +1132,21 @@ 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);
>   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);
>   void cik_prefetch_TC_L2_async(struct si_context *sctx, struct pipe_resource *buf,
>   			      uint64_t offset, unsigned size);
>   void cik_emit_prefetch_L2(struct si_context *sctx, bool vertex_stage_only);
> +void si_test_gds(struct si_context *sctx);
>   void si_init_cp_dma_functions(struct si_context *sctx);
>   
>   /* si_debug.c */
>   void si_save_cs(struct radeon_winsys *ws, struct radeon_cmdbuf *cs,
>   		struct radeon_saved_cs *saved, bool get_buffer_list);
>   void si_clear_saved_cs(struct radeon_saved_cs *saved);
>   void si_destroy_saved_cs(struct si_saved_cs *scs);
>   void si_auto_log_cs(void *data, struct u_log_context *log);
>   void si_log_hw_flush(struct si_context *sctx);
>   void si_log_draw_state(struct si_context *sctx, struct u_log_context *log);
> 


More information about the mesa-dev mailing list