jumper, split store_f16 into to_half, store4
Pretty much the same deal as the last CL going the other direction: split store_f16 into to_half() and store4(). Platforms that had fused strategies here get a little less optimal, but the code's easier to follow, maintain, and reuse. Also adds widen_cast() to encapsulate the fairly common pattern of expanding one of our logical vector types (e.g. 8-byte U16) up to the width of the physical vector type (e.g. 16-byte __m128i). This operation is deeply understood by Clang, and often is a no-op. I could make bit_cast() do this, but it seems clearer to have two names. Change-Id: I7ba5bb4746acfcaa6d486379f67e07baee3820b2 Reviewed-on: https://skia-review.googlesource.com/11204 Reviewed-by: Herb Derby <herb@google.com> Commit-Queue: Mike Klein <mtklein@chromium.org>
This commit is contained in:
Mike Klein
Skia Commit-Bot
parent
744808823f
commit
95f53be005
@ -3139,18 +3139,19 @@ _sk_load_f16_vfp4:
|
||||
HIDDEN _sk_store_f16_vfp4
|
||||
.globl _sk_store_f16_vfp4
|
||||
_sk_store_f16_vfp4:
|
||||
.long 0xeef00b41 // vmov.f64 d16, d1
|
||||
.long 0xeef03b42 // vmov.f64 d19, d2
|
||||
.long 0xf2631113 // vorr d17, d3, d3
|
||||
.long 0xf2602110 // vorr d18, d0, d0
|
||||
.long 0xf3fa00a1 // vtrn.32 d16, d17
|
||||
.long 0xf3f61620 // vcvt.f16.f32 d17, q8
|
||||
.long 0xf3fa20a3 // vtrn.32 d18, d19
|
||||
.long 0xf2630113 // vorr d16, d3, d3
|
||||
.long 0xe5913000 // ldr r3, [r1]
|
||||
.long 0xf3f60622 // vcvt.f16.f32 d16, q9
|
||||
.long 0xf2612111 // vorr d18, d1, d1
|
||||
.long 0xf3f67620 // vcvt.f16.f32 d23, q8
|
||||
.long 0xe5933000 // ldr r3, [r3]
|
||||
.long 0xf3f66602 // vcvt.f16.f32 d22, q1
|
||||
.long 0xe0833180 // add r3, r3, r0, lsl #3
|
||||
.long 0xf443084f // vst2.16 {d16-d17}, [r3]
|
||||
.long 0xf3f65622 // vcvt.f16.f32 d21, q9
|
||||
.long 0xf3f64600 // vcvt.f16.f32 d20, q0
|
||||
.long 0xf22211b2 // vorr d1, d18, d18
|
||||
.long 0xf22031b0 // vorr d3, d16, d16
|
||||
.long 0xf4c3470d // vst4.16 {d20[0],d21[0],d22[0],d23[0]}, [r3]!
|
||||
.long 0xf4c3474f // vst4.16 {d20[1],d21[1],d22[1],d23[1]}, [r3]
|
||||
.long 0xe2813008 // add r3, r1, #8
|
||||
.long 0xe591c004 // ldr ip, [r1, #4]
|
||||
.long 0xe1a01003 // mov r1, r3
|
||||
@ -3193,7 +3194,6 @@ _sk_clamp_y_vfp4:
|
||||
.long 0xf26218a1 // vadd.i32 d17, d18, d17
|
||||
.long 0xf2201fa1 // vmin.f32 d1, d16, d17
|
||||
.long 0xe12fff1c // bx ip
|
||||
.long 0xe320f000 // nop {0}
|
||||
|
||||
HIDDEN _sk_repeat_x_vfp4
|
||||
.globl _sk_repeat_x_vfp4
|
||||
@ -6907,7 +6907,7 @@ _sk_lerp_565_avx:
|
||||
.byte 255 // (bad)
|
||||
.byte 255 // (bad)
|
||||
.byte 255 // (bad)
|
||||
.byte 233,255,255,255,225 // jmpq ffffffffe2001208 <_sk_linear_gradient_2stops_avx+0xffffffffe1ffeb0f>
|
||||
.byte 233,255,255,255,225 // jmpq ffffffffe2001208 <_sk_linear_gradient_2stops_avx+0xffffffffe1ffeb13>
|
||||
.byte 255 // (bad)
|
||||
.byte 255 // (bad)
|
||||
.byte 255 // (bad)
|
||||
@ -7876,32 +7876,32 @@ _sk_store_f16_avx:
|
||||
.byte 196,67,125,25,202,1 // vextractf128 $0x1,%ymm9,%xmm10
|
||||
.byte 196,193,41,114,210,13 // vpsrld $0xd,%xmm10,%xmm10
|
||||
.byte 196,193,49,114,209,13 // vpsrld $0xd,%xmm9,%xmm9
|
||||
.byte 197,60,89,217 // vmulps %ymm1,%ymm8,%ymm11
|
||||
.byte 196,66,49,43,202 // vpackusdw %xmm10,%xmm9,%xmm9
|
||||
.byte 197,60,89,209 // vmulps %ymm1,%ymm8,%ymm10
|
||||
.byte 196,67,125,25,211,1 // vextractf128 $0x1,%ymm10,%xmm11
|
||||
.byte 196,193,33,114,211,13 // vpsrld $0xd,%xmm11,%xmm11
|
||||
.byte 196,193,41,114,210,13 // vpsrld $0xd,%xmm10,%xmm10
|
||||
.byte 196,66,41,43,211 // vpackusdw %xmm11,%xmm10,%xmm10
|
||||
.byte 197,60,89,218 // vmulps %ymm2,%ymm8,%ymm11
|
||||
.byte 196,67,125,25,220,1 // vextractf128 $0x1,%ymm11,%xmm12
|
||||
.byte 196,193,25,114,212,13 // vpsrld $0xd,%xmm12,%xmm12
|
||||
.byte 196,193,33,114,211,13 // vpsrld $0xd,%xmm11,%xmm11
|
||||
.byte 197,60,89,234 // vmulps %ymm2,%ymm8,%ymm13
|
||||
.byte 196,67,125,25,238,1 // vextractf128 $0x1,%ymm13,%xmm14
|
||||
.byte 196,193,9,114,214,13 // vpsrld $0xd,%xmm14,%xmm14
|
||||
.byte 196,193,17,114,213,13 // vpsrld $0xd,%xmm13,%xmm13
|
||||
.byte 196,66,33,43,220 // vpackusdw %xmm12,%xmm11,%xmm11
|
||||
.byte 197,60,89,195 // vmulps %ymm3,%ymm8,%ymm8
|
||||
.byte 196,67,125,25,199,1 // vextractf128 $0x1,%ymm8,%xmm15
|
||||
.byte 196,193,1,114,215,13 // vpsrld $0xd,%xmm15,%xmm15
|
||||
.byte 196,67,125,25,196,1 // vextractf128 $0x1,%ymm8,%xmm12
|
||||
.byte 196,193,25,114,212,13 // vpsrld $0xd,%xmm12,%xmm12
|
||||
.byte 196,193,57,114,208,13 // vpsrld $0xd,%xmm8,%xmm8
|
||||
.byte 196,193,33,115,251,2 // vpslldq $0x2,%xmm11,%xmm11
|
||||
