[Mesa-dev] [PATCH V3 17/22] i965/gen9: Add XY_FAST_COPY_BLT support to intelEmitCopyBlit()
Anuj Phogat
anuj.phogat at gmail.com
Wed Jun 3 20:11:11 PDT 2015
On Wed, Jun 3, 2015 at 3:25 PM, Ben Widawsky <ben at bwidawsk.net> wrote:
> On Tue, Jun 02, 2015 at 04:04:07PM -0700, Anuj Phogat wrote:
>> This patch enables using XY_FAST_COPY_BLT only for Yf/Ys tiled buffers.
>> It can be later turned on for other tiling patterns (X,Y) too.
>>
>> V3: Flush in between sequential fast copy blits.
>> Fix src/dst alignment requirements.
>> Make can_fast_copy_blit() helper.
>> Use ffs(), is_power_of_two()
>> Move overlap computation inside intel_miptree_blit().
>> Use XY_FAST prefix for the macros.
>>
>> Signed-off-by: Anuj Phogat <anuj.phogat at gmail.com>
>> Cc: Ben Widawsky <ben at bwidawsk.net>
>> ---
>> src/mesa/drivers/dri/i965/intel_blit.c | 306 +++++++++++++++++++++++----
>> src/mesa/drivers/dri/i965/intel_blit.h | 2 +
>> src/mesa/drivers/dri/i965/intel_copy_image.c | 2 +
>> src/mesa/drivers/dri/i965/intel_reg.h | 36 ++++
>> 4 files changed, 299 insertions(+), 47 deletions(-)
>>
>> diff --git a/src/mesa/drivers/dri/i965/intel_blit.c b/src/mesa/drivers/dri/i965/intel_blit.c
>> index cd89f9d..8f0b48b 100644
>> --- a/src/mesa/drivers/dri/i965/intel_blit.c
>> +++ b/src/mesa/drivers/dri/i965/intel_blit.c
>> @@ -43,6 +43,23 @@
>>
>> #define FILE_DEBUG_FLAG DEBUG_BLIT
>>
>> +#define SET_TILING_XY_FAST_COPY_BLT(tiling, tr_mode, type) \
>> +({ \
>> + switch (tiling) { \
>> + case I915_TILING_X: \
>> + CMD |= type ## _TILED_X; \
>> + break; \
>> + case I915_TILING_Y: \
>> + if (tr_mode == INTEL_MIPTREE_TRMODE_YS) \
>> + CMD |= type ## _TILED_64K; \
>> + else \
>> + CMD |= type ## _TILED_Y; \
>> + break; \
>> + default: \
>> + unreachable("not reached"); \
>> + } \
>> +})
>> +
>> static void
>> intel_miptree_set_alpha_to_one(struct brw_context *brw,
>> struct intel_mipmap_tree *mt,
>> @@ -75,6 +92,10 @@ static uint32_t
>> br13_for_cpp(int cpp)
>> {
>> switch (cpp) {
>> + case 16:
>> + return BR13_32323232;
>> + case 8:
>> + return BR13_16161616;
>> case 4:
>> return BR13_8888;
>> case 2:
>> @@ -86,6 +107,81 @@ br13_for_cpp(int cpp)
>> }
>> }
>>
>> +static uint32_t
>> +get_tr_horizontal_align(uint32_t tr_mode, uint32_t cpp, bool is_src) {
>> + /*Alignment tables for YF/YS tiled surfaces. */
>> + const uint32_t align_2d_yf[] = {64, 64, 32, 32, 16};
>> + const uint32_t align_2d_ys[] = {256, 256, 128, 128, 64};
>> + const uint32_t bpp = cpp * 8;
>> + uint32_t align;
>> + int i = 0;
>> +
>> + if (tr_mode == INTEL_MIPTREE_TRMODE_NONE)
>> + return 0;
>> +
>> + /* Compute array index. */
>> + assert (bpp >= 8 && bpp <= 128 && is_power_of_two(bpp));
>> + i = ffs(bpp / 8) - 1;
>> +
>> + if (tr_mode == INTEL_MIPTREE_TRMODE_YF)
>> + align = align_2d_yf[i];
>> + else
>> + align = align_2d_ys[i];
>> +
>> + switch(align) {
>> + case 512:
>> + return is_src ? XY_FAST_SRC_H_ALIGN_512 : XY_FAST_DST_H_ALIGN_512;
>> + case 256:
>> + return is_src ? XY_FAST_SRC_H_ALIGN_256 : XY_FAST_DST_H_ALIGN_256;
>> + case 128:
>> + return is_src ? XY_FAST_SRC_H_ALIGN_128 : XY_FAST_DST_H_ALIGN_128;
>> + case 64:
>> + return is_src ? XY_FAST_SRC_H_ALIGN_64 : XY_FAST_DST_H_ALIGN_64;
>> + case 32:
>> + /* XY_FAST_COPY_BLT doesn't support horizontal alignment of 16. */
>> + case 16:
>> + return is_src ? XY_FAST_SRC_H_ALIGN_32 : XY_FAST_DST_H_ALIGN_32;
>> + default:
>> + unreachable("not reached");
>> + }
>> +}
>
> I'd make this smaller, but I think my brain works differently than most. I don't
> see anything wrong with this code though, so it's up to you.
