Alex might remember this one. Good news is that I think I fixed VCN which was
the reason we had to back it out last time round. Bad news is that I added
GFX8 and GFX9 to the series so maybe I broke something new.
In short, the essence of the series is to replace a pattern of:
ib->ptr[ib->length_dw++] = bytes - 1;
ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
ib->ptr[ib->length_dw++] = lower_32_bits(src);
ib->ptr[ib->length_dw++] = upper_32_bits(src);
ib->ptr[ib->length_dw++] = lower_32_bits(pe);
ib->ptr[ib->length_dw++] = upper_32_bits(pe);
With:
*ptr++ = bytes - 1;
*ptr++ = 0; /* src/dst endian swap */
*ptr++ = lower_32_bits(src);
*ptr++ = upper_32_bits(src);
*ptr++ = lower_32_bits(pe);
*ptr++ = upper_32_bits(pe);
ib->length_dw = ptr - ib->ptr;
As bloat-o-meter below shows, it is a nice compaction win for PTE updates and
buffer migrations. Maybe it's not huge but then again a cacheline or two less
code to execute, and cache to store it in here and there is not bad. Also
important is that I find the source code pattern easier to read but for that I
guess it could be a matter of taste.
add/remove: 0/1 grow/shrink: 3/74 up/down: 47/-9705 (-9658)
Function old new delta
cik_sdma_ring_pad_ib 120 151 +31
si_dma_ring_pad_ib 36 44 +8
sdma_v7_1_ring_pad_ib 118 126 +8
sdma_v7_0_ring_pad_ib 118 117 -1
sdma_v6_0_ring_pad_ib 118 117 -1
sdma_v5_2_ring_pad_ib 118 117 -1
sdma_v5_0_ring_pad_ib 118 117 -1
sdma_v4_4_2_ring_pad_ib 118 117 -1
sdma_v4_0_ring_pad_ib 118 117 -1
sdma_v2_4_ring_pad_ib 118 117 -1
sdma_v3_0_ring_pad_ib 118 115 -3
si_dma_emit_fill_buffer 108 71 -37
amdgpu_vcn_dec_sw_send_msg 816 777 -39
si_dma_vm_write_pte 158 115 -43
si_dma_vm_copy_pte 137 87 -50
si_dma_emit_copy_buffer 134 84 -50
sgpr_init_compute_shader 224 168 -56
sdma_v3_0_vm_write_pte 163 102 -61
sdma_v2_4_vm_write_pte 163 102 -61
cik_sdma_vm_write_pte 163 102 -61
sdma_v7_1_vm_write_pte 168 105 -63
sdma_v7_1_emit_fill_buffer 119 56 -63
sdma_v7_0_vm_write_pte 168 105 -63
sdma_v7_0_emit_fill_buffer 119 56 -63
sdma_v6_0_vm_write_pte 168 105 -63
sdma_v6_0_emit_fill_buffer 119 56 -63
sdma_v5_2_vm_write_pte 168 105 -63
sdma_v5_2_emit_fill_buffer 119 56 -63
sdma_v5_0_vm_write_pte 168 105 -63
sdma_v5_0_emit_fill_buffer 119 56 -63
sdma_v4_4_2_vm_write_pte 168 105 -63
sdma_v4_4_2_emit_fill_buffer 119 56 -63
sdma_v4_0_vm_write_pte 168 105 -63
sdma_v4_0_emit_fill_buffer 119 56 -63
sdma_v3_0_emit_fill_buffer 116 53 -63
sdma_v2_4_emit_fill_buffer 116 53 -63
cik_sdma_emit_fill_buffer 116 53 -63
amdgpu_vcn_dec_send_msg 596 513 -83
sdma_v7_1_emit_copy_buffer 169 76 -93
sdma_v6_0_emit_copy_buffer 169 76 -93
sdma_v5_2_emit_copy_buffer 169 76 -93
sdma_v5_0_emit_copy_buffer 169 76 -93
sdma_v4_4_2_emit_copy_buffer 169 76 -93
sdma_v4_0_emit_copy_buffer 169 76 -93
sdma_v3_0_vm_copy_pte 158 64 -94
sdma_v3_0_emit_copy_buffer 155 61 -94
sdma_v2_4_vm_copy_pte 158 64 -94
sdma_v2_4_emit_copy_buffer 155 61 -94
cik_sdma_vm_copy_pte 158 64 -94
cik_sdma_emit_copy_buffer 155 61 -94
sdma_v7_1_vm_copy_pte 163 68 -95
sdma_v6_0_vm_copy_pte 163 68 -95
sdma_v5_2_vm_copy_pte 163 68 -95
sdma_v5_0_vm_copy_pte 163 68 -95