.byte 196,65,33,235,201 // vpor %xmm9,%xmm11,%xmm9
|
||||
.byte 196,193,33,115,252,2 // vpslldq $0x2,%xmm12,%xmm11
|
||||
.byte 196,65,33,235,226 // vpor %xmm10,%xmm11,%xmm12
|
||||
.byte 196,193,57,115,248,2 // vpslldq $0x2,%xmm8,%xmm8
|
||||
.byte 196,65,57,235,197 // vpor %xmm13,%xmm8,%xmm8
|
||||
.byte 196,193,41,115,255,2 // vpslldq $0x2,%xmm15,%xmm10
|
||||
.byte 196,65,41,235,238 // vpor %xmm14,%xmm10,%xmm13
|
||||
.byte 196,65,49,98,216 // vpunpckldq %xmm8,%xmm9,%xmm11
|
||||
.byte 196,65,49,106,208 // vpunpckhdq %xmm8,%xmm9,%xmm10
|
||||
.byte 196,65,25,98,205 // vpunpckldq %xmm13,%xmm12,%xmm9
|
||||
.byte 196,65,25,106,197 // vpunpckhdq %xmm13,%xmm12,%xmm8
|
||||
.byte 196,66,57,43,196 // vpackusdw %xmm12,%xmm8,%xmm8
|
||||
.byte 196,65,49,97,226 // vpunpcklwd %xmm10,%xmm9,%xmm12
|
||||
.byte 196,65,49,105,234 // vpunpckhwd %xmm10,%xmm9,%xmm13
|
||||
.byte 196,65,33,97,200 // vpunpcklwd %xmm8,%xmm11,%xmm9
|
||||
.byte 196,65,33,105,192 // vpunpckhwd %xmm8,%xmm11,%xmm8
|
||||
.byte 196,65,25,98,217 // vpunpckldq %xmm9,%xmm12,%xmm11
|
||||
.byte 196,65,25,106,209 // vpunpckhdq %xmm9,%xmm12,%xmm10
|
||||
.byte 196,65,17,98,200 // vpunpckldq %xmm8,%xmm13,%xmm9
|
||||
.byte 196,65,17,106,192 // vpunpckhdq %xmm8,%xmm13,%xmm8
|
||||
.byte 72,133,201 // test %rcx,%rcx
|
||||
.byte 117,31 // jne 2142 <_sk_store_f16_avx+0xd6>
|
||||
.byte 117,31 // jne 213e <_sk_store_f16_avx+0xd2>
|
||||
.byte 196,65,120,17,28,248 // vmovups %xmm11,(%r8,%rdi,8)
|
||||
.byte 196,65,120,17,84,248,16 // vmovups %xmm10,0x10(%r8,%rdi,8)
|
||||
.byte 196,65,120,17,76,248,32 // vmovups %xmm9,0x20(%r8,%rdi,8)
|
||||
@ -7910,22 +7910,22 @@ _sk_store_f16_avx:
|
||||
.byte 255,224 // jmpq *%rax
|
||||
.byte 196,65,121,214,28,248 // vmovq %xmm11,(%r8,%rdi,8)
|
||||
.byte 72,131,249,1 // cmp $0x1,%rcx
|
||||
.byte 116,240 // je 213e <_sk_store_f16_avx+0xd2>
|
||||
.byte 116,240 // je 213a <_sk_store_f16_avx+0xce>
|
||||
.byte 196,65,121,23,92,248,8 // vmovhpd %xmm11,0x8(%r8,%rdi,8)
|
||||
.byte 72,131,249,3 // cmp $0x3,%rcx
|
||||
.byte 114,227 // jb 213e <_sk_store_f16_avx+0xd2>
|
||||
.byte 114,227 // jb 213a <_sk_store_f16_avx+0xce>
|
||||
.byte 196,65,121,214,84,248,16 // vmovq %xmm10,0x10(%r8,%rdi,8)
|
||||
.byte 116,218 // je 213e <_sk_store_f16_avx+0xd2>
|
||||
.byte 116,218 // je 213a <_sk_store_f16_avx+0xce>
|
||||
.byte 196,65,121,23,84,248,24 // vmovhpd %xmm10,0x18(%r8,%rdi,8)
|
||||
.byte 72,131,249,5 // cmp $0x5,%rcx
|
||||
.byte 114,205 // jb 213e <_sk_store_f16_avx+0xd2>
|
||||
.byte 114,205 // jb 213a <_sk_store_f16_avx+0xce>
|
||||
.byte 196,65,121,214,76,248,32 // vmovq %xmm9,0x20(%r8,%rdi,8)
|
||||
.byte 116,196 // je 213e <_sk_store_f16_avx+0xd2>
|
||||
.byte 116,196 // je 213a <_sk_store_f16_avx+0xce>
|
||||
.byte 196,65,121,23,76,248,40 // vmovhpd %xmm9,0x28(%r8,%rdi,8)
|
||||
.byte 72,131,249,7 // cmp $0x7,%rcx
|
||||
.byte 114,183 // jb 213e <_sk_store_f16_avx+0xd2>
|
||||
.byte 114,183 // jb 213a <_sk_store_f16_avx+0xce>
|
||||
.byte 196,65,121,214,68,248,48 // vmovq %xmm8,0x30(%r8,%rdi,8)
|
||||
.byte 235,174 // jmp 213e <_sk_store_f16_avx+0xd2>
|
||||
.byte 235,174 // jmp 213a <_sk_store_f16_avx+0xce>
|
||||
|
||||
HIDDEN _sk_store_f32_avx
|
||||
.globl _sk_store_f32_avx
|
||||
@ -7942,7 +7942,7 @@ _sk_store_f32_avx:
|
||||
.byte 196,65,37,20,196 // vunpcklpd %ymm12,%ymm11,%ymm8
|
||||
.byte 196,65,37,21,220 // vunpckhpd %ymm12,%ymm11,%ymm11
|
||||
.byte 72,133,201 // test %rcx,%rcx
|
||||
.byte 117,55 // jne 21fd <_sk_store_f32_avx+0x6d>
|
||||
.byte 117,55 // jne 21f9 <_sk_store_f32_avx+0x6d>
|
||||
.byte 196,67,45,24,225,1 // vinsertf128 $0x1,%xmm9,%ymm10,%ymm12
|
||||
.byte 196,67,61,24,235,1 // vinsertf128 $0x1,%xmm11,%ymm8,%ymm13
|
||||
.byte 196,67,45,6,201,49 // vperm2f128 $0x31,%ymm9,%ymm10,%ymm9
|
||||
@ -7955,22 +7955,22 @@ _sk_store_f32_avx:
|
||||
.byte 255,224 // jmpq *%rax
|
||||
.byte 196,65,121,17,20,128 // vmovupd %xmm10,(%r8,%rax,4)
|
||||
.byte 72,131,249,1 // cmp $0x1,%rcx
|
||||
.byte 116,240 // je 21f9 <_sk_store_f32_avx+0x69>
|
||||
.byte 116,240 // je 21f5 <_sk_store_f32_avx+0x69>
|
||||
.byte 196,65,121,17,76,128,16 // vmovupd %xmm9,0x10(%r8,%rax,4)
|
||||
.byte 72,131,249,3 // cmp $0x3,%rcx
|
||||
.byte 114,227 // jb 21f9 <_sk_store_f32_avx+0x69>
|
||||
.byte 114,227 // jb 21f5 <_sk_store_f32_avx+0x69>
|
||||
.byte 196,65,121,17,68,128,32 // vmovupd %xmm8,0x20(%r8,%rax,4)
|
||||
.byte 116,218 // je 21f9 <_sk_store_f32_avx+0x69>
|
||||
.byte 116,218 // je 21f5 <_sk_store_f32_avx+0x69>
|
||||
.byte 196,65,121,17,92,128,48 // vmovupd %xmm11,0x30(%r8,%rax,4)
|
||||
.byte 72,131,249,5 // cmp $0x5,%rcx
|
||||
.byte 114,205 // jb 21f9 <_sk_store_f32_avx+0x69>