>
I agree. It'll be nice to make it shorter.
> int shift = is_src ? 17 : 10
>
> Then either use your table, or:
> /* XY_FAST_COPY_BLT doesn't support horizontal alignment of 16. */
> if (align == 16)
> align = 32;
>
> return (ffs(align) - 6) << shift;
>
>> +
>> +static uint32_t
>> +get_tr_vertical_align(uint32_t tr_mode, uint32_t cpp, bool is_src) {
>> + /* Vertical alignment tables for YF/YS tiled surfaces. */
>> + const unsigned align_2d_yf[] = {64, 32, 32, 16, 16};
>> + const unsigned align_2d_ys[] = {256, 128, 128, 64, 64};
>> + const uint32_t bpp = cpp * 8;
>> + uint32_t align;
>> + int i = 0;
>> +
>> + if (tr_mode == INTEL_MIPTREE_TRMODE_NONE)
>> + return 0;
>> +
>> + /* Compute array index. */
>> + assert (bpp >= 8 && bpp <= 128 && is_power_of_two(bpp));
>> + i = ffs(bpp / 8) - 1;
>> +
>> + if (tr_mode == INTEL_MIPTREE_TRMODE_YF)
>> + align = align_2d_yf[i];
>> + else
>> + align = align_2d_ys[i];
>> +
>> + switch(align) {
>> + case 256:
>> + return is_src ? XY_FAST_SRC_V_ALIGN_256 : XY_FAST_DST_V_ALIGN_256;
>> + case 128:
>> + return is_src ? XY_FAST_SRC_V_ALIGN_128 : XY_FAST_DST_V_ALIGN_128;
>> + case 64:
>> + /* XY_FAST_COPY_BLT doesn't support vertical alignments of 16 and 32. */
>> + case 32:
>> + case 16:
>> + return is_src ? XY_FAST_SRC_V_ALIGN_64 : XY_FAST_DST_V_ALIGN_64;
>> + default:
>> + unreachable("not reached");
>> + }
>> +}
>> +
>> /**
>> * Emits the packet for switching the blitter from X to Y tiled or back.
>> *
>> @@ -248,9 +344,11 @@ intel_miptree_blit(struct brw_context *brw,
>> src_pitch,
>> src_mt->bo, src_mt->offset,
>> src_mt->tiling,
>> + src_mt->tr_mode,
>> dst_mt->pitch,
>> dst_mt->bo, dst_mt->offset,
>> dst_mt->tiling,
>> + dst_mt->tr_mode,
>> src_x, src_y,
>> dst_x, dst_y,
>> width, height,
>> @@ -282,6 +380,63 @@ alignment_valid(struct brw_context *brw, unsigned offset, uint32_t tiling)
>> return true;
>> }
>>
>> +static bool
>> +can_fast_copy_blit(struct brw_context *brw,
>> + int16_t src_x, uintptr_t src_offset, uint32_t src_pitch,
>> + uint32_t src_tiling, uint32_t src_tr_mode,
>> + int16_t dst_x, uintptr_t dst_offset, uint32_t dst_pitch,
>> + uint32_t dst_tiling, uint32_t dst_tr_mode,
>> + int16_t w, uint32_t cpp)
>> +{
>> + const bool dst_tiling_none = dst_tiling == I915_TILING_NONE;
>> + const bool src_tiling_none = src_tiling == I915_TILING_NONE;
>> +
>> + if (brw->gen < 9)
>> + return false;
>> +
>> + /* Check for an overlap. */
>> + if ((src_x < dst_x && (src_x + w) > dst_x) ||
>> + (src_x > dst_x && (dst_x + w) > src_x))
>> + return false;
>
> Can you remind me why Y isn't a factor in determining overlap?