sdma_v4_4_2_vm_copy_pte 163 68 -95
sdma_v4_0_vm_copy_pte 163 68 -95
sdma_v7_0_vm_copy_pte 183 75 -108
sdma_v7_0_emit_copy_buffer 317 202 -115
si_dma_vm_set_pte_pde 338 214 -124
sdma_v7_1_vm_set_pte_pde 218 72 -146
sdma_v7_0_vm_set_pte_pde 218 72 -146
sdma_v6_0_vm_set_pte_pde 218 72 -146
sdma_v5_2_vm_set_pte_pde 218 72 -146
sdma_v5_0_vm_set_pte_pde 218 72 -146
sdma_v4_4_2_vm_set_pte_pde 218 72 -146
sdma_v4_0_vm_set_pte_pde 218 72 -146
sdma_v3_0_vm_set_pte_pde 215 69 -146
sdma_v2_4_vm_set_pte_pde 215 69 -146
cik_sdma_vm_set_pte_pde 215 69 -146
amdgpu_vcn_unified_ring_ib_header 172 - -172
amdgpu_vce_get_destroy_msg 785 497 -288
gfx_v9_4_2_run_shader.constprop 885 519 -366
amdgpu_vcn_enc_ring_test_ib 2044 1520 -524
uvd_v6_0_enc_ring_test_ib 1466 928 -538
amdgpu_vce_ring_test_ib 1363 822 -541
uvd_v7_0_enc_ring_test_ib 1466 904 -562
gfx_v8_0_late_init 2295 1533 -762
gfx_v9_0_late_init 3443 2612 -831
Total: Before=10682414, After=10672756, chg -0.09%
Cc: Timur Kristóf <[email protected]>
Cc: Alex Deucher <[email protected]>
Tvrtko Ursulin (18):
drm/amdgpu: More compact VCN IB emission
drm/amdgpu: More compact VCE IB emission
drm/amdgpu: More compact UVD 6 IB emission
drm/amdgpu: More compact UVD 7 IB emission
drm/amdgpu: More compact SI SDMA emission
drm/amdgpu: More compact CIK SDMA IB emission
drm/amdgpu: More compact GFX8 IB emission
drm/amdgpu: More compact GFX 9.4.2 IB emission
drm/amdgpu: More compact GFX9 IB emission
drm/amdgpu: More compact SDMA 2.4 IB emission
drm/amdgpu: More compact SDMA 3.0 IB emission
drm/amdgpu: More compact SDMA 4.0 IB emission
drm/amdgpu: More compact SDMA 4.4.2 IB emission
drm/amdgpu: More compact SDMA 5.0 IB emission
drm/amdgpu: More compact SDMA 5.2 IB emission
drm/amdgpu: More compact SDMA 6.0 IB emission
drm/amdgpu: More compact SDMA 7.0 IB emission
drm/amdgpu: More compact SDMA 7.1 IB emission
drivers/gpu/drm/amd/amdgpu/amdgpu_vce.c | 96 ++++++++--------
drivers/gpu/drm/amd/amdgpu/amdgpu_vcn.c | 134 +++++++++++++----------
drivers/gpu/drm/amd/amdgpu/cik_sdma.c | 105 +++++++++++-------
drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c | 106 +++++++++---------
drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c | 115 +++++++++----------
drivers/gpu/drm/amd/amdgpu/gfx_v9_4_2.c | 51 ++++-----
drivers/gpu/drm/amd/amdgpu/sdma_v2_4.c | 108 ++++++++++--------
drivers/gpu/drm/amd/amdgpu/sdma_v3_0.c | 108 ++++++++++--------
drivers/gpu/drm/amd/amdgpu/sdma_v4_0.c | 109 ++++++++++--------
drivers/gpu/drm/amd/amdgpu/sdma_v4_4_2.c | 108 ++++++++++--------
drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c | 106 +++++++++++-------
drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c | 110 +++++++++++--------
drivers/gpu/drm/amd/amdgpu/sdma_v6_0.c | 110 +++++++++++--------
drivers/gpu/drm/amd/amdgpu/sdma_v7_0.c | 119 +++++++++++---------
drivers/gpu/drm/amd/amdgpu/sdma_v7_1.c | 102 ++++++++++-------
drivers/gpu/drm/amd/amdgpu/si_dma.c | 84 ++++++++------
drivers/gpu/drm/amd/amdgpu/uvd_v6_0.c | 72 ++++++------
drivers/gpu/drm/amd/amdgpu/uvd_v7_0.c | 72 ++++++------
18 files changed, 1038 insertions(+), 777 deletions(-)
--
2.55.0