|
||||
.byte 114,205 // jb 21f5 <_sk_store_f32_avx+0x69>
|
||||
.byte 196,67,125,25,84,128,64,1 // vextractf128 $0x1,%ymm10,0x40(%r8,%rax,4)
|
||||
.byte 116,195 // je 21f9 <_sk_store_f32_avx+0x69>
|
||||
.byte 116,195 // je 21f5 <_sk_store_f32_avx+0x69>
|
||||
.byte 196,67,125,25,76,128,80,1 // vextractf128 $0x1,%ymm9,0x50(%r8,%rax,4)
|
||||
.byte 72,131,249,7 // cmp $0x7,%rcx
|
||||
.byte 114,181 // jb 21f9 <_sk_store_f32_avx+0x69>
|
||||
.byte 114,181 // jb 21f5 <_sk_store_f32_avx+0x69>
|
||||
.byte 196,67,125,25,68,128,96,1 // vextractf128 $0x1,%ymm8,0x60(%r8,%rax,4)
|
||||
.byte 235,171 // jmp 21f9 <_sk_store_f32_avx+0x69>
|
||||
.byte 235,171 // jmp 21f5 <_sk_store_f32_avx+0x69>
|
||||
|
||||
HIDDEN _sk_clamp_x_avx
|
||||
.globl _sk_clamp_x_avx
|
||||
@ -10038,27 +10038,29 @@ _sk_store_f16_sse41:
|
||||
.byte 72,139,0 // mov (%rax),%rax
|
||||
.byte 185,0,0,128,7 // mov $0x7800000,%ecx
|
||||
.byte 102,68,15,110,193 // movd %ecx,%xmm8
|
||||
.byte 102,69,15,112,192,0 // pshufd $0x0,%xmm8,%xmm8
|
||||
.byte 102,69,15,111,200 // movdqa %xmm8,%xmm9
|
||||
.byte 68,15,89,200 // mulps %xmm0,%xmm9
|
||||
.byte 102,65,15,114,209,13 // psrld $0xd,%xmm9
|
||||
.byte 102,69,15,111,208 // movdqa %xmm8,%xmm10
|
||||
.byte 102,69,15,112,200,0 // pshufd $0x0,%xmm8,%xmm9
|
||||
.byte 102,69,15,111,193 // movdqa %xmm9,%xmm8
|
||||
.byte 68,15,89,192 // mulps %xmm0,%xmm8
|
||||
.byte 102,65,15,114,208,13 // psrld $0xd,%xmm8
|
||||
.byte 102,69,15,56,43,192 // packusdw %xmm8,%xmm8
|
||||
.byte 102,69,15,111,209 // movdqa %xmm9,%xmm10
|
||||
.byte 68,15,89,209 // mulps %xmm1,%xmm10
|
||||
.byte 102,65,15,114,210,13 // psrld $0xd,%xmm10
|
||||
.byte 102,69,15,111,216 // movdqa %xmm8,%xmm11
|
||||
.byte 102,69,15,56,43,210 // packusdw %xmm10,%xmm10
|
||||
.byte 102,69,15,111,217 // movdqa %xmm9,%xmm11
|
||||
.byte 68,15,89,218 // mulps %xmm2,%xmm11
|
||||
.byte 102,65,15,114,211,13 // psrld $0xd,%xmm11
|
||||
.byte 68,15,89,195 // mulps %xmm3,%xmm8
|
||||
.byte 102,65,15,114,208,13 // psrld $0xd,%xmm8
|
||||
.byte 102,65,15,115,250,2 // pslldq $0x2,%xmm10
|
||||
.byte 102,69,15,235,209 // por %xmm9,%xmm10
|
||||
.byte 102,65,15,115,248,2 // pslldq $0x2,%xmm8
|
||||
.byte 102,69,15,235,195 // por %xmm11,%xmm8
|
||||
.byte 102,69,15,111,202 // movdqa %xmm10,%xmm9
|
||||
.byte 102,69,15,98,200 // punpckldq %xmm8,%xmm9
|
||||
.byte 102,69,15,56,43,219 // packusdw %xmm11,%xmm11
|
||||
.byte 68,15,89,203 // mulps %xmm3,%xmm9
|
||||
.byte 102,65,15,114,209,13 // psrld $0xd,%xmm9
|
||||
.byte 102,69,15,56,43,201 // packusdw %xmm9,%xmm9
|
||||
.byte 102,69,15,97,194 // punpcklwd %xmm10,%xmm8
|
||||
.byte 102,69,15,97,217 // punpcklwd %xmm9,%xmm11
|
||||
.byte 102,69,15,111,200 // movdqa %xmm8,%xmm9
|
||||
.byte 102,69,15,98,203 // punpckldq %xmm11,%xmm9
|
||||
.byte 243,68,15,127,12,248 // movdqu %xmm9,(%rax,%rdi,8)
|
||||
.byte 102,69,15,106,208 // punpckhdq %xmm8,%xmm10
|
||||
.byte 243,68,15,127,84,248,16 // movdqu %xmm10,0x10(%rax,%rdi,8)
|
||||
.byte 102,69,15,106,195 // punpckhdq %xmm11,%xmm8
|
||||
.byte 243,68,15,127,68,248,16 // movdqu %xmm8,0x10(%rax,%rdi,8)
|
||||
.byte 72,173 // lods %ds:(%rsi),%rax
|
||||
.byte 255,224 // jmpq *%rax
|
||||
|
||||
@ -12263,27 +12265,37 @@ _sk_store_f16_sse2:
|
||||
.byte 72,139,0 // mov (%rax),%rax
|
||||
.byte 185,0,0,128,7 // mov $0x7800000,%ecx
|
||||
.byte 102,68,15,110,193 // movd %ecx,%xmm8
|
||||
.byte 102,69,15,112,192,0 // pshufd $0x0,%xmm8,%xmm8
|
||||
.byte 102,69,15,111,200 // movdqa %xmm8,%xmm9
|
||||
.byte 68,15,89,200 // mulps %xmm0,%xmm9
|
||||
.byte 102,65,15,114,209,13 // psrld $0xd,%xmm9
|
||||
.byte 102,69,15,111,208 // movdqa %xmm8,%xmm10
|
||||
.byte 102,69,15,112,200,0 // pshufd $0x0,%xmm8,%xmm9
|
||||
.byte 102,69,15,111,193 // movdqa %xmm9,%xmm8
|
||||
.byte 68,15,89,192 // mulps %xmm0,%xmm8
|
||||
.byte 102,65,15,114,208,13 // psrld $0xd,%xmm8
|
||||
.byte 102,65,15,114,240,16 // pslld $0x10,%xmm8
|
||||
.byte 102,65,15,114,224,16 // psrad $0x10,%xmm8
|
||||
.byte 102,69,15,107,192 // packssdw %xmm8,%xmm8
|
||||
.byte 102,69,15,111,209 // movdqa %xmm9,%xmm10
|
||||
.byte 68,15,89,209 // mulps %xmm1,%xmm10
|
||||
.byte 102,65,15,114,210,13 // psrld $0xd,%xmm10
|
||||
.byte 102,69,15,111,216 // movdqa %xmm8,%xmm11
|
||||
.byte 102,65,15,114,242,16 // pslld $0x10,%xmm10
|
||||
.byte 102,65,15,114,226,16 // psrad $0x10,%xmm10
|
||||
.byte 102,69,15,107,210 // packssdw %xmm10,%xmm10
|
||||
.byte 102,69,15,111,217 // movdqa %xmm9,%xmm11
|
||||
.byte 68,15,89,218 // mulps %xmm2,%xmm11
|
||||
.byte 102,65,15,114,211,13 // psrld $0xd,%xmm11
|
||||
.byte 68,15,89,195 // mulps %xmm3,%xmm8
|
||||