>
This condition is incomplete :(. I'll fix this in V4 by using existing function
regions_overlap() to do the check here.
>> +
>> + /* Enable fast copy blit only if the surfaces are Yf/Ys tiled.
>> + * FIXME: Based on performance data, remove this condition later to
>> + * enable for all types of surfaces.
>> + */
>> + if (src_tr_mode == INTEL_MIPTREE_TRMODE_NONE &&
>> + dst_tr_mode == INTEL_MIPTREE_TRMODE_NONE)
>> + return false;
>> +
>> + /* For all surface types buffers must be cacheline-aligned. */
>> + if ((dst_offset | src_offset) & 63)
>> + return false;
>> +
>> + /* Color depth greater than 128 bits not supported. */
>> + if (cpp > 16)
>> + return false;
>> +
>> + /* For Fast Copy Blits the pitch cannot be a negative number. So, bit 15
>> + * of the destination pitch must be zero.
>> + */
>> + if ((src_pitch >> 15 & 1) != 0 || (dst_pitch >> 15 & 1) != 0)
>> + return false;
>> +
>> + /* For Linear surfaces, the pitch has to be an OWord (16byte) multiple. */
>> + if ((src_tiling_none && src_pitch % 16 != 0) ||
>> + (dst_tiling_none && dst_pitch % 16 != 0))
>> + return false;
>> +
>> + /* For Tiled surfaces, the pitch has to be a multiple of the Tile width
>> + * (X direction width of the Tile). This means the pitch value will
>> + * always be Cache Line aligned (64byte multiple).
>> + */
>> + if ((!dst_tiling_none && dst_pitch % 64 != 0) ||
>> + (!src_tiling_none && src_pitch % 64 != 0))
>> + return false;
>> +
>> + return true;
>> +}
>> +
>> /* Copy BitBlt
>> */
>> bool
>> @@ -291,10 +446,12 @@ intelEmitCopyBlit(struct brw_context *brw,
>> drm_intel_bo *src_buffer,
>> GLuint src_offset,
>> uint32_t src_tiling,
>> + uint32_t src_tr_mode,
>> GLshort dst_pitch,
>> drm_intel_bo *dst_buffer,
>> GLuint dst_offset,
>> uint32_t dst_tiling,
>> + uint32_t dst_tr_mode,
>> GLshort src_x, GLshort src_y,
>> GLshort dst_x, GLshort dst_y,
>> GLshort w, GLshort h,
>> @@ -306,11 +463,7 @@ intelEmitCopyBlit(struct brw_context *brw,
>> drm_intel_bo *aper_array[3];
>> bool dst_y_tiled = dst_tiling == I915_TILING_Y;
>> bool src_y_tiled = src_tiling == I915_TILING_Y;
>> -
>> - if (!alignment_valid(brw, dst_offset, dst_tiling))
>> - return false;
>> - if (!alignment_valid(brw, src_offset, src_tiling))
>> - return false;
>> + bool use_fast_copy_blit = false;
>>
>> if ((dst_y_tiled || src_y_tiled) && brw->gen < 6)
>> return false;
>> @@ -342,52 +495,107 @@ intelEmitCopyBlit(struct brw_context *brw,
>> src_buffer, src_pitch, src_offset, src_x, src_y,
>> dst_buffer, dst_pitch, dst_offset, dst_x, dst_y, w, h);
>>
>> - /* Blit pitch must be dword-aligned. Otherwise, the hardware appears to drop
>> - * the low bits. Offsets must be naturally aligned.