.byte 102,65,15,114,208,13 // psrld $0xd,%xmm8
|
||||
.byte 102,65,15,115,250,2 // pslldq $0x2,%xmm10
|
||||
.byte 102,69,15,235,209 // por %xmm9,%xmm10
|
||||
.byte 102,65,15,115,248,2 // pslldq $0x2,%xmm8
|
||||
.byte 102,69,15,235,195 // por %xmm11,%xmm8
|
||||
.byte 102,69,15,111,202 // movdqa %xmm10,%xmm9
|
||||
.byte 102,69,15,98,200 // punpckldq %xmm8,%xmm9
|
||||
.byte 102,65,15,114,243,16 // pslld $0x10,%xmm11
|
||||
.byte 102,65,15,114,227,16 // psrad $0x10,%xmm11
|
||||
.byte 102,69,15,107,219 // packssdw %xmm11,%xmm11
|
||||
.byte 68,15,89,203 // mulps %xmm3,%xmm9
|
||||
.byte 102,65,15,114,209,13 // psrld $0xd,%xmm9
|
||||
.byte 102,65,15,114,241,16 // pslld $0x10,%xmm9
|
||||
.byte 102,65,15,114,225,16 // psrad $0x10,%xmm9
|
||||
.byte 102,69,15,107,201 // packssdw %xmm9,%xmm9
|
||||
.byte 102,69,15,97,194 // punpcklwd %xmm10,%xmm8
|
||||
.byte 102,69,15,97,217 // punpcklwd %xmm9,%xmm11
|
||||
.byte 102,69,15,111,200 // movdqa %xmm8,%xmm9
|
||||
.byte 102,69,15,98,203 // punpckldq %xmm11,%xmm9
|
||||
.byte 243,68,15,127,12,248 // movdqu %xmm9,(%rax,%rdi,8)
|
||||
.byte 102,69,15,106,208 // punpckhdq %xmm8,%xmm10
|
||||
.byte 243,68,15,127,84,248,16 // movdqu %xmm10,0x10(%rax,%rdi,8)
|
||||
.byte 102,69,15,106,195 // punpckhdq %xmm11,%xmm8
|
||||
.byte 243,68,15,127,68,248,16 // movdqu %xmm8,0x10(%rax,%rdi,8)
|
||||
.byte 72,173 // lods %ds:(%rsi),%rax
|
||||
.byte 255,224 // jmpq *%rax
|
||||
|
||||
|
||||
@ -4286,32 +4286,32 @@ _sk_store_f16_avx LABEL PROC
|
||||
DB 196,67,125,25,202,1 ; vextractf128 $0x1,%ymm9,%xmm10
|
||||
DB 196,193,41,114,210,13 ; vpsrld $0xd,%xmm10,%xmm10
|
||||
DB 196,193,49,114,209,13 ; vpsrld $0xd,%xmm9,%xmm9
|
||||
DB 197,60,89,217 ; vmulps %ymm1,%ymm8,%ymm11
|
||||
DB 196,66,49,43,202 ; vpackusdw %xmm10,%xmm9,%xmm9
|
||||
DB 197,60,89,209 ; vmulps %ymm1,%ymm8,%ymm10
|
||||
DB 196,67,125,25,211,1 ; vextractf128 $0x1,%ymm10,%xmm11
|
||||
DB 196,193,33,114,211,13 ; vpsrld $0xd,%xmm11,%xmm11
|
||||
DB 196,193,41,114,210,13 ; vpsrld $0xd,%xmm10,%xmm10
|
||||
DB 196,66,41,43,211 ; vpackusdw %xmm11,%xmm10,%xmm10
|
||||
DB 197,60,89,218 ; vmulps %ymm2,%ymm8,%ymm11
|
||||
DB 196,67,125,25,220,1 ; vextractf128 $0x1,%ymm11,%xmm12
|
||||
DB 196,193,25,114,212,13 ; vpsrld $0xd,%xmm12,%xmm12
|
||||
DB 196,193,33,114,211,13 ; vpsrld $0xd,%xmm11,%xmm11
|
||||
DB 197,60,89,234 ; vmulps %ymm2,%ymm8,%ymm13
|
||||
DB 196,67,125,25,238,1 ; vextractf128 $0x1,%ymm13,%xmm14
|
||||
DB 196,193,9,114,214,13 ; vpsrld $0xd,%xmm14,%xmm14
|
||||
DB 196,193,17,114,213,13 ; vpsrld $0xd,%xmm13,%xmm13
|
||||
DB 196,66,33,43,220 ; vpackusdw %xmm12,%xmm11,%xmm11
|
||||
DB 197,60,89,195 ; vmulps %ymm3,%ymm8,%ymm8
|
||||
DB 196,67,125,25,199,1 ; vextractf128 $0x1,%ymm8,%xmm15
|
||||
DB 196,193,1,114,215,13 ; vpsrld $0xd,%xmm15,%xmm15
|
||||
DB 196,67,125,25,196,1 ; vextractf128 $0x1,%ymm8,%xmm12
|
||||
DB 196,193,25,114,212,13 ; vpsrld $0xd,%xmm12,%xmm12
|
||||
DB 196,193,57,114,208,13 ; vpsrld $0xd,%xmm8,%xmm8
|
||||
DB 196,193,33,115,251,2 ; vpslldq $0x2,%xmm11,%xmm11
|
||||
DB 196,65,33,235,201 ; vpor %xmm9,%xmm11,%xmm9
|
||||
DB 196,193,33,115,252,2 ; vpslldq $0x2,%xmm12,%xmm11
|
||||
DB 196,65,33,235,226 ; vpor %xmm10,%xmm11,%xmm12
|
||||
DB 196,193,57,115,248,2 ; vpslldq $0x2,%xmm8,%xmm8
|
||||
DB 196,65,57,235,197 ; vpor %xmm13,%xmm8,%xmm8
|
||||
DB 196,193,41,115,255,2 ; vpslldq $0x2,%xmm15,%xmm10
|
||||
DB 196,65,41,235,238 ; vpor %xmm14,%xmm10,%xmm13
|
||||
DB 196,65,49,98,216 ; vpunpckldq %xmm8,%xmm9,%xmm11
|
||||
DB 196,65,49,106,208 ; vpunpckhdq %xmm8,%xmm9,%xmm10
|
||||
DB 196,65,25,98,205 ; vpunpckldq %xmm13,%xmm12,%xmm9
|
||||
DB 196,65,25,106,197 ; vpunpckhdq %xmm13,%xmm12,%xmm8
|
||||
DB 196,66,57,43,196 ; vpackusdw %xmm12,%xmm8,%xmm8
|
||||
DB 196,65,49,97,226 ; vpunpcklwd %xmm10,%xmm9,%xmm12
|
||||
DB 196,65,49,105,234 ; vpunpckhwd %xmm10,%xmm9,%xmm13
|
||||
DB 196,65,33,97,200 ; vpunpcklwd %xmm8,%xmm11,%xmm9
|
||||
DB 196,65,33,105,192 ; vpunpckhwd %xmm8,%xmm11,%xmm8
|
||||
DB 196,65,25,98,217 ; vpunpckldq %xmm9,%xmm12,%xmm11
|
||||
DB 196,65,25,106,209 ; vpunpckhdq %xmm9,%xmm12,%xmm10
|
||||
DB 196,65,17,98,200 ; vpunpckldq %xmm8,%xmm13,%xmm9
|
||||
DB 196,65,17,106,192 ; vpunpckhdq %xmm8,%xmm13,%xmm8
|
||||
DB 72,133,201 ; test %rcx,%rcx
|
||||
DB 117,31 ; jne 21de <_sk_store_f16_avx+0xd6>
|
||||
DB 117,31 ; jne 21da <_sk_store_f16_avx+0xd2>
|
||||
DB 196,65,120,17,28,248 ; vmovups %xmm11,(%r8,%rdi,8)
|
||||
DB 196,65,120,17,84,248,16 ; vmovups %xmm10,0x10(%r8,%rdi,8)
|
||||
DB 196,65,120,17,76,248,32 ; vmovups %xmm9,0x20(%r8,%rdi,8)
|
||||
@ -4320,22 +4320,22 @@ _sk_store_f16_avx LABEL PROC
|
||||
DB 255,224 ; jmpq *%rax
|
||||