>> - */
>> - if (src_pitch % 4 != 0 || src_offset % cpp != 0 ||
>> - dst_pitch % 4 != 0 || dst_offset % cpp != 0)
>> - return false;
>> -
>> - /* For big formats (such as floating point), do the copy using 16 or 32bpp
>> - * and multiply the coordinates.
>> - */
>> - if (cpp > 4) {
>> - if (cpp % 4 == 2) {
>> - dst_x *= cpp / 2;
>> - dst_x2 *= cpp / 2;
>> - src_x *= cpp / 2;
>> - cpp = 2;
>> - } else {
>> - assert(cpp % 4 == 0);
>> - dst_x *= cpp / 4;
>> - dst_x2 *= cpp / 4;
>> - src_x *= cpp / 4;
>> - cpp = 4;
>> + use_fast_copy_blit = can_fast_copy_blit(brw,
>> + src_x, src_offset, src_pitch,
>> + src_tiling, src_tr_mode,
>> + dst_x, dst_offset, dst_pitch,
>> + dst_tiling, dst_tr_mode,
>> + w, cpp);
>> + assert(use_fast_copy_blit ||
>> + (src_tr_mode == INTEL_MIPTREE_TRMODE_NONE &&
>> + dst_tr_mode == INTEL_MIPTREE_TRMODE_NONE));
>> +
>> + if (use_fast_copy_blit) {
>> + /* When two sequential fast copy blits have different source surfaces,
>> + * but their destinations refer to the same destination surfaces and
>> + * therefore destinations overlap it is imperative that a flush be
>> + * inserted between the two blits.
>> + *
>> + * FIXME: Figure out a way to avoid flushing when not required.
>> + */
>> + intel_batchbuffer_emit_mi_flush(brw);
>> +
>> + assert(cpp <= 16);
>> + BR13 = br13_for_cpp(cpp);
>> +
>> + if (src_tr_mode == INTEL_MIPTREE_TRMODE_YF)
>> + BR13 |= XY_FAST_SRC_TRMODE_YF;
>> +
>> + if (dst_tr_mode == INTEL_MIPTREE_TRMODE_YF)
>> + BR13 |= XY_FAST_DST_TRMODE_YF;
>> +
>> + CMD = XY_FAST_COPY_BLT_CMD;
>> +
>> + if (dst_tiling != I915_TILING_NONE) {
>> + SET_TILING_XY_FAST_COPY_BLT(dst_tiling, dst_tr_mode, XY_FAST_DST);
>> + /* Pitch value should be specified as a number of Dwords. */
>> + dst_pitch /= 4;
>> + }
>> + if (src_tiling != I915_TILING_NONE) {
>> + SET_TILING_XY_FAST_COPY_BLT(src_tiling, src_tr_mode, XY_FAST_SRC);
>> + /* Pitch value should be specified as a number of Dwords. */
>> + src_pitch /= 4;
>> }
>> - }
>>
>> - BR13 = br13_for_cpp(cpp) | translate_raster_op(logic_op) << 16;
>> + CMD |= get_tr_horizontal_align(src_tr_mode, cpp, true /* is_src */);
>> + CMD |= get_tr_vertical_align(src_tr_mode, cpp, true /* is_src */);
>>
>> - switch (cpp) {
>> - case 1:
>> - case 2:
>> - CMD = XY_SRC_COPY_BLT_CMD;
>> - break;
>> - case 4:
>> - CMD = XY_SRC_COPY_BLT_CMD | XY_BLT_WRITE_ALPHA | XY_BLT_WRITE_RGB;
>> - break;
>> - default:
>> - return false;
>> - }
>> + CMD |= get_tr_horizontal_align(dst_tr_mode, cpp, false /* is_src */);
>> + CMD |= get_tr_vertical_align(dst_tr_mode, cpp, false /* is_src */);
>>
>> - if (dst_tiling != I915_TILING_NONE) {
>> - CMD |= XY_DST_TILED;
>> - dst_pitch /= 4;
>> - }
>> - if (src_tiling != I915_TILING_NONE) {
>> - CMD |= XY_SRC_TILED;
>> - src_pitch /= 4;
>> + } else {
>> + /* For big formats (such as floating point), do the copy using 16 or
>> + * 32bpp and multiply the coordinates.