DB 196,65,121,214,28,248 ; vmovq %xmm11,(%r8,%rdi,8)
|
||||
DB 72,131,249,1 ; cmp $0x1,%rcx
|
||||
DB 116,240 ; je 21da <_sk_store_f16_avx+0xd2>
|
||||
DB 116,240 ; je 21d6 <_sk_store_f16_avx+0xce>
|
||||
DB 196,65,121,23,92,248,8 ; vmovhpd %xmm11,0x8(%r8,%rdi,8)
|
||||
DB 72,131,249,3 ; cmp $0x3,%rcx
|
||||
DB 114,227 ; jb 21da <_sk_store_f16_avx+0xd2>
|
||||
DB 114,227 ; jb 21d6 <_sk_store_f16_avx+0xce>
|
||||
DB 196,65,121,214,84,248,16 ; vmovq %xmm10,0x10(%r8,%rdi,8)
|
||||
DB 116,218 ; je 21da <_sk_store_f16_avx+0xd2>
|
||||
DB 116,218 ; je 21d6 <_sk_store_f16_avx+0xce>
|
||||
DB 196,65,121,23,84,248,24 ; vmovhpd %xmm10,0x18(%r8,%rdi,8)
|
||||
DB 72,131,249,5 ; cmp $0x5,%rcx
|
||||
DB 114,205 ; jb 21da <_sk_store_f16_avx+0xd2>
|
||||
DB 114,205 ; jb 21d6 <_sk_store_f16_avx+0xce>
|
||||
DB 196,65,121,214,76,248,32 ; vmovq %xmm9,0x20(%r8,%rdi,8)
|
||||
DB 116,196 ; je 21da <_sk_store_f16_avx+0xd2>
|
||||
DB 116,196 ; je 21d6 <_sk_store_f16_avx+0xce>
|
||||
DB 196,65,121,23,76,248,40 ; vmovhpd %xmm9,0x28(%r8,%rdi,8)
|
||||
DB 72,131,249,7 ; cmp $0x7,%rcx
|
||||
DB 114,183 ; jb 21da <_sk_store_f16_avx+0xd2>
|
||||
DB 114,183 ; jb 21d6 <_sk_store_f16_avx+0xce>
|
||||
DB 196,65,121,214,68,248,48 ; vmovq %xmm8,0x30(%r8,%rdi,8)
|
||||
DB 235,174 ; jmp 21da <_sk_store_f16_avx+0xd2>
|
||||
DB 235,174 ; jmp 21d6 <_sk_store_f16_avx+0xce>
|
||||
|
||||
PUBLIC _sk_store_f32_avx
|
||||
_sk_store_f32_avx LABEL PROC
|
||||
@ -4351,7 +4351,7 @@ _sk_store_f32_avx LABEL PROC
|
||||
DB 196,65,37,20,196 ; vunpcklpd %ymm12,%ymm11,%ymm8
|
||||
DB 196,65,37,21,220 ; vunpckhpd %ymm12,%ymm11,%ymm11
|
||||
DB 72,133,201 ; test %rcx,%rcx
|
||||
DB 117,55 ; jne 2299 <_sk_store_f32_avx+0x6d>
|
||||
DB 117,55 ; jne 2295 <_sk_store_f32_avx+0x6d>
|
||||
DB 196,67,45,24,225,1 ; vinsertf128 $0x1,%xmm9,%ymm10,%ymm12
|
||||
DB 196,67,61,24,235,1 ; vinsertf128 $0x1,%xmm11,%ymm8,%ymm13
|
||||
DB 196,67,45,6,201,49 ; vperm2f128 $0x31,%ymm9,%ymm10,%ymm9
|
||||
@ -4364,22 +4364,22 @@ _sk_store_f32_avx LABEL PROC
|
||||
DB 255,224 ; jmpq *%rax
|
||||
DB 196,65,121,17,20,128 ; vmovupd %xmm10,(%r8,%rax,4)
|
||||
DB 72,131,249,1 ; cmp $0x1,%rcx
|
||||
DB 116,240 ; je 2295 <_sk_store_f32_avx+0x69>
|
||||
DB 116,240 ; je 2291 <_sk_store_f32_avx+0x69>
|
||||
DB 196,65,121,17,76,128,16 ; vmovupd %xmm9,0x10(%r8,%rax,4)
|
||||
DB 72,131,249,3 ; cmp $0x3,%rcx
|
||||
DB 114,227 ; jb 2295 <_sk_store_f32_avx+0x69>
|
||||
DB 114,227 ; jb 2291 <_sk_store_f32_avx+0x69>
|
||||
DB 196,65,121,17,68,128,32 ; vmovupd %xmm8,0x20(%r8,%rax,4)
|
||||
DB 116,218 ; je 2295 <_sk_store_f32_avx+0x69>
|
||||
DB 116,218 ; je 2291 <_sk_store_f32_avx+0x69>
|
||||
DB 196,65,121,17,92,128,48 ; vmovupd %xmm11,0x30(%r8,%rax,4)
|
||||
DB 72,131,249,5 ; cmp $0x5,%rcx
|
||||
DB 114,205 ; jb 2295 <_sk_store_f32_avx+0x69>
|
||||
DB 114,205 ; jb 2291 <_sk_store_f32_avx+0x69>
|
||||
DB 196,67,125,25,84,128,64,1 ; vextractf128 $0x1,%ymm10,0x40(%r8,%rax,4)
|
||||
DB 116,195 ; je 2295 <_sk_store_f32_avx+0x69>
|
||||
DB 116,195 ; je 2291 <_sk_store_f32_avx+0x69>
|
||||
DB 196,67,125,25,76,128,80,1 ; vextractf128 $0x1,%ymm9,0x50(%r8,%rax,4)
|
||||
DB 72,131,249,7 ; cmp $0x7,%rcx
|
||||
DB 114,181 ; jb 2295 <_sk_store_f32_avx+0x69>
|
||||
DB 114,181 ; jb 2291 <_sk_store_f32_avx+0x69>
|
||||
DB 196,67,125,25,68,128,96,1 ; vextractf128 $0x1,%ymm8,0x60(%r8,%rax,4)
|
||||
DB 235,171 ; jmp 2295 <_sk_store_f32_avx+0x69>
|
||||
DB 235,171 ; jmp 2291 <_sk_store_f32_avx+0x69>
|
||||
|
||||
PUBLIC _sk_clamp_x_avx
|
||||
_sk_clamp_x_avx LABEL PROC
|
||||
@ -6412,27 +6412,29 @@ _sk_store_f16_sse41 LABEL PROC
|
||||
DB 72,139,0 ; mov (%rax),%rax
|
||||
DB 185,0,0,128,7 ; mov $0x7800000,%ecx
|
||||
DB 102,68,15,110,193 ; movd %ecx,%xmm8
|
||||
DB 102,69,15,112,192,0 ; pshufd $0x0,%xmm8,%xmm8
|
||||
DB 102,69,15,111,200 ; movdqa %xmm8,%xmm9
|
||||
DB 68,15,89,200 ; mulps %xmm0,%xmm9
|
||||
DB 102,65,15,114,209,13 ; psrld $0xd,%xmm9
|
||||
DB 102,69,15,111,208 ; movdqa %xmm8,%xmm10
|
||||
DB 102,69,15,112,200,0 ; pshufd $0x0,%xmm8,%xmm9
|
||||
DB 102,69,15,111,193 ; movdqa %xmm9,%xmm8
|
||||
DB 68,15,89,192 ; mulps %xmm0,%xmm8
|
||||
DB 102,65,15,114,208,13 ; psrld $0xd,%xmm8
|
||||
DB 102,69,15,56,43,192 ; packusdw %xmm8,%xmm8
|
||||
DB 102,69,15,111,209 ; movdqa %xmm9,%xmm10
|
||||
DB 68,15,89,209 ; mulps %xmm1,%xmm10
|
||||
DB 102,65,15,114,210,13 ; psrld $0xd,%xmm10
|
||||
DB 102,69,15,111,216 ; movdqa %xmm8,%xmm11
|
||||
DB 102,69,15,56,43,210 ; packusdw %xmm10,%xmm10
|
||||
DB 102,69,15,111,217 ; movdqa %xmm9,%xmm11
|
||||
DB 68,15,89,218 ; mulps %xmm2,%xmm11
|
||||
DB 102,65,15,114,211,13 ; psrld $0xd,%xmm11
|
||||
DB 68,15,89,195 ; mulps %xmm3,%xmm8
|
||||
DB 102,65,15,114,208,13 ; psrld $0xd,%xmm8
|
||||
DB 102,65,15,115,250,2 ; pslldq $0x2,%xmm10
|
||||
DB 102,69,15,235,209 ; por %xmm9,%xmm10
|
||||
DB 102,65,15,115,248,2 ; pslldq $0x2,%xmm8
|
||||
DB 102,69,15,235,195 ; por %xmm11,%xmm8
|
||||
DB 102,69,15,111,202 ; movdqa %xmm10,%xmm9
|
||||
DB 102,69,15,98,200 ; punpckldq %xmm8,%xmm9
|
||||
DB 102,69,15,56,43,219 ; packusdw %xmm11,%xmm11
|
||||
DB 68,15,89,203 ; mulps %xmm3,%xmm9
|
||||
DB 102,65,15,114,209,13 ; psrld $0xd,%xmm9
|
||||
DB 102,69,15,56,43,201 ; packusdw %xmm9,%xmm9
|
||||
DB 102,69,15,97,194 ; punpcklwd %xmm10,%xmm8
|
||||
DB 102,69,15,97,217 ; punpcklwd %xmm9,%xmm11
|
||||
DB 102,69,15,111,200 ; movdqa %xmm8,%xmm9
|
||||
DB 102,69,15,98,203 ; punpckldq %xmm11,%xmm9
|
||||
DB 243,68,15,127,12,248 ; movdqu %xmm9,(%rax,%rdi,8)
|
||||
DB 102,69,15,106,208 ; punpckhdq %xmm8,%xmm10
|
||||
DB 243,68,15,127,84,248,16 ; movdqu %xmm10,0x10(%rax,%rdi,8)
|
||||
DB 102,69,15,106,195 ; punpckhdq %xmm11,%xmm8
|
||||
DB 243,68,15,127,68,248,16 ; movdqu %xmm8,0x10(%rax,%rdi,8)
|
||||
DB 72,173 ; lods %ds:(%rsi),%rax
|
||||
DB 255,224 ; jmpq *%rax
|
||||
|
||||
@ -8599,27 +8601,37 @@ _sk_store_f16_sse2 LABEL PROC
|
||||
DB 72,139,0 ; mov (%rax),%rax
|
||||
DB 185,0,0,128,7 ; mov $0x7800000,%ecx
|
||||
DB 102,68,15,110,193 ; movd %ecx,%xmm8
|
||||
DB 102,69,15,112,192,0 ; pshufd $0x0,%xmm8,%xmm8
|
||||
DB 102,69,15,111,200 ; movdqa %xmm8,%xmm9
|
||||
DB 68,15,89,200 ; mulps %xmm0,%xmm9
|
||||
DB 102,65,15,114,209,13 ; psrld $0xd,%xmm9
|
||||
DB 102,69,15,111,208 ; movdqa %xmm8,%xmm10
|
||||
DB 102,69,15,112,200,0 ; pshufd $0x0,%xmm8,%xmm9
|
||||
DB 102,69,15,111,193 ; movdqa %xmm9,%xmm8
|
||||
DB 68,15,89,192 ; mulps %xmm0,%xmm8
|
||||
DB 102,65,15,114,208,13 ; psrld $0xd,%xmm8
|
||||
DB 102,65,15,114,240,16 ; pslld $0x10,%xmm8
|
||||
DB 102,65,15,114,224,16 ; psrad $0x10,%xmm8
|
||||
DB 102,69,15,107,192 ; packssdw %xmm8,%xmm8
|
||||
DB 102,69,15,111,209 ; movdqa %xmm9,%xmm10
|
||||
DB 68,15,89,209 ; mulps %xmm1,%xmm10
|
||||
DB 102,65,15,114,210,13 ; psrld $0xd,%xmm10
|
||||
DB 102,69,15,111,216 ; movdqa %xmm8,%xmm11
|
||||
DB 102,65,15,114,242,16 ; pslld $0x10,%xmm10
|
||||
DB 102,65,15,114,226,16 ; psrad $0x10,%xmm10
|
||||
DB 102,69,15,107,210 ; packssdw %xmm10,%xmm10
|
||||
DB 102,69,15,111,217 ; movdqa %xmm9,%xmm11
|
||||
DB 68,15,89,218 ; mulps %xmm2,%xmm11
|
||||
DB 102,65,15,114,211,13 ; psrld $0xd,%xmm11
|
||||
DB 68,15,89,195 ; mulps %xmm3,%xmm8
|
||||
DB 102,65,15,114,208,13 ; psrld $0xd,%xmm8
|
||||
DB 102,65,15,115,250,2 ; pslldq $0x2,%xmm10
|
||||
DB 102,69,15,235,209 ; por %xmm9,%xmm10
|
||||
DB 102,65,15,115,248,2 ; pslldq $0x2,%xmm8
|
||||
DB 102,69,15,235,195 ; por %xmm11,%xmm8
|
||||
DB 102,69,15,111,202 ; movdqa %xmm10,%xmm9
|
||||
DB 102,69,15,98,200 ; punpckldq %xmm8,%xmm9
|
||||
DB 102,65,15,114,243,16 ; pslld $0x10,%xmm11
|
||||
DB 102,65,15,114,227,16 ; psrad $0x10,%xmm11
|
||||
DB 102,69,15,107,219 ; packssdw %xmm11,%xmm11
|
||||
DB 68,15,89,203 ; mulps %xmm3,%xmm9
|
||||
DB 102,65,15,114,209,13 ; psrld $0xd,%xmm9
|
||||
DB 102,65,15,114,241,16 ; pslld $0x10,%xmm9
|
||||
DB 102,65,15,114,225,16 ; psrad $0x10,%xmm9
|
||||
DB 102,69,15,107,201 ; packssdw %xmm9,%xmm9
|
||||
DB 102,69,15,97,194 ; punpcklwd %xmm10,%xmm8
|
||||
DB 102,69,15,97,217 ; punpcklwd %xmm9,%xmm11
|
||||
DB 102,69,15,111,200 ; movdqa %xmm8,%xmm9
|
||||
DB 102,69,15,98,203 ; punpckldq %xmm11,%xmm9
|
||||
DB 243,68,15,127,12,248 ; movdqu %xmm9,(%rax,%rdi,8)
|
||||
DB 102,69,15,106,208 ; punpckhdq %xmm8,%xmm10
|
||||
DB 243,68,15,127,84,248,16 ; movdqu %xmm10,0x10(%rax,%rdi,8)
|
||||
DB 102,69,15,106,195 ; punpckhdq %xmm11,%xmm8
|
||||
DB 243,68,15,127,68,248,16 ; movdqu %xmm8,0x10(%rax,%rdi,8)
|
||||
DB 72,173 ; lods %ds:(%rsi),%rax
|
||||
DB 255,224 ; jmpq *%rax
|
||||
|
||||
|
||||
@ -28,6 +28,14 @@ SI Dst bit_cast(const Src& src) {
|
||||
return unaligned_load<Dst>(&src);
|
||||
}
|
||||
|
||||
template <typename Dst, typename Src>
|
||||
SI Dst widen_cast(const Src& src) {
|
||||
static_assert(sizeof(Dst) > sizeof(Src), "");
|
||||
Dst dst;
|
||||
memcpy(&dst, &src, sizeof(Src));
|
||||
return dst;
|
||||
}
|
||||
|
||||
// A couple functions for embedding constants directly into code,
|
||||
// so that no .const or .literal4 section is created.
|
||||
SI int C(int x) {
|
||||
|
||||
@ -634,117 +634,10 @@ STAGE(load_f16) {
|
||||
STAGE(store_f16) {
|
||||
auto ptr = *(uint64_t**)ctx + x;
|
||||
|
||||
#if !defined(JUMPER)
|
||||
auto float_to_half = [&](F f) {
|
||||
return bit_cast<U32>(f * bit_cast<F>(U32(0x07800000_i))) // Fix up the exponent,
|
||||
>> 13; // then line up the mantissa.