>> + */
>> + if (cpp > 4) {
>> + if (cpp % 4 == 2) {
>> + dst_x *= cpp / 2;
>> + dst_x2 *= cpp / 2;
>> + src_x *= cpp / 2;
>> + cpp = 2;
>> + } else {
>> + assert(cpp % 4 == 0);
>> + dst_x *= cpp / 4;
>> + dst_x2 *= cpp / 4;
>> + src_x *= cpp / 4;
>> + cpp = 4;
>> + }
>> + }
>> +
>> + if (!alignment_valid(brw, dst_offset, dst_tiling))
>> + return false;
>> + if (!alignment_valid(brw, src_offset, src_tiling))
>> + return false;
>> +
>> + /* Blit pitch must be dword-aligned. Otherwise, the hardware appears to drop
>> + * the low bits. Offsets must be naturally aligned.
>> + */
>> + if (src_pitch % 4 != 0 || src_offset % cpp != 0 ||
>> + dst_pitch % 4 != 0 || dst_offset % cpp != 0)
>> + return false;
>> +
>> + assert(cpp <= 4);
>> + BR13 = br13_for_cpp(cpp) | translate_raster_op(logic_op) << 16;
>> + switch (cpp) {
>> + case 1:
>> + case 2:
>> + CMD = XY_SRC_COPY_BLT_CMD;
>> + break;
>> + case 4:
>> + CMD = XY_SRC_COPY_BLT_CMD | XY_BLT_WRITE_ALPHA | XY_BLT_WRITE_RGB;
>> + break;
>> + default:
>> + return false;
>> + }
>> +
>> + if (dst_tiling != I915_TILING_NONE) {
>> + CMD |= XY_DST_TILED;
>> + dst_pitch /= 4;
>> + }
>> + if (src_tiling != I915_TILING_NONE) {
>> + CMD |= XY_SRC_TILED;
>> + src_pitch /= 4;
>> + }
>> }
>>
>> if (dst_y2 <= dst_y || dst_x2 <= dst_x) {
>> @@ -545,7 +753,9 @@ intel_emit_linear_blit(struct brw_context *brw,
>> dst_x = dst_offset % 64;
>> ok = intelEmitCopyBlit(brw, 1,
>> pitch, src_bo, src_offset - src_x, I915_TILING_NONE,
>> + INTEL_MIPTREE_TRMODE_NONE,
>> pitch, dst_bo, dst_offset - dst_x, I915_TILING_NONE,
>> + INTEL_MIPTREE_TRMODE_NONE,
>> src_x, 0, /* src x/y */
>> dst_x, 0, /* dst x/y */
>> pitch, height, /* w, h */
>> @@ -564,7 +774,9 @@ intel_emit_linear_blit(struct brw_context *brw,
>> if (size != 0) {
>> ok = intelEmitCopyBlit(brw, 1,
>> pitch, src_bo, src_offset - src_x, I915_TILING_NONE,
>> + INTEL_MIPTREE_TRMODE_NONE,
>> pitch, dst_bo, dst_offset - dst_x, I915_TILING_NONE,
>> + INTEL_MIPTREE_TRMODE_NONE,
>> src_x, 0, /* src x/y */
>> dst_x, 0, /* dst x/y */
>> size, 1, /* w, h */
>> diff --git a/src/mesa/drivers/dri/i965/intel_blit.h b/src/mesa/drivers/dri/i965/intel_blit.h
>> index f563939..4f4bbc5 100644
>> --- a/src/mesa/drivers/dri/i965/intel_blit.h
>> +++ b/src/mesa/drivers/dri/i965/intel_blit.h
>> @@ -37,10 +37,12 @@ intelEmitCopyBlit(struct brw_context *brw,
>> drm_intel_bo *src_buffer,
>> GLuint src_offset,
>> uint32_t src_tiling,
>> + uint32_t src_tr_mode,
>> GLshort dst_pitch,
>> drm_intel_bo *dst_buffer,
>> GLuint dst_offset,
>> uint32_t dst_tiling,
>> + uint32_t dst_tr_mode,
>> GLshort srcx, GLshort srcy,