|
||||
};
|
||||
auto rgba = (int16_t*)ptr;
|
||||
rgba[0] = float_to_half(r);
|
||||
rgba[1] = float_to_half(g);
|
||||
rgba[2] = float_to_half(b);
|
||||
rgba[3] = float_to_half(a);
|
||||
#elif defined(__aarch64__)
|
||||
float16x4x4_t halfs = {{
|
||||
vcvt_f16_f32(r),
|
||||
vcvt_f16_f32(g),
|
||||
vcvt_f16_f32(b),
|
||||
vcvt_f16_f32(a),
|
||||
}};
|
||||
vst4_f16((float16_t*)ptr, halfs);
|
||||
#elif defined(__arm__)
|
||||
float16x4x2_t rb_ga = {{
|
||||
vcvt_f16_f32(float32x4_t{r[0], b[0], r[1], b[1]}),
|
||||
vcvt_f16_f32(float32x4_t{g[0], a[0], g[1], a[1]}),
|
||||
}};
|
||||
vst2_f16((float16_t*)ptr, rb_ga);
|
||||
#elif defined(__AVX2__) && defined(__FMA__) && defined(__F16C__)
|
||||
auto R = _mm256_cvtps_ph(r, _MM_FROUND_CUR_DIRECTION),
|
||||
G = _mm256_cvtps_ph(g, _MM_FROUND_CUR_DIRECTION),
|
||||
B = _mm256_cvtps_ph(b, _MM_FROUND_CUR_DIRECTION),
|
||||
A = _mm256_cvtps_ph(a, _MM_FROUND_CUR_DIRECTION);
|
||||
|
||||
auto rg0123 = _mm_unpacklo_epi16(R, G), // r0 g0 r1 g1 r2 g2 r3 g3
|
||||
rg4567 = _mm_unpackhi_epi16(R, G), // r4 g4 r5 g5 r6 g6 r7 g7
|
||||
ba0123 = _mm_unpacklo_epi16(B, A),
|
||||
ba4567 = _mm_unpackhi_epi16(B, A);
|
||||
|
||||
auto _01 = _mm_unpacklo_epi32(rg0123, ba0123),
|
||||
_23 = _mm_unpackhi_epi32(rg0123, ba0123),
|
||||
_45 = _mm_unpacklo_epi32(rg4567, ba4567),
|
||||
_67 = _mm_unpackhi_epi32(rg4567, ba4567);
|
||||
|
||||
if (__builtin_expect(tail,0)) {
|
||||
auto dst = (double*)ptr;
|
||||
if (tail > 0) { _mm_storel_pd(dst+0, _01); }
|
||||
if (tail > 1) { _mm_storeh_pd(dst+1, _01); }
|
||||
if (tail > 2) { _mm_storel_pd(dst+2, _23); }
|
||||
if (tail > 3) { _mm_storeh_pd(dst+3, _23); }
|
||||
if (tail > 4) { _mm_storel_pd(dst+4, _45); }
|
||||
if (tail > 5) { _mm_storeh_pd(dst+5, _45); }
|
||||
if (tail > 6) { _mm_storel_pd(dst+6, _67); }
|
||||
} else {
|
||||
_mm_storeu_si128((__m128i*)ptr + 0, _01);
|
||||
_mm_storeu_si128((__m128i*)ptr + 1, _23);
|
||||
_mm_storeu_si128((__m128i*)ptr + 2, _45);
|
||||
_mm_storeu_si128((__m128i*)ptr + 3, _67);
|
||||
}
|
||||
#elif defined(__AVX__)
|
||||
auto float_to_half = [&](F f) {
|
||||
return bit_cast<U32>(f * bit_cast<F>(U32(0x07800000_i))) // Fix up the exponent,
|
||||
>> 13; // then line up the mantissa.
|
||||
};
|
||||
U32 R = float_to_half(r),
|
||||
G = float_to_half(g),
|
||||
B = float_to_half(b),
|
||||
A = float_to_half(a);
|
||||
auto r0123 = _mm256_extractf128_si256(R, 0),
|
||||
r4567 = _mm256_extractf128_si256(R, 1),
|
||||
g0123 = _mm256_extractf128_si256(G, 0),
|
||||
g4567 = _mm256_extractf128_si256(G, 1),
|
||||
b0123 = _mm256_extractf128_si256(B, 0),
|
||||
b4567 = _mm256_extractf128_si256(B, 1),
|
||||
a0123 = _mm256_extractf128_si256(A, 0),
|
||||
a4567 = _mm256_extractf128_si256(A, 1);
|
||||
auto rg0123 = r0123 | _mm_slli_si128(g0123,2),
|
||||
rg4567 = r4567 | _mm_slli_si128(g4567,2),
|
||||
ba0123 = b0123 | _mm_slli_si128(a0123,2),
|
||||
ba4567 = b4567 | _mm_slli_si128(a4567,2);
|
||||
|
||||
auto _01 = _mm_unpacklo_epi32(rg0123, ba0123),
|
||||
_23 = _mm_unpackhi_epi32(rg0123, ba0123),
|
||||
_45 = _mm_unpacklo_epi32(rg4567, ba4567),
|
||||
_67 = _mm_unpackhi_epi32(rg4567, ba4567);
|
||||
|
||||
if (__builtin_expect(tail,0)) {
|
||||
auto dst = (double*)ptr;
|
||||
if (tail > 0) { _mm_storel_pd(dst+0, _01); }
|
||||
if (tail > 1) { _mm_storeh_pd(dst+1, _01); }
|
||||
if (tail > 2) { _mm_storel_pd(dst+2, _23); }
|
||||
if (tail > 3) { _mm_storeh_pd(dst+3, _23); }
|
||||
if (tail > 4) { _mm_storel_pd(dst+4, _45); }
|
||||
if (tail > 5) { _mm_storeh_pd(dst+5, _45); }
|
||||
if (tail > 6) { _mm_storel_pd(dst+6, _67); }
|
||||
} else {
|
||||
_mm_storeu_si128((__m128i*)ptr + 0, _01);
|
||||
_mm_storeu_si128((__m128i*)ptr + 1, _23);
|
||||
_mm_storeu_si128((__m128i*)ptr + 2, _45);
|
||||
_mm_storeu_si128((__m128i*)ptr + 3, _67);
|
||||
}
|
||||
#elif defined(__SSE2__)
|
||||
auto float_to_half = [&](F f) {
|
||||
return bit_cast<U32>(f * bit_cast<F>(U32(0x07800000_i))) // Fix up the exponent,
|
||||
>> 13; // then line up the mantissa.
|
||||
};
|
||||
U32 R = float_to_half(r),
|
||||
G = float_to_half(g),
|
||||
B = float_to_half(b),
|
||||
A = float_to_half(a);
|
||||
U32 rg = R | _mm_slli_si128(G,2),
|
||||
ba = B | _mm_slli_si128(A,2);
|
||||
_mm_storeu_si128((__m128i*)ptr + 0, _mm_unpacklo_epi32(rg, ba));
|
||||
_mm_storeu_si128((__m128i*)ptr + 1, _mm_unpackhi_epi32(rg, ba));
|
||||
#endif
|
||||
store4(ptr,tail, to_half(r)
|
||||
, to_half(g)
|
||||
, to_half(b)
|
||||
, to_half(a));
|
||||
}
|
||||
|
||||
STAGE(store_f32) {
|
||||
|
||||
@ -48,12 +48,23 @@
|
||||
*b = ptr[2];
|
||||
*a = ptr[3];
|
||||
}
|
||||
SI void store4(void* vptr, size_t tail, U16 r, U16 g, U16 b, U16 a) {
|
||||
auto ptr = (uint16_t*)vptr;
|
||||
ptr[0] = r;
|
||||
ptr[1] = g;
|
||||
ptr[2] = b;
|
||||
ptr[3] = a;
|
||||
}
|
||||
|
||||
SI F from_half(U16 h) {
|
||||
if ((int16_t)h < 0x0400) { h = 0; } // Flush denorm and negative to zero.
|
||||
return bit_cast<F>(h << 13) // Line up the mantissa,
|
||||
* bit_cast<F>(U32(0x77800000)); // then fix up the exponent.
|
||||
}
|
||||
SI U16 to_half(F f) {
|
||||
return bit_cast<U32>(f * bit_cast<F>(U32(0x07800000_i))) // Fix up the exponent,
|
||||
>> 13; // then line up the mantissa.
|
||||
}
|
||||
|
||||
#elif defined(__aarch64__)
|
||||
#include <arm_neon.h>
|
||||
@ -88,11 +99,14 @@
|
||||
*b = rgba.val[2];
|
||||
*a = rgba.val[3];
|
||||
}
|
||||
|
||||
SI F from_half(U16 h) {
|
||||
return vcvt_f32_f16(h);
|
||||
SI void store4(void* ptr, size_t tail, U16 r, U16 g, U16 b, U16 a) {
|
||||
uint16x4x4_t rgba = {{r,g,b,a}};
|
||||
vst4_u16((uint16_t*)ptr, rgba);
|
||||
}
|
||||
|
||||
SI F from_half(U16 h) { return vcvt_f32_f16(h); }
|
||||
SI U16 to_half(F f) { return vcvt_f16_f32(f); }
|
||||
|
||||
#elif defined(__arm__)
|
||||
#if defined(__thumb2__) || !defined(__ARM_ARCH_7A__) || !defined(__ARM_VFPV4__)
|
||||
#error On ARMv7, compile with -march=armv7-a -mfpu=neon-vfp4, without -mthumb.
|
||||
@ -135,12 +149,27 @@
|
||||
*b = unaligned_load<U16>(rgba.val+2);
|
||||
*a = unaligned_load<U16>(rgba.val+3);
|
||||
}
|
||||
SI void store4(void* vptr, size_t tail, U16 r, U16 g, U16 b, U16 a) {
|
||||
auto ptr = (uint16_t*)vptr;
|
||||
uint16x4x4_t rgba = {{
|
||||
widen_cast<uint16x4_t>(r),
|
||||
widen_cast<uint16x4_t>(g),
|
||||
widen_cast<uint16x4_t>(b),
|
||||
widen_cast<uint16x4_t>(a),
|
||||
}};
|
||||
vst4_lane_u16(ptr + 0, rgba, 0);
|
||||
vst4_lane_u16(ptr + 4, rgba, 1);
|
||||
}
|
||||
|
||||
SI F from_half(U16 h) {
|
||||
uint16x4_t v;
|
||||
memcpy(&v, &h, sizeof(h));
|
||||
auto v = widen_cast<uint16x4_t>(h);
|
||||
return vget_low_f32(vcvt_f32_f16(v));
|
||||
}
|
||||
SI U16 to_half(F f) {
|
||||
auto v = widen_cast<float32x4_t>(f);
|
||||
uint16x4_t h = vcvt_f16_f32(v);
|
||||
return unaligned_load<U16>(&h);
|
||||
}
|
||||
|
||||
#elif defined(__AVX__)
|
||||
#include <immintrin.h>
|
||||
@ -222,6 +251,33 @@
|
||||
*b = _mm_unpacklo_epi64(ba0123, ba4567);
|
||||
*a = _mm_unpackhi_epi64(ba0123, ba4567);
|
||||
}
|
||||
SI void store4(void* ptr, size_t tail, U16 r, U16 g, U16 b, U16 a) {
|
||||
auto rg0123 = _mm_unpacklo_epi16(r, g), // r0 g0 r1 g1 r2 g2 r3 g3
|
||||
rg4567 = _mm_unpackhi_epi16(r, g), // r4 g4 r5 g5 r6 g6 r7 g7
|
||||
ba0123 = _mm_unpacklo_epi16(b, a),
|
||||
ba4567 = _mm_unpackhi_epi16(b, a);
|
||||
|
||||
auto _01 = _mm_unpacklo_epi32(rg0123, ba0123),
|
||||
_23 = _mm_unpackhi_epi32(rg0123, ba0123),
|
||||
_45 = _mm_unpacklo_epi32(rg4567, ba4567),
|
||||
_67 = _mm_unpackhi_epi32(rg4567, ba4567);
|
||||
|
||||
if (__builtin_expect(tail,0)) {
|
||||
auto dst = (double*)ptr;
|
||||
if (tail > 0) { _mm_storel_pd(dst+0, _01); }
|
||||
if (tail > 1) { _mm_storeh_pd(dst+1, _01); }
|
||||
if (tail > 2) { _mm_storel_pd(dst+2, _23); }
|
||||
if (tail > 3) { _mm_storeh_pd(dst+3, _23); }
|
||||
if (tail > 4) { _mm_storel_pd(dst+4, _45); }
|
||||
if (tail > 5) { _mm_storeh_pd(dst+5, _45); }
|
||||
if (tail > 6) { _mm_storel_pd(dst+6, _67); }
|
||||
} else {
|
||||
_mm_storeu_si128((__m128i*)ptr + 0, _01);
|
||||
_mm_storeu_si128((__m128i*)ptr + 1, _23);
|
||||
_mm_storeu_si128((__m128i*)ptr + 2, _45);
|
||||
_mm_storeu_si128((__m128i*)ptr + 3, _67);
|
||||
}
|
||||
}
|
||||
|
||||
SI F from_half(U16 h) {
|
||||
#if defined(__AVX2__)
|
||||
@ -237,6 +293,14 @@
|
||||
* bit_cast<F>(U32(0x77800000_i)); // then fix up the exponent.
|
||||
#endif
|
||||
}
|
||||
SI U16 to_half(F f) {
|
||||
#if defined(__AVX2__)
|
||||
return _mm256_cvtps_ph(f, _MM_FROUND_CUR_DIRECTION);
|
||||
#else
|
||||
return pack(bit_cast<U32>(f * bit_cast<F>(U32(0x07800000_i))) // Fix up the exponent,
|
||||
>> 13); // then line up the mantissa.
|
||||
#endif
|
||||
}
|
||||
|
||||
#elif defined(__SSE2__)
|
||||
#include <immintrin.h>
|
||||
@ -266,8 +330,7 @@
|
||||
return unaligned_load<U16>(&p); // We have two copies. Return (the lower) one.
|
||||
}
|
||||
SI U8 pack(U16 v) {
|
||||
__m128i r;
|
||||
memcpy(&r, &v, sizeof(v));
|
||||
auto r = widen_cast<__m128i>(v);
|
||||
r = _mm_packus_epi16(r,r);
|
||||
return unaligned_load<U8>(&r);
|
||||
}
|
||||
@ -302,10 +365,15 @@
|
||||
*b = unaligned_load<U16>((uint16_t*)&ba + 0);
|
||||
*a = unaligned_load<U16>((uint16_t*)&ba + 4);
|
||||
}
|
||||
SI void store4(const void* ptr, size_t tail, U16 r, U16 g, U16 b, U16 a) {
|
||||
auto rg = _mm_unpacklo_epi16(widen_cast<__m128i>(r), widen_cast<__m128i>(g)),
|
||||
ba = _mm_unpacklo_epi16(widen_cast<__m128i>(b), widen_cast<__m128i>(a));
|
||||
_mm_storeu_si128((__m128i*)ptr + 0, _mm_unpacklo_epi32(rg, ba));
|
||||
_mm_storeu_si128((__m128i*)ptr + 1, _mm_unpackhi_epi32(rg, ba));
|
||||
}
|
||||
|
||||
SI F from_half(U16 h) {
|
||||
__m128i v;
|
||||
memcpy(&v, &h, sizeof(h));
|
||||
auto v = widen_cast<__m128i>(h);
|
||||
|
||||
// Same deal as AVX: flush denorms and negatives to zero.
|
||||
v = _mm_andnot_si128(_mm_cmplt_epi16(v, _mm_set1_epi32(0x04000400_i)), v);
|
||||
@ -314,6 +382,10 @@
|
||||
return bit_cast<F>(w << 13) // Line up the mantissa,
|
||||
* bit_cast<F>(U32(0x77800000_i)); // then fix up the exponent.
|
||||
}
|
||||
SI U16 to_half(F f) {
|
||||
return pack(bit_cast<U32>(f * bit_cast<F>(U32(0x07800000_i))) // Fix up the exponent,
|
||||
>> 13); // then line up the mantissa.
|
||||
}
|
||||
#endif
|
||||
|
||||
// We need to be a careful with casts.
|
||||
|
||||
Loading…
Reference in New Issue
Block a user