>> GLshort dstx, GLshort dsty,
>> GLshort w, GLshort h,
>> diff --git a/src/mesa/drivers/dri/i965/intel_copy_image.c b/src/mesa/drivers/dri/i965/intel_copy_image.c
>> index f4c7eff..3706704 100644
>> --- a/src/mesa/drivers/dri/i965/intel_copy_image.c
>> +++ b/src/mesa/drivers/dri/i965/intel_copy_image.c
>> @@ -126,9 +126,11 @@ copy_image_with_blitter(struct brw_context *brw,
>> src_mt->pitch,
>> src_mt->bo, src_mt->offset,
>> src_mt->tiling,
>> + src_mt->tr_mode,
>> dst_mt->pitch,
>> dst_mt->bo, dst_mt->offset,
>> dst_mt->tiling,
>> + dst_mt->tr_mode,
>> src_x, src_y,
>> dst_x, dst_y,
>> src_width, src_height,
>> diff --git a/src/mesa/drivers/dri/i965/intel_reg.h b/src/mesa/drivers/dri/i965/intel_reg.h
>> index bd14e18..c8ec56b 100644
>> --- a/src/mesa/drivers/dri/i965/intel_reg.h
>> +++ b/src/mesa/drivers/dri/i965/intel_reg.h
>> @@ -102,6 +102,8 @@
>>
>> #define XY_SRC_COPY_BLT_CMD (CMD_2D | (0x53 << 22))
>>
>> +#define XY_FAST_COPY_BLT_CMD (CMD_2D | (0x42 << 22))
>> +
>> #define XY_TEXT_IMMEDIATE_BLIT_CMD (CMD_2D | (0x31 << 22))
>> # define XY_TEXT_BYTE_PACKED (1 << 16)
>>
>> @@ -111,10 +113,44 @@
>> #define XY_SRC_TILED (1 << 15)
>> #define XY_DST_TILED (1 << 11)
>>
>> +/* BR00 */
>> +#define XY_FAST_SRC_TILED_64K (3 << 20)
>> +#define XY_FAST_SRC_TILED_Y (2 << 20)
>> +#define XY_FAST_SRC_TILED_X (1 << 20)
>> +
>> +#define XY_FAST_DST_TILED_64K (3 << 13)
>> +#define XY_FAST_DST_TILED_Y (2 << 13)
>> +#define XY_FAST_DST_TILED_X (1 << 13)
>> +
>> +#define XY_FAST_SRC_H_ALIGN_512 (4 << 17)
>> +#define XY_FAST_SRC_H_ALIGN_256 (3 << 17)
>> +#define XY_FAST_SRC_H_ALIGN_128 (2 << 17)
>> +#define XY_FAST_SRC_H_ALIGN_64 (1 << 17)
>> +#define XY_FAST_SRC_H_ALIGN_32 (0 << 17)
>> +
>> +#define XY_FAST_SRC_V_ALIGN_256 (2 << 15)
>> +#define XY_FAST_SRC_V_ALIGN_128 (1 << 15)
>> +#define XY_FAST_SRC_V_ALIGN_64 (0 << 15)
>> +
>> +#define XY_FAST_DST_H_ALIGN_512 (4 << 10)
>> +#define XY_FAST_DST_H_ALIGN_256 (3 << 10)
>> +#define XY_FAST_DST_H_ALIGN_128 (2 << 10)
>> +#define XY_FAST_DST_H_ALIGN_64 (1 << 10)
>> +#define XY_FAST_DST_H_ALIGN_32 (0 << 10)
>> +
>> +#define XY_FAST_DST_V_ALIGN_256 (2 << 8)
>> +#define XY_FAST_DST_V_ALIGN_128 (1 << 8)
>> +#define XY_FAST_DST_V_ALIGN_64 (0 << 8)
>> +
>> /* BR13 */
>> #define BR13_8 (0x0 << 24)
>> #define BR13_565 (0x1 << 24)
>> #define BR13_8888 (0x3 << 24)
>> +#define BR13_16161616 (0x4 << 24)
>> +#define BR13_32323232 (0x5 << 24)
>> +
>> +#define XY_FAST_SRC_TRMODE_YF (1 << 31)
>> +#define XY_FAST_DST_TRMODE_YF (1 << 30)
>>
>> /* Pipeline Statistics Counter Registers */
>> #define IA_VERTICES_COUNT 0x2310
More information about the mesa-dev
mailing list